Skip to contentSkip to conversation

KVM: iommu: Overhaul device posted IRQs support

PATCH Applied Sean Christopherson 9 participants 113 messages Jun 11, 2025 – Jan 24

Subsystems (11)
x86 KVM arm64 KVM x86 +7 more subsystems
lore
Share & tools
lore

Sean Christopherson

author maintainer
#0 lore raw

v3 cover KVM: iommu: Overhaul device posted IRQs support

Marc/Oliver,

Patch 1 is an arm64 fix that I'm guessing you'll want to grab for 6.16. Assuming that's the case, I'll make sure this series lands on top of kvm/master (or maybe an -rc?) at the appropriate point in time. Though if you can grab the patch sooner than later, that'd be super helpful :-)

Oh, and the other patches are of interest to arm64 are:

[PATCH v3 32/62] KVM: Don't WARN if updating IRQ bypass route fails
[PATCH v3 33/62] KVM: Fold kvm_arch_irqfd_route_changed() into kvm_arch_update_irqfd_routing()

In theory, I _think_ those could be moved earlier so that there aren't multi-arch patches buried in a massive x86-centric series, but I really don't want to try and re-disentangle x86's posted interrupt mess at this point.

TL;DR: Overhaul device posted interrupts in KVM and IOMMU, and AVIC in
general.

This applies on the series to add CONFIG_KVM_IOAPIC (and to kill irq_comm.c):

https://lore.kernel.org/all/[email protected]

Fix a variety of bugs related to device posted IRQs, especially on the AMD side, and clean up KVM's implementation (this series actually removes more code than it adds).

Batch #1 is new in this version, and consists of two aforementioned arm64 changes.

Batch #2 is mostly SVM specific:

- Cleans up various warts and bugs in the IRTE tracking
- Fixes AVIC to not reject large VMs (honor KVM's ABI)
- Wire up AVIC to enable_ipiv to support disabling IPI virtualization while still utilizing device posted interrupts, and to workaround erratum #1235.

Batch #3 overhauls the guts of IRQ bypass in KVM, and moves the vast majority of the logic to common x86; only the code that needs to communicate with the IOMMU is truly vendor specific.

Batch #4 is more SVM/AVIC cleanups that are made possible by batch #3.

Batch #5 adds WARNs and drops dead code after all the previous cleanups and fixes (I don't want to add the WARNs earlier; I don't see any point in adding WARNs in code that's known to be broken).

Batch #6 is yet more SVM/AVIC cleanups, with the specific goal of configuring IRTEs to generate GA log interrupts if and only if KVM actually needs a wake event.

v3:
- Rebase on kvm/next to pick up relevant arm64 irqfd routing changes, and account for arm64 as appropriate.
- Fix a suspiciously similar bug in arm64's version of kvm_arch_irqfd_route_changed().
- Add a patch to rename kvm_set_msi_irq() to kvm_msi_to_lapic_irq().

v2:
- https://lore.kernel.org/all/[email protected]
- Drop patches that were already merged.
- Move code into irq.c, not x86.c. [Paolo]
- Collect review/testing tags. [Sairaj, Vasant]
- Sqaush fixup for a comment that was added in the prior patch. [Sairaj]
- Rewrote the changelog for "Delete IRTE link from previous vCPU irrespective of new routing". [Sairaj]
- Actually drop "struct amd_svm_iommu_ir" and all usage in "Track per-vCPU IRTEs using kvm_kernel_irqfd structure" (the previous version was getting hilarious lucky with struct offsets). [Sairaj]
- Drop unused params from kvm_pi_update_irte() and pi_update_irte(). [Sairaj]
- Document the rules and behavior of amd_iommu_update_ga(). [Joerg]
- Fix a changelog typo. [Paolo]
- Document that GALogIntr isn't cached, i.e. can be safely updated without an invalidation. [Joao, Vasant]
- Rework avic_vcpu_{load,put}() to use an enumerated parameter instead of a series of booleans. [Paolo]
- Drop a redundant "&& new". [Francesco]
- Drop the *** DO NOT MERGE *** testing hack patches.

v1: https://lore.kernel.org/all/[email protected]

Maxim Levitsky (2):
KVM: SVM: Add enable_ipiv param, never set IsRunning if disabled
KVM: SVM: Disable (x2)AVIC IPI virtualization if CPU has erratum #1235

Sean Christopherson (60):
KVM: arm64: Explicitly treat routing entry type changes as changes
KVM: arm64: WARN if unmapping vLPI fails
KVM: Pass new routing entries and irqfd when updating IRTEs
KVM: SVM: Track per-vCPU IRTEs using kvm_kernel_irqfd structure
KVM: SVM: Delete IRTE link from previous vCPU before setting new IRTE iommu/amd: KVM: SVM: Delete now-unused cached/previous GA tag fields
KVM: SVM: Delete IRTE link from previous vCPU irrespective of new
routing
KVM: SVM: Drop pointless masking of default APIC base when setting
V_APIC_BAR
KVM: SVM: Drop pointless masking of kernel page pa's with AVIC HPA
masks
KVM: SVM: Add helper to deduplicate code for getting AVIC backing page
KVM: SVM: Drop vcpu_svm's pointless avic_backing_page field
KVM: SVM: Inhibit AVIC if ID is too big instead of rejecting vCPU
creation
KVM: SVM: Drop redundant check in AVIC code on ID during vCPU creation
KVM: SVM: Track AVIC tables as natively sized pointers, not "struct
pages"
KVM: SVM: Drop superfluous "cache" of AVIC Physical ID entry pointer
KVM: VMX: Move enable_ipiv knob to common x86
KVM: VMX: Suppress PI notifications whenever the vCPU is put
KVM: SVM: Add a comment to explain why avic_vcpu_blocking() ignores
IRQ blocking
iommu/amd: KVM: SVM: Use pi_desc_addr to derive ga_root_ptr iommu/amd: KVM: SVM: Pass NULL @vcpu_info to indicate "not guest mode"
KVM: SVM: Stop walking list of routing table entries when updating
IRTE
KVM: VMX: Stop walking list of routing table entries when updating
IRTE
KVM: SVM: Extract SVM specific code out of get_pi_vcpu_info()
KVM: x86: Move IRQ routing/delivery APIs from x86.c => irq.c
KVM: x86: Nullify irqfd->producer after updating IRTEs
KVM: x86: Dedup AVIC vs. PI code for identifying target vCPU
KVM: x86: Move posted interrupt tracepoint to common code
KVM: SVM: Clean up return handling in avic_pi_update_irte()
iommu: KVM: Split "struct vcpu_data" into separate AMD vs. Intel
structs
KVM: Don't WARN if updating IRQ bypass route fails
KVM: Fold kvm_arch_irqfd_route_changed() into
kvm_arch_update_irqfd_routing()
KVM: x86: Track irq_bypass_vcpu in common x86 code
KVM: x86: Skip IOMMU IRTE updates if there's no old or new vCPU being
targeted
KVM: x86: Don't update IRTE entries when old and new routes were !MSI
KVM: SVM: Revert IRTE to legacy mode if IOMMU doesn't provide IR
metadata
KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU iommu/amd: Document which IRTE fields amd_iommu_update_ga() can modify iommu/amd: KVM: SVM: Infer IsRun from validity of pCPU destination iommu/amd: Factor out helper for manipulating IRTE GA/CPU info iommu/amd: KVM: SVM: Set pCPU info in IRTE when setting vCPU affinity iommu/amd: KVM: SVM: Add IRTE metadata to affined vCPU's list if AVIC
is inhibited
KVM: SVM: Don't check for assigned device(s) when updating affinity
KVM: SVM: Don't check for assigned device(s) when activating AVIC
KVM: SVM: WARN if (de)activating guest mode in IOMMU fails
KVM: SVM: Process all IRTEs on affinity change even if one update
fails
KVM: SVM: WARN if updating IRTE GA fields in IOMMU fails
KVM: x86: Drop superfluous "has assigned device" check in
kvm_pi_update_irte()
KVM: x86: WARN if IRQ bypass isn't supported in kvm_pi_update_irte()
KVM: x86: WARN if IRQ bypass routing is updated without in-kernel
local APIC
KVM: SVM: WARN if ir_list is non-empty at vCPU free
KVM: x86: Decouple device assignment from IRQ bypass
KVM: VMX: WARN if VT-d Posted IRQs aren't possible when starting IRQ
bypass
KVM: SVM: Use vcpu_idx, not vcpu_id, for GA log tag/metadata iommu/amd: WARN if KVM calls GA IRTE helpers without virtual APIC
support
KVM: SVM: Fold avic_set_pi_irte_mode() into its sole caller
KVM: SVM: Don't check vCPU's blocking status when toggling AVIC on/off
KVM: SVM: Consolidate IRTE update when toggling AVIC on/off iommu/amd: KVM: SVM: Allow KVM to control need for GA log interrupts
KVM: SVM: Generate GA log IRQs only if the associated vCPUs is
blocking
KVM: x86: Rename kvm_set_msi_irq() => kvm_msi_to_lapic_irq()

 arch/arm64/kvm/arm.c                 |  19 +-
 arch/arm64/kvm/vgic/vgic-v4.c        |  10 +-
 arch/x86/include/asm/irq_remapping.h |  17 +-
 arch/x86/include/asm/kvm-x86-ops.h   |   2 +-
 arch/x86/include/asm/kvm_host.h      |  23 +-
 arch/x86/include/asm/svm.h           |  13 +-
 arch/x86/kvm/irq.c                   | 152 +++++-
 arch/x86/kvm/svm/avic.c              | 702 ++++++++++++---------------
 arch/x86/kvm/svm/svm.c               |   4 +
 arch/x86/kvm/svm/svm.h               |  32 +-
 arch/x86/kvm/trace.h                 |  19 +-
 arch/x86/kvm/vmx/capabilities.h      |   1 -
 arch/x86/kvm/vmx/main.c              |   2 +-
 arch/x86/kvm/vmx/posted_intr.c       | 140 ++----
 arch/x86/kvm/vmx/posted_intr.h       |  10 +-
 arch/x86/kvm/vmx/vmx.c               |   2 -
 arch/x86/kvm/x86.c                   |  90 +---
 drivers/iommu/amd/amd_iommu_types.h  |   1 -
 drivers/iommu/amd/iommu.c            | 125 +++--
 drivers/iommu/intel/irq_remapping.c  |  10 +-
 include/kvm/arm_vgic.h               |   2 +-
 include/linux/amd-iommu.h            |  25 +-
 include/linux/kvm_host.h             |   9 +-
 include/linux/kvm_irqfd.h            |   4 +
 virt/kvm/eventfd.c                   |  22 +-
 25 files changed, 691 insertions(+), 745 deletions(-)

base-commit: 06880162469d702d052e5d51b49a24e43f182af8

Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#1 lore raw

PATCH v3 1/62 KVM: arm64: Explicitly treat routing entry type changes as changes

Explicitly treat type differences as GSI routing changes, as comparing MSI data between two entries could get a false negative, e.g. if userspace changed the type but left the type-specific data as-

Note, the same bug was fixed in x86 by commit bcda70c56f3e ("KVM: x86:
Explicitly treat routing entry type changes as changes").

Fixes 4bf3693d36af ("KVM: arm64: Unmap vLPIs affected by changes to GSI routing information")
---
 arch/arm64/kvm/arm.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
arch/arm64/kvm/arm.c +2 -1
index de2b4e9c9f9f..38a91bb5d4c7 100644
@@ -2764,7 +2764,8 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
 bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
 				  struct kvm_kernel_irq_routing_entry *new)
 {
-	if (new->type != KVM_IRQ_ROUTING_MSI)
+	if (old->type != KVM_IRQ_ROUTING_MSI ||
+	    new->type != KVM_IRQ_ROUTING_MSI)
 		return true;
 
 	return memcmp(&old->msi, &new->msi, sizeof(new->msi));
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
Reviewed-by on v3 1/62
#2 lore raw

Re: [PATCH v3 01/62] KVM: arm64: Explicitly treat routing entry type changes as changes

Quoted 6 lines · Sean Christopherson

Explicitly treat type differences as GSI routing changes, as comparing MSI data between two entries could get a false negative, e.g. if userspace changed the type but left the type-specific data as-

Note, the same bug was fixed in x86 by commit bcda70c56f3e ("KVM: x86:
Explicitly treat routing entry type changes as changes").

Yeah, I'll let you guess where I got the idea from...

Quoted 2 lines
Fixes 4bf3693d36af ("KVM: arm64: Unmap vLPIs affected by changes to GSI routing information")

Thanks,
Oliver

maintainer Applied
#3 lore raw

Re: (subset) [PATCH v3 01/62] KVM: arm64: Explicitly treat routing entry type changes as changes

In reply to Sean Christopherson
Quoted 8 lines · Sean Christopherson

Explicitly treat type differences as GSI routing changes, as comparing MSI data between two entries could get a false negative, e.g. if userspace changed the type but left the type-specific data as-

Note, the same bug was fixed in x86 by commit bcda70c56f3e ("KVM: x86:
Explicitly treat routing entry type changes as changes").

[...]

Applied to fixes, thanks!

[01/62] KVM: arm64: Explicitly treat routing entry type changes as changes

        commit: 1fbe6861a6d9a942fb8ab8677ddf1ecb86b1af60

Cheers,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author maintainer
#4 lore raw

PATCH v3 2/62 KVM: arm64: WARN if unmapping vLPI fails

In reply to Sean Christopherson

WARN if unmapping a vLPI in kvm_vgic_v4_unset_forwarding() fails, as failure means an IRTE has likely been left in a bad state, i.e. IRQs won't go where they should. WARNing in arch code will eventually allow dropping all the plumbing a similar WARN in kvm_irq_routing_update(), and thus avoid having to plumb back a return code just to WARN.

---
 arch/arm64/kvm/vgic/vgic-v4.c | 1 +
 1 file changed, 1 insertion(+)
arch/arm64/kvm/vgic/vgic-v4.c +1 -0
index 193946108192..86e54cefc237 100644
@@ -546,6 +546,7 @@ int kvm_vgic_v4_unset_forwarding(struct kvm *kvm, int host_irq)
 		atomic_dec(&irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vlpi_count);
 		irq->hw = false;
 		ret = its_unmap_vlpi(host_irq);
+		WARN_ON_ONCE(ret);
 	}
 
 	raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
maintainer
#5 lore raw

Re: [PATCH v3 02/62] KVM: arm64: WARN if unmapping vLPI fails

Quoted 4 lines · Sean Christopherson

WARN if unmapping a vLPI in kvm_vgic_v4_unset_forwarding() fails, as failure means an IRTE has likely been left in a bad state, i.e. IRQs won't go where they should.

I have no idea what an IRTE is. But not having an VLPI mapping for an interrupt at the point where we're tearing down the forwarding is pretty benign. IRQs *still* go where they should, and we don't lose anything.

What it may mean is that userspace and the guest may have done things in the wrong order for KVM to accelerate interrupt delivery. That's silly, but also completely harmless.

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
author maintainer
#6 lore raw

Re: [PATCH v3 02/62] KVM: arm64: WARN if unmapping vLPI fails

Quoted 8 lines · Marc Zyngier
Quoted 4 lines · Sean Christopherson

WARN if unmapping a vLPI in kvm_vgic_v4_unset_forwarding() fails, as failure means an IRTE has likely been left in a bad state, i.e. IRQs won't go where they should.

I have no idea what an IRTE is.

Sorry, x86 IOMMU terminology (Interrupt Remapping Table Entry). I think the GIC terminology would be ITS entry? Or maybe ITS mapping?

Quoted 3 lines

But not having an VLPI mapping for an interrupt at the point where we're tearing down the forwarding is pretty benign. IRQs *still* go where they should, and we don't lose anything.

This is the code I'm trying to refer to:

static int its_vlpi_unmap(struct irq_data *d)

  {
	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
	u32 event = its_get_event_id(d);

	if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d))  <=== this shouldn't happen?
		return -EINVAL;

	/* Drop the virtual mapping */
	its_send_discard(its_dev, event);

	/* and restore the physical one */
	irqd_clr_forwarded_to_vcpu(d);
	its_send_mapti(its_dev, d->hwirq, event);
	lpi_update_config(d, 0xff, (lpi_prop_prio |
				    LPI_PROP_ENABLED |
				    LPI_PROP_GROUP1));

	/* Potentially unmap the VM from this ITS */
	its_unmap_vm(its_dev->its, its_dev->event_map.vm);

	/*
	 * Drop the refcount and make the device available again if
	 * this was the last VLPI.
	 */
	if (!--its_dev->event_map.nr_vlpis) {
		its_dev->event_map.vm = NULL;
		kfree(its_dev->event_map.vlpi_maps);
	}

	return 0;

}

When called from kvm_vgic_v4_unset_forwarding()

	if (irq->hw) {
		atomic_dec(&irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vlpi_count);
		irq->hw = false;
		ret = its_unmap_vlpi(host_irq);
	}

IIUC, irq->hw is set if and only if KVM has configured the IRQ to be fowarded directly to a vCPU. Based on this comment in its_map_vlpi():

	/*
	 * The host will never see that interrupt firing again, so it
	 * is vital that we don't do any lazy masking.
	 */

and this code in its_vlpi_map():

		/* Drop the physical mapping */
		its_send_discard(its_dev, event);

my understanding is that the associated physical IRQ will not be delivered to the host while the IRQ is being forwarded to a vCPU.

irq->hw should only become true for MSIs (I'm crossing my fingers that SGIs aren't in play here) if its_map_vlpi() succeeds:

	ret = its_map_vlpi(virq, &map);
	if (ret)
		goto out_unlock_irq;

	irq->hw		= true;
	irq->host_irq	= virq;
	atomic_inc(&map.vpe->vlpi_count);

and so if its_unmap_vlpi() fails in kvm_vgic_v4_unset_forwarding(), then from KVM's perspective, the worst case scenario is that an IRQ has been left in a forwarded state, i.e. the physical mapping hasn't been reinstalled.

KVM already WARNs if kvm_vgic_v4_unset_forwarding() fails when KVM is reacting to a routing change (this is the WARN I want to move into arch code so that kvm_arch_update_irqfd_routing() doesn't plumb a pointless error code up the stack):

		if (irqfd->producer &&
		    kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry)) {
			int ret = kvm_arch_update_irqfd_routing(
					irqfd->kvm, irqfd->producer->irq,
					irqfd->gsi, 1);
			WARN_ON(ret);
		}

It's only the kvm_arch_irq_bypass_del_producer() case where KVM doesn't WARN. If that fails, then the IRQ has potentially been left in a forwarded state, despite whatever driver "owns" the physical IRQ having removed its producer. E.g. if VFIO detaches its irqbypass producer and gives the device back to the host, then whatever is using the device in the host won't receive IRQs as expected.

Looking at this again, its_free_ite() also WARNs on its_unmap_vlpi() failure, so wouldn't it make sense to have its_unmap_vlpi() WARN if irq_set_vcpu_affinity() fails? The only possible failures are that the GIC doesn't have a v4 ITS (from its_irq_set_vcpu_affinity()):

	/* Need a v4 ITS */
	if (!is_v4(its_dev->its))
		return -EINVAL;

	guard(raw_spinlock)(&its_dev->event_map.vlpi_lock);

	/* Unmap request? */
	if (!info)
		return its_vlpi_unmap(d);

or that KVM has gotten out of sync with the GIC/ITS (from its_vlpi_unmap()):

	if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d))
		return -EINVAL;

All of those failure scenario seem like warnable offences when KVM thinks it has configured the IRQ to be forwarded to a vCPU.

E.g.

--

arch/arm64/kvm/vgic/vgic-its.c +1 -1
index 534049c7c94b..98630dae910d 100644
@@ -758,7 +758,7 @@ static void its_free_ite(struct kvm *kvm, struct its_ite *ite)
        if (irq) {
                scoped_guard(raw_spinlock_irqsave, &irq->irq_lock) {
                        if (irq->hw)
-                               WARN_ON(its_unmap_vlpi(ite->irq->host_irq));
+                               its_unmap_vlpi(ite->irq->host_irq);
 
                        irq->hw = false;
                }
arch/arm64/kvm/vgic/vgic-v4.c +2 -2
index 193946108192..911170d4a9c8 100644
@@ -545,10 +545,10 @@ int kvm_vgic_v4_unset_forwarding(struct kvm *kvm, int host_irq)
        if (irq->hw) {
                atomic_dec(&irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vlpi_count);
                irq->hw = false;
-               ret = its_unmap_vlpi(host_irq);
+               its_unmap_vlpi(host_irq);
        }
 
        raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
        vgic_put_irq(kvm, irq);
-       return ret;
+       return 0;
 }
drivers/irqchip/irq-gic-v4.c +2 -2
index 58c28895f8c4..8455b4a5fbb0 100644
@@ -342,10 +342,10 @@ int its_get_vlpi(int irq, struct its_vlpi_map *map)
        return irq_set_vcpu_affinity(irq, &info);
 }
 
-int its_unmap_vlpi(int irq)
+void its_unmap_vlpi(int irq)
 {
        irq_clear_status_flags(irq, IRQ_DISABLE_UNLAZY);
-       return irq_set_vcpu_affinity(irq, NULL);
+       WARN_ON_ONCE(irq_set_vcpu_affinity(irq, NULL));
 }
 
 int its_prop_update_vlpi(int irq, u8 config, bool inv)
include/linux/irqchip/arm-gic-v4.h +1 -1
index 7f1f11a5e4e4..0b0887099fd7 100644
@@ -146,7 +146,7 @@ int its_commit_vpe(struct its_vpe *vpe);
 int its_invall_vpe(struct its_vpe *vpe);
 int its_map_vlpi(int irq, struct its_vlpi_map *map);
 int its_get_vlpi(int irq, struct its_vlpi_map *map);
-int its_unmap_vlpi(int irq);
+void its_unmap_vlpi(int irq);
 int its_prop_update_vlpi(int irq, u8 config, bool inv);
 int its_prop_update_vsgi(int irq, u8 priority, bool group);
 
#7 lore raw

Re: [PATCH v3 02/62] KVM: arm64: WARN if unmapping vLPI fails

Quoted 4 lines · Sean Christopherson
Quoted 3 lines · Marc Zyngier
Quoted 1 line · Sean Christopherson
Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

WARN if unmapping a vLPI in kvm_vgic_v4_unset_forwarding() fails, as failure means an IRTE has likely been left in a bad state, i.e. IRQs won't go where they should.

I have no idea what an IRTE is.

Sorry, x86 IOMMU terminology (Interrupt Remapping Table Entry). I think the GIC terminology would be ITS entry? Or maybe ITS mapping?

We tend to just call it a 'vLPI mapping', which under the hood implies a couple other translations have been wired up as well (vPE + Device).

Quoted 3 lines
Quoted 3 lines

But not having an VLPI mapping for an interrupt at the point where we're tearing down the forwarding is pretty benign. IRQs *still* go where they should, and we don't lose anything.

The VM may not actually be getting torn down, though. The series of fixes [*] we took for 6.16 addressed games that VMMs might be playing on irqbypass for a live VM.

[*] https://lore.kernel.org/kvmarm/[email protected]/

Quoted 2 lines

All of those failure scenario seem like warnable offences when KVM thinks it has configured the IRQ to be forwarded to a vCPU.

I tend to agree here, especially considering how horribly fragile GICv4 has been in some systems. I know of a couple implementations where ITS command failures and/or unmapped MSIs are fatal for the entire machine. Debugging them has been a genuine pain in the ass.

WARN'ing when state tracking for vLPIs is out of whack would've made it a little easier.

Thanks,
Oliver

author maintainer
#8 lore raw

Re: [PATCH v3 02/62] KVM: arm64: WARN if unmapping vLPI fails

Quoted 15 lines · Oliver Upton
Quoted 4 lines · Sean Christopherson
Quoted 3 lines · Marc Zyngier

But not having an VLPI mapping for an interrupt at the point where we're tearing down the forwarding is pretty benign. IRQs *still* go where they should, and we don't lose anything.

The VM may not actually be getting torn down, though. The series of fixes [*] we took for 6.16 addressed games that VMMs might be playing on irqbypass for a live VM.

[*] https://lore.kernel.org/kvmarm/[email protected]/

Quoted 2 lines

All of those failure scenario seem like warnable offences when KVM thinks it has configured the IRQ to be forwarded to a vCPU.

Quoted 7 lines

I tend to agree here, especially considering how horribly fragile GICv4 has been in some systems. I know of a couple implementations where ITS command failures and/or unmapped MSIs are fatal for the entire machine. Debugging them has been a genuine pain in the ass.

WARN'ing when state tracking for vLPIs is out of whack would've made it a little easier.

Marc, does this look and read better?

I'd really, really like to get this sorted out asap, as it's the only thing blocking the series, and I want to get the series into linux-next early next week, before I go OOO for ~10 days.

--
From: Sean Christopherson <[email protected]>
Date: Thu, 12 Jun 2025 16:51:47 -0700
Subject: [PATCH] KVM: arm64: WARN if unmapping a vLPI fails in any path

When unmapping a vLPI, WARN if nullifying vCPU affinity fails, not just if failure occurs when freeing an ITE. If undoing vCPU affinity fails, then odds are very good that vLPI state tracking has has gotten out of whack, i.e. that KVM and the GIC disagree on the state of an IRQ/vLPI. At best, inconsistent state means there is a lurking bug/flaw somewhere. At worst, the inconsistency could eventually be fatal to the host, e.g. if an ITS command fails because KVM's view of things doesn't match reality/hardware.

Note, only the call from kvm_arch_irq_bypass_del_producer() by way of kvm_vgic_v4_unset_forwarding() doesn't already WARN. Common KVM's kvm_irq_routing_update() WARNs if kvm_arch_update_irqfd_routing() fails. For that path, if its_unmap_vlpi() fails in kvm_vgic_v4_unset_forwarding(), the only possible causes are that the GIC doesn't have a v4 ITS (from its_irq_set_vcpu_affinity()):

        /* Need a v4 ITS */
        if (!is_v4(its_dev->its))
                return -EINVAL;

        guard(raw_spinlock)(&its_dev->event_map.vlpi_lock);

        /* Unmap request? */
        if (!info)
                return its_vlpi_unmap(d);

or that KVM has gotten out of sync with the GIC/ITS (from its_vlpi_unmap()):

        if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d))
                return -EINVAL;

All of the above failure scenarios are warnable offences, as they should never occur absent a kernel/KVM bug.

---
 arch/arm64/kvm/vgic/vgic-its.c     | 2 +-
 arch/arm64/kvm/vgic/vgic-v4.c      | 4 ++--
 drivers/irqchip/irq-gic-v4.c       | 4 ++--
 include/linux/irqchip/arm-gic-v4.h | 2 +-
 4 files changed, 6 insertions(+), 6 deletions(-)
arch/arm64/kvm/vgic/vgic-its.c +1 -1
index 534049c7c94b..98630dae910d 100644
@@ -758,7 +758,7 @@ static void its_free_ite(struct kvm *kvm, struct its_ite *ite)
 	if (irq) {
 		scoped_guard(raw_spinlock_irqsave, &irq->irq_lock) {
 			if (irq->hw)
-				WARN_ON(its_unmap_vlpi(ite->irq->host_irq));
+				its_unmap_vlpi(ite->irq->host_irq);
 
 			irq->hw = false;
 		}
arch/arm64/kvm/vgic/vgic-v4.c +2 -2
index 193946108192..911170d4a9c8 100644
@@ -545,10 +545,10 @@ int kvm_vgic_v4_unset_forwarding(struct kvm *kvm, int host_irq)
 	if (irq->hw) {
 		atomic_dec(&irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vlpi_count);
 		irq->hw = false;
-		ret = its_unmap_vlpi(host_irq);
+		its_unmap_vlpi(host_irq);
 	}
 
 	raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
 	vgic_put_irq(kvm, irq);
-	return ret;
+	return 0;
 }
drivers/irqchip/irq-gic-v4.c +2 -2
index 58c28895f8c4..8455b4a5fbb0 100644
@@ -342,10 +342,10 @@ int its_get_vlpi(int irq, struct its_vlpi_map *map)
 	return irq_set_vcpu_affinity(irq, &info);
 }
 
-int its_unmap_vlpi(int irq)
+void its_unmap_vlpi(int irq)
 {
 	irq_clear_status_flags(irq, IRQ_DISABLE_UNLAZY);
-	return irq_set_vcpu_affinity(irq, NULL);
+	WARN_ON_ONCE(irq_set_vcpu_affinity(irq, NULL));
 }
 
 int its_prop_update_vlpi(int irq, u8 config, bool inv)
include/linux/irqchip/arm-gic-v4.h +1 -1
index 7f1f11a5e4e4..0b0887099fd7 100644
@@ -146,7 +146,7 @@ int its_commit_vpe(struct its_vpe *vpe);
 int its_invall_vpe(struct its_vpe *vpe);
 int its_map_vlpi(int irq, struct its_vlpi_map *map);
 int its_get_vlpi(int irq, struct its_vlpi_map *map);
-int its_unmap_vlpi(int irq);
+void its_unmap_vlpi(int irq);
 int its_prop_update_vlpi(int irq, u8 config, bool inv);
 int its_prop_update_vsgi(int irq, u8 priority, bool group);
 

base-commit: 4fc39a165c70a49991b7cc29be3a19eddcd9e5b9 --

maintainer
#9 lore raw

Re: [PATCH v3 02/62] KVM: arm64: WARN if unmapping vLPI fails

Quoted 35 lines · Sean Christopherson
Quoted 15 lines · Oliver Upton
Quoted 4 lines · Sean Christopherson
Quoted 3 lines · Marc Zyngier

But not having an VLPI mapping for an interrupt at the point where we're tearing down the forwarding is pretty benign. IRQs *still* go where they should, and we don't lose anything.

The VM may not actually be getting torn down, though. The series of fixes [*] we took for 6.16 addressed games that VMMs might be playing on irqbypass for a live VM.

[*] https://lore.kernel.org/kvmarm/[email protected]/

Quoted 2 lines

All of those failure scenario seem like warnable offences when KVM thinks it has configured the IRQ to be forwarded to a vCPU.

Quoted 7 lines

I tend to agree here, especially considering how horribly fragile GICv4 has been in some systems. I know of a couple implementations where ITS command failures and/or unmapped MSIs are fatal for the entire machine. Debugging them has been a genuine pain in the ass.

WARN'ing when state tracking for vLPIs is out of whack would've made it a little easier.

Marc, does this look and read better?

I'd really, really like to get this sorted out asap, as it's the only thing blocking the series, and I want to get the series into linux-next early next week, before I go OOO for ~10 days.

--
From: Sean Christopherson <[email protected]>
Date: Thu, 12 Jun 2025 16:51:47 -0700
Subject: [PATCH] KVM: arm64: WARN if unmapping a vLPI fails in any path

Quoted 7 lines

When unmapping a vLPI, WARN if nullifying vCPU affinity fails, not just if failure occurs when freeing an ITE.  If undoing vCPU affinity fails, then odds are very good that vLPI state tracking has has gotten out of whack, i.e. that KVM and the GIC disagree on the state of an IRQ/vLPI.  At best, inconsistent state means there is a lurking bug/flaw somewhere.  At worst, the inconsistency could eventually be fatal to the host, e.g. if an ITS command fails because KVM's view of things doesn't match reality/hardware.

Btw, we finally figured out the reason some machines were just going dark on kexec, with the new kernel being corrupted. It turns out the GIC is still scribbling on the vLPI Pending Table even after it isn't the vLPI Pending Table any more, and is now part of the new kernel's text.

In my queue I have a patch to call unmap_all_vpes() ∀ kvm on kexec, which makes the problem go away.

#10 lore raw

Re: [PATCH v3 02/62] KVM: arm64: WARN if unmapping vLPI fails

In reply to Sean Christopherson
Quoted 21 lines · Sean Christopherson
Quoted 20 lines · Oliver Upton
Quoted 4 lines · Sean Christopherson
Quoted 3 lines · Marc Zyngier

But not having an VLPI mapping for an interrupt at the point where we're tearing down the forwarding is pretty benign. IRQs *still* go where they should, and we don't lose anything.

The VM may not actually be getting torn down, though. The series of fixes [*] we took for 6.16 addressed games that VMMs might be playing on irqbypass for a live VM.

[*] https://lore.kernel.org/kvmarm/[email protected]/

Quoted 2 lines

All of those failure scenario seem like warnable offences when KVM thinks it has configured the IRQ to be forwarded to a vCPU.

I tend to agree here, especially considering how horribly fragile GICv4 has been in some systems. I know of a couple implementations where ITS command failures and/or unmapped MSIs are fatal for the entire machine. Debugging them has been a genuine pain in the ass.

Quoted 8 lines
Quoted 2 lines

WARN'ing when state tracking for vLPIs is out of whack would've made it a little easier.

Marc, does this look and read better?

I'd really, really like to get this sorted out asap, as it's the only thing blocking the series, and I want to get the series into linux-next early next week, before I go OOO for ~10 days.

Can you just send it out as a standalone patch? It's only tangientally related to the truckload of x86 stuff that I'd rather not pull in the event of conflict resolution.

Thanks,
Oliver

Quoted 110 lines

--
From: Sean Christopherson <[email protected]>
Date: Thu, 12 Jun 2025 16:51:47 -0700
Subject: [PATCH] KVM: arm64: WARN if unmapping a vLPI fails in any path

When unmapping a vLPI, WARN if nullifying vCPU affinity fails, not just if failure occurs when freeing an ITE. If undoing vCPU affinity fails, then odds are very good that vLPI state tracking has has gotten out of whack, i.e. that KVM and the GIC disagree on the state of an IRQ/vLPI. At best, inconsistent state means there is a lurking bug/flaw somewhere. At worst, the inconsistency could eventually be fatal to the host, e.g. if an ITS command fails because KVM's view of things doesn't match reality/hardware.

Note, only the call from kvm_arch_irq_bypass_del_producer() by way of kvm_vgic_v4_unset_forwarding() doesn't already WARN. Common KVM's kvm_irq_routing_update() WARNs if kvm_arch_update_irqfd_routing() fails. For that path, if its_unmap_vlpi() fails in kvm_vgic_v4_unset_forwarding(), the only possible causes are that the GIC doesn't have a v4 ITS (from its_irq_set_vcpu_affinity()):

        /* Need a v4 ITS */
        if (!is_v4(its_dev->its))
                return -EINVAL;

        guard(raw_spinlock)(&its_dev->event_map.vlpi_lock);

        /* Unmap request? */
        if (!info)
                return its_vlpi_unmap(d);

or that KVM has gotten out of sync with the GIC/ITS (from its_vlpi_unmap()):

        if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d))
                return -EINVAL;

All of the above failure scenarios are warnable offences, as they should never occur absent a kernel/KVM bug.

---
 arch/arm64/kvm/vgic/vgic-its.c     | 2 +-
 arch/arm64/kvm/vgic/vgic-v4.c      | 4 ++--
 drivers/irqchip/irq-gic-v4.c       | 4 ++--
 include/linux/irqchip/arm-gic-v4.h | 2 +-
 4 files changed, 6 insertions(+), 6 deletions(-)
arch/arm64/kvm/vgic/vgic-its.c +1 -1
index 534049c7c94b..98630dae910d 100644
@@ -758,7 +758,7 @@ static void its_free_ite(struct kvm *kvm, struct its_ite *ite)
 	if (irq) {
 		scoped_guard(raw_spinlock_irqsave, &irq->irq_lock) {
 			if (irq->hw)
-				WARN_ON(its_unmap_vlpi(ite->irq->host_irq));
+				its_unmap_vlpi(ite->irq->host_irq);
 
 			irq->hw = false;
 		}
arch/arm64/kvm/vgic/vgic-v4.c +2 -2
index 193946108192..911170d4a9c8 100644
@@ -545,10 +545,10 @@ int kvm_vgic_v4_unset_forwarding(struct kvm *kvm, int host_irq)
 	if (irq->hw) {
 		atomic_dec(&irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vlpi_count);
 		irq->hw = false;
-		ret = its_unmap_vlpi(host_irq);
+		its_unmap_vlpi(host_irq);
 	}
 
 	raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
 	vgic_put_irq(kvm, irq);
-	return ret;
+	return 0;
 }
drivers/irqchip/irq-gic-v4.c +2 -2
index 58c28895f8c4..8455b4a5fbb0 100644
@@ -342,10 +342,10 @@ int its_get_vlpi(int irq, struct its_vlpi_map *map)
 	return irq_set_vcpu_affinity(irq, &info);
 }
 
-int its_unmap_vlpi(int irq)
+void its_unmap_vlpi(int irq)
 {
 	irq_clear_status_flags(irq, IRQ_DISABLE_UNLAZY);
-	return irq_set_vcpu_affinity(irq, NULL);
+	WARN_ON_ONCE(irq_set_vcpu_affinity(irq, NULL));
 }
 
 int its_prop_update_vlpi(int irq, u8 config, bool inv)
include/linux/irqchip/arm-gic-v4.h +1 -1
index 7f1f11a5e4e4..0b0887099fd7 100644
@@ -146,7 +146,7 @@ int its_commit_vpe(struct its_vpe *vpe);
 int its_invall_vpe(struct its_vpe *vpe);
 int its_map_vlpi(int irq, struct its_vlpi_map *map);
 int its_get_vlpi(int irq, struct its_vlpi_map *map);
-int its_unmap_vlpi(int irq);
+void its_unmap_vlpi(int irq);
 int its_prop_update_vlpi(int irq, u8 config, bool inv);
 int its_prop_update_vsgi(int irq, u8 priority, bool group);
 

base-commit: 4fc39a165c70a49991b7cc29be3a19eddcd9e5b9 --

author maintainer
#11 lore raw

Re: [PATCH v3 02/62] KVM: arm64: WARN if unmapping vLPI fails

Quoted 25 lines · Oliver Upton
Quoted 24 lines · Sean Christopherson
Quoted 22 lines · Oliver Upton
Quoted 4 lines · Sean Christopherson
Quoted 3 lines · Marc Zyngier

But not having an VLPI mapping for an interrupt at the point where we're tearing down the forwarding is pretty benign. IRQs *still* go where they should, and we don't lose anything.

The VM may not actually be getting torn down, though. The series of fixes [*] we took for 6.16 addressed games that VMMs might be playing on irqbypass for a live VM.

[*] https://lore.kernel.org/kvmarm/[email protected]/

Quoted 2 lines

All of those failure scenario seem like warnable offences when KVM thinks it has configured the IRQ to be forwarded to a vCPU.

I tend to agree here, especially considering how horribly fragile GICv4 has been in some systems. I know of a couple implementations where ITS command failures and/or unmapped MSIs are fatal for the entire machine. Debugging them has been a genuine pain in the ass.

WARN'ing when state tracking for vLPIs is out of whack would've made it a little easier.

Quoted 8 lines
Quoted 5 lines

Marc, does this look and read better?

I'd really, really like to get this sorted out asap, as it's the only thing blocking the series, and I want to get the series into linux-next early next week, before I go OOO for ~10 days.

Can you just send it out as a standalone patch? It's only tangientally related to the truckload of x86 stuff

The issue is that "KVM: Don't WARN if updating IRQ bypass route fails" directly depends on both this patch and decent chunk of the x86 crud. I could probably trim some of the x86 crud by reshuffling patches around, but I can't get rid of it entirely.

Quoted 1 line

that I'd rather not pull in the event of conflict resolution.

LOL, why not? :-)

If I post it as a standalone patch, could you/Marc put it into a stable topic branch based on kvm/master? (kvm/master now has patch 1, yay!) Then I can create a topic branch for this mountain of stuff based on the arm64 topic branch.

#12 lore raw

Re: [PATCH v3 02/62] KVM: arm64: WARN if unmapping vLPI fails

Quoted 6 lines · Sean Christopherson
Quoted 2 lines · Oliver Upton

Can you just send it out as a standalone patch? It's only tangientally related to the truckload of x86 stuff

The issue is that "KVM: Don't WARN if updating IRQ bypass route fails" directly depends on both this patch

Eh? At worst we break bisection for the warn, which I'm sure we manage to commit similar high crimes on the regular.

Quoted 3 lines

If I post it as a standalone patch, could you/Marc put it into a stable topic branch based on kvm/master? (kvm/master now has patch 1, yay!) Then I can create a topic branch for this mountain of stuff based on the arm64 topic branch.

Ok, how about making the arm64 piece patch 1 in your series and you take the whole pile. If we need it, I'll bug you for a ref that only has the first change.

That ok?

Thanks,
Oliver

author maintainer
#13 lore raw

Re: [PATCH v3 02/62] KVM: arm64: WARN if unmapping vLPI fails

Quoted 8 lines · Oliver Upton
Quoted 3 lines · Sean Christopherson

If I post it as a standalone patch, could you/Marc put it into a stable topic branch based on kvm/master? (kvm/master now has patch 1, yay!) Then I can create a topic branch for this mountain of stuff based on the arm64 topic branch.

Ok, how about making the arm64 piece patch 1 in your series and you take the whole pile. If we need it, I'll bug you for a ref that only has the first change.

Any preference as to whether I formally post the last version, or if I apply it directly from this thread?

Quoted 1 line

That ok?

Ya, works for me. What's the going bribe rate for an ack these days? :-D

Acked-by
#14 lore raw

Re: [PATCH v3 02/62] KVM: arm64: WARN if unmapping vLPI fails

Quoted 6 lines · Sean Christopherson
Quoted 5 lines · Oliver Upton
Quoted 3 lines · Sean Christopherson

If I post it as a standalone patch, could you/Marc put it into a stable topic branch based on kvm/master? (kvm/master now has patch 1, yay!) Then I can create a topic branch for this mountain of stuff based on the arm64 topic branch.

Quoted 6 lines
Quoted 3 lines

Ok, how about making the arm64 piece patch 1 in your series and you take the whole pile. If we need it, I'll bug you for a ref that only has the first change.

Any preference as to whether I formally post the last version, or if I apply it directly from this thread?

Since I have this email on my phone I'm always in favor of not getting spammed :) This thread has all the discourse too, so likely a better artifact.

Quoted 3 lines
Quoted 1 line

That ok?

Ya, works for me. What's the going bribe rate for an ack these days? :-D

Usually about the cost of a pint ;-) For the arm64 bits:

Signature · 2 lines
Thanks,
Oliver
author maintainer
#15 lore raw

PATCH v3 3/62 KVM: Pass new routing entries and irqfd when updating IRTEs

In reply to Sean Christopherson

When updating IRTEs in response to a GSI routing or IRQ bypass change, pass the new/current routing information along with the associated irqfd. This will allow KVM x86 to harden, simplify, and deduplicate its code.

Since adding/removing a bypass producer is now conveniently protected with irqfds.lock, i.e. can't run concurrently with kvm_irq_routing_update(), use the routing information cached in the irqfd instead of looking up the information in the current GSI routing tables.

Opportunistically convert an existing printk() to pr_info() and put its string onto a single line (old code that strictly adhered to 80 chars).

---
 arch/arm64/kvm/arm.c            |  7 ++++---
 arch/x86/include/asm/kvm_host.h |  6 ++++--
 arch/x86/kvm/svm/avic.c         | 18 +++++++----------
 arch/x86/kvm/svm/svm.h          |  5 +++--
 arch/x86/kvm/vmx/posted_intr.c  | 19 ++++++++---------
 arch/x86/kvm/vmx/posted_intr.h  |  8 ++++++--
 arch/x86/kvm/x86.c              | 36 ++++++++++++++++++---------------
 include/linux/kvm_host.h        |  7 +++++--
 virt/kvm/eventfd.c              | 11 +++++-----
 9 files changed, 62 insertions(+), 55 deletions(-)
arch/arm64/kvm/arm.c +4 -3
index 38a91bb5d4c7..a9a39e0375f7 100644
@@ -2771,8 +2771,9 @@ bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
 	return memcmp(&old->msi, &new->msi, sizeof(new->msi));
 }
 
-int kvm_arch_update_irqfd_routing(struct kvm *kvm, unsigned int host_irq,
-				  uint32_t guest_irq, bool set)
+int kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
+				  struct kvm_kernel_irq_routing_entry *old,
+				  struct kvm_kernel_irq_routing_entry *new)
 {
 	/*
 	 * Remapping the vLPI requires taking the its_lock mutex to resolve
@@ -2781,7 +2782,7 @@ int kvm_arch_update_irqfd_routing(struct kvm *kvm, unsigned int host_irq,
 	 *
 	 * Unmap the vLPI and fall back to software LPI injection.
 	 */
-	return kvm_vgic_v4_unset_forwarding(kvm, host_irq);
+	return kvm_vgic_v4_unset_forwarding(irqfd->kvm, irqfd->producer->irq);
 }
 
 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
arch/x86/include/asm/kvm_host.h +4 -2
index 21ccb122ab76..2a6ef1398da7 100644
@@ -296,6 +296,7 @@ enum x86_intercept_stage;
  */
 #define KVM_APIC_PV_EOI_PENDING	1
 
+struct kvm_kernel_irqfd;
 struct kvm_kernel_irq_routing_entry;
 
 /*
@@ -1844,8 +1845,9 @@ struct kvm_x86_ops {
 	void (*vcpu_blocking)(struct kvm_vcpu *vcpu);
 	void (*vcpu_unblocking)(struct kvm_vcpu *vcpu);
 
-	int (*pi_update_irte)(struct kvm *kvm, unsigned int host_irq,
-			      uint32_t guest_irq, bool set);
+	int (*pi_update_irte)(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
+			      unsigned int host_irq, uint32_t guest_irq,
+			      struct kvm_kernel_irq_routing_entry *new);
 	void (*pi_start_assignment)(struct kvm *kvm);
 	void (*apicv_pre_state_restore)(struct kvm_vcpu *vcpu);
 	void (*apicv_post_state_restore)(struct kvm_vcpu *vcpu);
arch/x86/kvm/svm/avic.c +7 -11
index 7338879d1c0c..adacf00d6664 100644
@@ -18,6 +18,7 @@
 #include <linux/hashtable.h>
 #include <linux/amd-iommu.h>
 #include <linux/kvm_host.h>
+#include <linux/kvm_irqfd.h>
 
 #include <asm/irq_remapping.h>
 
@@ -885,21 +886,14 @@ get_pi_vcpu_info(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
 	return 0;
 }
 
-/*
- * avic_pi_update_irte - set IRTE for Posted-Interrupts
- *
- * @kvm: kvm
- * @host_irq: host irq of the interrupt
- * @guest_irq: gsi of the interrupt
- * @set: set or unset PI
- * returns 0 on success, < 0 on failure
- */
-int avic_pi_update_irte(struct kvm *kvm, unsigned int host_irq,
-			uint32_t guest_irq, bool set)
+int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
+			unsigned int host_irq, uint32_t guest_irq,
+			struct kvm_kernel_irq_routing_entry *new)
 {
 	struct kvm_kernel_irq_routing_entry *e;
 	struct kvm_irq_routing_table *irq_rt;
 	bool enable_remapped_mode = true;
+	bool set = !!new;
 	int idx, ret = 0;
 
 	if (!kvm_arch_has_assigned_device(kvm) || !kvm_arch_has_irq_bypass())
@@ -925,6 +919,8 @@ int avic_pi_update_irte(struct kvm *kvm, unsigned int host_irq,
 		if (e->type != KVM_IRQ_ROUTING_MSI)
 			continue;
 
+		WARN_ON_ONCE(new && memcmp(e, new, sizeof(*new)));
+
 		/**
 		 * Here, we setup with legacy mode in the following cases:
 		 * 1. When cannot target interrupt to a specific vcpu.
arch/x86/kvm/svm/svm.h +3 -2
index e6f3c6a153a0..b35fce30d923 100644
@@ -736,8 +736,9 @@ void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu);
 void avic_vcpu_put(struct kvm_vcpu *vcpu);
 void avic_apicv_post_state_restore(struct kvm_vcpu *vcpu);
 void avic_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu);
-int avic_pi_update_irte(struct kvm *kvm, unsigned int host_irq,
-			uint32_t guest_irq, bool set);
+int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
+			unsigned int host_irq, uint32_t guest_irq,
+			struct kvm_kernel_irq_routing_entry *new);
 void avic_vcpu_blocking(struct kvm_vcpu *vcpu);
 void avic_vcpu_unblocking(struct kvm_vcpu *vcpu);
 void avic_ring_doorbell(struct kvm_vcpu *vcpu);
arch/x86/kvm/vmx/posted_intr.c +8 -11
index 5c615e5845bf..110fb19848ab 100644
@@ -2,6 +2,7 @@
 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 
 #include <linux/kvm_host.h>
+#include <linux/kvm_irqfd.h>
 
 #include <asm/irq_remapping.h>
 #include <asm/cpu.h>
@@ -294,17 +295,9 @@ void vmx_pi_start_assignment(struct kvm *kvm)
 	kvm_make_all_cpus_request(kvm, KVM_REQ_UNBLOCK);
 }
 
-/*
- * vmx_pi_update_irte - set IRTE for Posted-Interrupts
- *
- * @kvm: kvm
- * @host_irq: host irq of the interrupt
- * @guest_irq: gsi of the interrupt
- * @set: set or unset PI
- * returns 0 on success, < 0 on failure
- */
-int vmx_pi_update_irte(struct kvm *kvm, unsigned int host_irq,
-		       uint32_t guest_irq, bool set)
+int vmx_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
+		       unsigned int host_irq, uint32_t guest_irq,
+		       struct kvm_kernel_irq_routing_entry *new)
 {
 	struct kvm_kernel_irq_routing_entry *e;
 	struct kvm_irq_routing_table *irq_rt;
@@ -312,6 +305,7 @@ int vmx_pi_update_irte(struct kvm *kvm, unsigned int host_irq,
 	struct kvm_lapic_irq irq;
 	struct kvm_vcpu *vcpu;
 	struct vcpu_data vcpu_info;
+	bool set = !!new;
 	int idx, ret = 0;
 
 	if (!vmx_can_use_vtd_pi(kvm))
@@ -329,6 +323,9 @@ int vmx_pi_update_irte(struct kvm *kvm, unsigned int host_irq,
 	hlist_for_each_entry(e, &irq_rt->map[guest_irq], link) {
 		if (e->type != KVM_IRQ_ROUTING_MSI)
 			continue;
+
+		WARN_ON_ONCE(new && memcmp(e, new, sizeof(*new)));
+
 		/*
 		 * VT-d PI cannot support posting multicast/broadcast
 		 * interrupts to a vCPU, we still use interrupt remapping
arch/x86/kvm/vmx/posted_intr.h +6 -2
index 80499ea0e674..a94afcb55f7f 100644
@@ -3,6 +3,9 @@
 #define __KVM_X86_VMX_POSTED_INTR_H
 
 #include <linux/bitmap.h>
+#include <linux/find.h>
+#include <linux/kvm_host.h>
+
 #include <asm/posted_intr.h>
 
 void vmx_vcpu_pi_load(struct kvm_vcpu *vcpu, int cpu);
@@ -11,8 +14,9 @@ void pi_wakeup_handler(void);
 void __init pi_init_cpu(int cpu);
 void pi_apicv_pre_state_restore(struct kvm_vcpu *vcpu);
 bool pi_has_pending_interrupt(struct kvm_vcpu *vcpu);
-int vmx_pi_update_irte(struct kvm *kvm, unsigned int host_irq,
-		       uint32_t guest_irq, bool set);
+int vmx_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
+		       unsigned int host_irq, uint32_t guest_irq,
+		       struct kvm_kernel_irq_routing_entry *new);
 void vmx_pi_start_assignment(struct kvm *kvm);
 
 static inline int pi_find_highest_vector(struct pi_desc *pi_desc)
arch/x86/kvm/x86.c +20 -16
index 28a20f0aa3dd..93711a5ef272 100644
@@ -13507,31 +13507,31 @@ int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
 	struct kvm_kernel_irqfd *irqfd =
 		container_of(cons, struct kvm_kernel_irqfd, consumer);
 	struct kvm *kvm = irqfd->kvm;
-	int ret;
+	int ret = 0;
 
 	kvm_arch_start_assignment(irqfd->kvm);
 
 	spin_lock_irq(&kvm->irqfds.lock);
 	irqfd->producer = prod;
 
-	ret = kvm_x86_call(pi_update_irte)(irqfd->kvm,
-					   prod->irq, irqfd->gsi, 1);
-	if (ret)
-		kvm_arch_end_assignment(irqfd->kvm);
-
+	if (irqfd->irq_entry.type == KVM_IRQ_ROUTING_MSI) {
+		ret = kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, prod->irq,
+						   irqfd->gsi, &irqfd->irq_entry);
+		if (ret)
+			kvm_arch_end_assignment(irqfd->kvm);
+	}
 	spin_unlock_irq(&kvm->irqfds.lock);
 
-
 	return ret;
 }
 
 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
 				      struct irq_bypass_producer *prod)
 {
-	int ret;
 	struct kvm_kernel_irqfd *irqfd =
 		container_of(cons, struct kvm_kernel_irqfd, consumer);
 	struct kvm *kvm = irqfd->kvm;
+	int ret;
 
 	WARN_ON(irqfd->producer != prod);
 
@@ -13544,11 +13544,13 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
 	spin_lock_irq(&kvm->irqfds.lock);
 	irqfd->producer = NULL;
 
-	ret = kvm_x86_call(pi_update_irte)(irqfd->kvm,
-					   prod->irq, irqfd->gsi, 0);
-	if (ret)
-		printk(KERN_INFO "irq bypass consumer (token %p) unregistration"
-		       " fails: %d\n", irqfd->consumer.token, ret);
+	if (irqfd->irq_entry.type == KVM_IRQ_ROUTING_MSI) {
+		ret = kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, prod->irq,
+						   irqfd->gsi, NULL);
+		if (ret)
+			pr_info("irq bypass consumer (token %p) unregistration fails: %d\n",
+				irqfd->consumer.token, ret);
+	}
 
 	spin_unlock_irq(&kvm->irqfds.lock);
 
@@ -13556,10 +13558,12 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
 	kvm_arch_end_assignment(irqfd->kvm);
 }
 
-int kvm_arch_update_irqfd_routing(struct kvm *kvm, unsigned int host_irq,
-				   uint32_t guest_irq, bool set)
+int kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
+				  struct kvm_kernel_irq_routing_entry *old,
+				  struct kvm_kernel_irq_routing_entry *new)
 {
-	return kvm_x86_call(pi_update_irte)(kvm, host_irq, guest_irq, set);
+	return kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, irqfd->producer->irq,
+					    irqfd->gsi, new);
 }
 
 bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
include/linux/kvm_host.h +5 -2
index 3b5575d0b574..a4160c1c0c6b 100644
@@ -2401,6 +2401,8 @@ struct kvm_vcpu *kvm_get_running_vcpu(void);
 struct kvm_vcpu * __percpu *kvm_get_running_vcpus(void);
 
 #if IS_ENABLED(CONFIG_HAVE_KVM_IRQ_BYPASS)
+struct kvm_kernel_irqfd;
+
 bool kvm_arch_has_irq_bypass(void);
 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *,
 			   struct irq_bypass_producer *);
@@ -2408,8 +2410,9 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *,
 			   struct irq_bypass_producer *);
 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *);
 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *);
-int kvm_arch_update_irqfd_routing(struct kvm *kvm, unsigned int host_irq,
-				  uint32_t guest_irq, bool set);
+int kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
+				  struct kvm_kernel_irq_routing_entry *old,
+				  struct kvm_kernel_irq_routing_entry *new);
 bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *,
 				  struct kvm_kernel_irq_routing_entry *);
 #endif /* CONFIG_HAVE_KVM_IRQ_BYPASS */
virt/kvm/eventfd.c +5 -6
index 11e5d1e3f12e..85581550dc8d 100644
@@ -285,9 +285,9 @@ void __attribute__((weak)) kvm_arch_irq_bypass_start(
 {
 }
 
-int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
-				struct kvm *kvm, unsigned int host_irq,
-				uint32_t guest_irq, bool set)
+int __weak kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
+					 struct kvm_kernel_irq_routing_entry *old,
+					 struct kvm_kernel_irq_routing_entry *new)
 {
 	return 0;
 }
@@ -619,9 +619,8 @@ void kvm_irq_routing_update(struct kvm *kvm)
 #if IS_ENABLED(CONFIG_HAVE_KVM_IRQ_BYPASS)
 		if (irqfd->producer &&
 		    kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry)) {
-			int ret = kvm_arch_update_irqfd_routing(
-					irqfd->kvm, irqfd->producer->irq,
-					irqfd->gsi, 1);
+			int ret = kvm_arch_update_irqfd_routing(irqfd, &old, &irqfd->irq_entry);
+
 			WARN_ON(ret);
 		}
 #endif
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#16 lore raw

PATCH v3 4/62 KVM: SVM: Track per-vCPU IRTEs using kvm_kernel_irqfd structure

In reply to Sean Christopherson

Track the IRTEs that are posting to an SVM vCPU via the associated irqfd structure and GSI routing instead of dynamically allocating a separate data structure. In addition to eliminating an atomic allocation, this will allow hoisting much of the IRTE update logic to common x86.

---
 arch/x86/kvm/svm/avic.c   | 71 +++++++++++++++------------------------
 arch/x86/kvm/svm/svm.h    | 10 +++---
 include/linux/kvm_irqfd.h |  3 ++
 3 files changed, 36 insertions(+), 48 deletions(-)
arch/x86/kvm/svm/avic.c +27 -44
index adacf00d6664..d33c01379421 100644
@@ -75,14 +75,6 @@ static bool next_vm_id_wrapped = 0;
 static DEFINE_SPINLOCK(svm_vm_data_hash_lock);
 bool x2avic_enabled;
 
-/*
- * This is a wrapper of struct amd_iommu_ir_data.
- */
-struct amd_svm_iommu_ir {
-	struct list_head node;	/* Used by SVM for per-vcpu ir_list */
-	void *data;		/* Storing pointer to struct amd_ir_data */
-};
-
 static void avic_activate_vmcb(struct vcpu_svm *svm)
 {
 	struct vmcb *vmcb = svm->vmcb01.ptr;
@@ -746,8 +738,8 @@ static int avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
 {
 	int ret = 0;
 	unsigned long flags;
-	struct amd_svm_iommu_ir *ir;
 	struct vcpu_svm *svm = to_svm(vcpu);
+	struct kvm_kernel_irqfd *irqfd;
 
 	if (!kvm_arch_has_assigned_device(vcpu->kvm))
 		return 0;
@@ -761,11 +753,11 @@ static int avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
 	if (list_empty(&svm->ir_list))
 		goto out;
 
-	list_for_each_entry(ir, &svm->ir_list, node) {
+	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list) {
 		if (activate)
-			ret = amd_iommu_activate_guest_mode(ir->data);
+			ret = amd_iommu_activate_guest_mode(irqfd->irq_bypass_data);
 		else
-			ret = amd_iommu_deactivate_guest_mode(ir->data);
+			ret = amd_iommu_deactivate_guest_mode(irqfd->irq_bypass_data);
 		if (ret)
 			break;
 	}
@@ -774,27 +766,30 @@ static int avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
 	return ret;
 }
 
-static void svm_ir_list_del(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
+static void svm_ir_list_del(struct vcpu_svm *svm,
+			    struct kvm_kernel_irqfd *irqfd,
+			    struct amd_iommu_pi_data *pi)
 {
 	unsigned long flags;
-	struct amd_svm_iommu_ir *cur;
+	struct kvm_kernel_irqfd *cur;
 
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
-	list_for_each_entry(cur, &svm->ir_list, node) {
-		if (cur->data != pi->ir_data)
+	list_for_each_entry(cur, &svm->ir_list, vcpu_list) {
+		if (cur->irq_bypass_data != pi->ir_data)
 			continue;
-		list_del(&cur->node);
-		kfree(cur);
+		if (WARN_ON_ONCE(cur != irqfd))
+			continue;
+		list_del(&irqfd->vcpu_list);
 		break;
 	}
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
 }
 
-static int svm_ir_list_add(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
+static int svm_ir_list_add(struct vcpu_svm *svm,
+			   struct kvm_kernel_irqfd *irqfd,
+			   struct amd_iommu_pi_data *pi)
 {
-	int ret = 0;
 	unsigned long flags;
-	struct amd_svm_iommu_ir *ir;
 	u64 entry;
 
 	if (WARN_ON_ONCE(!pi->ir_data))
@@ -811,25 +806,14 @@ static int svm_ir_list_add(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
 		struct kvm_vcpu *prev_vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
 		struct vcpu_svm *prev_svm;
 
-		if (!prev_vcpu) {
-			ret = -EINVAL;
-			goto out;
-		}
+		if (!prev_vcpu)
+			return -EINVAL;
 
 		prev_svm = to_svm(prev_vcpu);
-		svm_ir_list_del(prev_svm, pi);
+		svm_ir_list_del(prev_svm, irqfd, pi);
 	}
 
-	/**
-	 * Allocating new amd_iommu_pi_data, which will get
-	 * add to the per-vcpu ir_list.
-	 */
-	ir = kzalloc(sizeof(struct amd_svm_iommu_ir), GFP_ATOMIC | __GFP_ACCOUNT);
-	if (!ir) {
-		ret = -ENOMEM;
-		goto out;
-	}
-	ir->data = pi->ir_data;
+	irqfd->irq_bypass_data = pi->ir_data;
 
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
 
@@ -844,10 +828,9 @@ static int svm_ir_list_add(struct vcpu_svm *svm, struct amd_iommu_pi_data *pi)
 		amd_iommu_update_ga(entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK,
 				    true, pi->ir_data);
 
-	list_add(&ir->node, &svm->ir_list);
+	list_add(&irqfd->vcpu_list, &svm->ir_list);
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
-out:
-	return ret;
+	return 0;
 }
 
 /*
@@ -951,7 +934,7 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			 * scheduling information in IOMMU irte.
 			 */
 			if (!ret && pi.is_guest_mode)
-				svm_ir_list_add(svm, &pi);
+				svm_ir_list_add(svm, irqfd, &pi);
 		}
 
 		if (!ret && svm) {
@@ -992,7 +975,7 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 
 			vcpu = kvm_get_vcpu_by_id(kvm, id);
 			if (vcpu)
-				svm_ir_list_del(to_svm(vcpu), &pi);
+				svm_ir_list_del(to_svm(vcpu), irqfd, &pi);
 		}
 	}
 out:
@@ -1004,8 +987,8 @@ static inline int
 avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu, bool r)
 {
 	int ret = 0;
-	struct amd_svm_iommu_ir *ir;
 	struct vcpu_svm *svm = to_svm(vcpu);
+	struct kvm_kernel_irqfd *irqfd;
 
 	lockdep_assert_held(&svm->ir_list_lock);
 
@@ -1019,8 +1002,8 @@ avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu, bool r)
 	if (list_empty(&svm->ir_list))
 		return 0;
 
-	list_for_each_entry(ir, &svm->ir_list, node) {
-		ret = amd_iommu_update_ga(cpu, r, ir->data);
+	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list) {
+		ret = amd_iommu_update_ga(cpu, r, irqfd->irq_bypass_data);
 		if (ret)
 			return ret;
 	}
arch/x86/kvm/svm/svm.h +6 -4
index b35fce30d923..cc27877d69ae 100644
@@ -310,10 +310,12 @@ struct vcpu_svm {
 	u64 *avic_physical_id_cache;
 
 	/*
-	 * Per-vcpu list of struct amd_svm_iommu_ir:
-	 * This is used mainly to store interrupt remapping information used
-	 * when update the vcpu affinity. This avoids the need to scan for
-	 * IRTE and try to match ga_tag in the IOMMU driver.
+	 * Per-vCPU list of irqfds that are eligible to post IRQs directly to
+	 * the vCPU (a.k.a. device posted IRQs, a.k.a. IRQ bypass).  The list
+	 * is used to reconfigure IRTEs when the vCPU is loaded/put (to set the
+	 * target pCPU), when AVIC is toggled on/off (to (de)activate bypass),
+	 * and if the irqfd becomes ineligible for posting (to put the IRTE
+	 * back into remapped mode).
 	 */
 	struct list_head ir_list;
 	spinlock_t ir_list_lock;
include/linux/kvm_irqfd.h +3 -0
index 8ad43692e3bb..6510a48e62aa 100644
@@ -59,6 +59,9 @@ struct kvm_kernel_irqfd {
 	struct work_struct shutdown;
 	struct irq_bypass_consumer consumer;
 	struct irq_bypass_producer *producer;
+
+	struct list_head vcpu_list;
+	void *irq_bypass_data;
 };
 
 #endif /* __LINUX_KVM_IRQFD_H */
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#17 lore raw

PATCH v3 5/62 KVM: SVM: Delete IRTE link from previous vCPU before setting new IRTE

In reply to Sean Christopherson

Delete the previous per-vCPU IRTE link prior to modifying the IRTE. If forcing the IRTE back to remapped mode fails, the IRQ is already broken; keeping stale metadata won't change that, and the IOMMU should be sufficiently paranoid to sanitize the IRTE when the IRQ is freed and reallocated.

This will allow hoisting the vCPU tracking to common x86, which in turn will allow most of the IRTE update code to be deduplicated.

---
 arch/x86/kvm/svm/avic.c   | 60 +++++++++------------------------------
 include/linux/kvm_irqfd.h |  1 +
 2 files changed, 14 insertions(+), 47 deletions(-)
arch/x86/kvm/svm/avic.c +13 -47
index d33c01379421..ed7374f0bd5a 100644
@@ -766,23 +766,19 @@ static int avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
 	return ret;
 }
 
-static void svm_ir_list_del(struct vcpu_svm *svm,
-			    struct kvm_kernel_irqfd *irqfd,
-			    struct amd_iommu_pi_data *pi)
+static void svm_ir_list_del(struct kvm_kernel_irqfd *irqfd)
 {
+	struct kvm_vcpu *vcpu = irqfd->irq_bypass_vcpu;
 	unsigned long flags;
-	struct kvm_kernel_irqfd *cur;
 
-	spin_lock_irqsave(&svm->ir_list_lock, flags);
-	list_for_each_entry(cur, &svm->ir_list, vcpu_list) {
-		if (cur->irq_bypass_data != pi->ir_data)
-			continue;
-		if (WARN_ON_ONCE(cur != irqfd))
-			continue;
-		list_del(&irqfd->vcpu_list);
-		break;
-	}
-	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
+	if (!vcpu)
+		return;
+
+	spin_lock_irqsave(&to_svm(vcpu)->ir_list_lock, flags);
+	list_del(&irqfd->vcpu_list);
+	spin_unlock_irqrestore(&to_svm(vcpu)->ir_list_lock, flags);
+
+	irqfd->irq_bypass_vcpu = NULL;
 }
 
 static int svm_ir_list_add(struct vcpu_svm *svm,
@@ -795,24 +791,7 @@ static int svm_ir_list_add(struct vcpu_svm *svm,
 	if (WARN_ON_ONCE(!pi->ir_data))
 		return -EINVAL;
 
-	/**
-	 * In some cases, the existing irte is updated and re-set,
-	 * so we need to check here if it's already been * added
-	 * to the ir_list.
-	 */
-	if (pi->prev_ga_tag) {
-		struct kvm *kvm = svm->vcpu.kvm;
-		u32 vcpu_id = AVIC_GATAG_TO_VCPUID(pi->prev_ga_tag);
-		struct kvm_vcpu *prev_vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
-		struct vcpu_svm *prev_svm;
-
-		if (!prev_vcpu)
-			return -EINVAL;
-
-		prev_svm = to_svm(prev_vcpu);
-		svm_ir_list_del(prev_svm, irqfd, pi);
-	}
-
+	irqfd->irq_bypass_vcpu = &svm->vcpu;
 	irqfd->irq_bypass_data = pi->ir_data;
 
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
@@ -904,6 +883,8 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 
 		WARN_ON_ONCE(new && memcmp(e, new, sizeof(*new)));
 
+		svm_ir_list_del(irqfd);
+
 		/**
 		 * Here, we setup with legacy mode in the following cases:
 		 * 1. When cannot target interrupt to a specific vcpu.
@@ -962,21 +943,6 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		pi.prev_ga_tag = 0;
 		pi.is_guest_mode = false;
 		ret = irq_set_vcpu_affinity(host_irq, &pi);
-
-		/**
-		 * Check if the posted interrupt was previously
-		 * setup with the guest_mode by checking if the ga_tag
-		 * was cached. If so, we need to clean up the per-vcpu
-		 * ir_list.
-		 */
-		if (!ret && pi.prev_ga_tag) {
-			int id = AVIC_GATAG_TO_VCPUID(pi.prev_ga_tag);
-			struct kvm_vcpu *vcpu;
-
-			vcpu = kvm_get_vcpu_by_id(kvm, id);
-			if (vcpu)
-				svm_ir_list_del(to_svm(vcpu), irqfd, &pi);
-		}
 	}
 out:
 	srcu_read_unlock(&kvm->irq_srcu, idx);
include/linux/kvm_irqfd.h +1 -0
index 6510a48e62aa..361c07f4466d 100644
@@ -60,6 +60,7 @@ struct kvm_kernel_irqfd {
 	struct irq_bypass_consumer consumer;
 	struct irq_bypass_producer *producer;
 
+	struct kvm_vcpu *irq_bypass_vcpu;
 	struct list_head vcpu_list;
 	void *irq_bypass_data;
 };
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#18 lore raw

PATCH v3 6/62 iommu/amd: KVM: SVM: Delete now-unused cached/previous GA tag fields

In reply to Sean Christopherson

Delete the amd_ir_data.prev_ga_tag field now that all usage is superfluous.

---
 arch/x86/kvm/svm/avic.c             |  2 --
 drivers/iommu/amd/amd_iommu_types.h |  1 -
 drivers/iommu/amd/iommu.c           | 10 ----------
 include/linux/amd-iommu.h           |  2 +-
 4 files changed, 1 insertion(+), 14 deletions(-)
arch/x86/kvm/svm/avic.c +0 -2
index ed7374f0bd5a..4e8380d2f017 100644
@@ -938,9 +938,7 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		/**
 		 * Here, pi is used to:
 		 * - Tell IOMMU to use legacy mode for this interrupt.
-		 * - Retrieve ga_tag of prior interrupt remapping data.
 		 */
-		pi.prev_ga_tag = 0;
 		pi.is_guest_mode = false;
 		ret = irq_set_vcpu_affinity(host_irq, &pi);
 	}
drivers/iommu/amd/amd_iommu_types.h +0 -1
index 5089b58e528a..57a96f3e7b84 100644
@@ -1060,7 +1060,6 @@ struct irq_2_irte {
 };
 
 struct amd_ir_data {
-	u32 cached_ga_tag;
 	struct amd_iommu *iommu;
 	struct irq_2_irte irq_2_irte;
 	struct msi_msg msi_entry;
drivers/iommu/amd/iommu.c +0 -10
index f34209b08b4c..f23635b062f0 100644
@@ -3887,23 +3887,13 @@ static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
 	ir_data->cfg = irqd_cfg(data);
 	pi_data->ir_data = ir_data;
 
-	pi_data->prev_ga_tag = ir_data->cached_ga_tag;
 	if (pi_data->is_guest_mode) {
 		ir_data->ga_root_ptr = (pi_data->base >> 12);
 		ir_data->ga_vector = vcpu_pi_info->vector;
 		ir_data->ga_tag = pi_data->ga_tag;
 		ret = amd_iommu_activate_guest_mode(ir_data);
-		if (!ret)
-			ir_data->cached_ga_tag = pi_data->ga_tag;
 	} else {
 		ret = amd_iommu_deactivate_guest_mode(ir_data);
-
-		/*
-		 * This communicates the ga_tag back to the caller
-		 * so that it can do all the necessary clean up.
-		 */
-		if (!ret)
-			ir_data->cached_ga_tag = 0;
 	}
 
 	return ret;
include/linux/amd-iommu.h +1 -1
index 062fbd4c9b77..1f9b13d803c5 100644
@@ -19,8 +19,8 @@ struct amd_iommu;
  */
 struct amd_iommu_pi_data {
 	u32 ga_tag;
-	u32 prev_ga_tag;
 	u64 base;
+
 	bool is_guest_mode;
 	struct vcpu_data *vcpu_data;
 	void *ir_data;
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#19 lore raw

PATCH v3 7/62 KVM: SVM: Delete IRTE link from previous vCPU irrespective of new routing

In reply to Sean Christopherson

Delete the IRTE link from the previous vCPU irrespective of the new routing state, i.e. even if the IRTE won't be configured to post IRQs to a vCPU. Whether or not the new route is postable as no bearing on the *old* route. Failure to delete the link can result in KVM incorrectly updating the IRTE, e.g. if the "old" vCPU is scheduled in/out.

Fixes 411b44ba80ab ("svm: Implements update_pi_irte hook to setup posted interrupt")
---
 arch/x86/kvm/svm/avic.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
arch/x86/kvm/svm/avic.c +6 -2
index 4e8380d2f017..c981ce764b45 100644
@@ -861,6 +861,12 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 	if (!kvm_arch_has_assigned_device(kvm) || !kvm_arch_has_irq_bypass())
 		return 0;
 
+	/*
+	 * If the IRQ was affined to a different vCPU, remove the IRTE metadata
+	 * from the *previous* vCPU's list.
+	 */
+	svm_ir_list_del(irqfd);
+
 	pr_debug("SVM: %s: host_irq=%#x, guest_irq=%#x, set=%#x\n",
 		 __func__, host_irq, guest_irq, set);
 
@@ -883,8 +889,6 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 
 		WARN_ON_ONCE(new && memcmp(e, new, sizeof(*new)));
 
-		svm_ir_list_del(irqfd);
-
 		/**
 		 * Here, we setup with legacy mode in the following cases:
 		 * 1. When cannot target interrupt to a specific vcpu.
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#20 lore raw

PATCH v3 8/62 KVM: SVM: Drop pointless masking of default APIC base when setting V_APIC_BAR

In reply to Sean Christopherson

Drop VMCB_AVIC_APIC_BAR_MASK, it's just a regurgitation of the maximum theoretical 4KiB-aligned physical address, i.e. is not novel in any way, and its only usage is to mask the default APIC base, which is 4KiB aligned and (obviously) a legal physical address.

No functional change intended.

---
 arch/x86/include/asm/svm.h | 2 --
 arch/x86/kvm/svm/avic.c    | 2 +-
 2 files changed, 1 insertion(+), 3 deletions(-)
arch/x86/include/asm/svm.h +0 -2
index ad954a1a6656..89a666952b01 100644
@@ -260,8 +260,6 @@ struct __attribute__ ((__packed__)) vmcb_control_area {
 
 #define AVIC_DOORBELL_PHYSICAL_ID_MASK			GENMASK_ULL(11, 0)
 
-#define VMCB_AVIC_APIC_BAR_MASK				0xFFFFFFFFFF000ULL
-
 #define AVIC_UNACCEL_ACCESS_WRITE_MASK		1
 #define AVIC_UNACCEL_ACCESS_OFFSET_MASK		0xFF0
 #define AVIC_UNACCEL_ACCESS_VECTOR_MASK		0xFFFFFFFF
arch/x86/kvm/svm/avic.c +1 -1
index c981ce764b45..5344ae76c590 100644
@@ -244,7 +244,7 @@ void avic_init_vmcb(struct vcpu_svm *svm, struct vmcb *vmcb)
 	vmcb->control.avic_backing_page = bpa & AVIC_HPA_MASK;
 	vmcb->control.avic_logical_id = lpa & AVIC_HPA_MASK;
 	vmcb->control.avic_physical_id = ppa & AVIC_HPA_MASK;
-	vmcb->control.avic_vapic_bar = APIC_DEFAULT_PHYS_BASE & VMCB_AVIC_APIC_BAR_MASK;
+	vmcb->control.avic_vapic_bar = APIC_DEFAULT_PHYS_BASE;
 
 	if (kvm_apicv_activated(svm->vcpu.kvm))
 		avic_activate_vmcb(svm);
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
Reviewed-by on v3 8/62
#21 lore raw

Re: [PATCH v3 08/62] KVM: SVM: Drop pointless masking of default APIC base when setting V_APIC_BAR

Quoted 5 lines · Sean Christopherson

Drop VMCB_AVIC_APIC_BAR_MASK, it's just a regurgitation of the maximum theoretical 4KiB-aligned physical address, i.e. is not novel in any way, and its only usage is to mask the default APIC base, which is 4KiB aligned and (obviously) a legal physical address.

Quoted 8 lines

No functional change intended.

---
 arch/x86/include/asm/svm.h | 2 --
 arch/x86/kvm/svm/avic.c    | 2 +-
 2 files changed, 1 insertion(+), 3 deletions(-)

- Naveen

author maintainer
#22 lore raw

PATCH v3 9/62 KVM: SVM: Drop pointless masking of kernel page pa's with AVIC HPA masks

In reply to Sean Christopherson

Drop AVIC_HPA_MASK and all its users, the mask is just the 4KiB-aligned maximum theoretical physical address for x86-64 CPUs, as x86-64 is currently defined (going beyond PA52 would require an entirely new paging mode, which would arguably create a new, different architecture).

All usage in KVM masks the result of page_to_phys(), which on x86-64 is guaranteed to be 4KiB aligned and a legal physical address; if either of those requirements doesn't hold true, KVM has far bigger problems.

Drop masking the avic_backing_page with AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK for all the same reasons, but keep the macro even though it's unused in functional code. It's a distinct architectural define, and having the definition in software helps visualize the layout of an entry. And to be hyper-paranoid about MAXPA going beyond 52, add a compile-time assert to ensure the kernel's maximum supported physical address stays in bounds.

The unnecessary masking in avic_init_vmcb() also incorrectly assumes that SME's C-bit resides between bits 51:11; that holds true for current CPUs, but isn't required by AMD's architecture:

In some implementations, the bit used may be a physical address bit

Key word being "may".

Opportunistically use the GENMASK_ULL() version for AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK, which is far more readable than a set of repeating Fs.

---
 arch/x86/include/asm/svm.h |  4 +---
 arch/x86/kvm/svm/avic.c    | 18 ++++++++++--------
 2 files changed, 11 insertions(+), 11 deletions(-)
arch/x86/include/asm/svm.h +1 -3
index 89a666952b01..36f67c69ea66 100644
@@ -253,7 +253,7 @@ struct __attribute__ ((__packed__)) vmcb_control_area {
 #define AVIC_LOGICAL_ID_ENTRY_VALID_MASK		(1 << 31)
 
 #define AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK	GENMASK_ULL(11, 0)
-#define AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK	(0xFFFFFFFFFFULL << 12)
+#define AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK	GENMASK_ULL(51, 12)
 #define AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK		(1ULL << 62)
 #define AVIC_PHYSICAL_ID_ENTRY_VALID_MASK		(1ULL << 63)
 #define AVIC_PHYSICAL_ID_TABLE_SIZE_MASK		(0xFFULL)
@@ -288,8 +288,6 @@ enum avic_ipi_failure_cause {
 static_assert((AVIC_MAX_PHYSICAL_ID & AVIC_PHYSICAL_MAX_INDEX_MASK) == AVIC_MAX_PHYSICAL_ID);
 static_assert((X2AVIC_MAX_PHYSICAL_ID & AVIC_PHYSICAL_MAX_INDEX_MASK) == X2AVIC_MAX_PHYSICAL_ID);
 
-#define AVIC_HPA_MASK	~((0xFFFULL << 52) | 0xFFF)
-
 #define SVM_SEV_FEAT_SNP_ACTIVE				BIT(0)
 #define SVM_SEV_FEAT_RESTRICTED_INJECTION		BIT(3)
 #define SVM_SEV_FEAT_ALTERNATE_INJECTION		BIT(4)
arch/x86/kvm/svm/avic.c +10 -8
index 5344ae76c590..4b882148f2c0 100644
@@ -241,9 +241,9 @@ void avic_init_vmcb(struct vcpu_svm *svm, struct vmcb *vmcb)
 	phys_addr_t lpa = __sme_set(page_to_phys(kvm_svm->avic_logical_id_table_page));
 	phys_addr_t ppa = __sme_set(page_to_phys(kvm_svm->avic_physical_id_table_page));
 
-	vmcb->control.avic_backing_page = bpa & AVIC_HPA_MASK;
-	vmcb->control.avic_logical_id = lpa & AVIC_HPA_MASK;
-	vmcb->control.avic_physical_id = ppa & AVIC_HPA_MASK;
+	vmcb->control.avic_backing_page = bpa;
+	vmcb->control.avic_logical_id = lpa;
+	vmcb->control.avic_physical_id = ppa;
 	vmcb->control.avic_vapic_bar = APIC_DEFAULT_PHYS_BASE;
 
 	if (kvm_apicv_activated(svm->vcpu.kvm))
@@ -301,9 +301,12 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	if (!entry)
 		return -EINVAL;
 
-	new_entry = __sme_set((page_to_phys(svm->avic_backing_page) &
-			      AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK) |
-			      AVIC_PHYSICAL_ID_ENTRY_VALID_MASK);
+	/* Note, fls64() returns the bit position, +1. */
+	BUILD_BUG_ON(__PHYSICAL_MASK_SHIFT >
+		     fls64(AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK));
+
+	new_entry = __sme_set(page_to_phys(svm->avic_backing_page)) |
+		    AVIC_PHYSICAL_ID_ENTRY_VALID_MASK;
 	WRITE_ONCE(*entry, new_entry);
 
 	svm->avic_physical_id_cache = entry;
@@ -903,8 +906,7 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			enable_remapped_mode = false;
 
 			/* Try to enable guest_mode in IRTE */
-			pi.base = __sme_set(page_to_phys(svm->avic_backing_page) &
-					    AVIC_HPA_MASK);
+			pi.base = __sme_set(page_to_phys(svm->avic_backing_page));
 			pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
 						     svm->vcpu.vcpu_id);
 			pi.is_guest_mode = true;
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
Reviewed-by on v3 9/62
#23 lore raw

Re: [PATCH v3 09/62] KVM: SVM: Drop pointless masking of kernel page pa's with AVIC HPA masks

Quoted 1 line

KVM: SVM: Drop pointless masking of kernel page pa's with AVIC HPA masks

				      ^^^^^^^^^^^^^^^^

Just "physical addresses", or perhaps "HPAs"?

Quoted 21 lines · Sean Christopherson

Drop AVIC_HPA_MASK and all its users, the mask is just the 4KiB-aligned maximum theoretical physical address for x86-64 CPUs, as x86-64 is currently defined (going beyond PA52 would require an entirely new paging mode, which would arguably create a new, different architecture).

All usage in KVM masks the result of page_to_phys(), which on x86-64 is guaranteed to be 4KiB aligned and a legal physical address; if either of those requirements doesn't hold true, KVM has far bigger problems.

Drop masking the avic_backing_page with AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK for all the same reasons, but keep the macro even though it's unused in functional code. It's a distinct architectural define, and having the definition in software helps visualize the layout of an entry. And to be hyper-paranoid about MAXPA going beyond 52, add a compile-time assert to ensure the kernel's maximum supported physical address stays in bounds.

The unnecessary masking in avic_init_vmcb() also incorrectly assumes that SME's C-bit resides between bits 51:11; that holds true for current CPUs, but isn't required by AMD's architecture:

Quoted 7 lines

In some implementations, the bit used may be a physical address bit

Key word being "may".

Opportunistically use the GENMASK_ULL() version for AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK, which is far more readable than a set of repeating Fs.

Makes sense.

- Naveen

author maintainer
#24 lore raw

PATCH v3 10/62 KVM: SVM: Add helper to deduplicate code for getting AVIC backing page

In reply to Sean Christopherson

Add a helper to get the physical address of the AVIC backing page, both to deduplicate code and to prepare for getting the address directly from apic->regs, at which point it won't be all that obvious that the address in question is what SVM calls the AVIC backing page.

No functional change intended.

---
 arch/x86/kvm/svm/avic.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)
arch/x86/kvm/svm/avic.c +9 -5
index 4b882148f2c0..c36f7db9252e 100644
@@ -234,14 +234,18 @@ int avic_vm_init(struct kvm *kvm)
 	return err;
 }
 
+static phys_addr_t avic_get_backing_page_address(struct vcpu_svm *svm)
+{
+	return __sme_set(page_to_phys(svm->avic_backing_page));
+}
+
 void avic_init_vmcb(struct vcpu_svm *svm, struct vmcb *vmcb)
 {
 	struct kvm_svm *kvm_svm = to_kvm_svm(svm->vcpu.kvm);
-	phys_addr_t bpa = __sme_set(page_to_phys(svm->avic_backing_page));
 	phys_addr_t lpa = __sme_set(page_to_phys(kvm_svm->avic_logical_id_table_page));
 	phys_addr_t ppa = __sme_set(page_to_phys(kvm_svm->avic_physical_id_table_page));
 
-	vmcb->control.avic_backing_page = bpa;
+	vmcb->control.avic_backing_page = avic_get_backing_page_address(svm);
 	vmcb->control.avic_logical_id = lpa;
 	vmcb->control.avic_physical_id = ppa;
 	vmcb->control.avic_vapic_bar = APIC_DEFAULT_PHYS_BASE;
@@ -305,7 +309,7 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	BUILD_BUG_ON(__PHYSICAL_MASK_SHIFT >
 		     fls64(AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK));
 
-	new_entry = __sme_set(page_to_phys(svm->avic_backing_page)) |
+	new_entry = avic_get_backing_page_address(svm) |
 		    AVIC_PHYSICAL_ID_ENTRY_VALID_MASK;
 	WRITE_ONCE(*entry, new_entry);
 
@@ -845,7 +849,7 @@ get_pi_vcpu_info(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
 	pr_debug("SVM: %s: use GA mode for irq %u\n", __func__,
 		 irq.vector);
 	*svm = to_svm(vcpu);
-	vcpu_info->pi_desc_addr = __sme_set(page_to_phys((*svm)->avic_backing_page));
+	vcpu_info->pi_desc_addr = avic_get_backing_page_address(*svm);
 	vcpu_info->vector = irq.vector;
 
 	return 0;
@@ -906,7 +910,7 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			enable_remapped_mode = false;
 
 			/* Try to enable guest_mode in IRTE */
-			pi.base = __sme_set(page_to_phys(svm->avic_backing_page));
+			pi.base = avic_get_backing_page_address(svm);
 			pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
 						     svm->vcpu.vcpu_id);
 			pi.is_guest_mode = true;
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
Reviewed-by on v3 10/62
#25 lore raw

Re: [PATCH v3 10/62] KVM: SVM: Add helper to deduplicate code for getting AVIC backing page

Quoted 5 lines · Sean Christopherson

Add a helper to get the physical address of the AVIC backing page, both to deduplicate code and to prepare for getting the address directly from apic->regs, at which point it won't be all that obvious that the address in question is what SVM calls the AVIC backing page.

Quoted 8 lines

No functional change intended.

---
 arch/x86/kvm/svm/avic.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)

LGTM.

- Naveen

author maintainer
#26 lore raw

PATCH v3 11/62 KVM: SVM: Drop vcpu_svm's pointless avic_backing_page field

In reply to Sean Christopherson

Drop vcpu_svm's avic_backing_page pointer and instead grab the physical address of KVM's vAPIC page directly from the source. Getting a physical address from a kernel virtual address is not an expensive operation, and getting the physical address from a struct page is *more* expensive for CONFIG_SPARSEMEM=y kernels. Regardless, none of the paths that consume the address are hot paths, i.e. shaving cycles is not a priority.

Eliminating the "cache" means KVM doesn't have to worry about the cache being invalid, which will simplify a future fix when dealing with vCPU IDs that are too big.

WARN if KVM attempts to allocate a vCPU's AVIC backing page without an in-kernel local APIC. avic_init_vcpu() bails early if the APIC is not in-kernel, and KVM disallows enabling an in-kernel APIC after vCPUs have been created, i.e. it should be impossible to reach avic_init_backing_page() without the vAPIC being allocated.

---
 arch/x86/kvm/svm/avic.c | 6 ++----
 arch/x86/kvm/svm/svm.h  | 1 -
 2 files changed, 2 insertions(+), 5 deletions(-)
arch/x86/kvm/svm/avic.c +2 -4
index c36f7db9252e..ab228872a19b 100644
@@ -236,7 +236,7 @@ int avic_vm_init(struct kvm *kvm)
 
 static phys_addr_t avic_get_backing_page_address(struct vcpu_svm *svm)
 {
-	return __sme_set(page_to_phys(svm->avic_backing_page));
+	return __sme_set(__pa(svm->vcpu.arch.apic->regs));
 }
 
 void avic_init_vmcb(struct vcpu_svm *svm, struct vmcb *vmcb)
@@ -281,7 +281,7 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	    (id > X2AVIC_MAX_PHYSICAL_ID))
 		return -EINVAL;
 
-	if (!vcpu->arch.apic->regs)
+	if (WARN_ON_ONCE(!vcpu->arch.apic->regs))
 		return -EINVAL;
 
 	if (kvm_apicv_activated(vcpu->kvm)) {
@@ -298,8 +298,6 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 			return ret;
 	}
 
-	svm->avic_backing_page = virt_to_page(vcpu->arch.apic->regs);
-
 	/* Setting AVIC backing page address in the phy APIC ID table */
 	entry = avic_get_physical_id_entry(vcpu, id);
 	if (!entry)
arch/x86/kvm/svm/svm.h +0 -1
index cc27877d69ae..1585288200f4 100644
@@ -306,7 +306,6 @@ struct vcpu_svm {
 
 	u32 ldr_reg;
 	u32 dfr_reg;
-	struct page *avic_backing_page;
 	u64 *avic_physical_id_cache;
 
 	/*
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
Reviewed-by on v3 11/62
#27 lore raw

Re: [PATCH v3 11/62] KVM: SVM: Drop vcpu_svm's pointless avic_backing_page field

Quoted 17 lines · Sean Christopherson

Drop vcpu_svm's avic_backing_page pointer and instead grab the physical address of KVM's vAPIC page directly from the source. Getting a physical address from a kernel virtual address is not an expensive operation, and getting the physical address from a struct page is *more* expensive for CONFIG_SPARSEMEM=y kernels. Regardless, none of the paths that consume the address are hot paths, i.e. shaving cycles is not a priority.

Eliminating the "cache" means KVM doesn't have to worry about the cache being invalid, which will simplify a future fix when dealing with vCPU IDs that are too big.

WARN if KVM attempts to allocate a vCPU's AVIC backing page without an in-kernel local APIC. avic_init_vcpu() bails early if the APIC is not in-kernel, and KVM disallows enabling an in-kernel APIC after vCPUs have been created, i.e. it should be impossible to reach avic_init_backing_page() without the vAPIC being allocated.

Quoted 6 lines
---
 arch/x86/kvm/svm/avic.c | 6 ++----
 arch/x86/kvm/svm/svm.h  | 1 -
 2 files changed, 2 insertions(+), 5 deletions(-)

- Naveen

author maintainer
#28 lore raw

PATCH v3 12/62 KVM: SVM: Inhibit AVIC if ID is too big instead of rejecting vCPU creation

In reply to Sean Christopherson

Inhibit AVIC with a new "ID too big" flag if userspace creates a vCPU with an ID that is too big, but otherwise allow vCPU creation to succeed. Rejecting KVM_CREATE_VCPU with EINVAL violates KVM's ABI as KVM advertises that the max vCPU ID is 4095, but disallows creating vCPUs with IDs bigger than 254 (AVIC) or 511 (x2AVIC).

Alternatively, KVM could advertise an accurate value depending on which AVIC mode is in use, but that wouldn't really solve the underlying problem, e.g. would be a breaking change if KVM were to ever try and enable AVIC or x2AVIC by default.

---
 arch/x86/include/asm/kvm_host.h |  9 ++++++++-
 arch/x86/kvm/svm/avic.c         | 14 ++++++++++++--
 arch/x86/kvm/svm/svm.h          |  3 ++-
 3 files changed, 22 insertions(+), 4 deletions(-)
arch/x86/include/asm/kvm_host.h +8 -1
index 2a6ef1398da7..a9b709db7c59 100644
@@ -1314,6 +1314,12 @@ enum kvm_apicv_inhibit {
 	 */
 	APICV_INHIBIT_REASON_LOGICAL_ID_ALIASED,
 
+	/*
+	 * AVIC is disabled because the vCPU's APIC ID is beyond the max
+	 * supported by AVIC/x2AVIC, i.e. the vCPU is unaddressable.
+	 */
+	APICV_INHIBIT_REASON_PHYSICAL_ID_TOO_BIG,
+
 	NR_APICV_INHIBIT_REASONS,
 };
 
@@ -1332,7 +1338,8 @@ enum kvm_apicv_inhibit {
 	__APICV_INHIBIT_REASON(IRQWIN),			\
 	__APICV_INHIBIT_REASON(PIT_REINJ),		\
 	__APICV_INHIBIT_REASON(SEV),			\
-	__APICV_INHIBIT_REASON(LOGICAL_ID_ALIASED)
+	__APICV_INHIBIT_REASON(LOGICAL_ID_ALIASED),	\
+	__APICV_INHIBIT_REASON(PHYSICAL_ID_TOO_BIG)
 
 struct kvm_arch {
 	unsigned long n_used_mmu_pages;
arch/x86/kvm/svm/avic.c +12 -2
index ab228872a19b..f0a74b102c57 100644
@@ -277,9 +277,19 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	int id = vcpu->vcpu_id;
 	struct vcpu_svm *svm = to_svm(vcpu);
 
+	/*
+	 * Inhibit AVIC if the vCPU ID is bigger than what is supported by AVIC
+	 * hardware.  Immediately clear apicv_active, i.e. don't wait until the
+	 * KVM_REQ_APICV_UPDATE request is processed on the first KVM_RUN, as
+	 * avic_vcpu_load() expects to be called if and only if the vCPU has
+	 * fully initialized AVIC.
+	 */
 	if ((!x2avic_enabled && id > AVIC_MAX_PHYSICAL_ID) ||
-	    (id > X2AVIC_MAX_PHYSICAL_ID))
-		return -EINVAL;
+	    (id > X2AVIC_MAX_PHYSICAL_ID)) {
+		kvm_set_apicv_inhibit(vcpu->kvm, APICV_INHIBIT_REASON_PHYSICAL_ID_TOO_BIG);
+		vcpu->arch.apic->apicv_active = false;
+		return 0;
+	}
 
 	if (WARN_ON_ONCE(!vcpu->arch.apic->regs))
 		return -EINVAL;
arch/x86/kvm/svm/svm.h +2 -1
index 1585288200f4..71e3c003580e 100644
@@ -722,7 +722,8 @@ extern struct kvm_x86_nested_ops svm_nested_ops;
 	BIT(APICV_INHIBIT_REASON_PHYSICAL_ID_ALIASED) |	\
 	BIT(APICV_INHIBIT_REASON_APIC_ID_MODIFIED) |	\
 	BIT(APICV_INHIBIT_REASON_APIC_BASE_MODIFIED) |	\
-	BIT(APICV_INHIBIT_REASON_LOGICAL_ID_ALIASED)	\
+	BIT(APICV_INHIBIT_REASON_LOGICAL_ID_ALIASED) |	\
+	BIT(APICV_INHIBIT_REASON_PHYSICAL_ID_TOO_BIG)	\
 )
 
 bool avic_hardware_setup(void);
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
#29 lore raw

Re: [PATCH v3 12/62] KVM: SVM: Inhibit AVIC if ID is too big instead of rejecting vCPU creation

Quoted 58 lines · Sean Christopherson

Inhibit AVIC with a new "ID too big" flag if userspace creates a vCPU with an ID that is too big, but otherwise allow vCPU creation to succeed. Rejecting KVM_CREATE_VCPU with EINVAL violates KVM's ABI as KVM advertises that the max vCPU ID is 4095, but disallows creating vCPUs with IDs bigger than 254 (AVIC) or 511 (x2AVIC).

Alternatively, KVM could advertise an accurate value depending on which AVIC mode is in use, but that wouldn't really solve the underlying problem, e.g. would be a breaking change if KVM were to ever try and enable AVIC or x2AVIC by default.

---
 arch/x86/include/asm/kvm_host.h |  9 ++++++++-
 arch/x86/kvm/svm/avic.c         | 14 ++++++++++++--
 arch/x86/kvm/svm/svm.h          |  3 ++-
 3 files changed, 22 insertions(+), 4 deletions(-)
arch/x86/include/asm/kvm_host.h +8 -1
index 2a6ef1398da7..a9b709db7c59 100644
@@ -1314,6 +1314,12 @@ enum kvm_apicv_inhibit {
 	 */
 	APICV_INHIBIT_REASON_LOGICAL_ID_ALIASED,
 
+	/*
+	 * AVIC is disabled because the vCPU's APIC ID is beyond the max
+	 * supported by AVIC/x2AVIC, i.e. the vCPU is unaddressable.
+	 */
+	APICV_INHIBIT_REASON_PHYSICAL_ID_TOO_BIG,
+
 	NR_APICV_INHIBIT_REASONS,
 };
 
@@ -1332,7 +1338,8 @@ enum kvm_apicv_inhibit {
 	__APICV_INHIBIT_REASON(IRQWIN),			\
 	__APICV_INHIBIT_REASON(PIT_REINJ),		\
 	__APICV_INHIBIT_REASON(SEV),			\
-	__APICV_INHIBIT_REASON(LOGICAL_ID_ALIASED)
+	__APICV_INHIBIT_REASON(LOGICAL_ID_ALIASED),	\
+	__APICV_INHIBIT_REASON(PHYSICAL_ID_TOO_BIG)
 
 struct kvm_arch {
 	unsigned long n_used_mmu_pages;
arch/x86/kvm/svm/avic.c +3 -0
index ab228872a19b..f0a74b102c57 100644
@@ -277,9 +277,19 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	int id = vcpu->vcpu_id;
 	struct vcpu_svm *svm = to_svm(vcpu);
 
+	/*
+	 * Inhibit AVIC if the vCPU ID is bigger than what is supported by AVIC
+	 * hardware.  Immediately clear apicv_active, i.e. don't wait until the

arch/x86/kvm/svm/avic.c static int avic_init_backing_page(struct kvm_vcpu *vcpu) · line 290

Quoted 10 lines
arch/x86/kvm/svm/avic.c +7 -2
+	 * KVM_REQ_APICV_UPDATE request is processed on the first KVM_RUN, as
+	 * avic_vcpu_load() expects to be called if and only if the vCPU has
+	 * fully initialized AVIC.
+	 */
 	if ((!x2avic_enabled && id > AVIC_MAX_PHYSICAL_ID) ||
-	    (id > X2AVIC_MAX_PHYSICAL_ID))
-		return -EINVAL;
+	    (id > X2AVIC_MAX_PHYSICAL_ID)) {
+		kvm_set_apicv_inhibit(vcpu->kvm, APICV_INHIBIT_REASON_PHYSICAL_ID_TOO_BIG);
+		vcpu->arch.apic->apicv_active = false;

This bothers me a bit. kvm_create_lapic() does this:

          if (enable_apicv) {
                  apic->apicv_active = true;
                  kvm_make_request(KVM_REQ_APICV_UPDATE, vcpu);
	  }

But, setting apic->apicv_active to false here means KVM_REQ_APICV_UPDATE is going to be a no-op.

This does not look to be a big deal given that kvm_create_lapic() itself is called just a bit before svm_vcpu_create() (which calls the above function through avic_init_vcpu()) in kvm_arch_vcpu_create(), so there isn't that much done before apicv_active is toggled.

But, this made me wonder if introducing a kvm_x86_op to check and enable/disable apic->apicv_active in kvm_create_lapic() might be cleaner overall. Maybe even have it be the initialization point for APICv:
apicv_init(), so we can invoke avic_init_vcpu() right away?

- Naveen

author maintainer
#30 lore raw

Re: [PATCH v3 12/62] KVM: SVM: Inhibit AVIC if ID is too big instead of rejecting vCPU creation

Quoted 24 lines · Naveen N Rao
Quoted 11 lines · Sean Christopherson
arch/x86/kvm/svm/avic.c +3 -0
index ab228872a19b..f0a74b102c57 100644
@@ -277,9 +277,19 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	int id = vcpu->vcpu_id;
 	struct vcpu_svm *svm = to_svm(vcpu);
 
+	/*
+	 * Inhibit AVIC if the vCPU ID is bigger than what is supported by AVIC
+	 * hardware.  Immediately clear apicv_active, i.e. don't wait until the

arch/x86/kvm/svm/avic.c static int avic_init_backing_page(struct kvm_vcpu *vcpu) · line 290

Quoted 10 lines
arch/x86/kvm/svm/avic.c +7 -2
+	 * KVM_REQ_APICV_UPDATE request is processed on the first KVM_RUN, as
+	 * avic_vcpu_load() expects to be called if and only if the vCPU has
+	 * fully initialized AVIC.
+	 */
 	if ((!x2avic_enabled && id > AVIC_MAX_PHYSICAL_ID) ||
-	    (id > X2AVIC_MAX_PHYSICAL_ID))
-		return -EINVAL;
+	    (id > X2AVIC_MAX_PHYSICAL_ID)) {
+		kvm_set_apicv_inhibit(vcpu->kvm, APICV_INHIBIT_REASON_PHYSICAL_ID_TOO_BIG);
+		vcpu->arch.apic->apicv_active = false;

This bothers me a bit. kvm_create_lapic() does this:

Quoted 7 lines
          if (enable_apicv) {
                  apic->apicv_active = true;
                  kvm_make_request(KVM_REQ_APICV_UPDATE, vcpu);
	  }

But, setting apic->apicv_active to false here means KVM_REQ_APICV_UPDATE is going to be a no-op.

That's fine, KVM_REQ_APICV_UPDATE is a nop in other scenarios, too. I agree it's not ideal, but this is a rather extreme edge case and a one-time slow path, so I don't think it's worth doing anything special just to avoid KVM_REQ_APICV_UPDATE.

Quoted 8 lines

This does not look to be a big deal given that kvm_create_lapic() itself is called just a bit before svm_vcpu_create() (which calls the above function through avic_init_vcpu()) in kvm_arch_vcpu_create(), so there isn't that much done before apicv_active is toggled.

But, this made me wonder if introducing a kvm_x86_op to check and enable/disable apic->apicv_active in kvm_create_lapic() might be cleaner overall

Hmm, yes and no. I completely agree that clearing apicv_active in avic.c is all kinds of gross, but clearing apic->apicv_active in lapic.c to handle this particular scenario is just as problematic, because then avic_init_backing_page() would need to check kvm_vcpu_apicv_active() to determine whether or not to allocate the backing page. In a way, that's even worse, because setting apic->apicv_active by default is purely an optimization, i.e. leaving it %false _should_ work as well, it would just be suboptimal. But if AVIC were to key off apic->apicv_active, that could lead to KVM incorrectly skipping allocation of the AVIC backing page.

Quoted 2 lines

Maybe even have it be the initialization point for APICv:
apicv_init(), so we can invoke avic_init_vcpu() right away?

I mostly like this idea (actually, I take that back; see below), but VMX throws a big wrench in things. Unlike SVM, VMX doesn't have a singular "enable APICv" flag. Rather, KVM considers "APICv" to be the combination of APIC-register virtualization, virtual interrupt delivery, and posted interrupts.

Which is totally fine. The big wrench is that the are other features that interact with "APICv" and require allocations. E.g. the APIC access page isn't actually tied to enable_apicv, it's tied to yet another VMX feature, "virtualize APIC accesses" (not to be confused with APIC-register virtualization; don't blame me, I didn't create the control knobs/names).

As a result, KVM may need to allocate the APIC access page (not to be confused with the APIC *backing* page; again, don't blame me :-D) when enable_apicv=false, and even more confusingly, NOT allocate the APIC access page when enable_apicv=true.

	if (cpu_need_virtualize_apic_accesses(vcpu)) {  <=== not an enable_apic check, *sigh*
		err = kvm_alloc_apic_access_page(vcpu->kvm);
		if (err)
			goto free_vmcs;
	}

And for both SVM and VMX, IPI virtualization adds another wrinkle, in that the per-vCPU setup needs to fill an entry in a per-VM table. And filling that entry needs to happen at the very end of vCPU creation, e.g. so that the vCPU can't be targeted until KVM is ready to run the vCPU.

Ouch. And I'm pretty sure there's a use-after-free in the AVIC code. If svm_vcpu_alloc_msrpm() fails, the avic_physical_id_table[] will still have a pointer to freed vAPIC page.

Thus, invoking avic_init_vcpu() "right away" is unfortunately flat out unsafe (took me a while to realize that).

So while I 100% agree with your dislike of this patch, I think it's the least awful band-aid, at least in the short term.

Longer term, I'd definitely be in favor of cleaning up the related flows, but I think that's best done on top of this series, because I suspect it'll be somewhat invasive.

Acked-by Naveen N Rao (AMD)
#31 lore raw

Re: [PATCH v3 12/62] KVM: SVM: Inhibit AVIC if ID is too big instead of rejecting vCPU creation

Quoted 46 lines · Sean Christopherson
Quoted 24 lines · Naveen N Rao
Quoted 11 lines · Sean Christopherson
arch/x86/kvm/svm/avic.c +3 -0
index ab228872a19b..f0a74b102c57 100644
@@ -277,9 +277,19 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	int id = vcpu->vcpu_id;
 	struct vcpu_svm *svm = to_svm(vcpu);
 
+	/*
+	 * Inhibit AVIC if the vCPU ID is bigger than what is supported by AVIC
+	 * hardware.  Immediately clear apicv_active, i.e. don't wait until the

arch/x86/kvm/svm/avic.c static int avic_init_backing_page(struct kvm_vcpu *vcpu) · line 290

Quoted 10 lines
arch/x86/kvm/svm/avic.c +7 -2
+	 * KVM_REQ_APICV_UPDATE request is processed on the first KVM_RUN, as
+	 * avic_vcpu_load() expects to be called if and only if the vCPU has
+	 * fully initialized AVIC.
+	 */
 	if ((!x2avic_enabled && id > AVIC_MAX_PHYSICAL_ID) ||
-	    (id > X2AVIC_MAX_PHYSICAL_ID))
-		return -EINVAL;
+	    (id > X2AVIC_MAX_PHYSICAL_ID)) {
+		kvm_set_apicv_inhibit(vcpu->kvm, APICV_INHIBIT_REASON_PHYSICAL_ID_TOO_BIG);
+		vcpu->arch.apic->apicv_active = false;

This bothers me a bit. kvm_create_lapic() does this:

Quoted 7 lines
          if (enable_apicv) {
                  apic->apicv_active = true;
                  kvm_make_request(KVM_REQ_APICV_UPDATE, vcpu);
	  }

But, setting apic->apicv_active to false here means KVM_REQ_APICV_UPDATE is going to be a no-op.

That's fine, KVM_REQ_APICV_UPDATE is a nop in other scenarios, too. I agree it's not ideal, but this is a rather extreme edge case and a one-time slow path, so I don't think it's worth doing anything special just to avoid KVM_REQ_APICV_UPDATE.

Quoted 8 lines

This does not look to be a big deal given that kvm_create_lapic() is called just a bit before svm_vcpu_create() (which calls the above function through avic_init_vcpu()) in kvm_arch_vcpu_create(), so there isn't that much done before apicv_active is toggled.

But, this made me wonder if introducing a kvm_x86_op to check and enable/disable apic->apicv_active in kvm_create_lapic() might be cleaner overall

Quoted 9 lines

Hmm, yes and no. I completely agree that clearing apicv_active in avic.c is all kinds of gross, but clearing apic->apicv_active in lapic.c to handle this particular scenario is just as problematic, because then avic_init_backing_page() would need to check kvm_vcpu_apicv_active() to determine whether or not to allocate the backing page. In a way, that's even worse, because setting apic->apicv_active by default is purely an optimization, i.e. leaving it %false _should_ work as well, it would just be suboptimal. But if AVIC were to key off apic->apicv_active, that could
lead to KVM incorrectly skipping allocation of the AVIC backing page.

I understand your concern about key'ing off apic->apicv_active - that would definitely require a thorough audit and does add complexity to this.

However, as far as I can see, after your current series, we no longer maintain a pointer to the AVIC backing page, but instead rely on the lapic-allocated page.

Were you referring to the APIC access page though? That is behind kvm_apicv_activated() today, which looks to be problematic if there are inhibits set during vcpu_create() and if those can be unset later? Shouldn't we be allocating the apic access page unconditionally here?

Quoted 19 lines
Quoted 3 lines

Maybe even have it be the initialization point for APICv:
apicv_init(), so we can invoke avic_init_vcpu() right away?

I mostly like this idea (actually, I take that back; see below), but VMX throws a big wrench in things. Unlike SVM, VMX doesn't have a singular "enable APICv" flag. Rather, KVM considers "APICv" to be the combination of APIC-register virtualization, virtual interrupt delivery, and posted interrupts.

Which is totally fine. The big wrench is that the are other features that interact with "APICv" and require allocations. E.g. the APIC access page isn't actually tied to enable_apicv, it's tied to yet another VMX feature, "virtualize APIC accesses" (not to be confused with APIC-register virtualization; don't blame me, I didn't create the control knobs/names).

As a result, KVM may need to allocate the APIC access page (not to be confused with the APIC *backing* page; again, don't blame me :-D) when enable_apicv=false, and even more confusingly, NOT allocate the APIC access page when enable_apicv=true.

Quoted 5 lines
	if (cpu_need_virtualize_apic_accesses(vcpu)) {  <=== not an enable_apic check, *sigh*
		err = kvm_alloc_apic_access_page(vcpu->kvm);
		if (err)
			goto free_vmcs;
	}

Thanks for the background and the details there -- I am going to need some time to go unpack all of that :)

Quoted 9 lines

And for both SVM and VMX, IPI virtualization adds another wrinkle, in that the per-vCPU setup needs to fill an entry in a per-VM table. And filling that entry needs to happen at the very end of vCPU creation, e.g. so that the vCPU can't be targeted until KVM is ready to run the vCPU.

Ouch. And I'm pretty sure there's a use-after-free in the AVIC code. If svm_vcpu_alloc_msrpm() fails, the avic_physical_id_table[] will still have a pointer to freed vAPIC page.

Oops.

Quoted 10 lines

Thus, invoking avic_init_vcpu() "right away" is unfortunately flat out unsafe (took me a while to realize that).

So while I 100% agree with your dislike of this patch, I think it's the least awful band-aid, at least in the short term.

Longer term, I'd definitely be in favor of cleaning up the related flows, but I think that's best done on top of this series, because I suspect it'll be somewhat invasive.

Sure, that makes sense. For this patch:

Thanks,
Naveen

author maintainer
#32 lore raw

Re: [PATCH v3 12/62] KVM: SVM: Inhibit AVIC if ID is too big instead of rejecting vCPU creation

Quoted 15 lines · Naveen N Rao
Quoted 9 lines · Sean Christopherson

Hmm, yes and no. I completely agree that clearing apicv_active in avic.c is all kinds of gross, but clearing apic->apicv_active in lapic.c to handle this particular scenario is just as problematic, because then avic_init_backing_page() would need to check kvm_vcpu_apicv_active() to determine whether or not to allocate the backing page. In a way, that's even worse, because setting apic->apicv_active by default is purely an optimization, i.e. leaving it %false _should_ work as well, it would just be suboptimal. But if AVIC were to key off apic->apicv_active, that could lead to KVM incorrectly skipping allocation of the AVIC backing page.

I understand your concern about key'ing off apic->apicv_active - that would definitely require a thorough audit and does add complexity to this.

Quoted 5 lines

However, as far as I can see, after your current series, we no longer maintain a pointer to the AVIC backing page, but instead rely on the lapic-allocated page.

Were you referring to the APIC access page though?

Gah, yes. I was hyper aware of the two things when typing up the response, and still managed to screw up. *sigh* :-)

Quoted 3 lines

That is behind kvm_apicv_activated() today, which looks to be problematic if there are inhibits set during vcpu_create() and if those can be unset later? Shouldn't we be allocating the apic access page unconditionally here?

In theory, yes. In practice, this guards against an unnecessary allocation for SEV+ guests (see APICV_INHIBIT_REASON_SEV).

That said, I completely agree that checking kvm_apicv_activated() is weird and sketchy. Hopefully that can be cleaned up, too (but after this series).

author maintainer
#33 lore raw

PATCH v3 13/62 KVM: SVM: Drop redundant check in AVIC code on ID during vCPU creation

In reply to Sean Christopherson

Drop avic_get_physical_id_entry()'s compatibility check on the incoming ID, as its sole caller, avic_init_backing_page(), performs the exact same check. Drop avic_get_physical_id_entry() entirely as the only remaining functionality is getting the address of the Physical ID table, and accessing the array without an immediate bounds check is kludgy.

Opportunistically add a compile-time assertion to ensure the vcpu_id can't result in a bounds overflow, e.g. if KVM (really) messed up a maximum physical ID #define, as well as run-time assertions so that a NULL pointer dereference is morphed into a safer WARN().

No functional change intended.

---
 arch/x86/kvm/svm/avic.c | 37 +++++++++++++++----------------------
 1 file changed, 15 insertions(+), 22 deletions(-)
arch/x86/kvm/svm/avic.c +15 -22
index f0a74b102c57..948bab48083b 100644
@@ -256,26 +256,12 @@ void avic_init_vmcb(struct vcpu_svm *svm, struct vmcb *vmcb)
 		avic_deactivate_vmcb(svm);
 }
 
-static u64 *avic_get_physical_id_entry(struct kvm_vcpu *vcpu,
-				       unsigned int index)
-{
-	u64 *avic_physical_id_table;
-	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
-
-	if ((!x2avic_enabled && index > AVIC_MAX_PHYSICAL_ID) ||
-	    (index > X2AVIC_MAX_PHYSICAL_ID))
-		return NULL;
-
-	avic_physical_id_table = page_address(kvm_svm->avic_physical_id_table_page);
-
-	return &avic_physical_id_table[index];
-}
-
 static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 {
-	u64 *entry, new_entry;
-	int id = vcpu->vcpu_id;
+	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	struct vcpu_svm *svm = to_svm(vcpu);
+	u32 id = vcpu->vcpu_id;
+	u64 *table, new_entry;
 
 	/*
 	 * Inhibit AVIC if the vCPU ID is bigger than what is supported by AVIC
@@ -291,6 +277,9 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 		return 0;
 	}
 
+	BUILD_BUG_ON((AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE ||
+		     (X2AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE);
+
 	if (WARN_ON_ONCE(!vcpu->arch.apic->regs))
 		return -EINVAL;
 
@@ -309,9 +298,7 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	}
 
 	/* Setting AVIC backing page address in the phy APIC ID table */
-	entry = avic_get_physical_id_entry(vcpu, id);
-	if (!entry)
-		return -EINVAL;
+	table = page_address(kvm_svm->avic_physical_id_table_page);
 
 	/* Note, fls64() returns the bit position, +1. */
 	BUILD_BUG_ON(__PHYSICAL_MASK_SHIFT >
@@ -319,9 +306,9 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 
 	new_entry = avic_get_backing_page_address(svm) |
 		    AVIC_PHYSICAL_ID_ENTRY_VALID_MASK;
-	WRITE_ONCE(*entry, new_entry);
+	WRITE_ONCE(table[id], new_entry);
 
-	svm->avic_physical_id_cache = entry;
+	svm->avic_physical_id_cache = &table[id];
 
 	return 0;
 }
@@ -1004,6 +991,9 @@ void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	if (WARN_ON(h_physical_id & ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK))
 		return;
 
+	if (WARN_ON_ONCE(!svm->avic_physical_id_cache))
+		return;
+
 	/*
 	 * No need to update anything if the vCPU is blocking, i.e. if the vCPU
 	 * is being scheduled in after being preempted.  The CPU entries in the
@@ -1044,6 +1034,9 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 
 	lockdep_assert_preemption_disabled();
 
+	if (WARN_ON_ONCE(!svm->avic_physical_id_cache))
+		return;
+
 	/*
 	 * Note, reading the Physical ID entry outside of ir_list_lock is safe
 	 * as only the pCPU that has loaded (or is loading) the vCPU is allowed
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
Reviewed-by on v3 13/62
#34 lore raw

Re: [PATCH v3 13/62] KVM: SVM: Drop redundant check in AVIC code on ID during vCPU creation

Quoted 12 lines · Sean Christopherson

static int avic_init_backing_page(struct kvm_vcpu *vcpu) {

arch/x86/kvm/svm/avic.c +3 -2
-	u64 *entry, new_entry;
-	int id = vcpu->vcpu_id;
+	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	struct vcpu_svm *svm = to_svm(vcpu);
+	u32 id = vcpu->vcpu_id;
+	u64 *table, new_entry;
 
 	/*
 	 * Inhibit AVIC if the vCPU ID is bigger than what is supported by AVIC

arch/x86/kvm/svm/avic.c static int avic_init_backing_page(struct kvm_vcpu *vcpu) · line 281

Quoted 5 lines
arch/x86/kvm/svm/avic.c +2 -0
@@ -291,6 +277,9 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 		return 0;
 	}
 
+	BUILD_BUG_ON((AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE ||
+		     (X2AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE);
						    ^^^^^^^^^^^^^^

Renaming new_entry to just 'entry' and using sizeof(entry) makes this more readable for me. Otherwise, for this patch:

As an aside, there are a few static asserts to validate some of the related macros. Can this also be a static_assert(), or is there is reason to prefer BUILD_BUG_ON()?

- Naveen

author maintainer
#35 lore raw

Re: [PATCH v3 13/62] KVM: SVM: Drop redundant check in AVIC code on ID during vCPU creation

Quoted 16 lines · Naveen N Rao
Quoted 15 lines · Sean Christopherson

static int avic_init_backing_page(struct kvm_vcpu *vcpu) {

arch/x86/kvm/svm/avic.c +3 -2
-	u64 *entry, new_entry;
-	int id = vcpu->vcpu_id;
+	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	struct vcpu_svm *svm = to_svm(vcpu);
+	u32 id = vcpu->vcpu_id;
+	u64 *table, new_entry;
 
 	/*
 	 * Inhibit AVIC if the vCPU ID is bigger than what is supported by AVIC
@@ -291,6 +277,9 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 		return 0;
 	}
 
Quoted 5 lines

arch/x86/kvm/svm/avic.c static int avic_init_backing_page(struct kvm_vcpu *vcpu) · line 281

Quoted 2 lines
arch/x86/kvm/svm/avic.c +2 -0
+	BUILD_BUG_ON((AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE ||
+		     (X2AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE);
						    ^^^^^^^^^^^^^^

Renaming new_entry to just 'entry' and using sizeof(entry) makes this more readable for me.

Good call, though I think it makes sense to do that on top so as to minimize the churn in this patch. I'll post a patch, unless you want the honors?

Quoted 6 lines

Otherwise, for this patch:

As an aside, there are a few static asserts to validate some of the related macros. Can this also be a static_assert(), or is there is reason to prefer BUILD_BUG_ON()?

For this particular assertion, static_assert() would be fine. That said, BUILD_BUG_ON() is slightly preferred in this context.

The advantage of BUILD_BUG_ON() is that it works so long as the condition is compile-time constant, whereas static_assert() requires the condition to an integer constant expression. E.g. BUILD_BUG_ON() can be used so long as the condition is eventually resolved to a constant, whereas static_assert() has stricter requirements.

E.g. the fls64() assert below is fully resolved at compile time, but isn't a purely constant expression, i.e. that one *needs* to be BUILD_BUG_ON().

--
arch/x86/kvm/svm/avic.c: In function ‘avic_init_backing_page’:
arch/x86/kvm/svm/avic.c:293:45: error: expression in static assertion is not constant
293 | static_assert(__PHYSICAL_MASK_SHIFT <=
include/linux/build_bug.h:78:56: note: in definition of macro ‘__static_assert’
78 | #define __static_assert(expr, msg, ...) _Static_assert(expr, msg)
| ^~~~
arch/x86/kvm/svm/avic.c:293:9: note: in expansion of macro ‘static_assert’

  293 |         static_assert(__PHYSICAL_MASK_SHIFT <=
      |         ^~~~~~~~~~~~~

make[5]: *** [scripts/Makefile.build:203: arch/x86/kvm/svm/avic.o] Error 1 --

The downside of BUILD_BUG_ON() is that it can't be used at global scope, i.e. needs to be called from a function.

As a result, when adding an assertion in a function, using BUILD_BUG_ON() is slightly preferred, because it's less likely to break in the future. E.g. if X2AVIC_MAX_PHYSICAL_ID were changed to something that is a compile-time constant, but for whatever reason isn't a pure integer constant.

#36 lore raw

Re: [PATCH v3 13/62] KVM: SVM: Drop redundant check in AVIC code on ID during vCPU creation

Quoted 17 lines · Sean Christopherson
Quoted 16 lines · Naveen N Rao
Quoted 15 lines · Sean Christopherson

static int avic_init_backing_page(struct kvm_vcpu *vcpu) {

arch/x86/kvm/svm/avic.c +3 -2
-	u64 *entry, new_entry;
-	int id = vcpu->vcpu_id;
+	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	struct vcpu_svm *svm = to_svm(vcpu);
+	u32 id = vcpu->vcpu_id;
+	u64 *table, new_entry;
 
 	/*
 	 * Inhibit AVIC if the vCPU ID is bigger than what is supported by AVIC
@@ -291,6 +277,9 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 		return 0;
 	}
 
Quoted 8 lines
Quoted 5 lines

arch/x86/kvm/svm/avic.c static int avic_init_backing_page(struct kvm_vcpu *vcpu) · line 281

Quoted 2 lines
arch/x86/kvm/svm/avic.c +2 -0
+	BUILD_BUG_ON((AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE ||
+		     (X2AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE);
						    ^^^^^^^^^^^^^^

Renaming new_entry to just 'entry' and using sizeof(entry) makes this more readable for me.

Good call, though I think it makes sense to do that on top so as to minimize the churn in this patch. I'll post a patch, unless you want the honors?

Not at all, please feel free to add a patch (or not, given that this will be a trivial change).

Quoted 33 lines
Quoted 7 lines

Otherwise, for this patch:

As an aside, there are a few static asserts to validate some of the related macros. Can this also be a static_assert(), or is there is reason to prefer BUILD_BUG_ON()?

For this particular assertion, static_assert() would be fine. That said, BUILD_BUG_ON() is slightly preferred in this context.

The advantage of BUILD_BUG_ON() is that it works so long as the condition is compile-time constant, whereas static_assert() requires the condition to an integer constant expression. E.g. BUILD_BUG_ON() can be used so long as the condition is eventually resolved to a constant, whereas static_assert() has stricter requirements.

E.g. the fls64() assert below is fully resolved at compile time, but isn't a purely constant expression, i.e. that one *needs* to be BUILD_BUG_ON().

--
arch/x86/kvm/svm/avic.c: In function ‘avic_init_backing_page’:
arch/x86/kvm/svm/avic.c:293:45: error: expression in static assertion is not constant
293 | static_assert(__PHYSICAL_MASK_SHIFT <=
include/linux/build_bug.h:78:56: note: in definition of macro ‘__static_assert’
78 | #define __static_assert(expr, msg, ...) _Static_assert(expr, msg)
| ^~~~
arch/x86/kvm/svm/avic.c:293:9: note: in expansion of macro ‘static_assert’

  293 |         static_assert(__PHYSICAL_MASK_SHIFT <=
      |         ^~~~~~~~~~~~~

make[5]: *** [scripts/Makefile.build:203: arch/x86/kvm/svm/avic.o] Error 1 --

Quoted 7 lines

The downside of BUILD_BUG_ON() is that it can't be used at global scope, i.e. needs to be called from a function.

As a result, when adding an assertion in a function, using BUILD_BUG_ON() is slightly preferred, because it's less likely to break in the future. E.g. if X2AVIC_MAX_PHYSICAL_ID were changed to something that is a compile-time constant, but for whatever reason isn't a pure integer constant.

Understood, thanks for the explanation.

- Naveen

author maintainer
#37 lore raw

PATCH v3 14/62 KVM: SVM: Track AVIC tables as natively sized pointers, not "struct pages"

In reply to Sean Christopherson

Allocate and track AVIC's logical and physical tables as u32 and u64 pointers respectively, as managing the pages as "struct page" pointers adds an almost absurd amount of boilerplate and complexity. E.g. with page_address() out of the way, svm->avic_physical_id_cache becomes completely superfluous, and will be removed in a future cleanup.

No functional change intended.

---
 arch/x86/kvm/svm/avic.c | 49 ++++++++++++++---------------------------
 arch/x86/kvm/svm/svm.h  |  4 ++--
 2 files changed, 18 insertions(+), 35 deletions(-)
arch/x86/kvm/svm/avic.c +16 -33
index 948bab48083b..bf18b0b643d9 100644
@@ -172,10 +172,8 @@ void avic_vm_destroy(struct kvm *kvm)
 	if (!enable_apicv)
 		return;
 
-	if (kvm_svm->avic_logical_id_table_page)
-		__free_page(kvm_svm->avic_logical_id_table_page);
-	if (kvm_svm->avic_physical_id_table_page)
-		__free_page(kvm_svm->avic_physical_id_table_page);
+	free_page((unsigned long)kvm_svm->avic_logical_id_table);
+	free_page((unsigned long)kvm_svm->avic_physical_id_table);
 
 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
 	hash_del(&kvm_svm->hnode);
@@ -188,27 +186,19 @@ int avic_vm_init(struct kvm *kvm)
 	int err = -ENOMEM;
 	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
 	struct kvm_svm *k2;
-	struct page *p_page;
-	struct page *l_page;
 	u32 vm_id;
 
 	if (!enable_apicv)
 		return 0;
 
-	/* Allocating physical APIC ID table (4KB) */
-	p_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
-	if (!p_page)
+	kvm_svm->avic_physical_id_table = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT);
+	if (!kvm_svm->avic_physical_id_table)
 		goto free_avic;
 
-	kvm_svm->avic_physical_id_table_page = p_page;
-
-	/* Allocating logical APIC ID table (4KB) */
-	l_page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO);
-	if (!l_page)
+	kvm_svm->avic_logical_id_table = (void *)get_zeroed_page(GFP_KERNEL_ACCOUNT);
+	if (!kvm_svm->avic_logical_id_table)
 		goto free_avic;
 
-	kvm_svm->avic_logical_id_table_page = l_page;
-
 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
  again:
 	vm_id = next_vm_id = (next_vm_id + 1) & AVIC_VM_ID_MASK;
@@ -242,12 +232,10 @@ static phys_addr_t avic_get_backing_page_address(struct vcpu_svm *svm)
 void avic_init_vmcb(struct vcpu_svm *svm, struct vmcb *vmcb)
 {
 	struct kvm_svm *kvm_svm = to_kvm_svm(svm->vcpu.kvm);
-	phys_addr_t lpa = __sme_set(page_to_phys(kvm_svm->avic_logical_id_table_page));
-	phys_addr_t ppa = __sme_set(page_to_phys(kvm_svm->avic_physical_id_table_page));
 
 	vmcb->control.avic_backing_page = avic_get_backing_page_address(svm);
-	vmcb->control.avic_logical_id = lpa;
-	vmcb->control.avic_physical_id = ppa;
+	vmcb->control.avic_logical_id = __sme_set(__pa(kvm_svm->avic_logical_id_table));
+	vmcb->control.avic_physical_id = __sme_set(__pa(kvm_svm->avic_physical_id_table));
 	vmcb->control.avic_vapic_bar = APIC_DEFAULT_PHYS_BASE;
 
 	if (kvm_apicv_activated(svm->vcpu.kvm))
@@ -261,7 +249,7 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	struct vcpu_svm *svm = to_svm(vcpu);
 	u32 id = vcpu->vcpu_id;
-	u64 *table, new_entry;
+	u64 new_entry;
 
 	/*
 	 * Inhibit AVIC if the vCPU ID is bigger than what is supported by AVIC
@@ -277,8 +265,8 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 		return 0;
 	}
 
-	BUILD_BUG_ON((AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE ||
-		     (X2AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE);
+	BUILD_BUG_ON((AVIC_MAX_PHYSICAL_ID + 1) * sizeof(new_entry) > PAGE_SIZE ||
+		     (X2AVIC_MAX_PHYSICAL_ID + 1) * sizeof(new_entry) > PAGE_SIZE);
 
 	if (WARN_ON_ONCE(!vcpu->arch.apic->regs))
 		return -EINVAL;
@@ -297,18 +285,16 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 			return ret;
 	}
 
-	/* Setting AVIC backing page address in the phy APIC ID table */
-	table = page_address(kvm_svm->avic_physical_id_table_page);
-
 	/* Note, fls64() returns the bit position, +1. */
 	BUILD_BUG_ON(__PHYSICAL_MASK_SHIFT >
 		     fls64(AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK));
 
+	/* Setting AVIC backing page address in the phy APIC ID table */
 	new_entry = avic_get_backing_page_address(svm) |
 		    AVIC_PHYSICAL_ID_ENTRY_VALID_MASK;
-	WRITE_ONCE(table[id], new_entry);
+	WRITE_ONCE(kvm_svm->avic_physical_id_table[id], new_entry);
 
-	svm->avic_physical_id_cache = &table[id];
+	svm->avic_physical_id_cache = &kvm_svm->avic_physical_id_table[id];
 
 	return 0;
 }
@@ -442,7 +428,7 @@ static int avic_kick_target_vcpus_fast(struct kvm *kvm, struct kvm_lapic *source
 		if (apic_x2apic_mode(source))
 			avic_logical_id_table = NULL;
 		else
-			avic_logical_id_table = page_address(kvm_svm->avic_logical_id_table_page);
+			avic_logical_id_table = kvm_svm->avic_logical_id_table;
 
 		/*
 		 * AVIC is inhibited if vCPUs aren't mapped 1:1 with logical
@@ -544,7 +530,6 @@ unsigned long avic_vcpu_get_apicv_inhibit_reasons(struct kvm_vcpu *vcpu)
 static u32 *avic_get_logical_id_entry(struct kvm_vcpu *vcpu, u32 ldr, bool flat)
 {
 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
-	u32 *logical_apic_id_table;
 	u32 cluster, index;
 
 	ldr = GET_APIC_LOGICAL_ID(ldr);
@@ -565,9 +550,7 @@ static u32 *avic_get_logical_id_entry(struct kvm_vcpu *vcpu, u32 ldr, bool flat)
 		return NULL;
 	index += (cluster << 2);
 
-	logical_apic_id_table = (u32 *) page_address(kvm_svm->avic_logical_id_table_page);
-
-	return &logical_apic_id_table[index];
+	return &kvm_svm->avic_logical_id_table[index];
 }
 
 static void avic_ldr_write(struct kvm_vcpu *vcpu, u8 g_physical_id, u32 ldr)
arch/x86/kvm/svm/svm.h +2 -2
index 71e3c003580e..ec5d77d42a49 100644
@@ -123,8 +123,8 @@ struct kvm_svm {
 
 	/* Struct members for AVIC */
 	u32 avic_vm_id;
-	struct page *avic_logical_id_table_page;
-	struct page *avic_physical_id_table_page;
+	u32 *avic_logical_id_table;
+	u64 *avic_physical_id_table;
 	struct hlist_node hnode;
 
 	struct kvm_sev_info sev_info;
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
Acked-by on v3 14/62
#38 lore raw

Re: [PATCH v3 14/62] KVM: SVM: Track AVIC tables as natively sized pointers, not "struct pages"

arch/x86/kvm/svm/avic.c static int avic_init_backing_page(struct kvm_vcpu · line 270

Quoted 9 lines · Sean Christopherson
arch/x86/kvm/svm/avic.c +2 -2
@@ -277,8 +265,8 @@ static int avic_init_backing_page(struct kvm_vcpu 
 *vcpu)
 		return 0;
 	}
 
-	BUILD_BUG_ON((AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE ||
-		     (X2AVIC_MAX_PHYSICAL_ID + 1) * sizeof(*table) > PAGE_SIZE);
+	BUILD_BUG_ON((AVIC_MAX_PHYSICAL_ID + 1) * sizeof(new_entry) > PAGE_SIZE ||
+		     (X2AVIC_MAX_PHYSICAL_ID + 1) * sizeof(new_entry) > PAGE_SIZE);

Ah, you change to 'new_entry' here. Would be good to rename it to just 'entry', but that's a minor nit. Ignore my comment on the previous patch.

For this patch:

- Naveen

author maintainer
#39 lore raw

PATCH v3 15/62 KVM: SVM: Drop superfluous "cache" of AVIC Physical ID entry pointer

In reply to Sean Christopherson

Drop the vCPU's pointer to its AVIC Physical ID entry, and simply index the table directly. Caching a pointer address is completely unnecessary for performance, and while the field technically caches the result of the pointer calculation, it's all too easy to misinterpret the name and think that the field somehow caches the _data_ in the table.

No functional change intended.

---
 arch/x86/kvm/svm/avic.c | 27 +++++++++++++++------------
 arch/x86/kvm/svm/svm.h  |  1 -
 2 files changed, 15 insertions(+), 13 deletions(-)
arch/x86/kvm/svm/avic.c +15 -12
index bf18b0b643d9..0c0be274d29e 100644
@@ -294,8 +294,6 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 		    AVIC_PHYSICAL_ID_ENTRY_VALID_MASK;
 	WRITE_ONCE(kvm_svm->avic_physical_id_table[id], new_entry);
 
-	svm->avic_physical_id_cache = &kvm_svm->avic_physical_id_table[id];
-
 	return 0;
 }
 
@@ -770,13 +768,16 @@ static int svm_ir_list_add(struct vcpu_svm *svm,
 			   struct kvm_kernel_irqfd *irqfd,
 			   struct amd_iommu_pi_data *pi)
 {
+	struct kvm_vcpu *vcpu = &svm->vcpu;
+	struct kvm *kvm = vcpu->kvm;
+	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
 	unsigned long flags;
 	u64 entry;
 
 	if (WARN_ON_ONCE(!pi->ir_data))
 		return -EINVAL;
 
-	irqfd->irq_bypass_vcpu = &svm->vcpu;
+	irqfd->irq_bypass_vcpu = vcpu;
 	irqfd->irq_bypass_data = pi->ir_data;
 
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
@@ -787,7 +788,7 @@ static int svm_ir_list_add(struct vcpu_svm *svm,
 	 * will update the pCPU info when the vCPU awkened and/or scheduled in.
 	 * See also avic_vcpu_load().
 	 */
-	entry = READ_ONCE(*(svm->avic_physical_id_cache));
+	entry = READ_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id]);
 	if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
 		amd_iommu_update_ga(entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK,
 				    true, pi->ir_data);
@@ -964,17 +965,18 @@ avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu, bool r)
 
 void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 {
-	u64 entry;
+	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	int h_physical_id = kvm_cpu_get_apicid(cpu);
 	struct vcpu_svm *svm = to_svm(vcpu);
 	unsigned long flags;
+	u64 entry;
 
 	lockdep_assert_preemption_disabled();
 
 	if (WARN_ON(h_physical_id & ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK))
 		return;
 
-	if (WARN_ON_ONCE(!svm->avic_physical_id_cache))
+	if (WARN_ON_ONCE(vcpu->vcpu_id * sizeof(entry) >= PAGE_SIZE))
 		return;
 
 	/*
@@ -996,14 +998,14 @@ void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	 */
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
 
-	entry = READ_ONCE(*(svm->avic_physical_id_cache));
+	entry = READ_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id]);
 	WARN_ON_ONCE(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK);
 
 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
 	entry |= (h_physical_id & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK);
 	entry |= AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
 
-	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
+	WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
 	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id, true);
 
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
@@ -1011,13 +1013,14 @@ void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 
 void avic_vcpu_put(struct kvm_vcpu *vcpu)
 {
-	u64 entry;
+	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	struct vcpu_svm *svm = to_svm(vcpu);
 	unsigned long flags;
+	u64 entry;
 
 	lockdep_assert_preemption_disabled();
 
-	if (WARN_ON_ONCE(!svm->avic_physical_id_cache))
+	if (WARN_ON_ONCE(vcpu->vcpu_id * sizeof(entry) >= PAGE_SIZE))
 		return;
 
 	/*
@@ -1027,7 +1030,7 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 	 * can't be scheduled out and thus avic_vcpu_{put,load}() can't run
 	 * recursively.
 	 */
-	entry = READ_ONCE(*(svm->avic_physical_id_cache));
+	entry = READ_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id]);
 
 	/* Nothing to do if IsRunning == '0' due to vCPU blocking. */
 	if (!(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK))
@@ -1046,7 +1049,7 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 	avic_update_iommu_vcpu_affinity(vcpu, -1, 0);
 
 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
-	WRITE_ONCE(*(svm->avic_physical_id_cache), entry);
+	WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
 
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
 
arch/x86/kvm/svm/svm.h +0 -1
index ec5d77d42a49..f225d0bed152 100644
@@ -306,7 +306,6 @@ struct vcpu_svm {
 
 	u32 ldr_reg;
 	u32 dfr_reg;
-	u64 *avic_physical_id_cache;
 
 	/*
 	 * Per-vCPU list of irqfds that are eligible to post IRQs directly to
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
Reviewed-by on v3 15/62
#40 lore raw

Re: [PATCH v3 15/62] KVM: SVM: Drop superfluous "cache" of AVIC Physical ID entry pointer

Quoted 8 lines · Sean Christopherson

Drop the vCPU's pointer to its AVIC Physical ID entry, and simply index the table directly. Caching a pointer address is completely unnecessary for performance, and while the field technically caches the result of the pointer calculation, it's all too easy to misinterpret the name and think that the field somehow caches the _data_ in the table.

No functional change intended.

Quoted 7 lines
---
 arch/x86/kvm/svm/avic.c | 27 +++++++++++++++------------
 arch/x86/kvm/svm/svm.h  |  1 -
 2 files changed, 15 insertions(+), 13 deletions(-)

- Naveen

author maintainer
#41 lore raw

PATCH v3 16/62 KVM: VMX: Move enable_ipiv knob to common x86

In reply to Sean Christopherson

Move enable_ipiv to common x86 so that it can be reused by SVM to control IPI virtualization when AVIC is enabled. SVM doesn't actually provide a way to truly disable IPI virtualization, but KVM can get close enough by skipping the necessary table programming.

---
 arch/x86/include/asm/kvm_host.h | 1 +
 arch/x86/kvm/vmx/capabilities.h | 1 -
 arch/x86/kvm/vmx/vmx.c          | 2 --
 arch/x86/kvm/x86.c              | 3 +++
 4 files changed, 4 insertions(+), 3 deletions(-)
arch/x86/include/asm/kvm_host.h +1 -0
index a9b709db7c59..cba82d7a701d 100644
@@ -1950,6 +1950,7 @@ struct kvm_arch_async_pf {
 extern u32 __read_mostly kvm_nr_uret_msrs;
 extern bool __read_mostly allow_smaller_maxphyaddr;
 extern bool __read_mostly enable_apicv;
+extern bool __read_mostly enable_ipiv;
 extern bool __read_mostly enable_device_posted_irqs;
 extern struct kvm_x86_ops kvm_x86_ops;
 
arch/x86/kvm/vmx/capabilities.h +0 -1
index cb6588238f46..5316c27f6099 100644
@@ -15,7 +15,6 @@ extern bool __read_mostly enable_ept;
 extern bool __read_mostly enable_unrestricted_guest;
 extern bool __read_mostly enable_ept_ad_bits;
 extern bool __read_mostly enable_pml;
-extern bool __read_mostly enable_ipiv;
 extern int __read_mostly pt_mode;
 
 #define PT_MODE_SYSTEM		0
arch/x86/kvm/vmx/vmx.c +0 -2
index 9ff00ae9f05a..f79604bc0127 100644
@@ -112,8 +112,6 @@ static bool __read_mostly fasteoi = 1;
 module_param(fasteoi, bool, 0444);
 
 module_param(enable_apicv, bool, 0444);
-
-bool __read_mostly enable_ipiv = true;
 module_param(enable_ipiv, bool, 0444);
 
 module_param(enable_device_posted_irqs, bool, 0444);
arch/x86/kvm/x86.c +3 -0
index 93711a5ef272..f1b0dbce9a8b 100644
@@ -226,6 +226,9 @@ EXPORT_SYMBOL_GPL(allow_smaller_maxphyaddr);
 bool __read_mostly enable_apicv = true;
 EXPORT_SYMBOL_GPL(enable_apicv);
 
+bool __read_mostly enable_ipiv = true;
+EXPORT_SYMBOL_GPL(enable_ipiv);
+
 bool __read_mostly enable_device_posted_irqs = true;
 EXPORT_SYMBOL_GPL(enable_device_posted_irqs);
 
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#42 lore raw

PATCH v3 17/62 KVM: SVM: Add enable_ipiv param, never set IsRunning if disabled

In reply to Sean Christopherson

From: Maxim Levitsky <[email protected]>

Let userspace "disable" IPI virtualization for AVIC via the enable_ipiv module param, by never setting IsRunning. SVM doesn't provide a way to disable IPI virtualization in hardware, but by ensuring CPUs never see IsRunning=1, every IPI in the guest (except for self-IPIs) will generate a VM-Exit.

To avoid setting the real IsRunning bit, while still allowing KVM to use each vCPU's entry to update GA log entries, simply maintain a shadow of the entry, without propagating IsRunning updates to the real table when IPI virtualization is disabled.

Providing a way to effectively disable IPI virtualization will allow KVM to safely enable AVIC on hardware that is susceptible to erratum #1235, which causes hardware to sometimes fail to detect that the IsRunning bit has been cleared by software.

Note, the table _must_ be fully populated, as broadcast IPIs skip invalid entries, i.e. won't generate VM-Exit if every entry is invalid, and so simply pointing the VMCB at a common dummy table won't work.

Alternatively, KVM could allocate a shadow of the entire table, but that'd be a waste of 4KiB since the per-vCPU entry doesn't actually consume an additional 8 bytes of memory (vCPU structures are large enough that they are backed by order-N pages).

[sean: keep "entry" variables, reuse enable_ipiv, split from erratum]
---
 arch/x86/kvm/svm/avic.c | 32 ++++++++++++++++++++++++++------
 arch/x86/kvm/svm/svm.c  |  2 ++
 arch/x86/kvm/svm/svm.h  |  8 ++++++++
 3 files changed, 36 insertions(+), 6 deletions(-)
arch/x86/kvm/svm/avic.c +26 -6
index 0c0be274d29e..48c737e1200a 100644
@@ -292,6 +292,13 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	/* Setting AVIC backing page address in the phy APIC ID table */
 	new_entry = avic_get_backing_page_address(svm) |
 		    AVIC_PHYSICAL_ID_ENTRY_VALID_MASK;
+	svm->avic_physical_id_entry = new_entry;
+
+	/*
+	 * Initialize the real table, as vCPUs must have a valid entry in order
+	 * for broadcast IPIs to function correctly (broadcast IPIs ignore
+	 * invalid entries, i.e. aren't guaranteed to generate a VM-Exit).
+	 */
 	WRITE_ONCE(kvm_svm->avic_physical_id_table[id], new_entry);
 
 	return 0;
@@ -769,8 +776,6 @@ static int svm_ir_list_add(struct vcpu_svm *svm,
 			   struct amd_iommu_pi_data *pi)
 {
 	struct kvm_vcpu *vcpu = &svm->vcpu;
-	struct kvm *kvm = vcpu->kvm;
-	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
 	unsigned long flags;
 	u64 entry;
 
@@ -788,7 +793,7 @@ static int svm_ir_list_add(struct vcpu_svm *svm,
 	 * will update the pCPU info when the vCPU awkened and/or scheduled in.
 	 * See also avic_vcpu_load().
 	 */
-	entry = READ_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id]);
+	entry = svm->avic_physical_id_entry;
 	if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
 		amd_iommu_update_ga(entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK,
 				    true, pi->ir_data);
@@ -998,14 +1003,26 @@ void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	 */
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
 
-	entry = READ_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id]);
+	entry = svm->avic_physical_id_entry;
 	WARN_ON_ONCE(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK);
 
 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
 	entry |= (h_physical_id & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK);
 	entry |= AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
 
+	svm->avic_physical_id_entry = entry;
+
+	/*
+	 * If IPI virtualization is disabled, clear IsRunning when updating the
+	 * actual Physical ID table, so that the CPU never sees IsRunning=1.
+	 * Keep the APIC ID up-to-date in the entry to minimize the chances of
+	 * things going sideways if hardware peeks at the ID.
+	 */
+	if (!enable_ipiv)
+		entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
+
 	WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
+
 	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id, true);
 
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
@@ -1030,7 +1047,7 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 	 * can't be scheduled out and thus avic_vcpu_{put,load}() can't run
 	 * recursively.
 	 */
-	entry = READ_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id]);
+	entry = svm->avic_physical_id_entry;
 
 	/* Nothing to do if IsRunning == '0' due to vCPU blocking. */
 	if (!(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK))
@@ -1049,7 +1066,10 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 	avic_update_iommu_vcpu_affinity(vcpu, -1, 0);
 
 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
-	WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
+	svm->avic_physical_id_entry = entry;
+
+	if (enable_ipiv)
+		WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
 
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
 
arch/x86/kvm/svm/svm.c +2 -0
index 0ad1a6d4fb6d..56d11f7b4bef 100644
@@ -231,6 +231,7 @@ module_param(tsc_scaling, int, 0444);
  */
 static bool avic;
 module_param(avic, bool, 0444);
+module_param(enable_ipiv, bool, 0444);
 
 module_param(enable_device_posted_irqs, bool, 0444);
 
@@ -5594,6 +5595,7 @@ static __init int svm_hardware_setup(void)
 	enable_apicv = avic = avic && avic_hardware_setup();
 
 	if (!enable_apicv) {
+		enable_ipiv = false;
 		svm_x86_ops.vcpu_blocking = NULL;
 		svm_x86_ops.vcpu_unblocking = NULL;
 		svm_x86_ops.vcpu_get_apicv_inhibit_reasons = NULL;
arch/x86/kvm/svm/svm.h +8 -0
index f225d0bed152..939ff0e35a2b 100644
@@ -307,6 +307,14 @@ struct vcpu_svm {
 	u32 ldr_reg;
 	u32 dfr_reg;
 
+	/* This is essentially a shadow of the vCPU's actual entry in the
+	 * Physical ID table that is programmed into the VMCB, i.e. that is
+	 * seen by the CPU.  If IPI virtualization is disabled, IsRunning is
+	 * only ever set in the shadow, i.e. is never propagated to the "real"
+	 * table, so that hardware never sees IsRunning=1.
+	 */
+	u64 avic_physical_id_entry;
+
 	/*
 	 * Per-vCPU list of irqfds that are eligible to post IRQs directly to
 	 * the vCPU (a.k.a. device posted IRQs, a.k.a. IRQ bypass).  The list
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
#43 lore raw

Re: [PATCH v3 17/62] KVM: SVM: Add enable_ipiv param, never set IsRunning if disabled

Quoted 7 lines · Sean Christopherson

From: Maxim Levitsky <[email protected]>

Let userspace "disable" IPI virtualization for AVIC via the enable_ipiv module param, by never setting IsRunning. SVM doesn't provide a way to disable IPI virtualization in hardware, but by ensuring CPUs never see IsRunning=1, every IPI in the guest (except for self-IPIs) will generate a VM-Exit.

I think this is good to have regardless of the erratum. Not sure about VMX, but does it make sense to intercept writes to the self-ipi MSR as well?

Quoted 103 lines

To avoid setting the real IsRunning bit, while still allowing KVM to use each vCPU's entry to update GA log entries, simply maintain a shadow of the entry, without propagating IsRunning updates to the real table when IPI virtualization is disabled.

Providing a way to effectively disable IPI virtualization will allow KVM to safely enable AVIC on hardware that is susceptible to erratum #1235, which causes hardware to sometimes fail to detect that the IsRunning bit has been cleared by software.

Note, the table _must_ be fully populated, as broadcast IPIs skip invalid entries, i.e. won't generate VM-Exit if every entry is invalid, and so simply pointing the VMCB at a common dummy table won't work.

Alternatively, KVM could allocate a shadow of the entire table, but that'd be a waste of 4KiB since the per-vCPU entry doesn't actually consume an additional 8 bytes of memory (vCPU structures are large enough that they are backed by order-N pages).

[sean: keep "entry" variables, reuse enable_ipiv, split from erratum]
---
 arch/x86/kvm/svm/avic.c | 32 ++++++++++++++++++++++++++------
 arch/x86/kvm/svm/svm.c  |  2 ++
 arch/x86/kvm/svm/svm.h  |  8 ++++++++
 3 files changed, 36 insertions(+), 6 deletions(-)
arch/x86/kvm/svm/avic.c +22 -5
index 0c0be274d29e..48c737e1200a 100644
@@ -292,6 +292,13 @@ static int avic_init_backing_page(struct kvm_vcpu *vcpu)
 	/* Setting AVIC backing page address in the phy APIC ID table */
 	new_entry = avic_get_backing_page_address(svm) |
 		    AVIC_PHYSICAL_ID_ENTRY_VALID_MASK;
+	svm->avic_physical_id_entry = new_entry;
+
+	/*
+	 * Initialize the real table, as vCPUs must have a valid entry in order
+	 * for broadcast IPIs to function correctly (broadcast IPIs ignore
+	 * invalid entries, i.e. aren't guaranteed to generate a VM-Exit).
+	 */
 	WRITE_ONCE(kvm_svm->avic_physical_id_table[id], new_entry);
 
 	return 0;
@@ -769,8 +776,6 @@ static int svm_ir_list_add(struct vcpu_svm *svm,
 			   struct amd_iommu_pi_data *pi)
 {
 	struct kvm_vcpu *vcpu = &svm->vcpu;
-	struct kvm *kvm = vcpu->kvm;
-	struct kvm_svm *kvm_svm = to_kvm_svm(kvm);
 	unsigned long flags;
 	u64 entry;
 
@@ -788,7 +793,7 @@ static int svm_ir_list_add(struct vcpu_svm *svm,
 	 * will update the pCPU info when the vCPU awkened and/or scheduled in.
 	 * See also avic_vcpu_load().
 	 */
-	entry = READ_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id]);
+	entry = svm->avic_physical_id_entry;
 	if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
 		amd_iommu_update_ga(entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK,
 				    true, pi->ir_data);
@@ -998,14 +1003,26 @@ void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	 */
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
 
-	entry = READ_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id]);
+	entry = svm->avic_physical_id_entry;
 	WARN_ON_ONCE(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK);
 
 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
 	entry |= (h_physical_id & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK);
 	entry |= AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
 
+	svm->avic_physical_id_entry = entry;
+
+	/*
+	 * If IPI virtualization is disabled, clear IsRunning when updating the
+	 * actual Physical ID table, so that the CPU never sees IsRunning=1.
+	 * Keep the APIC ID up-to-date in the entry to minimize the chances of
+	 * things going sideways if hardware peeks at the ID.
+	 */
+	if (!enable_ipiv)
+		entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
+
 	WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
+
 	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id, true);
 
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
@@ -1030,7 +1047,7 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 	 * can't be scheduled out and thus avic_vcpu_{put,load}() can't run
 	 * recursively.
 	 */
-	entry = READ_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id]);
+	entry = svm->avic_physical_id_entry;
 
 	/* Nothing to do if IsRunning == '0' due to vCPU blocking. */
 	if (!(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK))

arch/x86/kvm/svm/avic.c void avic_vcpu_put(struct kvm_vcpu *vcpu) · line 1072

Quoted 8 lines
arch/x86/kvm/svm/avic.c +4 -1
@@ -1049,7 +1066,10 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 	avic_update_iommu_vcpu_affinity(vcpu, -1, 0);
 
 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
-	WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
+	svm->avic_physical_id_entry = entry;
+
+	if (enable_ipiv)
+		WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);

If enable_ipiv is false, then isRunning bit will never be set and we would have bailed out earlier. So, the check for enable_ipiv can be dropped here (or converted into an assert).

- Naveen

#44 lore raw

Re: [PATCH v3 17/62] KVM: SVM: Add enable_ipiv param, never set IsRunning if disabled

Quoted 13 lines · Naveen N Rao
Quoted 12 lines · Sean Christopherson
arch/x86/kvm/svm/avic.c +1 -1
@@ -1030,7 +1047,7 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 	 * can't be scheduled out and thus avic_vcpu_{put,load}() can't run
 	 * recursively.
 	 */
-	entry = READ_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id]);
+	entry = svm->avic_physical_id_entry;
 
 	/* Nothing to do if IsRunning == '0' due to vCPU blocking. */
 	if (!(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK))
@@ -1049,7 +1066,10 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 	avic_update_iommu_vcpu_affinity(vcpu, -1, 0);
 
Quoted 10 lines

arch/x86/kvm/svm/avic.c void avic_vcpu_put(struct kvm_vcpu *vcpu) · line 1072

Quoted 6 lines
arch/x86/kvm/svm/avic.c +4 -1
 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
-	WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
+	svm->avic_physical_id_entry = entry;
+
+	if (enable_ipiv)
+		WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);

If enable_ipiv is false, then isRunning bit will never be set and we would have bailed out earlier. So, the check for enable_ipiv can be dropped here (or converted into an assert).

Ignore this, I got this wrong, sorry. The earlier check is against the local copy of the physical ID table entry, which will indeed have isRunning set so this is all good.

- Naveen

author maintainer
#45 lore raw

Re: [PATCH v3 17/62] KVM: SVM: Add enable_ipiv param, never set IsRunning if disabled

In reply to Naveen N Rao
Quoted 11 lines · Naveen N Rao
Quoted 7 lines · Sean Christopherson

From: Maxim Levitsky <[email protected]>

Let userspace "disable" IPI virtualization for AVIC via the enable_ipiv module param, by never setting IsRunning. SVM doesn't provide a way to disable IPI virtualization in hardware, but by ensuring CPUs never see IsRunning=1, every IPI in the guest (except for self-IPIs) will generate a VM-Exit.

I think this is good to have regardless of the erratum. Not sure about VMX, but does it make sense to intercept writes to the self-ipi MSR as well?

That doesn't work for AVIC, i.e. if the guest is MMIO to access the virtual APIC.

Regardless, I don't see any reason to manually intercept self-IPIs when IPI virtualization is disabled. AFAIK, there's no need to do so for correctness, and Intel's self-IPI virtualization isn't tied to IPI virtualization either. Self-IPI virtualization is enabled by virtual interrupt delivery, which in turn is enabled by KVM when enable_apicv is true:

Self-IPI virtualization occurs only if the “virtual-interrupt delivery” VM-execution control is 1.

Reviewed-by
#46 lore raw

Re: [PATCH v3 17/62] KVM: SVM: Add enable_ipiv param, never set IsRunning if disabled

Quoted 10 lines · Sean Christopherson
Quoted 9 lines · Naveen N Rao
Quoted 7 lines · Sean Christopherson

From: Maxim Levitsky <[email protected]>

Let userspace "disable" IPI virtualization for AVIC via the enable_ipiv module param, by never setting IsRunning. SVM doesn't provide a way to disable IPI virtualization in hardware, but by ensuring CPUs never see IsRunning=1, every IPI in the guest (except for self-IPIs) will generate a VM-Exit.

Quoted 4 lines
Quoted 2 lines

I think this is good to have regardless of the erratum. Not sure about VMX, but does it make sense to intercept writes to the self-ipi MSR as well?

That doesn't work for AVIC, i.e. if the guest is MMIO to access the virtual APIC.

Right, I was thinking about the Self-IPI MSR, but the ICR will also need to be intercepted.

Quoted 9 lines

Regardless, I don't see any reason to manually intercept self-IPIs when IPI virtualization is disabled. AFAIK, there's no need to do so for correctness, and Intel's self-IPI virtualization isn't tied to IPI virtualization either. Self-IPI virtualization is enabled by virtual interrupt delivery, which in turn is enabled by KVM when enable_apicv is true:

Self-IPI virtualization occurs only if the “virtual-interrupt delivery” VM-execution control is 1.

Excellent, for this patch:

Thanks,
Naveen

author maintainer
#47 lore raw

PATCH v3 18/62 KVM: SVM: Disable (x2)AVIC IPI virtualization if CPU has erratum #1235

In reply to Sean Christopherson

From: Maxim Levitsky <[email protected]>

Disable IPI virtualization on AMD Family 17h CPUs (Zen2 and Zen1), as hardware doesn't reliably detect changes to the 'IsRunning' bit during ICR write emulation, and might fail to VM-Exit on the sending vCPU, if IsRunning was recently cleared.

The absence of the VM-Exit leads to KVM not waking (or triggering nested VM-Exit) of the target vCPU(s) of the IPI, which can lead to hung vCPUs, unbounded delays in L2 execution, etc.

To workaround the erratum, simply disable IPI virtualization, which prevents KVM from setting IsRunning and thus eliminates the race where hardware sees a stale IsRunning=1. As a result, all ICR writes (except when "Self" shorthand is used) will VM-Exit and therefore be correctly emulated by KVM.

Disabling IPI virtualization does carry a performance penalty, but benchmarkng shows that enabling AVIC without IPI virtualization is still much better than not using AVIC at all, because AVIC still accelerates posted interrupts and the receiving end of the IPIs.

Note, when virtualizaing Self-IPIs, the CPU skips reading the physical ID table and updates the vIRR directly (because the vCPU is by definition actively running), i.e. Self-IPI isn't susceptible to the erratum *and* is still accelerated by hardware.

[sean: rebase, massage changelog, disallow user override]
---
 arch/x86/kvm/svm/avic.c | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/x86/kvm/svm/avic.c +8 -0
index 48c737e1200a..bf8b59556373 100644
@@ -1187,6 +1187,14 @@ bool avic_hardware_setup(void)
 	if (x2avic_enabled)
 		pr_info("x2AVIC enabled\n");
 
+	/*
+	 * Disable IPI virtualization for AMD Family 17h CPUs (Zen1 and Zen2)
+	 * due to erratum 1235, which results in missed GA log events and thus
+	 * missed wake events for blocking vCPUs due to the CPU failing to see
+	 * a software update to clear IsRunning.
+	 */
+	enable_ipiv = enable_ipiv && boot_cpu_data.x86 != 0x17;
+
 	amd_iommu_register_ga_log_notifier(&avic_ga_log_notifier);
 
 	return true;
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
Acked-by on v3 18/62
#48 lore raw

Re: [PATCH v3 18/62] KVM: SVM: Disable (x2)AVIC IPI virtualization if CPU has erratum #1235

Quoted 9 lines · Sean Christopherson

From: Maxim Levitsky <[email protected]>

Disable IPI virtualization on AMD Family 17h CPUs (Zen2 and Zen1), as hardware doesn't reliably detect changes to the 'IsRunning' bit during ICR write emulation, and might fail to VM-Exit on the sending vCPU, if IsRunning was recently cleared.

The absence of the VM-Exit leads to KVM not waking (or triggering nested VM-Exit) of the target vCPU(s) of the IPI, which can lead to hung vCPUs,

^^^^^^^^^^^
VM-Exit of)

Quoted 8 lines

unbounded delays in L2 execution, etc.

To workaround the erratum, simply disable IPI virtualization, which prevents KVM from setting IsRunning and thus eliminates the race where hardware sees a stale IsRunning=1. As a result, all ICR writes (except when "Self" shorthand is used) will VM-Exit and therefore be correctly emulated by KVM.

Quoted 6 lines

Disabling IPI virtualization does carry a performance penalty, but benchmarkng shows that enabling AVIC without IPI virtualization is still much better than not using AVIC at all, because AVIC still accelerates posted interrupts and the receiving end of the IPIs.

Note, when virtualizaing Self-IPIs, the CPU skips reading the physical ID

	     ^^^^^^^^^^^^^
	     virtualizing
Quoted 16 lines

table and updates the vIRR directly (because the vCPU is by definition actively running), i.e. Self-IPI isn't susceptible to the erratum *and* is still accelerated by hardware.

[sean: rebase, massage changelog, disallow user override]
---
 arch/x86/kvm/svm/avic.c | 8 ++++++++
 1 file changed, 8 insertions(+)
arch/x86/kvm/svm/avic.c +0 -0
index 48c737e1200a..bf8b59556373 100644

arch/x86/kvm/svm/avic.c bool avic_hardware_setup(void) · line 1192

Quoted 6 lines
arch/x86/kvm/svm/avic.c +3 -0
@@ -1187,6 +1187,14 @@ bool avic_hardware_setup(void)
 	if (x2avic_enabled)
 		pr_info("x2AVIC enabled\n");
 
+	/*
+	 * Disable IPI virtualization for AMD Family 17h CPUs (Zen1 and Zen2)
+	 * due to erratum 1235, which results in missed GA log events and thus
							^^^^^^^^^^^^^

Not sure I understand the reference to GA log events here -- those are only for device interrupts and not IPIs.

arch/x86/kvm/svm/avic.c bool avic_hardware_setup(void) · line 1197

Quoted 5 lines
arch/x86/kvm/svm/avic.c +5 -0
+	 * missed wake events for blocking vCPUs due to the CPU failing to see
+	 * a software update to clear IsRunning.
+	 */
+	enable_ipiv = enable_ipiv && boot_cpu_data.x86 != 0x17;
+

Apart from the above, this LGTM.

- Naveen

author maintainer
#49 lore raw

Re: [PATCH v3 18/62] KVM: SVM: Disable (x2)AVIC IPI virtualization if CPU has erratum #1235

Quoted 9 lines · Naveen N Rao
Quoted 8 lines · Sean Christopherson
arch/x86/kvm/svm/avic.c +0 -0
index 48c737e1200a..bf8b59556373 100644
@@ -1187,6 +1187,14 @@ bool avic_hardware_setup(void)
 	if (x2avic_enabled)
 		pr_info("x2AVIC enabled\n");
 
Quoted 6 lines

arch/x86/kvm/svm/avic.c bool avic_hardware_setup(void) · line 1192

Quoted 3 lines
arch/x86/kvm/svm/avic.c +3 -0
+	/*
+	 * Disable IPI virtualization for AMD Family 17h CPUs (Zen1 and Zen2)
+	 * due to erratum 1235, which results in missed GA log events and thus
							^^^^^^^^^^^^^

Not sure I understand the reference to GA log events here -- those are only for device interrupts and not IPIs.

Doh, you're absolutely right. I'll fix to this when applying:

/*
* Disable IPI virtualization for AMD Family 17h CPUs (Zen1 and Zen2)
* due to erratum 1235, which results in missed VM-Exits on the sender
* and thus missed wake events for blocking vCPUs due to the CPU
* failing to see a software update to clear IsRunning. */

author maintainer
#50 lore raw

PATCH v3 19/62 KVM: VMX: Suppress PI notifications whenever the vCPU is put

In reply to Sean Christopherson

Suppress posted interrupt notifications (set PID.SN=1) whenever the vCPU is put, i.e. unloaded, not just when the vCPU is preempted, as KVM doesn't do anything in response to a notification IRQ that arrives in the host, nor does KVM rely on the Outstanding Notification (PID.ON) flag when the vCPU is unloaded. And, the cost of scanning the PIR to manually set PID.ON when loading the vCPU is quite small, especially relative to the cost of loading (and unloading) a vCPU.

On the flip side, leaving SN clear means a notification for the vCPU will result in a spurious IRQ for the pCPU, even if vCPU task is scheduled out, running in userspace, etc. Even worse, if the pCPU is running a different vCPU, the spurious IRQ could trigger posted interrupt processing for the wrong vCPU, which is technically a violation of the architecture, as setting bits in PIR aren't supposed to be propagated to the vIRR until a notification IRQ is received.

The saving grace of the current behavior is that hardware sends notification interrupts if and only if PID.ON=0, i.e. only the first posted interrupt for a vCPU will trigger a spurious IRQ (for each window where the vCPU is unloaded).

Ideally, KVM would suppress notifications before enabling IRQs in the VM-Exit, but KVM relies on PID.ON as an indicator that there is a posted interrupt pending in PIR, e.g. in vmx_sync_pir_to_irr(), and sadly there is no way to ask hardware to set PID.ON, but not generate an interrupt. That could be solved by using pi_has_pending_interrupt() instead of checking only PID.ON, but it's not at all clear that would be a performance win, as KVM would end up scanning the entire PIR whenever an interrupt isn't pending.

And long term, the spurious IRQ window, i.e. where a vCPU is loaded with IRQs enabled, can effectively be made smaller for hot paths by moving performance critical VM-Exit handlers into the fastpath, i.e. by never enabling IRQs for hot path VM-Exits.

---
 arch/x86/kvm/vmx/posted_intr.c | 29 ++++++++++++++++-------------
 1 file changed, 16 insertions(+), 13 deletions(-)
arch/x86/kvm/vmx/posted_intr.c +16 -13
index 110fb19848ab..d4826a6b674f 100644
@@ -73,13 +73,10 @@ void vmx_vcpu_pi_load(struct kvm_vcpu *vcpu, int cpu)
 	/*
 	 * If the vCPU wasn't on the wakeup list and wasn't migrated, then the
 	 * full update can be skipped as neither the vector nor the destination
-	 * needs to be changed.
+	 * needs to be changed.  Clear SN even if there is no assigned device,
+	 * again for simplicity.
 	 */
 	if (pi_desc->nv != POSTED_INTR_WAKEUP_VECTOR && vcpu->cpu == cpu) {
-		/*
-		 * Clear SN if it was set due to being preempted.  Again, do
-		 * this even if there is no assigned device for simplicity.
-		 */
 		if (pi_test_and_clear_sn(pi_desc))
 			goto after_clear_sn;
 		return;
@@ -225,17 +222,23 @@ void vmx_vcpu_pi_put(struct kvm_vcpu *vcpu)
 	if (!vmx_needs_pi_wakeup(vcpu))
 		return;
 
-	if (kvm_vcpu_is_blocking(vcpu) &&
+	/*
+	 * If the vCPU is blocking with IRQs enabled and ISN'T being preempted,
+	 * enable the wakeup handler so that notification IRQ wakes the vCPU as
+	 * expected.  There is no need to enable the wakeup handler if the vCPU
+	 * is preempted between setting its wait state and manually scheduling
+	 * out, as the task is still runnable, i.e. doesn't need a wake event
+	 * from KVM to be scheduled in.
+	 *
+	 * If the wakeup handler isn't being enabled, Suppress Notifications as
+	 * the cost of propagating PIR.IRR to PID.ON is negligible compared to
+	 * the cost of a spurious IRQ, and vCPU put/load is a slow path.
+	 */
+	if (!vcpu->preempted && kvm_vcpu_is_blocking(vcpu) &&
 	    ((is_td_vcpu(vcpu) && tdx_interrupt_allowed(vcpu)) ||
 	     (!is_td_vcpu(vcpu) && !vmx_interrupt_blocked(vcpu))))
 		pi_enable_wakeup_handler(vcpu);
-
-	/*
-	 * Set SN when the vCPU is preempted.  Note, the vCPU can both be seen
-	 * as blocking and preempted, e.g. if it's preempted between setting
-	 * its wait state and manually scheduling out.
-	 */
-	if (vcpu->preempted)
+	else
 		pi_set_sn(pi_desc);
 }
 
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#51 lore raw

PATCH v3 20/62 KVM: SVM: Add a comment to explain why avic_vcpu_blocking() ignores IRQ blocking

In reply to Sean Christopherson

Add a comment to explain why KVM clears IsRunning when putting a vCPU, even though leaving IsRunning=1 would be ok from a functional perspective. Per Maxim's experiments, a misbehaving VM could spam the AVIC doorbell so fast as to induce a 50%+ loss in performance.

---
 arch/x86/kvm/svm/avic.c | 31 ++++++++++++++++++-------------
 1 file changed, 18 insertions(+), 13 deletions(-)
arch/x86/kvm/svm/avic.c +18 -13
index bf8b59556373..3cf929ac117f 100644
@@ -1121,19 +1121,24 @@ void avic_vcpu_blocking(struct kvm_vcpu *vcpu)
 	if (!kvm_vcpu_apicv_active(vcpu))
 		return;
 
-       /*
-        * Unload the AVIC when the vCPU is about to block, _before_
-        * the vCPU actually blocks.
-        *
-        * Any IRQs that arrive before IsRunning=0 will not cause an
-        * incomplete IPI vmexit on the source, therefore vIRR will also
-        * be checked by kvm_vcpu_check_block() before blocking.  The
-        * memory barrier implicit in set_current_state orders writing
-        * IsRunning=0 before reading the vIRR.  The processor needs a
-        * matching memory barrier on interrupt delivery between writing
-        * IRR and reading IsRunning; the lack of this barrier might be
-        * the cause of errata #1235).
-        */
+	/*
+	 * Unload the AVIC when the vCPU is about to block, _before_ the vCPU
+	 * actually blocks.
+	 *
+	 * Note, any IRQs that arrive before IsRunning=0 will not cause an
+	 * incomplete IPI vmexit on the source; kvm_vcpu_check_block() handles
+	 * this by checking vIRR one last time before blocking.  The memory
+	 * barrier implicit in set_current_state orders writing IsRunning=0
+	 * before reading the vIRR.  The processor needs a matching memory
+	 * barrier on interrupt delivery between writing IRR and reading
+	 * IsRunning; the lack of this barrier might be the cause of errata #1235).
+	 *
+	 * Clear IsRunning=0 even if guest IRQs are disabled, i.e. even if KVM
+	 * doesn't need to detect events for scheduling purposes.  The doorbell
+	 * used to signal running vCPUs cannot be blocked, i.e. will perturb the
+	 * CPU and cause noisy neighbor problems if the VM is sending interrupts
+	 * to the vCPU while it's scheduled out.
+	 */
 	avic_vcpu_put(vcpu);
 }
 
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
Acked-by on v3 20/62
#52 lore raw

Re: [PATCH v3 20/62] KVM: SVM: Add a comment to explain why avic_vcpu_blocking() ignores IRQ blocking

Quoted 37 lines · Sean Christopherson

Add a comment to explain why KVM clears IsRunning when putting a vCPU, even though leaving IsRunning=1 would be ok from a functional perspective. Per Maxim's experiments, a misbehaving VM could spam the AVIC doorbell so fast as to induce a 50%+ loss in performance.

---
 arch/x86/kvm/svm/avic.c | 31 ++++++++++++++++++-------------
 1 file changed, 18 insertions(+), 13 deletions(-)
arch/x86/kvm/svm/avic.c +4 -13
index bf8b59556373..3cf929ac117f 100644
@@ -1121,19 +1121,24 @@ void avic_vcpu_blocking(struct kvm_vcpu *vcpu)
 	if (!kvm_vcpu_apicv_active(vcpu))
 		return;
 
-       /*
-        * Unload the AVIC when the vCPU is about to block, _before_
-        * the vCPU actually blocks.
-        *
-        * Any IRQs that arrive before IsRunning=0 will not cause an
-        * incomplete IPI vmexit on the source, therefore vIRR will also
-        * be checked by kvm_vcpu_check_block() before blocking.  The
-        * memory barrier implicit in set_current_state orders writing
-        * IsRunning=0 before reading the vIRR.  The processor needs a
-        * matching memory barrier on interrupt delivery between writing
-        * IRR and reading IsRunning; the lack of this barrier might be
-        * the cause of errata #1235).
-        */
+	/*
+	 * Unload the AVIC when the vCPU is about to block, _before_ the vCPU
+	 * actually blocks.
+	 *

arch/x86/kvm/svm/avic.c void avic_vcpu_blocking(struct kvm_vcpu *vcpu) · line 1137

Quoted 10 lines
arch/x86/kvm/svm/avic.c +10 -0
+	 * Note, any IRQs that arrive before IsRunning=0 will not cause an
+	 * incomplete IPI vmexit on the source; kvm_vcpu_check_block() handles
+	 * this by checking vIRR one last time before blocking.  The memory
+	 * barrier implicit in set_current_state orders writing IsRunning=0
+	 * before reading the vIRR.  The processor needs a matching memory
+	 * barrier on interrupt delivery between writing IRR and reading
+	 * IsRunning; the lack of this barrier might be the cause of errata #1235).
+	 *
+	 * Clear IsRunning=0 even if guest IRQs are disabled, i.e. even if KVM
+	 * doesn't need to detect events for scheduling purposes.  The doorbell

Nit: just IsRunning (you can drop the =0 part).

Trying to understand the significance of IRQs being disabled here. Is that a path KVM tries to optimize? Theoretically, it looks like we want to clear IsRunning regardless of whether the vCPU is blocked so as to prevent the guest from spamming the host with AVIC doorbells -- compared to always keeping IsRunning set so as to speed up VM entry/exit.

arch/x86/kvm/svm/avic.c void avic_vcpu_blocking(struct kvm_vcpu *vcpu) · line 1143

Quoted 6 lines
arch/x86/kvm/svm/avic.c +4 -0
+	 * used to signal running vCPUs cannot be blocked, i.e. will perturb the
+	 * CPU and cause noisy neighbor problems if the VM is sending interrupts
+	 * to the vCPU while it's scheduled out.
+	 */
 	avic_vcpu_put(vcpu);
 }

Otherwise, this LGTM.

Thanks,
Naveen

author maintainer
#53 lore raw

Re: [PATCH v3 20/62] KVM: SVM: Add a comment to explain why avic_vcpu_blocking() ignores IRQ blocking

Quoted 42 lines · Naveen N Rao
Quoted 41 lines · Sean Christopherson

Add a comment to explain why KVM clears IsRunning when putting a vCPU, even though leaving IsRunning=1 would be ok from a functional perspective. Per Maxim's experiments, a misbehaving VM could spam the AVIC doorbell so fast as to induce a 50%+ loss in performance.

---
 arch/x86/kvm/svm/avic.c | 31 ++++++++++++++++++-------------
 1 file changed, 18 insertions(+), 13 deletions(-)
arch/x86/kvm/svm/avic.c +8 -13
index bf8b59556373..3cf929ac117f 100644
@@ -1121,19 +1121,24 @@ void avic_vcpu_blocking(struct kvm_vcpu *vcpu)
 	if (!kvm_vcpu_apicv_active(vcpu))
 		return;
 
-       /*
-        * Unload the AVIC when the vCPU is about to block, _before_
-        * the vCPU actually blocks.
-        *
-        * Any IRQs that arrive before IsRunning=0 will not cause an
-        * incomplete IPI vmexit on the source, therefore vIRR will also
-        * be checked by kvm_vcpu_check_block() before blocking.  The
-        * memory barrier implicit in set_current_state orders writing
-        * IsRunning=0 before reading the vIRR.  The processor needs a
-        * matching memory barrier on interrupt delivery between writing
-        * IRR and reading IsRunning; the lack of this barrier might be
-        * the cause of errata #1235).
-        */
+	/*
+	 * Unload the AVIC when the vCPU is about to block, _before_ the vCPU
+	 * actually blocks.
+	 *
+	 * Note, any IRQs that arrive before IsRunning=0 will not cause an
+	 * incomplete IPI vmexit on the source; kvm_vcpu_check_block() handles
+	 * this by checking vIRR one last time before blocking.  The memory
+	 * barrier implicit in set_current_state orders writing IsRunning=0
Quoted 8 lines

arch/x86/kvm/svm/avic.c void avic_vcpu_blocking(struct kvm_vcpu *vcpu) · line 1137

Quoted 6 lines
arch/x86/kvm/svm/avic.c +6 -0
+	 * before reading the vIRR.  The processor needs a matching memory
+	 * barrier on interrupt delivery between writing IRR and reading
+	 * IsRunning; the lack of this barrier might be the cause of errata #1235).
+	 *
+	 * Clear IsRunning=0 even if guest IRQs are disabled, i.e. even if KVM
+	 * doesn't need to detect events for scheduling purposes.  The doorbell

Nit: just IsRunning (you can drop the =0 part).

Hmm, not really. It could be:

	/* Note, any IRQs that arrive while IsRunning is set will not cause an

or

	/* Note, any IRQs that arrive while IsRunning=1 will not cause an

but that's just regurgitating the spec. The slightly more interesting scenario that's being described here is what will happen if an IRQ arrives _just_ before the below code toggle IsRunning from 1 => 0.

Quoted 2 lines

Trying to understand the significance of IRQs being disabled here. Is that a path KVM tries to optimize?

Yep. KVM doesn't need a notification for the undelivered (virtual) IRQ, because it won't be handled by the vCPU until the vCPU enables IRQs, and thus it's not a valid wake event for the vCPU.

So, *if* spurious doorbells didn't affect performance or functionality, then ideally KVM would leave IsRunning=1, e.g. so that the IOMMU doesn't need to generate GA log events, and so that other vCPUs aren't forced to VM-Exit when sending an IPI. Unfortunately, spurious doorbells are quite intrusive, and so KVM "needs" to clear IsRunning.

Quoted 4 lines

Theoretically, it looks like we want to clear IsRunning regardless of whether the vCPU is blocked so as to prevent the guest from spamming the host with AVIC doorbells -- compared to always keeping IsRunning set so as to speed up VM entry/exit.

Yep, exactly.

Quoted 14 lines

arch/x86/kvm/svm/avic.c void avic_vcpu_blocking(struct kvm_vcpu *vcpu) · line 1143

Quoted 6 lines
arch/x86/kvm/svm/avic.c +4 -0
+	 * used to signal running vCPUs cannot be blocked, i.e. will perturb the
+	 * CPU and cause noisy neighbor problems if the VM is sending interrupts
+	 * to the vCPU while it's scheduled out.
+	 */
 	avic_vcpu_put(vcpu);
 }

Otherwise, this LGTM.

Thanks,
Naveen

#54 lore raw

Re: [PATCH v3 20/62] KVM: SVM: Add a comment to explain why avic_vcpu_blocking() ignores IRQ blocking

Quoted 56 lines · Sean Christopherson
Quoted 42 lines · Naveen N Rao
Quoted 41 lines · Sean Christopherson

Add a comment to explain why KVM clears IsRunning when putting a vCPU, even though leaving IsRunning=1 would be ok from a functional perspective. Per Maxim's experiments, a misbehaving VM could spam the AVIC doorbell so fast as to induce a 50%+ loss in performance.

---
 arch/x86/kvm/svm/avic.c | 31 ++++++++++++++++++-------------
 1 file changed, 18 insertions(+), 13 deletions(-)
arch/x86/kvm/svm/avic.c +8 -13
index bf8b59556373..3cf929ac117f 100644
@@ -1121,19 +1121,24 @@ void avic_vcpu_blocking(struct kvm_vcpu *vcpu)
 	if (!kvm_vcpu_apicv_active(vcpu))
 		return;
 
-       /*
-        * Unload the AVIC when the vCPU is about to block, _before_
-        * the vCPU actually blocks.
-        *
-        * Any IRQs that arrive before IsRunning=0 will not cause an
-        * incomplete IPI vmexit on the source, therefore vIRR will also
-        * be checked by kvm_vcpu_check_block() before blocking.  The
-        * memory barrier implicit in set_current_state orders writing
-        * IsRunning=0 before reading the vIRR.  The processor needs a
-        * matching memory barrier on interrupt delivery between writing
-        * IRR and reading IsRunning; the lack of this barrier might be
-        * the cause of errata #1235).
-        */
+	/*
+	 * Unload the AVIC when the vCPU is about to block, _before_ the vCPU
+	 * actually blocks.
+	 *
+	 * Note, any IRQs that arrive before IsRunning=0 will not cause an
+	 * incomplete IPI vmexit on the source; kvm_vcpu_check_block() handles
+	 * this by checking vIRR one last time before blocking.  The memory
+	 * barrier implicit in set_current_state orders writing IsRunning=0
Quoted 8 lines

arch/x86/kvm/svm/avic.c void avic_vcpu_blocking(struct kvm_vcpu *vcpu) · line 1137

Quoted 6 lines
arch/x86/kvm/svm/avic.c +6 -0
+	 * before reading the vIRR.  The processor needs a matching memory
+	 * barrier on interrupt delivery between writing IRR and reading
+	 * IsRunning; the lack of this barrier might be the cause of errata #1235).
+	 *
+	 * Clear IsRunning=0 even if guest IRQs are disabled, i.e. even if KVM
+	 * doesn't need to detect events for scheduling purposes.  The doorbell

Nit: just IsRunning (you can drop the =0 part).

Hmm, not really. It could be:

	/* Note, any IRQs that arrive while IsRunning is set will not cause an
Quoted 7 lines

or

	/* Note, any IRQs that arrive while IsRunning=1 will not cause an

but that's just regurgitating the spec. The slightly more interesting scenario that's being described here is what will happen if an IRQ arrives _just_ before the below code toggle IsRunning from 1 => 0.

Oh, you're right. I was referring to the last paragraph starting with "Clear IsRunning=0 even if", which to me feels like a double negation. It reads better if it is "Clear IsRunning even if ..."

Quoted 8 lines
Quoted 3 lines

Trying to understand the significance of IRQs being disabled here. Is that a path KVM tries to optimize?

Yep. KVM doesn't need a notification for the undelivered (virtual) IRQ, because it won't be handled by the vCPU until the vCPU enables IRQs, and thus it's not a valid wake event for the vCPU.

Quoted 5 lines

So, *if* spurious doorbells didn't affect performance or functionality, then ideally KVM would leave IsRunning=1, e.g. so that the IOMMU doesn't need to generate GA log events, and so that other vCPUs aren't forced to VM-Exit when sending an IPI. Unfortunately, spurious doorbells are quite intrusive, and so KVM "needs" to clear IsRunning.

Ok, got that. kvm_arch_vcpu_blocking() looks like it exists precisely for this reason.

Thanks,
Naveen

author maintainer
#55 lore raw

PATCH v3 21/62 iommu/amd: KVM: SVM: Use pi_desc_addr to derive ga_root_ptr

In reply to Sean Christopherson

Use vcpu_data.pi_desc_addr instead of amd_iommu_pi_data.base to get the GA root pointer. KVM is the only source of amd_iommu_pi_data.base, and KVM's one and only path for writing amd_iommu_pi_data.base computes the exact same value for vcpu_data.pi_desc_addr and amd_iommu_pi_data.base, and fills amd_iommu_pi_data.base if and only if vcpu_data.pi_desc_addr is valid, i.e. amd_iommu_pi_data.base is fully redundant.

---
 arch/x86/kvm/svm/avic.c   | 7 +++++--
 drivers/iommu/amd/iommu.c | 2 +-
 include/linux/amd-iommu.h | 2 --
 3 files changed, 6 insertions(+), 5 deletions(-)
arch/x86/kvm/svm/avic.c +5 -2
index 3cf929ac117f..461300bc5608 100644
@@ -893,8 +893,11 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 
 			enable_remapped_mode = false;
 
-			/* Try to enable guest_mode in IRTE */
-			pi.base = avic_get_backing_page_address(svm);
+			/*
+			 * Try to enable guest_mode in IRTE.  Note, the address
+			 * of the vCPU's AVIC backing page is passed to the
+			 * IOMMU via vcpu_info->pi_desc_addr.
+			 */
 			pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
 						     svm->vcpu.vcpu_id);
 			pi.is_guest_mode = true;
drivers/iommu/amd/iommu.c +1 -1
index f23635b062f0..512167f7aef4 100644
@@ -3888,7 +3888,7 @@ static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
 	pi_data->ir_data = ir_data;
 
 	if (pi_data->is_guest_mode) {
-		ir_data->ga_root_ptr = (pi_data->base >> 12);
+		ir_data->ga_root_ptr = (vcpu_pi_info->pi_desc_addr >> 12);
 		ir_data->ga_vector = vcpu_pi_info->vector;
 		ir_data->ga_tag = pi_data->ga_tag;
 		ret = amd_iommu_activate_guest_mode(ir_data);
include/linux/amd-iommu.h +0 -2
index 1f9b13d803c5..deeefc92a5cf 100644
@@ -19,8 +19,6 @@ struct amd_iommu;
  */
 struct amd_iommu_pi_data {
 	u32 ga_tag;
-	u64 base;
-
 	bool is_guest_mode;
 	struct vcpu_data *vcpu_data;
 	void *ir_data;
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#56 lore raw

PATCH v3 22/62 iommu/amd: KVM: SVM: Pass NULL @vcpu_info to indicate "not guest mode"

In reply to Sean Christopherson

Pass NULL to amd_ir_set_vcpu_affinity() to communicate "don't post to a vCPU" now that there's no need to communicate information back to KVM about the previous vCPU (KVM does its own tracking).

---
 arch/x86/kvm/svm/avic.c   | 16 ++++------------
 drivers/iommu/amd/iommu.c | 10 +++++++---
 2 files changed, 11 insertions(+), 15 deletions(-)
arch/x86/kvm/svm/avic.c +4 -12
index 461300bc5608..6260bf3697ba 100644
@@ -927,18 +927,10 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		}
 	}
 
-	ret = 0;
-	if (enable_remapped_mode) {
-		/* Use legacy mode in IRTE */
-		struct amd_iommu_pi_data pi;
-
-		/**
-		 * Here, pi is used to:
-		 * - Tell IOMMU to use legacy mode for this interrupt.
-		 */
-		pi.is_guest_mode = false;
-		ret = irq_set_vcpu_affinity(host_irq, &pi);
-	}
+	if (enable_remapped_mode)
+		ret = irq_set_vcpu_affinity(host_irq, NULL);
+	else
+		ret = 0;
 out:
 	srcu_read_unlock(&kvm->irq_srcu, idx);
 	return ret;
drivers/iommu/amd/iommu.c +7 -3
index 512167f7aef4..5141507587e1 100644
@@ -3864,7 +3864,6 @@ static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
 {
 	int ret;
 	struct amd_iommu_pi_data *pi_data = vcpu_info;
-	struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
 	struct amd_ir_data *ir_data = data->chip_data;
 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
 	struct iommu_dev_data *dev_data;
@@ -3885,9 +3884,14 @@ static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
 		return -EINVAL;
 
 	ir_data->cfg = irqd_cfg(data);
-	pi_data->ir_data = ir_data;
 
-	if (pi_data->is_guest_mode) {
+	if (pi_data) {
+		struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
+
+		pi_data->ir_data = ir_data;
+
+		WARN_ON_ONCE(!pi_data->is_guest_mode);
+
 		ir_data->ga_root_ptr = (vcpu_pi_info->pi_desc_addr >> 12);
 		ir_data->ga_vector = vcpu_pi_info->vector;
 		ir_data->ga_tag = pi_data->ga_tag;
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#57 lore raw

PATCH v3 23/62 KVM: SVM: Stop walking list of routing table entries when updating IRTE

In reply to Sean Christopherson

Now that KVM explicitly passes the new/current GSI routing to pi_update_irte(), simply use the provided routing entry and stop walking the routing table to find that entry. KVM, via setup_routing_entry() and sanity checked by kvm_get_msi_route(), disallows having a GSI configured to trigger multiple MSIs.

I.e. this is subtly a glorified nop, as KVM allows at most one MSI per GSI, the for-loop can only ever process one entry, and that entry is the new/current entry (see the WARN_ON_ONCE() added by "KVM: x86: Pass new routing entries and irqfd when updating IRTEs" to ensure @new matches the entry found in the routing table).

---
 arch/x86/kvm/svm/avic.c | 109 ++++++++++++++++------------------------
 1 file changed, 44 insertions(+), 65 deletions(-)
arch/x86/kvm/svm/avic.c +44 -65
index 6260bf3697ba..a83769bb8123 100644
@@ -843,11 +843,10 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			unsigned int host_irq, uint32_t guest_irq,
 			struct kvm_kernel_irq_routing_entry *new)
 {
-	struct kvm_kernel_irq_routing_entry *e;
-	struct kvm_irq_routing_table *irq_rt;
 	bool enable_remapped_mode = true;
-	bool set = !!new;
-	int idx, ret = 0;
+	struct vcpu_data vcpu_info;
+	struct vcpu_svm *svm = NULL;
+	int ret = 0;
 
 	if (!kvm_arch_has_assigned_device(kvm) || !kvm_arch_has_irq_bypass())
 		return 0;
@@ -859,72 +858,53 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 	svm_ir_list_del(irqfd);
 
 	pr_debug("SVM: %s: host_irq=%#x, guest_irq=%#x, set=%#x\n",
-		 __func__, host_irq, guest_irq, set);
+		 __func__, host_irq, guest_irq, !!new);
 
-	idx = srcu_read_lock(&kvm->irq_srcu);
-	irq_rt = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu);
+	/**
+	 * Here, we setup with legacy mode in the following cases:
+	 * 1. When cannot target interrupt to a specific vcpu.
+	 * 2. Unsetting posted interrupt.
+	 * 3. APIC virtualization is disabled for the vcpu.
+	 * 4. IRQ has incompatible delivery mode (SMI, INIT, etc)
+	 */
+	if (new && new->type == KVM_IRQ_ROUTING_MSI &&
+	    !get_pi_vcpu_info(kvm, new, &vcpu_info, &svm) &&
+	    kvm_vcpu_apicv_active(&svm->vcpu)) {
+		struct amd_iommu_pi_data pi;
 
-	if (guest_irq >= irq_rt->nr_rt_entries ||
-		hlist_empty(&irq_rt->map[guest_irq])) {
-		pr_warn_once("no route for guest_irq %u/%u (broken user space?)\n",
-			     guest_irq, irq_rt->nr_rt_entries);
-		goto out;
-	}
+		enable_remapped_mode = false;
 
-	hlist_for_each_entry(e, &irq_rt->map[guest_irq], link) {
-		struct vcpu_data vcpu_info;
-		struct vcpu_svm *svm = NULL;
-
-		if (e->type != KVM_IRQ_ROUTING_MSI)
-			continue;
-
-		WARN_ON_ONCE(new && memcmp(e, new, sizeof(*new)));
+		/*
+		 * Try to enable guest_mode in IRTE.  Note, the address
+		 * of the vCPU's AVIC backing page is passed to the
+		 * IOMMU via vcpu_info->pi_desc_addr.
+		 */
+		pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
+					     svm->vcpu.vcpu_id);
+		pi.is_guest_mode = true;
+		pi.vcpu_data = &vcpu_info;
+		ret = irq_set_vcpu_affinity(host_irq, &pi);
 
 		/**
-		 * Here, we setup with legacy mode in the following cases:
-		 * 1. When cannot target interrupt to a specific vcpu.
-		 * 2. Unsetting posted interrupt.
-		 * 3. APIC virtualization is disabled for the vcpu.
-		 * 4. IRQ has incompatible delivery mode (SMI, INIT, etc)
+		 * Here, we successfully setting up vcpu affinity in
+		 * IOMMU guest mode. Now, we need to store the posted
+		 * interrupt information in a per-vcpu ir_list so that
+		 * we can reference to them directly when we update vcpu
+		 * scheduling information in IOMMU irte.
 		 */
-		if (!get_pi_vcpu_info(kvm, e, &vcpu_info, &svm) && set &&
-		    kvm_vcpu_apicv_active(&svm->vcpu)) {
-			struct amd_iommu_pi_data pi;
-
-			enable_remapped_mode = false;
-
-			/*
-			 * Try to enable guest_mode in IRTE.  Note, the address
-			 * of the vCPU's AVIC backing page is passed to the
-			 * IOMMU via vcpu_info->pi_desc_addr.
-			 */
-			pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
-						     svm->vcpu.vcpu_id);
-			pi.is_guest_mode = true;
-			pi.vcpu_data = &vcpu_info;
-			ret = irq_set_vcpu_affinity(host_irq, &pi);
-
-			/**
-			 * Here, we successfully setting up vcpu affinity in
-			 * IOMMU guest mode. Now, we need to store the posted
-			 * interrupt information in a per-vcpu ir_list so that
-			 * we can reference to them directly when we update vcpu
-			 * scheduling information in IOMMU irte.
-			 */
-			if (!ret && pi.is_guest_mode)
-				svm_ir_list_add(svm, irqfd, &pi);
-		}
-
-		if (!ret && svm) {
-			trace_kvm_pi_irte_update(host_irq, svm->vcpu.vcpu_id,
-						 e->gsi, vcpu_info.vector,
-						 vcpu_info.pi_desc_addr, set);
-		}
-
-		if (ret < 0) {
-			pr_err("%s: failed to update PI IRTE\n", __func__);
-			goto out;
-		}
+		if (!ret)
+			ret = svm_ir_list_add(svm, irqfd, &pi);
+	}
+
+	if (!ret && svm) {
+		trace_kvm_pi_irte_update(host_irq, svm->vcpu.vcpu_id,
+					 guest_irq, vcpu_info.vector,
+					 vcpu_info.pi_desc_addr, !!new);
+	}
+
+	if (ret < 0) {
+		pr_err("%s: failed to update PI IRTE\n", __func__);
+		goto out;
 	}
 
 	if (enable_remapped_mode)
@@ -932,7 +912,6 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 	else
 		ret = 0;
 out:
-	srcu_read_unlock(&kvm->irq_srcu, idx);
 	return ret;
 }
 
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#58 lore raw

PATCH v3 24/62 KVM: VMX: Stop walking list of routing table entries when updating IRTE

In reply to Sean Christopherson

Now that KVM provides the to-be-updated routing entry, stop walking the routing table to find that entry. KVM, via setup_routing_entry() and sanity checked by kvm_get_msi_route(), disallows having a GSI configured to trigger multiple MSIs, i.e. the for-loop can only process one entry.

---
 arch/x86/kvm/vmx/posted_intr.c | 100 +++++++++++----------------------
 1 file changed, 33 insertions(+), 67 deletions(-)
arch/x86/kvm/vmx/posted_intr.c +33 -67
index d4826a6b674f..e59eae11f476 100644
@@ -302,78 +302,44 @@ int vmx_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		       unsigned int host_irq, uint32_t guest_irq,
 		       struct kvm_kernel_irq_routing_entry *new)
 {
-	struct kvm_kernel_irq_routing_entry *e;
-	struct kvm_irq_routing_table *irq_rt;
-	bool enable_remapped_mode = true;
 	struct kvm_lapic_irq irq;
 	struct kvm_vcpu *vcpu;
 	struct vcpu_data vcpu_info;
-	bool set = !!new;
-	int idx, ret = 0;
 
 	if (!vmx_can_use_vtd_pi(kvm))
 		return 0;
 
-	idx = srcu_read_lock(&kvm->irq_srcu);
-	irq_rt = srcu_dereference(kvm->irq_routing, &kvm->irq_srcu);
-	if (guest_irq >= irq_rt->nr_rt_entries ||
-	    hlist_empty(&irq_rt->map[guest_irq])) {
-		pr_warn_once("no route for guest_irq %u/%u (broken user space?)\n",
-			     guest_irq, irq_rt->nr_rt_entries);
-		goto out;
-	}
-
-	hlist_for_each_entry(e, &irq_rt->map[guest_irq], link) {
-		if (e->type != KVM_IRQ_ROUTING_MSI)
-			continue;
-
-		WARN_ON_ONCE(new && memcmp(e, new, sizeof(*new)));
-
-		/*
-		 * VT-d PI cannot support posting multicast/broadcast
-		 * interrupts to a vCPU, we still use interrupt remapping
-		 * for these kind of interrupts.
-		 *
-		 * For lowest-priority interrupts, we only support
-		 * those with single CPU as the destination, e.g. user
-		 * configures the interrupts via /proc/irq or uses
-		 * irqbalance to make the interrupts single-CPU.
-		 *
-		 * We will support full lowest-priority interrupt later.
-		 *
-		 * In addition, we can only inject generic interrupts using
-		 * the PI mechanism, refuse to route others through it.
-		 */
-
-		kvm_set_msi_irq(kvm, e, &irq);
-		if (!kvm_intr_is_single_vcpu(kvm, &irq, &vcpu) ||
-		    !kvm_irq_is_postable(&irq))
-			continue;
-
-		vcpu_info.pi_desc_addr = __pa(vcpu_to_pi_desc(vcpu));
-		vcpu_info.vector = irq.vector;
-
-		trace_kvm_pi_irte_update(host_irq, vcpu->vcpu_id, e->gsi,
-				vcpu_info.vector, vcpu_info.pi_desc_addr, set);
-
-		if (!set)
-			continue;
-
-		enable_remapped_mode = false;
-
-		ret = irq_set_vcpu_affinity(host_irq, &vcpu_info);
-		if (ret < 0) {
-			printk(KERN_INFO "%s: failed to update PI IRTE\n",
-					__func__);
-			goto out;
-		}
-	}
-
-	if (enable_remapped_mode)
-		ret = irq_set_vcpu_affinity(host_irq, NULL);
-
-	ret = 0;
-out:
-	srcu_read_unlock(&kvm->irq_srcu, idx);
-	return ret;
+	/*
+	 * VT-d PI cannot support posting multicast/broadcast
+	 * interrupts to a vCPU, we still use interrupt remapping
+	 * for these kind of interrupts.
+	 *
+	 * For lowest-priority interrupts, we only support
+	 * those with single CPU as the destination, e.g. user
+	 * configures the interrupts via /proc/irq or uses
+	 * irqbalance to make the interrupts single-CPU.
+	 *
+	 * We will support full lowest-priority interrupt later.
+	 *
+	 * In addition, we can only inject generic interrupts using
+	 * the PI mechanism, refuse to route others through it.
+	 */
+	if (!new || new->type != KVM_IRQ_ROUTING_MSI)
+		goto do_remapping;
+
+	kvm_set_msi_irq(kvm, new, &irq);
+
+	if (!kvm_intr_is_single_vcpu(kvm, &irq, &vcpu) ||
+	    !kvm_irq_is_postable(&irq))
+		goto do_remapping;
+
+	vcpu_info.pi_desc_addr = __pa(vcpu_to_pi_desc(vcpu));
+	vcpu_info.vector = irq.vector;
+
+	trace_kvm_pi_irte_update(host_irq, vcpu->vcpu_id, guest_irq,
+				 vcpu_info.vector, vcpu_info.pi_desc_addr, true);
+
+	return irq_set_vcpu_affinity(host_irq, &vcpu_info);
+do_remapping:
+	return irq_set_vcpu_affinity(host_irq, NULL);
 }
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#59 lore raw

PATCH v3 25/62 KVM: SVM: Extract SVM specific code out of get_pi_vcpu_info()

In reply to Sean Christopherson

Genericize SVM's get_pi_vcpu_info() so that it can be shared with VMX. The only SVM specific information it provides is the AVIC back page, and that can be trivially retrieved by its sole caller.

No functional change intended.

---
 arch/x86/kvm/svm/avic.c | 25 ++++++++++++-------------
 1 file changed, 12 insertions(+), 13 deletions(-)
arch/x86/kvm/svm/avic.c +12 -13
index a83769bb8123..3bbd565dcd0f 100644
@@ -816,14 +816,14 @@ static int svm_ir_list_add(struct vcpu_svm *svm,
  */
 static int
 get_pi_vcpu_info(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
-		 struct vcpu_data *vcpu_info, struct vcpu_svm **svm)
+		 struct vcpu_data *vcpu_info, struct kvm_vcpu **vcpu)
 {
 	struct kvm_lapic_irq irq;
-	struct kvm_vcpu *vcpu = NULL;
+	*vcpu = NULL;
 
 	kvm_set_msi_irq(kvm, e, &irq);
 
-	if (!kvm_intr_is_single_vcpu(kvm, &irq, &vcpu) ||
+	if (!kvm_intr_is_single_vcpu(kvm, &irq, vcpu) ||
 	    !kvm_irq_is_postable(&irq)) {
 		pr_debug("SVM: %s: use legacy intr remap mode for irq %u\n",
 			 __func__, irq.vector);
@@ -832,8 +832,6 @@ get_pi_vcpu_info(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
 
 	pr_debug("SVM: %s: use GA mode for irq %u\n", __func__,
 		 irq.vector);
-	*svm = to_svm(vcpu);
-	vcpu_info->pi_desc_addr = avic_get_backing_page_address(*svm);
 	vcpu_info->vector = irq.vector;
 
 	return 0;
@@ -845,7 +843,7 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 {
 	bool enable_remapped_mode = true;
 	struct vcpu_data vcpu_info;
-	struct vcpu_svm *svm = NULL;
+	struct kvm_vcpu *vcpu = NULL;
 	int ret = 0;
 
 	if (!kvm_arch_has_assigned_device(kvm) || !kvm_arch_has_irq_bypass())
@@ -868,19 +866,20 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 	 * 4. IRQ has incompatible delivery mode (SMI, INIT, etc)
 	 */
 	if (new && new->type == KVM_IRQ_ROUTING_MSI &&
-	    !get_pi_vcpu_info(kvm, new, &vcpu_info, &svm) &&
-	    kvm_vcpu_apicv_active(&svm->vcpu)) {
+	    !get_pi_vcpu_info(kvm, new, &vcpu_info, &vcpu) &&
+	    kvm_vcpu_apicv_active(vcpu)) {
 		struct amd_iommu_pi_data pi;
 
 		enable_remapped_mode = false;
 
+		vcpu_info.pi_desc_addr = avic_get_backing_page_address(to_svm(vcpu));
+
 		/*
 		 * Try to enable guest_mode in IRTE.  Note, the address
 		 * of the vCPU's AVIC backing page is passed to the
 		 * IOMMU via vcpu_info->pi_desc_addr.
 		 */
-		pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
-					     svm->vcpu.vcpu_id);
+		pi.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id, vcpu->vcpu_id);
 		pi.is_guest_mode = true;
 		pi.vcpu_data = &vcpu_info;
 		ret = irq_set_vcpu_affinity(host_irq, &pi);
@@ -893,11 +892,11 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		 * scheduling information in IOMMU irte.
 		 */
 		if (!ret)
-			ret = svm_ir_list_add(svm, irqfd, &pi);
+			ret = svm_ir_list_add(to_svm(vcpu), irqfd, &pi);
 	}
 
-	if (!ret && svm) {
-		trace_kvm_pi_irte_update(host_irq, svm->vcpu.vcpu_id,
+	if (!ret && vcpu) {
+		trace_kvm_pi_irte_update(host_irq, vcpu->vcpu_id,
 					 guest_irq, vcpu_info.vector,
 					 vcpu_info.pi_desc_addr, !!new);
 	}
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#60 lore raw

PATCH v3 26/62 KVM: x86: Move IRQ routing/delivery APIs from x86.c => irq.c

In reply to Sean Christopherson

Move a bunch of IRQ routing and delivery APIs from x86.c to irq.c. x86.c has grown quite fat, and irq.c is the perfect landing spot.

Opportunistically rewrite kvm_arch_irq_bypass_del_producer()'s comment, as the existing comment has several typos and is rather confusing.

---
 arch/x86/kvm/irq.c | 88 ++++++++++++++++++++++++++++++++++++++++++++++
 arch/x86/kvm/x86.c | 87 ---------------------------------------------
 2 files changed, 88 insertions(+), 87 deletions(-)
arch/x86/kvm/irq.c +88 -0
index a0b1499baf6e..b9d3ec72037c 100644
@@ -11,6 +11,7 @@
 
 #include <linux/export.h>
 #include <linux/kvm_host.h>
+#include <linux/kvm_irqfd.h>
 
 #include "hyperv.h"
 #include "ioapic.h"
@@ -332,6 +333,18 @@ int kvm_arch_set_irq_inatomic(struct kvm_kernel_irq_routing_entry *e,
 	return -EWOULDBLOCK;
 }
 
+int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_event,
+			bool line_status)
+{
+	if (!irqchip_in_kernel(kvm))
+		return -ENXIO;
+
+	irq_event->status = kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID,
+					irq_event->irq, irq_event->level,
+					line_status);
+	return 0;
+}
+
 bool kvm_arch_can_set_irq_routing(struct kvm *kvm)
 {
 	return irqchip_in_kernel(kvm);
@@ -495,6 +508,81 @@ void kvm_arch_irq_routing_update(struct kvm *kvm)
 		kvm_make_scan_ioapic_request(kvm);
 }
 
+int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
+				      struct irq_bypass_producer *prod)
+{
+	struct kvm_kernel_irqfd *irqfd =
+		container_of(cons, struct kvm_kernel_irqfd, consumer);
+	struct kvm *kvm = irqfd->kvm;
+	int ret = 0;
+
+	kvm_arch_start_assignment(irqfd->kvm);
+
+	spin_lock_irq(&kvm->irqfds.lock);
+	irqfd->producer = prod;
+
+	if (irqfd->irq_entry.type == KVM_IRQ_ROUTING_MSI) {
+		ret = kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, prod->irq,
+						   irqfd->gsi, &irqfd->irq_entry);
+		if (ret)
+			kvm_arch_end_assignment(irqfd->kvm);
+	}
+	spin_unlock_irq(&kvm->irqfds.lock);
+
+	return ret;
+}
+
+void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
+				      struct irq_bypass_producer *prod)
+{
+	struct kvm_kernel_irqfd *irqfd =
+		container_of(cons, struct kvm_kernel_irqfd, consumer);
+	struct kvm *kvm = irqfd->kvm;
+	int ret;
+
+	WARN_ON(irqfd->producer != prod);
+
+	/*
+	 * If the producer of an IRQ that is currently being posted to a vCPU
+	 * is unregistered, change the associated IRTE back to remapped mode as
+	 * the IRQ has been released (or repurposed) by the device driver, i.e.
+	 * KVM must relinquish control of the IRTE.
+	 */
+	spin_lock_irq(&kvm->irqfds.lock);
+	irqfd->producer = NULL;
+
+	if (irqfd->irq_entry.type == KVM_IRQ_ROUTING_MSI) {
+		ret = kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, prod->irq,
+						   irqfd->gsi, NULL);
+		if (ret)
+			pr_info("irq bypass consumer (token %p) unregistration fails: %d\n",
+				irqfd->consumer.token, ret);
+	}
+
+	spin_unlock_irq(&kvm->irqfds.lock);
+
+
+	kvm_arch_end_assignment(irqfd->kvm);
+}
+
+int kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
+				  struct kvm_kernel_irq_routing_entry *old,
+				  struct kvm_kernel_irq_routing_entry *new)
+{
+	return kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, irqfd->producer->irq,
+					    irqfd->gsi, new);
+}
+
+bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
+				  struct kvm_kernel_irq_routing_entry *new)
+{
+	if (old->type != KVM_IRQ_ROUTING_MSI ||
+	    new->type != KVM_IRQ_ROUTING_MSI)
+		return true;
+
+	return !!memcmp(&old->msi, &new->msi, sizeof(new->msi));
+}
+
 #ifdef CONFIG_KVM_IOAPIC
 #define IOAPIC_ROUTING_ENTRY(irq) \
 	{ .gsi = irq, .type = KVM_IRQ_ROUTING_IRQCHIP,	\
arch/x86/kvm/x86.c +0 -87

99 lines not on this page · raw message

Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#61 lore raw

PATCH v3 27/62 KVM: x86: Nullify irqfd->producer after updating IRTEs

In reply to Sean Christopherson

Nullify irqfd->producer (when it's going away) _after_ updating IRTEs so that the producer can be queried during the update.

---
 arch/x86/kvm/irq.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
arch/x86/kvm/irq.c +1 -2
index b9d3ec72037c..09f7a5cdca7d 100644
@@ -549,7 +549,6 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
 	 * KVM must relinquish control of the IRTE.
 	 */
 	spin_lock_irq(&kvm->irqfds.lock);
-	irqfd->producer = NULL;
 
 	if (irqfd->irq_entry.type == KVM_IRQ_ROUTING_MSI) {
 		ret = kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, prod->irq,
@@ -558,10 +557,10 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
 			pr_info("irq bypass consumer (token %p) unregistration fails: %d\n",
 				irqfd->consumer.token, ret);
 	}
+	irqfd->producer = NULL;
 
 	spin_unlock_irq(&kvm->irqfds.lock);
 
-
 	kvm_arch_end_assignment(irqfd->kvm);
 }
 
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#62 lore raw

PATCH v3 28/62 KVM: x86: Dedup AVIC vs. PI code for identifying target vCPU

In reply to Sean Christopherson

Hoist the logic for identifying the target vCPU for a posted interrupt into common x86. The code is functionally identical between Intel and AMD.

---
 arch/x86/include/asm/kvm_host.h |  2 +-
 arch/x86/kvm/irq.c              | 45 +++++++++++++++---
 arch/x86/kvm/svm/avic.c         | 82 ++++++++-------------------------
 arch/x86/kvm/svm/svm.h          |  2 +-
 arch/x86/kvm/vmx/posted_intr.c  | 55 ++++++----------------
 arch/x86/kvm/vmx/posted_intr.h  |  2 +-
 6 files changed, 75 insertions(+), 113 deletions(-)
arch/x86/include/asm/kvm_host.h +1 -1
index cba82d7a701d..c722adfedd96 100644
@@ -1854,7 +1854,7 @@ struct kvm_x86_ops {
 
 	int (*pi_update_irte)(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			      unsigned int host_irq, uint32_t guest_irq,
-			      struct kvm_kernel_irq_routing_entry *new);
+			      struct kvm_vcpu *vcpu, u32 vector);
 	void (*pi_start_assignment)(struct kvm *kvm);
 	void (*apicv_pre_state_restore)(struct kvm_vcpu *vcpu);
 	void (*apicv_post_state_restore)(struct kvm_vcpu *vcpu);
arch/x86/kvm/irq.c +39 -6

69 lines not on this page · raw message

arch/x86/kvm/svm/avic.c +19 -63

123 lines not on this page · raw message

arch/x86/kvm/svm/svm.h +1 -1
index 939ff0e35a2b..b5cd1927b009 100644
@@ -747,7 +747,7 @@ void avic_apicv_post_state_restore(struct kvm_vcpu *vcpu);
 void avic_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu);
 int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			unsigned int host_irq, uint32_t guest_irq,
-			struct kvm_kernel_irq_routing_entry *new);
+			struct kvm_vcpu *vcpu, u32 vector);
 void avic_vcpu_blocking(struct kvm_vcpu *vcpu);
 void avic_vcpu_unblocking(struct kvm_vcpu *vcpu);
 void avic_ring_doorbell(struct kvm_vcpu *vcpu);
arch/x86/kvm/vmx/posted_intr.c +14 -41

60 lines not on this page · raw message

arch/x86/kvm/vmx/posted_intr.h +1 -1
index a94afcb55f7f..94ed66ea6249 100644
@@ -16,7 +16,7 @@ void pi_apicv_pre_state_restore(struct kvm_vcpu *vcpu);
 bool pi_has_pending_interrupt(struct kvm_vcpu *vcpu);
 int vmx_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		       unsigned int host_irq, uint32_t guest_irq,
-		       struct kvm_kernel_irq_routing_entry *new);
+		       struct kvm_vcpu *vcpu, u32 vector);
 void vmx_pi_start_assignment(struct kvm *kvm);
 
 static inline int pi_find_highest_vector(struct pi_desc *pi_desc)
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#63 lore raw

PATCH v3 29/62 KVM: x86: Move posted interrupt tracepoint to common code

In reply to Sean Christopherson

Move the pi_irte_update tracepoint to common x86, and call it whenever the IRTE is modified. Tracing only the modifications that result in an IRQ being posted to a vCPU makes the tracepoint useless for debugging.

Drop the vendor specific address; plumbing that into common code isn't worth the trouble, as the address is meaningless without a whole pile of other information that isn't provided in any tracepoint.

---
 arch/x86/kvm/irq.c             | 12 ++++++++++--
 arch/x86/kvm/svm/avic.c        |  6 ------
 arch/x86/kvm/trace.h           | 19 +++++++------------
 arch/x86/kvm/vmx/posted_intr.c |  3 ---
 arch/x86/kvm/x86.c             |  1 -
 5 files changed, 17 insertions(+), 24 deletions(-)
arch/x86/kvm/irq.c +10 -2
index 5948aba9fdc0..ec13d1d7706d 100644
@@ -511,9 +511,11 @@ void kvm_arch_irq_routing_update(struct kvm *kvm)
 static int kvm_pi_update_irte(struct kvm_kernel_irqfd *irqfd,
 			      struct kvm_kernel_irq_routing_entry *entry)
 {
+	unsigned int host_irq = irqfd->producer->irq;
 	struct kvm *kvm = irqfd->kvm;
 	struct kvm_vcpu *vcpu = NULL;
 	struct kvm_lapic_irq irq;
+	int r;
 
 	if (!irqchip_in_kernel(kvm) ||
 	    !kvm_arch_has_irq_bypass() ||
@@ -540,8 +542,13 @@ static int kvm_pi_update_irte(struct kvm_kernel_irqfd *irqfd,
 			vcpu = NULL;
 	}
 
-	return kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, irqfd->producer->irq,
-					    irqfd->gsi, vcpu, irq.vector);
+	r = kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, host_irq, irqfd->gsi,
+					 vcpu, irq.vector);
+	if (r)
+		return r;
+
+	trace_kvm_pi_irte_update(host_irq, vcpu, irqfd->gsi, irq.vector, !!vcpu);
+	return 0;
 }
 
 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
@@ -595,6 +602,7 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
 
 	spin_unlock_irq(&kvm->irqfds.lock);
 
+
 	kvm_arch_end_assignment(irqfd->kvm);
 }
 
arch/x86/kvm/svm/avic.c +0 -6
index 14a1544af192..d8d50b8f14bb 100644
@@ -815,9 +815,6 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 	 */
 	svm_ir_list_del(irqfd);
 
-	pr_debug("SVM: %s: host_irq=%#x, guest_irq=%#x, set=%#x\n",
-		 __func__, host_irq, guest_irq, !!vcpu);
-
 	/**
 	 * Here, we setup with legacy mode in the following cases:
 	 * 1. When cannot target interrupt to a specific vcpu.
@@ -853,9 +850,6 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		 */
 		if (!ret)
 			ret = svm_ir_list_add(to_svm(vcpu), irqfd, &pi);
-
-		trace_kvm_pi_irte_update(host_irq, vcpu->vcpu_id, guest_irq,
-					 vector, vcpu_info.pi_desc_addr, true);
 	} else {
 		ret = irq_set_vcpu_affinity(host_irq, NULL);
 	}
arch/x86/kvm/trace.h +7 -12
index ababdba2c186..57d79fd31df0 100644
@@ -1176,37 +1176,32 @@ TRACE_EVENT(kvm_smm_transition,
  * Tracepoint for VT-d posted-interrupts and AMD-Vi Guest Virtual APIC.
  */
 TRACE_EVENT(kvm_pi_irte_update,
-	TP_PROTO(unsigned int host_irq, unsigned int vcpu_id,
-		 unsigned int gsi, unsigned int gvec,
-		 u64 pi_desc_addr, bool set),
-	TP_ARGS(host_irq, vcpu_id, gsi, gvec, pi_desc_addr, set),
+	TP_PROTO(unsigned int host_irq, struct kvm_vcpu *vcpu,
+		 unsigned int gsi, unsigned int gvec, bool set),
+	TP_ARGS(host_irq, vcpu, gsi, gvec, set),
 
 	TP_STRUCT__entry(
 		__field(	unsigned int,	host_irq	)
-		__field(	unsigned int,	vcpu_id		)
+		__field(	int,		vcpu_id		)
 		__field(	unsigned int,	gsi		)
 		__field(	unsigned int,	gvec		)
-		__field(	u64,		pi_desc_addr	)
 		__field(	bool,		set		)
 	),
 
 	TP_fast_assign(
 		__entry->host_irq	= host_irq;
-		__entry->vcpu_id	= vcpu_id;
+		__entry->vcpu_id	= vcpu ? vcpu->vcpu_id : -1;
 		__entry->gsi		= gsi;
 		__entry->gvec		= gvec;
-		__entry->pi_desc_addr	= pi_desc_addr;
 		__entry->set		= set;
 	),
 
-	TP_printk("PI is %s for irq %u, vcpu %u, gsi: 0x%x, "
-		  "gvec: 0x%x, pi_desc_addr: 0x%llx",
+	TP_printk("PI is %s for irq %u, vcpu %d, gsi: 0x%x, gvec: 0x%x",
 		  __entry->set ? "enabled and being updated" : "disabled",
 		  __entry->host_irq,
 		  __entry->vcpu_id,
 		  __entry->gsi,
-		  __entry->gvec,
-		  __entry->pi_desc_addr)
+		  __entry->gvec)
 );
 
 /*
arch/x86/kvm/vmx/posted_intr.c +0 -3
index 3de767c5d6b2..687ffde3b61c 100644
@@ -308,9 +308,6 @@ int vmx_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			.vector = vector,
 		};
 
-		trace_kvm_pi_irte_update(host_irq, vcpu->vcpu_id, guest_irq,
-					 vcpu_info.vector, vcpu_info.pi_desc_addr, true);
-
 		return irq_set_vcpu_affinity(host_irq, &vcpu_info);
 	} else {
 		return irq_set_vcpu_affinity(host_irq, NULL);
arch/x86/kvm/x86.c +0 -1
index 5ba5e28532de..d3571c484790 100644
@@ -13891,7 +13891,6 @@ EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_nested_intercepts);
 EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_write_tsc_offset);
 EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_ple_window_update);
 EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_pml_full);
-EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_pi_irte_update);
 EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_avic_unaccelerated_access);
 EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_avic_incomplete_ipi);
 EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_avic_ga_log);
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#64 lore raw

PATCH v3 30/62 KVM: SVM: Clean up return handling in avic_pi_update_irte()

In reply to Sean Christopherson

Clean up the return paths for avic_pi_update_irte() now that the refactoring dust has settled.

Opportunistically drop the pr_err() on IRTE update failures. Logging that a failure occurred without _any_ context is quite useless.

---
 arch/x86/kvm/svm/avic.c | 20 +++++---------------
 1 file changed, 5 insertions(+), 15 deletions(-)
arch/x86/kvm/svm/avic.c +5 -15
index d8d50b8f14bb..a0f3cdd2ea3f 100644
@@ -807,8 +807,6 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			unsigned int host_irq, uint32_t guest_irq,
 			struct kvm_vcpu *vcpu, u32 vector)
 {
-	int ret = 0;
-
 	/*
 	 * If the IRQ was affined to a different vCPU, remove the IRTE metadata
 	 * from the *previous* vCPU's list.
@@ -838,8 +836,11 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			.is_guest_mode = true,
 			.vcpu_data = &vcpu_info,
 		};
+		int ret;
 
 		ret = irq_set_vcpu_affinity(host_irq, &pi);
+		if (ret)
+			return ret;
 
 		/**
 		 * Here, we successfully setting up vcpu affinity in
@@ -848,20 +849,9 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		 * we can reference to them directly when we update vcpu
 		 * scheduling information in IOMMU irte.
 		 */
-		if (!ret)
-			ret = svm_ir_list_add(to_svm(vcpu), irqfd, &pi);
-	} else {
-		ret = irq_set_vcpu_affinity(host_irq, NULL);
+		return svm_ir_list_add(to_svm(vcpu), irqfd, &pi);
 	}
-
-	if (ret < 0) {
-		pr_err("%s: failed to update PI IRTE\n", __func__);
-		goto out;
-	}
-
-	ret = 0;
-out:
-	return ret;
+	return irq_set_vcpu_affinity(host_irq, NULL);
 }
 
 static inline int
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#65 lore raw

PATCH v3 31/62 iommu: KVM: Split "struct vcpu_data" into separate AMD vs. Intel structs

In reply to Sean Christopherson

Split the vcpu_data structure that serves as a handoff from KVM to IOMMU drivers into vendor specific structures. Overloading a single structure makes the code hard to read and maintain, is *very* misleading as it suggests that mixing vendors is actually supported, and bastardizing Intel's posted interrupt descriptor address when AMD's IOMMU already has its own structure is quite unnecessary.

---
 arch/x86/include/asm/irq_remapping.h | 15 ++++++++++++++-
 arch/x86/kvm/svm/avic.c              | 21 ++++++++-------------
 arch/x86/kvm/vmx/posted_intr.c       |  4 ++--
 drivers/iommu/amd/iommu.c            | 12 ++++--------
 drivers/iommu/intel/irq_remapping.c  | 10 +++++-----
 include/linux/amd-iommu.h            | 12 ------------
 6 files changed, 33 insertions(+), 41 deletions(-)
arch/x86/include/asm/irq_remapping.h +14 -1
index 5036f13ab69f..2dbc9cb61c2f 100644
@@ -26,7 +26,20 @@ enum {
 	IRQ_REMAP_X2APIC_MODE,
 };
 
-struct vcpu_data {
+/*
+ * This is mainly used to communicate information back-and-forth
+ * between SVM and IOMMU for setting up and tearing down posted
+ * interrupt
+ */
+struct amd_iommu_pi_data {
+	u64 vapic_addr;		/* Physical address of the vCPU's vAPIC. */
+	u32 ga_tag;
+	u32 vector;		/* Guest vector of the interrupt */
+	bool is_guest_mode;
+	void *ir_data;
+};
+
+struct intel_iommu_pi_data {
 	u64 pi_desc_addr;	/* Physical address of PI Descriptor */
 	u32 vector;		/* Guest vector of the interrupt */
 };
arch/x86/kvm/svm/avic.c +8 -13
index a0f3cdd2ea3f..6085a629c5e6 100644
@@ -822,23 +822,18 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 	 */
 	if (vcpu && kvm_vcpu_apicv_active(vcpu)) {
 		/*
-		 * Try to enable guest_mode in IRTE.  Note, the address
-		 * of the vCPU's AVIC backing page is passed to the
-		 * IOMMU via vcpu_info->pi_desc_addr.
+		 * Try to enable guest_mode in IRTE.
 		 */
-		struct vcpu_data vcpu_info = {
-			.pi_desc_addr = avic_get_backing_page_address(to_svm(vcpu)),
-			.vector = vector,
-		};
-
-		struct amd_iommu_pi_data pi = {
-			.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id, vcpu->vcpu_id),
+		struct amd_iommu_pi_data pi_data = {
+			.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
+					     vcpu->vcpu_id),
 			.is_guest_mode = true,
-			.vcpu_data = &vcpu_info,
+			.vapic_addr = avic_get_backing_page_address(to_svm(vcpu)),
+			.vector = vector,
 		};
 		int ret;
 
-		ret = irq_set_vcpu_affinity(host_irq, &pi);
+		ret = irq_set_vcpu_affinity(host_irq, &pi_data);
 		if (ret)
 			return ret;
 
@@ -849,7 +844,7 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		 * we can reference to them directly when we update vcpu
 		 * scheduling information in IOMMU irte.
 		 */
-		return svm_ir_list_add(to_svm(vcpu), irqfd, &pi);
+		return svm_ir_list_add(to_svm(vcpu), irqfd, &pi_data);
 	}
 	return irq_set_vcpu_affinity(host_irq, NULL);
 }
arch/x86/kvm/vmx/posted_intr.c +2 -2
index 687ffde3b61c..3a23c30f73cb 100644
@@ -303,12 +303,12 @@ int vmx_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		       struct kvm_vcpu *vcpu, u32 vector)
 {
 	if (vcpu) {
-		struct vcpu_data vcpu_info = {
+		struct intel_iommu_pi_data pi_data = {
 			.pi_desc_addr = __pa(vcpu_to_pi_desc(vcpu)),
 			.vector = vector,
 		};
 
-		return irq_set_vcpu_affinity(host_irq, &vcpu_info);
+		return irq_set_vcpu_affinity(host_irq, &pi_data);
 	} else {
 		return irq_set_vcpu_affinity(host_irq, NULL);
 	}
drivers/iommu/amd/iommu.c +4 -8
index 5141507587e1..36749efcc781 100644
@@ -3860,10 +3860,10 @@ int amd_iommu_deactivate_guest_mode(void *data)
 }
 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
 
-static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
+static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *info)
 {
 	int ret;
-	struct amd_iommu_pi_data *pi_data = vcpu_info;
+	struct amd_iommu_pi_data *pi_data = info;
 	struct amd_ir_data *ir_data = data->chip_data;
 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
 	struct iommu_dev_data *dev_data;
@@ -3886,14 +3886,10 @@ static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info)
 	ir_data->cfg = irqd_cfg(data);
 
 	if (pi_data) {
-		struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data;
-
 		pi_data->ir_data = ir_data;
 
-		WARN_ON_ONCE(!pi_data->is_guest_mode);
-
-		ir_data->ga_root_ptr = (vcpu_pi_info->pi_desc_addr >> 12);
-		ir_data->ga_vector = vcpu_pi_info->vector;
+		ir_data->ga_root_ptr = (pi_data->vapic_addr >> 12);
+		ir_data->ga_vector = pi_data->vector;
 		ir_data->ga_tag = pi_data->ga_tag;
 		ret = amd_iommu_activate_guest_mode(ir_data);
 	} else {
drivers/iommu/intel/irq_remapping.c +5 -5
index 3bc2a03cceca..6165bb919520 100644
@@ -1244,10 +1244,10 @@ static void intel_ir_compose_msi_msg(struct irq_data *irq_data,
 static int intel_ir_set_vcpu_affinity(struct irq_data *data, void *info)
 {
 	struct intel_ir_data *ir_data = data->chip_data;
-	struct vcpu_data *vcpu_pi_info = info;
+	struct intel_iommu_pi_data *pi_data = info;
 
 	/* stop posting interrupts, back to the default mode */
-	if (!vcpu_pi_info) {
+	if (!pi_data) {
 		__intel_ir_reconfigure_irte(data, true);
 	} else {
 		struct irte irte_pi;
@@ -1265,10 +1265,10 @@ static int intel_ir_set_vcpu_affinity(struct irq_data *data, void *info)
 		/* Update the posted mode fields */
 		irte_pi.p_pst = 1;
 		irte_pi.p_urgent = 0;
-		irte_pi.p_vector = vcpu_pi_info->vector;
-		irte_pi.pda_l = (vcpu_pi_info->pi_desc_addr >>
+		irte_pi.p_vector = pi_data->vector;
+		irte_pi.pda_l = (pi_data->pi_desc_addr >>
 				(32 - PDA_LOW_BIT)) & ~(-1UL << PDA_LOW_BIT);
-		irte_pi.pda_h = (vcpu_pi_info->pi_desc_addr >> 32) &
+		irte_pi.pda_h = (pi_data->pi_desc_addr >> 32) &
 				~(-1UL << PDA_HIGH_BIT);
 
 		ir_data->irq_2_iommu.posted_vcpu = true;
include/linux/amd-iommu.h +0 -12
index deeefc92a5cf..99b4fa9a0296 100644
@@ -12,18 +12,6 @@
 
 struct amd_iommu;
 
-/*
- * This is mainly used to communicate information back-and-forth
- * between SVM and IOMMU for setting up and tearing down posted
- * interrupt
- */
-struct amd_iommu_pi_data {
-	u32 ga_tag;
-	bool is_guest_mode;
-	struct vcpu_data *vcpu_data;
-	void *ir_data;
-};
-
 #ifdef CONFIG_AMD_IOMMU
 
 struct task_struct;
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#66 lore raw

PATCH v3 32/62 KVM: Don't WARN if updating IRQ bypass route fails

In reply to Sean Christopherson

Don't bother WARNing if updating an IRTE route fails now that vendor code provides much more precise WARNs. The generic WARN doesn't provide enough information to actually debug the problem, and has obviously done nothing to surface the myriad bugs in KVM x86's implementation.

Drop all of the associated return code plumbing that existed just so that common KVM could WARN.

---
 arch/arm64/kvm/arm.c          |  6 +++---
 arch/arm64/kvm/vgic/vgic-v4.c | 11 ++++-------
 arch/x86/kvm/irq.c            |  8 ++++----
 include/kvm/arm_vgic.h        |  2 +-
 include/linux/kvm_host.h      |  6 +++---
 virt/kvm/eventfd.c            | 15 ++++++---------
 6 files changed, 21 insertions(+), 27 deletions(-)
arch/arm64/kvm/arm.c +3 -3
index a9a39e0375f7..94fb2f096a20 100644
@@ -2771,9 +2771,9 @@ bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
 	return memcmp(&old->msi, &new->msi, sizeof(new->msi));
 }
 
-int kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
-				  struct kvm_kernel_irq_routing_entry *old,
-				  struct kvm_kernel_irq_routing_entry *new)
+void kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
+				   struct kvm_kernel_irq_routing_entry *old,
+				   struct kvm_kernel_irq_routing_entry *new)
 {
 	/*
 	 * Remapping the vLPI requires taking the its_lock mutex to resolve
arch/arm64/kvm/vgic/vgic-v4.c +4 -7
index 86e54cefc237..5e59a3f5d4c9 100644
@@ -527,29 +527,26 @@ static struct vgic_irq *__vgic_host_irq_get_vlpi(struct kvm *kvm, int host_irq)
 	return NULL;
 }
 
-int kvm_vgic_v4_unset_forwarding(struct kvm *kvm, int host_irq)
+void kvm_vgic_v4_unset_forwarding(struct kvm *kvm, int host_irq)
 {
 	struct vgic_irq *irq;
 	unsigned long flags;
-	int ret = 0;
 
 	if (!vgic_supports_direct_msis(kvm))
-		return 0;
+		return;
 
 	irq = __vgic_host_irq_get_vlpi(kvm, host_irq);
 	if (!irq)
-		return 0;
+		return;
 
 	raw_spin_lock_irqsave(&irq->irq_lock, flags);
 	WARN_ON(irq->hw && irq->host_irq != host_irq);
 	if (irq->hw) {
 		atomic_dec(&irq->target_vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vlpi_count);
 		irq->hw = false;
-		ret = its_unmap_vlpi(host_irq);
-		WARN_ON_ONCE(ret);
+		WARN_ON_ONCE(its_unmap_vlpi(host_irq));
 	}
 
 	raw_spin_unlock_irqrestore(&irq->irq_lock, flags);
 	vgic_put_irq(kvm, irq);
-	return ret;
 }
arch/x86/kvm/irq.c +4 -4
index ec13d1d7706d..adc250ca1171 100644
@@ -606,11 +606,11 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
 	kvm_arch_end_assignment(irqfd->kvm);
 }
 
-int kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
-				  struct kvm_kernel_irq_routing_entry *old,
-				  struct kvm_kernel_irq_routing_entry *new)
+void kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
+				   struct kvm_kernel_irq_routing_entry *old,
+				   struct kvm_kernel_irq_routing_entry *new)
 {
-	return kvm_pi_update_irte(irqfd, new);
+	kvm_pi_update_irte(irqfd, new);
 }
 
 bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
include/kvm/arm_vgic.h +1 -1
index 4a34f7f0a864..b2a04481de1a 100644
@@ -434,7 +434,7 @@ struct kvm_kernel_irq_routing_entry;
 int kvm_vgic_v4_set_forwarding(struct kvm *kvm, int irq,
 			       struct kvm_kernel_irq_routing_entry *irq_entry);
 
-int kvm_vgic_v4_unset_forwarding(struct kvm *kvm, int host_irq);
+void kvm_vgic_v4_unset_forwarding(struct kvm *kvm, int host_irq);
 
 int vgic_v4_load(struct kvm_vcpu *vcpu);
 void vgic_v4_commit(struct kvm_vcpu *vcpu);
include/linux/kvm_host.h +3 -3
index a4160c1c0c6b..8e74ac0f90b1 100644
@@ -2410,9 +2410,9 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *,
 			   struct irq_bypass_producer *);
 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *);
 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *);
-int kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
-				  struct kvm_kernel_irq_routing_entry *old,
-				  struct kvm_kernel_irq_routing_entry *new);
+void kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
+				   struct kvm_kernel_irq_routing_entry *old,
+				   struct kvm_kernel_irq_routing_entry *new);
 bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *,
 				  struct kvm_kernel_irq_routing_entry *);
 #endif /* CONFIG_HAVE_KVM_IRQ_BYPASS */
virt/kvm/eventfd.c +6 -9
index 85581550dc8d..a4f80fe8a5f3 100644
@@ -285,11 +285,11 @@ void __attribute__((weak)) kvm_arch_irq_bypass_start(
 {
 }
 
-int __weak kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
-					 struct kvm_kernel_irq_routing_entry *old,
-					 struct kvm_kernel_irq_routing_entry *new)
+void __weak kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
+					  struct kvm_kernel_irq_routing_entry *old,
+					  struct kvm_kernel_irq_routing_entry *new)
 {
-	return 0;
+
 }
 
 bool __attribute__((weak)) kvm_arch_irqfd_route_changed(
@@ -618,11 +618,8 @@ void kvm_irq_routing_update(struct kvm *kvm)
 
 #if IS_ENABLED(CONFIG_HAVE_KVM_IRQ_BYPASS)
 		if (irqfd->producer &&
-		    kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry)) {
-			int ret = kvm_arch_update_irqfd_routing(irqfd, &old, &irqfd->irq_entry);
-
-			WARN_ON(ret);
-		}
+		    kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry))
+			kvm_arch_update_irqfd_routing(irqfd, &old, &irqfd->irq_entry);
 #endif
 	}
 
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#67 lore raw

PATCH v3 33/62 KVM: Fold kvm_arch_irqfd_route_changed() into kvm_arch_update_irqfd_routing()

In reply to Sean Christopherson

Fold kvm_arch_irqfd_route_changed() into kvm_arch_update_irqfd_routing(). Calling arch code to know whether or not to call arch code is absurd.

---
 arch/arm64/kvm/arm.c     | 15 +++++----------
 arch/x86/kvm/irq.c       | 15 +++++----------
 include/linux/kvm_host.h |  2 --
 virt/kvm/eventfd.c       | 10 +---------
 4 files changed, 11 insertions(+), 31 deletions(-)
arch/arm64/kvm/arm.c +5 -10
index 94fb2f096a20..04f2116927b1 100644
@@ -2761,20 +2761,15 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
 	kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq);
 }
 
-bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
-				  struct kvm_kernel_irq_routing_entry *new)
-{
-	if (old->type != KVM_IRQ_ROUTING_MSI ||
-	    new->type != KVM_IRQ_ROUTING_MSI)
-		return true;
-
-	return memcmp(&old->msi, &new->msi, sizeof(new->msi));
-}
-
 void kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
 				   struct kvm_kernel_irq_routing_entry *old,
 				   struct kvm_kernel_irq_routing_entry *new)
 {
+	if (old->type == KVM_IRQ_ROUTING_MSI &&
+	    new->type == KVM_IRQ_ROUTING_MSI &&
+	    !memcmp(&old->msi, &new->msi, sizeof(new->msi)))
+		return;
+
 	/*
 	 * Remapping the vLPI requires taking the its_lock mutex to resolve
 	 * the new translation. We're in spinlock land at this point, so no
arch/x86/kvm/irq.c +5 -10
index adc250ca1171..48134aebb541 100644
@@ -610,19 +610,14 @@ void kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
 				   struct kvm_kernel_irq_routing_entry *old,
 				   struct kvm_kernel_irq_routing_entry *new)
 {
+	if (old->type == KVM_IRQ_ROUTING_MSI &&
+	    new->type == KVM_IRQ_ROUTING_MSI &&
+	    !memcmp(&old->msi, &new->msi, sizeof(new->msi)))
+		return;
+
 	kvm_pi_update_irte(irqfd, new);
 }
 
-bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
-				  struct kvm_kernel_irq_routing_entry *new)
-{
-	if (old->type != KVM_IRQ_ROUTING_MSI ||
-	    new->type != KVM_IRQ_ROUTING_MSI)
-		return true;
-
-	return !!memcmp(&old->msi, &new->msi, sizeof(new->msi));
-}
-
 #ifdef CONFIG_KVM_IOAPIC
 #define IOAPIC_ROUTING_ENTRY(irq) \
 	{ .gsi = irq, .type = KVM_IRQ_ROUTING_IRQCHIP,	\
include/linux/kvm_host.h +0 -2
index 8e74ac0f90b1..fb9ec06aa807 100644
@@ -2413,8 +2413,6 @@ void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *);
 void kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
 				   struct kvm_kernel_irq_routing_entry *old,
 				   struct kvm_kernel_irq_routing_entry *new);
-bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *,
-				  struct kvm_kernel_irq_routing_entry *);
 #endif /* CONFIG_HAVE_KVM_IRQ_BYPASS */
 
 #ifdef CONFIG_HAVE_KVM_INVALID_WAKEUPS
virt/kvm/eventfd.c +1 -9
index a4f80fe8a5f3..defc2c04d241 100644
@@ -291,13 +291,6 @@ void __weak kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
 {
 
 }
-
-bool __attribute__((weak)) kvm_arch_irqfd_route_changed(
-				struct kvm_kernel_irq_routing_entry *old,
-				struct kvm_kernel_irq_routing_entry *new)
-{
-	return true;
-}
 #endif
 
 static int
@@ -617,8 +610,7 @@ void kvm_irq_routing_update(struct kvm *kvm)
 		irqfd_update(kvm, irqfd);
 
 #if IS_ENABLED(CONFIG_HAVE_KVM_IRQ_BYPASS)
-		if (irqfd->producer &&
-		    kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry))
+		if (irqfd->producer)
 			kvm_arch_update_irqfd_routing(irqfd, &old, &irqfd->irq_entry);
 #endif
 	}
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#69 lore raw

PATCH v3 34/62 KVM: x86: Track irq_bypass_vcpu in common x86 code

In reply to Sean Christopherson

Track the vCPU that is being targeted for IRQ bypass, a.k.a. for a posted IRQ, in common x86 code. This will allow for additional consolidation of the SVM and VMX code.

---
 arch/x86/kvm/irq.c      | 7 ++++++-
 arch/x86/kvm/svm/avic.c | 4 ----
 2 files changed, 6 insertions(+), 5 deletions(-)
arch/x86/kvm/irq.c +6 -1
index 48134aebb541..6447ea518d01 100644
@@ -544,8 +544,13 @@ static int kvm_pi_update_irte(struct kvm_kernel_irqfd *irqfd,
 
 	r = kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, host_irq, irqfd->gsi,
 					 vcpu, irq.vector);
-	if (r)
+	if (r) {
+		WARN_ON_ONCE(irqfd->irq_bypass_vcpu && !vcpu);
+		irqfd->irq_bypass_vcpu = NULL;
 		return r;
+	}
+
+	irqfd->irq_bypass_vcpu = vcpu;
 
 	trace_kvm_pi_irte_update(host_irq, vcpu, irqfd->gsi, irq.vector, !!vcpu);
 	return 0;
arch/x86/kvm/svm/avic.c +0 -4
index 6085a629c5e6..97b747e82012 100644
@@ -767,22 +767,18 @@ static void svm_ir_list_del(struct kvm_kernel_irqfd *irqfd)
 	spin_lock_irqsave(&to_svm(vcpu)->ir_list_lock, flags);
 	list_del(&irqfd->vcpu_list);
 	spin_unlock_irqrestore(&to_svm(vcpu)->ir_list_lock, flags);
-
-	irqfd->irq_bypass_vcpu = NULL;
 }
 
 static int svm_ir_list_add(struct vcpu_svm *svm,
 			   struct kvm_kernel_irqfd *irqfd,
 			   struct amd_iommu_pi_data *pi)
 {
-	struct kvm_vcpu *vcpu = &svm->vcpu;
 	unsigned long flags;
 	u64 entry;
 
 	if (WARN_ON_ONCE(!pi->ir_data))
 		return -EINVAL;
 
-	irqfd->irq_bypass_vcpu = vcpu;
 	irqfd->irq_bypass_data = pi->ir_data;
 
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#70 lore raw

PATCH v3 35/62 KVM: x86: Skip IOMMU IRTE updates if there's no old or new vCPU being targeted

In reply to Sean Christopherson

Don't "reconfigure" an IRTE into host controlled mode when it's already in the state, i.e. if KVM's GSI routing changes but the IRQ wasn't and still isn't being posted to a vCPU.

---
 arch/x86/kvm/irq.c | 3 +++
 1 file changed, 3 insertions(+)
arch/x86/kvm/irq.c +3 -0
index 6447ea518d01..43e85ebc0d5b 100644
@@ -542,6 +542,9 @@ static int kvm_pi_update_irte(struct kvm_kernel_irqfd *irqfd,
 			vcpu = NULL;
 	}
 
+	if (!irqfd->irq_bypass_vcpu && !vcpu)
+		return 0;
+
 	r = kvm_x86_call(pi_update_irte)(irqfd, irqfd->kvm, host_irq, irqfd->gsi,
 					 vcpu, irq.vector);
 	if (r) {
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#71 lore raw

PATCH v3 36/62 KVM: x86: Don't update IRTE entries when old and new routes were !MSI

In reply to Sean Christopherson

Skip the entirety of IRTE updates on a GSI routing change if neither the old nor the new routing is for an MSI, i.e. if the neither routing setup allows for posting to a vCPU. If the IRTE isn't already host controlled, KVM has bigger problems.

---
 arch/x86/kvm/irq.c | 4 ++++
 1 file changed, 4 insertions(+)
arch/x86/kvm/irq.c +4 -0
index 43e85ebc0d5b..4119c1e880e7 100644
@@ -618,6 +618,10 @@ void kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
 				   struct kvm_kernel_irq_routing_entry *old,
 				   struct kvm_kernel_irq_routing_entry *new)
 {
+	if (new->type != KVM_IRQ_ROUTING_MSI &&
+	    old->type != KVM_IRQ_ROUTING_MSI)
+		return;
+
 	if (old->type == KVM_IRQ_ROUTING_MSI &&
 	    new->type == KVM_IRQ_ROUTING_MSI &&
 	    !memcmp(&old->msi, &new->msi, sizeof(new->msi)))
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#72 lore raw

PATCH v3 37/62 KVM: SVM: Revert IRTE to legacy mode if IOMMU doesn't provide IR metadata

In reply to Sean Christopherson

Revert the IRTE back to remapping mode if the AMD IOMMU driver mucks up and doesn't provide the necessary metadata. Returning an error up the stack without actually handling the error is useless and confusing.

---
 arch/x86/kvm/svm/avic.c | 23 +++++++++++++++--------
 1 file changed, 15 insertions(+), 8 deletions(-)
arch/x86/kvm/svm/avic.c +15 -8
index 97b747e82012..f1e9f0dd43e8 100644
@@ -769,16 +769,13 @@ static void svm_ir_list_del(struct kvm_kernel_irqfd *irqfd)
 	spin_unlock_irqrestore(&to_svm(vcpu)->ir_list_lock, flags);
 }
 
-static int svm_ir_list_add(struct vcpu_svm *svm,
-			   struct kvm_kernel_irqfd *irqfd,
-			   struct amd_iommu_pi_data *pi)
+static void svm_ir_list_add(struct vcpu_svm *svm,
+			    struct kvm_kernel_irqfd *irqfd,
+			    struct amd_iommu_pi_data *pi)
 {
 	unsigned long flags;
 	u64 entry;
 
-	if (WARN_ON_ONCE(!pi->ir_data))
-		return -EINVAL;
-
 	irqfd->irq_bypass_data = pi->ir_data;
 
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
@@ -796,7 +793,6 @@ static int svm_ir_list_add(struct vcpu_svm *svm,
 
 	list_add(&irqfd->vcpu_list, &svm->ir_list);
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
-	return 0;
 }
 
 int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
@@ -833,6 +829,16 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		if (ret)
 			return ret;
 
+		/*
+		 * Revert to legacy mode if the IOMMU didn't provide metadata
+		 * for the IRTE, which KVM needs to keep the IRTE up-to-date,
+		 * e.g. if the vCPU is migrated or AVIC is disabled.
+		 */
+		if (WARN_ON_ONCE(!pi_data.ir_data)) {
+			irq_set_vcpu_affinity(host_irq, NULL);
+			return -EIO;
+		}
+
 		/**
 		 * Here, we successfully setting up vcpu affinity in
 		 * IOMMU guest mode. Now, we need to store the posted
@@ -840,7 +846,8 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		 * we can reference to them directly when we update vcpu
 		 * scheduling information in IOMMU irte.
 		 */
-		return svm_ir_list_add(to_svm(vcpu), irqfd, &pi_data);
+		svm_ir_list_add(to_svm(vcpu), irqfd, &pi_data);
+		return 0;
 	}
 	return irq_set_vcpu_affinity(host_irq, NULL);
 }
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#73 lore raw

PATCH v3 38/62 KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU

In reply to Sean Christopherson

Now that svm_ir_list_add() isn't overloaded with all manner of weird things, fold it into avic_pi_update_irte(), and more importantly take ir_list_lock across the irq_set_vcpu_affinity() calls to ensure the info that's shoved into the IRTE is fresh. While preemption (and IRQs) is disabled on the task performing the IRTE update, thanks to irqfds.lock, that task doesn't hold the vCPU's mutex, i.e. preemption being disabled is irrelevant.

---
 arch/x86/kvm/svm/avic.c | 55 +++++++++++++++++------------------------
 1 file changed, 22 insertions(+), 33 deletions(-)
arch/x86/kvm/svm/avic.c +22 -33
index f1e9f0dd43e8..4747fb09aca4 100644
@@ -769,32 +769,6 @@ static void svm_ir_list_del(struct kvm_kernel_irqfd *irqfd)
 	spin_unlock_irqrestore(&to_svm(vcpu)->ir_list_lock, flags);
 }
 
-static void svm_ir_list_add(struct vcpu_svm *svm,
-			    struct kvm_kernel_irqfd *irqfd,
-			    struct amd_iommu_pi_data *pi)
-{
-	unsigned long flags;
-	u64 entry;
-
-	irqfd->irq_bypass_data = pi->ir_data;
-
-	spin_lock_irqsave(&svm->ir_list_lock, flags);
-
-	/*
-	 * Update the target pCPU for IOMMU doorbells if the vCPU is running.
-	 * If the vCPU is NOT running, i.e. is blocking or scheduled out, KVM
-	 * will update the pCPU info when the vCPU awkened and/or scheduled in.
-	 * See also avic_vcpu_load().
-	 */
-	entry = svm->avic_physical_id_entry;
-	if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
-		amd_iommu_update_ga(entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK,
-				    true, pi->ir_data);
-
-	list_add(&irqfd->vcpu_list, &svm->ir_list);
-	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
-}
-
 int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			unsigned int host_irq, uint32_t guest_irq,
 			struct kvm_vcpu *vcpu, u32 vector)
@@ -823,8 +797,18 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			.vapic_addr = avic_get_backing_page_address(to_svm(vcpu)),
 			.vector = vector,
 		};
+		struct vcpu_svm *svm = to_svm(vcpu);
+		u64 entry;
 		int ret;
 
+		/*
+		 * Prevent the vCPU from being scheduled out or migrated until
+		 * the IRTE is updated and its metadata has been added to the
+		 * list of IRQs being posted to the vCPU, to ensure the IRTE
+		 * isn't programmed with stale pCPU/IsRunning information.
+		 */
+		guard(spinlock_irqsave)(&svm->ir_list_lock);
+
 		ret = irq_set_vcpu_affinity(host_irq, &pi_data);
 		if (ret)
 			return ret;
@@ -839,14 +823,19 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			return -EIO;
 		}
 
-		/**
-		 * Here, we successfully setting up vcpu affinity in
-		 * IOMMU guest mode. Now, we need to store the posted
-		 * interrupt information in a per-vcpu ir_list so that
-		 * we can reference to them directly when we update vcpu
-		 * scheduling information in IOMMU irte.
+		/*
+		 * Update the target pCPU for IOMMU doorbells if the vCPU is
+		 * running.  If the vCPU is NOT running, i.e. is blocking or
+		 * scheduled out, KVM will update the pCPU info when the vCPU
+		 * is awakened and/or scheduled in.  See also avic_vcpu_load().
 		 */
-		svm_ir_list_add(to_svm(vcpu), irqfd, &pi_data);
+		entry = svm->avic_physical_id_entry;
+		if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
+			amd_iommu_update_ga(entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK,
+					    true, pi_data.ir_data);
+
+		irqfd->irq_bypass_data = pi_data.ir_data;
+		list_add(&irqfd->vcpu_list, &svm->ir_list);
 		return 0;
 	}
 	return irq_set_vcpu_affinity(host_irq, NULL);
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
#74 lore raw

Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU

Quoted 37 lines · Sean Christopherson

Now that svm_ir_list_add() isn't overloaded with all manner of weird things, fold it into avic_pi_update_irte(), and more importantly take ir_list_lock across the irq_set_vcpu_affinity() calls to ensure the info that's shoved into the IRTE is fresh. While preemption (and IRQs) is disabled on the task performing the IRTE update, thanks to irqfds.lock, that task doesn't hold the vCPU's mutex, i.e. preemption being disabled is irrelevant.

---
 arch/x86/kvm/svm/avic.c | 55 +++++++++++++++++------------------------
 1 file changed, 22 insertions(+), 33 deletions(-)
arch/x86/kvm/svm/avic.c +0 -16
index f1e9f0dd43e8..4747fb09aca4 100644
@@ -769,32 +769,6 @@ static void svm_ir_list_del(struct kvm_kernel_irqfd *irqfd)
 	spin_unlock_irqrestore(&to_svm(vcpu)->ir_list_lock, flags);
 }
 
-static void svm_ir_list_add(struct vcpu_svm *svm,
-			    struct kvm_kernel_irqfd *irqfd,
-			    struct amd_iommu_pi_data *pi)
-{
-	unsigned long flags;
-	u64 entry;
-
-	irqfd->irq_bypass_data = pi->ir_data;
-
-	spin_lock_irqsave(&svm->ir_list_lock, flags);
-
-	/*
-	 * Update the target pCPU for IOMMU doorbells if the vCPU is running.
-	 * If the vCPU is NOT running, i.e. is blocking or scheduled out, KVM
-	 * will update the pCPU info when the vCPU awkened and/or scheduled in.
-	 * See also avic_vcpu_load().

arch/x86/kvm/svm/avic.c static void svm_ir_list_del(struct kvm_kernel_irqfd *irqfd) · line 797

Quoted 10 lines
arch/x86/kvm/svm/avic.c +0 -10
-	 */
-	entry = svm->avic_physical_id_entry;
-	if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
-		amd_iommu_update_ga(entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK,
-				    true, pi->ir_data);
-
-	list_add(&irqfd->vcpu_list, &svm->ir_list);
-	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
-}
-

There are a few comments in avic_vcpu_load() and avic_vcpu_put() which still refer to svm_ir_list_add(). Would be good to update those.

- Naveen

#75 lore raw

Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU

In reply to Sean Christopherson
Quoted 27 lines · Sean Christopherson

Now that svm_ir_list_add() isn't overloaded with all manner of weird things, fold it into avic_pi_update_irte(), and more importantly take ir_list_lock across the irq_set_vcpu_affinity() calls to ensure the info that's shoved into the IRTE is fresh. While preemption (and IRQs) is disabled on the task performing the IRTE update, thanks to irqfds.lock, that task doesn't hold the vCPU's mutex, i.e. preemption being disabled is irrelevant.

---
 arch/x86/kvm/svm/avic.c | 55 +++++++++++++++++------------------------
 1 file changed, 22 insertions(+), 33 deletions(-)
arch/x86/kvm/svm/avic.c +2 -0
index f1e9f0dd43e8..4747fb09aca4 100644
 
 int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			unsigned int host_irq, uint32_t guest_irq,
 			struct kvm_vcpu *vcpu, u32 vector)
@@ -823,8 +797,18 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			.vapic_addr = avic_get_backing_page_address(to_svm(vcpu)),
 			.vector = vector,
 		};
+		struct vcpu_svm *svm = to_svm(vcpu);
+		u64 entry;

arch/x86/kvm/svm/avic.c int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm, · line 811

Quoted 10 lines
arch/x86/kvm/svm/avic.c +8 -0
 		int ret;
 
+		/*
+		 * Prevent the vCPU from being scheduled out or migrated until
+		 * the IRTE is updated and its metadata has been added to the
+		 * list of IRQs being posted to the vCPU, to ensure the IRTE
+		 * isn't programmed with stale pCPU/IsRunning information.
+		 */
+		guard(spinlock_irqsave)(&svm->ir_list_lock);
+

Hi,

I’m seeing a lockdep warning about a possible circular locking dependency involving svm->ir_list_lock and irq_desc_lock when using AMD SVM with AVIC enabled and a VFIO passthrough device, on 6.19-rc2.

Environment
===========

- Kernel: 6.19.0-rc2
- QEMU: 10.1.94
- CPU: AMD EPYC 9965
- Modules involved: kvm_amd, kvm, vfio_pci, vfio, irqbypass, mlx5_core
- Workload: QEMU guest with an mlx5 PCI device passed through.

Lockdep warning
===============

The warning is:

======================================================
WARNING: possible circular locking dependency detected
6.19.0-rc2 #20 Tainted: G E
------------------------------------------------------
CPU 58/KVM/28597 is trying to acquire lock:
ff12c47d4b1f34c0 (&irq_desc_lock_class){-.-.}-{2:2}, at: __irq_get_desc_lock+0x58/0xa0

but task is already holding lock:
ff12c49b28552110 (&svm->ir_list_lock){....}-{2:2}, at: avic_pi_update_irte+0x147/0x270 [kvm_amd]

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

    -> #3 (&svm->ir_list_lock){....}-{2:2}:
         _raw_spin_lock_irqsave+0x4e/0xb0
         __avic_vcpu_put+0x7a/0x150 [kvm_amd]
         avic_vcpu_put+0x50/0x70 [kvm_amd]
         svm_vcpu_put+0x38/0x70 [kvm_amd]
         kvm_arch_vcpu_put+0x21b/0x330 [kvm]
         kvm_sched_out+0x62/0x90 [kvm]
         __schedule+0x8d3/0x1d10
         __cond_resched+0x5c/0x80
         __mutex_lock+0x83/0x10b0
         mutex_lock_nested+0x1b/0x30
         kvm_hv_set_msr_common+0x199/0x12a0 [kvm]
         kvm_set_msr_common+0x468/0x1310 [kvm]
         svm_set_msr+0x645/0x730 [kvm_amd]
         __kvm_set_msr+0xa3/0x2f0 [kvm]
         kvm_set_msr_ignored_check+0x23/0x1b0 [kvm]
         do_set_msr+0x76/0xd0 [kvm]
         msr_io+0xbe/0x1c0 [kvm]
         kvm_arch_vcpu_ioctl+0x700/0x2090 [kvm]
         kvm_vcpu_ioctl+0x632/0xc60 [kvm]
         __x64_sys_ioctl+0xa5/0x100
         x64_sys_call+0x1243/0x26b0
         do_syscall_64+0x93/0x1470
         entry_SYSCALL_64_after_hwframe+0x76/0x7e

    -> #2 (&rq->__lock){-.-.}-{2:2}:
         _raw_spin_lock_nested+0x32/0x80
         raw_spin_rq_lock_nested+0x22/0xa0
         task_rq_lock+0x5f/0x150
         cgroup_move_task+0x46/0x110
         css_set_move_task+0xe1/0x240
         cgroup_post_fork+0x98/0x2d0
         copy_process+0x1ea8/0x2330
         kernel_clone+0xa7/0x440
         user_mode_thread+0x63/0x90
         rest_init+0x28/0x200
         start_kernel+0xae0/0xcd0
         x86_64_start_reservations+0x18/0x30
         x86_64_start_kernel+0xfd/0x150
         common_startup_64+0x13e/0x141

    -> #1 (&p->pi_lock){-.-.}-{2:2}:
         _raw_spin_lock_irqsave+0x4e/0xb0
         try_to_wake_up+0x59/0xaa0
         wake_up_process+0x15/0x30
         irq_do_set_affinity+0x145/0x270
         irq_set_affinity_locked+0x172/0x250
         irq_set_affinity+0x47/0x80
         write_irq_affinity.isra.0+0xfe/0x120
         irq_affinity_proc_write+0x1d/0x30
         proc_reg_write+0x69/0xa0
         vfs_write+0x110/0x560
         ksys_write+0x77/0x100
         __x64_sys_write+0x19/0x30
         x64_sys_call+0x79/0x26b0
         do_syscall_64+0x93/0x1470
         entry_SYSCALL_64_after_hwframe+0x76/0x7e

    -> #0 (&irq_desc_lock_class){-.-.}-{2:2}:
         __lock_acquire+0x1595/0x2640
         lock_acquire+0xc4/0x2c0
         _raw_spin_lock_irqsave+0x4e/0xb0
         __irq_get_desc_lock+0x58/0xa0
         irq_set_vcpu_affinity+0x4a/0x100
         avic_pi_update_irte+0x170/0x270 [kvm_amd]
         kvm_pi_update_irte+0xea/0x220 [kvm]
         kvm_arch_irq_bypass_add_producer+0x9b/0xb0 [kvm]
         __connect+0x5f/0x100 [irqbypass]
         irq_bypass_register_producer+0xe4/0xb90 [irqbypass]
         vfio_msi_set_vector_signal+0x1b0/0x330 [vfio_pci_core]
         vfio_msi_set_block+0x5a/0xd0 [vfio_pci_core]
         vfio_pci_set_msi_trigger+0x19e/0x260 [vfio_pci_core]
         vfio_pci_set_irqs_ioctl+0x46/0x140 [vfio_pci_core]
         vfio_pci_core_ioctl+0x6ea/0xc20 [vfio_pci_core]
         vfio_device_fops_unl_ioctl+0xb1/0x9d0 [vfio]
         __x64_sys_ioctl+0xa5/0x100
         x64_sys_call+0x1243/0x26b0
         do_syscall_64+0x93/0x1470
         entry_SYSCALL_64_after_hwframe+0x76/0x7e
  Chain exists of:
    &irq_desc_lock_class --> &rq->__lock --> &svm->ir_list_lock

  Possible unsafe locking scenario:

CPU0 CPU1
---- ----

   lock(&svm->ir_list_lock);
                                      lock(&rq->__lock);
                                      lock(&svm->ir_list_lock);

lock(&irq_desc_lock_class);

        *** DEADLOCK ***

At the point of the warning, the following locks are held:

  #0: &vdev->igate           (vfio_pci_core)
  #1: lock#10                (irqbypass)
  #2: &kvm->irqfds.lock      (kvm)
  #3: &svm->ir_list_lock     (kvm_amd)

and the stack backtrace has:

__irq_get_desc_lock
irq_set_vcpu_affinity

  avic_pi_update_irte               [kvm_amd]
  kvm_pi_update_irte                [kvm]
  kvm_arch_irq_bypass_add_producer  [kvm]
  __connect                         [irqbypass]
  irq_bypass_register_producer      [irqbypass]
  vfio_msi_set_vector_signal        [vfio_pci_core]
  vfio_pci_set_irqs_ioctl           [vfio_pci_core]
  vfio_pci_core_ioctl               [vfio_pci_core]
  vfio_device_fops_unl_ioctl        [vfio]

__x64_sys_ioctl
x64_sys_call
do_syscall_64
entry_SYSCALL_64_after_hwframe

So lockdep sees:

&irq_desc_lock_class -> &rq->__lock -> &svm->ir_list_lock

while avic_pi_update_irte() currently holds svm->ir_list_lock and then takes irq_desc_lock via irq_set_vcpu_affinity(), which creates the potential inversion.

Reproduction
============

Host:

- AMD EPYC + AVIC enabled
- Kernel 6.19-rc2 with lockdep
- VFIO passthrough of an mlx5 device (mlx5_core loaded)

Launch command with Passthrough device, with AVIC mode.

The warning triggers when enabling MSI/MSI-X for the passthrough device from the guest, i.e. via VFIO ioctl on the host that goes through irq_bypass and eventually calls avic_pi_update_irte().

I can reproduce this reliably when starting the guest with the VFIO device assigned after every host reboot.

Questions
=========

- Is this lockdep warning expected/benign in this code path, or does it indicate a real potential deadlock between svm->ir_list_lock and irq_desc_lock with AVIC + irq_bypass + VFIO?

- If this is considered a real issue, is the expected direction to:
* change the locking around avic_pi_update_irte()/svm->ir_list_lock, or
* adjust how irq_bypass / VFIO interacts with vCPU affinity updates on AMD/AVIC, or
* annotate the locking somehow if lockdep is over-reporting here?

I’m happy to:

- Provide my full .config
- Share the exact QEMU command line
- Test any proposed patches or instrumentation

Thanks,
Ankit Soni
[email protected]

Quoted 3 lines
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
maintainer
#76 lore raw

possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

Quoted 30 lines · Ankit Soni
   ======================================================
   WARNING: possible circular locking dependency detected
   6.19.0-rc2 #20 Tainted: G            E
   ------------------------------------------------------
   CPU 58/KVM/28597 is trying to acquire lock:
     ff12c47d4b1f34c0 (&irq_desc_lock_class){-.-.}-{2:2}, at: __irq_get_desc_lock+0x58/0xa0

     but task is already holding lock:
     ff12c49b28552110 (&svm->ir_list_lock){....}-{2:2}, at: avic_pi_update_irte+0x147/0x270 [kvm_amd]

     which lock already depends on the new lock.

   Chain exists of:
     &irq_desc_lock_class --> &rq->__lock --> &svm->ir_list_lock

   Possible unsafe locking scenario:

         CPU0                            CPU1
         ----                            ----
    lock(&svm->ir_list_lock);
                                       lock(&rq->__lock);
                                       lock(&svm->ir_list_lock);
    lock(&irq_desc_lock_class);

         *** DEADLOCK ***

So lockdep sees:

   &irq_desc_lock_class -> &rq->__lock -> &svm->ir_list_lock
Quoted 7 lines

while avic_pi_update_irte() currently holds svm->ir_list_lock and then takes irq_desc_lock via irq_set_vcpu_affinity(), which creates the potential inversion.

- Is this lockdep warning expected/benign in this code path, or does it indicate a real potential deadlock between svm->ir_list_lock and irq_desc_lock with AVIC + irq_bypass + VFIO?

I'd treat it as a potential (if unlikely) deadlock:

(a) irq_set_thread_affinity triggers the scheduler via wake_up_process, while irq_desc->lock is taken

(b) the scheduler calls into KVM with rq_lock taken, and KVM uses ir_list_lock within __avic_vcpu_load/__avic_vcpu_put

(c) KVM wants to block scheduling for a while and uses ir_list_lock for that purpose, but then takes irq_set_vcpu_affinity takes irq_desc->lock.

I don't think there's an alternative choice of lock for (c); and there's no easy way to pull the irq_desc->lock out of the IRQ subsystem--in fact the stickiness of the situation comes from rq->rq_lock and irq_desc->lock being both internal and not leaf.

Of the three, the most sketchy is (a); notably, __setup_irq() calls wake_up_process outside desc->lock. Therefore I'd like so much to treat it as a kernel/irq/ bug; and the simplest (perhaps too simple...) fix is to drop the wake_up_process(). The only cost is extra latency on the next interrupt after an affinity change.

kernel/irq/manage.c +2 -6
index 8b1b4c8a4f54..fc135bd079a4 100644
@@ -189,14 +189,10 @@ static void irq_set_thread_affinity(struct irq_desc *desc)
 	struct irqaction *action;
 
 	for_each_action_of_desc(desc, action) {
-		if (action->thread) {
+		if (action->thread)
 			set_bit(IRQTF_AFFINITY, &action->thread_flags);
-			wake_up_process(action->thread);
-		}
-		if (action->secondary && action->secondary->thread) {
+		if (action->secondary && action->secondary->thread)
 			set_bit(IRQTF_AFFINITY, &action->secondary->thread_flags);
-			wake_up_process(action->secondary->thread);
-		}
 	}
 }
 

Marc, what do you think?

Paolo

author maintainer
#77 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

Quoted 50 lines · Paolo Bonzini
Quoted 30 lines · Ankit Soni
   ======================================================
   WARNING: possible circular locking dependency detected
   6.19.0-rc2 #20 Tainted: G            E
   ------------------------------------------------------
   CPU 58/KVM/28597 is trying to acquire lock:
     ff12c47d4b1f34c0 (&irq_desc_lock_class){-.-.}-{2:2}, at: __irq_get_desc_lock+0x58/0xa0

     but task is already holding lock:
     ff12c49b28552110 (&svm->ir_list_lock){....}-{2:2}, at: avic_pi_update_irte+0x147/0x270 [kvm_amd]

     which lock already depends on the new lock.

   Chain exists of:
     &irq_desc_lock_class --> &rq->__lock --> &svm->ir_list_lock

   Possible unsafe locking scenario:

         CPU0                            CPU1
         ----                            ----
    lock(&svm->ir_list_lock);
                                       lock(&rq->__lock);
                                       lock(&svm->ir_list_lock);
    lock(&irq_desc_lock_class);

         *** DEADLOCK ***

So lockdep sees:

   &irq_desc_lock_class -> &rq->__lock -> &svm->ir_list_lock
Quoted 7 lines

while avic_pi_update_irte() currently holds svm->ir_list_lock and then takes irq_desc_lock via irq_set_vcpu_affinity(), which creates the potential inversion.

- Is this lockdep warning expected/benign in this code path, or does it indicate a real potential deadlock between svm->ir_list_lock and irq_desc_lock with AVIC + irq_bypass + VFIO?

I'd treat it as a potential (if unlikely) deadlock:

(a) irq_set_thread_affinity triggers the scheduler via wake_up_process, while irq_desc->lock is taken

(b) the scheduler calls into KVM with rq_lock taken, and KVM uses ir_list_lock within __avic_vcpu_load/__avic_vcpu_put

(c) KVM wants to block scheduling for a while and uses ir_list_lock for that purpose, but then takes irq_set_vcpu_affinity takes irq_desc->lock.

Quoted 6 lines

I don't think there's an alternative choice of lock for (c); and there's no easy way to pull the irq_desc->lock out of the IRQ subsystem--in fact the stickiness of the situation comes from rq->rq_lock and irq_desc->lock being both internal and not leaf.

Of the three, the most sketchy is (a);

Maybe the most unnecessary, but I think there's a pretty strong argument that
(d) is the most sketchy:

(d) KVM takes svm->ir_list_lock around the call to irq_set_vcpu_affinity()

Quoted 4 lines

notably, __setup_irq() calls wake_up_process outside desc->lock. Therefore I'd like so much to treat it as a kernel/irq/ bug; and the simplest (perhaps too simple...) fix is to drop the wake_up_process(). The only cost is extra latency on the next interrupt after an affinity change.

Alternatively, what if we rework the KVM<=>IOMMU exchange to decouple updating the IRTE from binding the metadata to the vCPU? KVM already has the necessary exports to do "out-of-band" updates due to the AVIC architecture requiring IRTE updates on scheduling changes.

It's a bit wonky (and not yet tested), but I like the idea of making svm->ir_list_lock a leaf lock so that we don't end up with a game of whack-a-mole, e.g. if something in the IRQ subsystem changes in the future.

---
 arch/x86/include/asm/irq_remapping.h |  3 --
 arch/x86/kvm/svm/avic.c              | 78 ++++++++++++++++++----------
 drivers/iommu/amd/iommu.c            | 24 +++------
 3 files changed, 57 insertions(+), 48 deletions(-)
arch/x86/include/asm/irq_remapping.h +0 -3
index 4e55d1755846..1426ecd09943 100644
@@ -35,9 +35,6 @@ struct amd_iommu_pi_data {
 	u64 vapic_addr;		/* Physical address of the vCPU's vAPIC. */
 	u32 ga_tag;
 	u32 vector;		/* Guest vector of the interrupt */
-	int cpu;
-	bool ga_log_intr;
-	bool is_guest_mode;
 	void *ir_data;
 };
 
arch/x86/kvm/svm/avic.c +50 -28

101 lines not on this page · raw message

drivers/iommu/amd/iommu.c +7 -17
index 5d45795c367a..855c6309900c 100644
@@ -3970,7 +3970,6 @@ EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
 
 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *info)
 {
-	int ret;
 	struct amd_iommu_pi_data *pi_data = info;
 	struct amd_ir_data *ir_data = data->chip_data;
 	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
@@ -3993,25 +3992,16 @@ static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *info)
 
 	ir_data->cfg = irqd_cfg(data);
 
-	if (pi_data) {
-		pi_data->ir_data = ir_data;
+	if (!pi_data)
+		return amd_iommu_deactivate_guest_mode(ir_data);
 
-		ir_data->ga_root_ptr = (pi_data->vapic_addr >> 12);
-		ir_data->ga_vector = pi_data->vector;
-		ir_data->ga_tag = pi_data->ga_tag;
-		if (pi_data->is_guest_mode)
-			ret = amd_iommu_activate_guest_mode(ir_data, pi_data->cpu,
-							    pi_data->ga_log_intr);
-		else
-			ret = amd_iommu_deactivate_guest_mode(ir_data);
-	} else {
-		ret = amd_iommu_deactivate_guest_mode(ir_data);
-	}
-
-	return ret;
+	pi_data->ir_data = ir_data;
+	ir_data->ga_root_ptr = (pi_data->vapic_addr >> 12);
+	ir_data->ga_vector = pi_data->vector;
+	ir_data->ga_tag = pi_data->ga_tag;
+	return 0;
 }
 
-
 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
 			       struct amd_ir_data *ir_data,
 			       struct irq_2_irte *irte_info,

base-commit: 9448598b22c50c8a5bb77a9103e2d49f134c9578 --

maintainer
#78 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

Quoted 22 lines · Sean Christopherson
Quoted 21 lines · Paolo Bonzini
Quoted 3 lines · Ankit Soni

- Is this lockdep warning expected/benign in this code path, or does it indicate a real potential deadlock between svm->ir_list_lock and irq_desc_lock with AVIC + irq_bypass + VFIO?

I'd treat it as a potential (if unlikely) deadlock:

(a) irq_set_thread_affinity triggers the scheduler via wake_up_process, while irq_desc->lock is taken

(b) the scheduler calls into KVM with rq_lock taken, and KVM uses ir_list_lock within __avic_vcpu_load/__avic_vcpu_put

(c) KVM wants to block scheduling for a while and uses ir_list_lock for that purpose, but then irq_set_vcpu_affinity takes irq_desc->lock.

I don't think there's an alternative choice of lock for (c); and there's no easy way to pull the irq_desc->lock out of the IRQ subsystem--in fact the stickiness of the situation comes from rq->rq_lock and irq_desc->lock being both internal and not leaf.

Quoted 6 lines
Quoted 1 line

Of the three, the most sketchy is (a);

Maybe the most unnecessary, but I think there's a pretty strong argument that
(d) is the most sketchy:

(d) KVM takes svm->ir_list_lock around the call to irq_set_vcpu_affinity()

That's (c). :)

I called irq_set_thread_affinity() sketchy, because it's a core kernel subsystem that takes a pretty common lock, and directly calls a function that takes another very common lock. KVM's deadlock scenario seems like a relatively natural occurrence, though in this case it can be fixed outside kernel/irq as well.

Quoted 9 lines
Quoted 4 lines

notably, __setup_irq() calls wake_up_process outside desc->lock. Therefore I'd like so much to treat it as a kernel/irq/ bug; and the simplest (perhaps too simple...) fix is to drop the wake_up_process(). The only cost is extra latency on the next interrupt after an affinity change.

Alternatively, what if we rework the KVM<=>IOMMU exchange to decouple updating the IRTE from binding the metadata to the vCPU? KVM already has the necessary exports to do "out-of-band" updates due to the AVIC architecture requiring IRTE updates on scheduling changes.

In fact this was actually my first idea, exactly because it makes svm->ir_list_lock a leaf lock!

I threw it away because it makes amd_ir_set_vcpu_affinity() weird, passing back the ir_data but not really doing anything else. Basically its role becomes little more than violate abstractions, which seemed wrong. On the other hand, drivers/iommu is already very much tied to the KVM vendor modules (in particular avic.c already calls amd_iommu_{,de}activate_guest_mode), so who am I to judge what the IOMMU driver does.

Paolo

Quoted 181 lines

It's a bit wonky (and not yet tested), but I like the idea of making svm->ir_list_lock a leaf lock so that we don't end up with a game of whack-a-mole, e.g. if something in the IRQ subsystem changes in the future.

---
  arch/x86/include/asm/irq_remapping.h |  3 --
  arch/x86/kvm/svm/avic.c              | 78 ++++++++++++++++++----------
  drivers/iommu/amd/iommu.c            | 24 +++------
  3 files changed, 57 insertions(+), 48 deletions(-)
arch/x86/include/asm/irq_remapping.h +0 -3
index 4e55d1755846..1426ecd09943 100644
@@ -35,9 +35,6 @@ struct amd_iommu_pi_data {
  	u64 vapic_addr;		/* Physical address of the vCPU's vAPIC. */
  	u32 ga_tag;
  	u32 vector;		/* Guest vector of the interrupt */
-	int cpu;
-	bool ga_log_intr;
-	bool is_guest_mode;
  	void *ir_data;
  };
 
arch/x86/kvm/svm/avic.c +50 -28
index 6b77b2033208..0f4f353c7db6 100644
@@ -868,6 +868,51 @@ static void svm_ir_list_del(struct kvm_kernel_irqfd *irqfd)
  	raw_spin_unlock_irqrestore(&to_svm(vcpu)->ir_list_lock, flags);
  }
 
+static int avic_pi_add_irte(struct kvm_kernel_irqfd *irqfd, void *ir_data,
+			    struct kvm_vcpu *vcpu)
+{
+	struct vcpu_svm *svm = to_svm(vcpu);
+	int r;
+
+	/*
+	 * Prevent the vCPU from being scheduled out or migrated until the IRTE
+	 * is updated and its metadata has been added to the list of IRQs being
+	 * posted to the vCPU, to ensure the IRTE isn't programmed with stale
+	 * pCPU/IsRunning information.
+	 */
+	guard(raw_spinlock_irqsave)(&svm->ir_list_lock);
+
+	if (kvm_vcpu_apicv_active(vcpu)) {
+		u64 entry = svm->avic_physical_id_entry;
+		bool ga_log_intr;
+		int cpu;
+
+		/*
+		 * Update the target pCPU for IOMMU doorbells if the vCPU is
+		 * running.  If the vCPU is NOT running, i.e. is blocking or
+		 * scheduled out, KVM will update the pCPU info when the vCPU
+		 * is awakened and/or scheduled in.  See also avic_vcpu_load().
+		 */
+		if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK) {
+			cpu = entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
+			ga_log_intr = false;
+		} else {
+			cpu = -1;
+			ga_log_intr = entry & AVIC_PHYSICAL_ID_ENTRY_GA_LOG_INTR;
+		}
+		r = amd_iommu_activate_guest_mode(ir_data, cpu, ga_log_intr);
+	} else {
+		r = amd_iommu_deactivate_guest_mode(ir_data);
+	}
+
+	if (r)
+		return r;
+
+	irqfd->irq_bypass_data = ir_data;
+	list_add(&irqfd->vcpu_list, &svm->ir_list);
+	return 0;
+}
+
  int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
  			unsigned int host_irq, uint32_t guest_irq,
  			struct kvm_vcpu *vcpu, u32 vector)
@@ -888,36 +933,11 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
  		struct amd_iommu_pi_data pi_data = {
  			.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
  					     vcpu->vcpu_idx),
-			.is_guest_mode = kvm_vcpu_apicv_active(vcpu),
  			.vapic_addr = avic_get_backing_page_address(to_svm(vcpu)),
  			.vector = vector,
  		};
-		struct vcpu_svm *svm = to_svm(vcpu);
-		u64 entry;
  		int ret;
 
-		/*
-		 * Prevent the vCPU from being scheduled out or migrated until
-		 * the IRTE is updated and its metadata has been added to the
-		 * list of IRQs being posted to the vCPU, to ensure the IRTE
-		 * isn't programmed with stale pCPU/IsRunning information.
-		 */
-		guard(raw_spinlock_irqsave)(&svm->ir_list_lock);
-
-		/*
-		 * Update the target pCPU for IOMMU doorbells if the vCPU is
-		 * running.  If the vCPU is NOT running, i.e. is blocking or
-		 * scheduled out, KVM will update the pCPU info when the vCPU
-		 * is awakened and/or scheduled in.  See also avic_vcpu_load().
-		 */
-		entry = svm->avic_physical_id_entry;
-		if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK) {
-			pi_data.cpu = entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
-		} else {
-			pi_data.cpu = -1;
-			pi_data.ga_log_intr = entry & AVIC_PHYSICAL_ID_ENTRY_GA_LOG_INTR;
-		}
-
  		ret = irq_set_vcpu_affinity(host_irq, &pi_data);
  		if (ret)
  			return ret;
@@ -932,9 +952,11 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
  			return -EIO;
  		}
 
-		irqfd->irq_bypass_data = pi_data.ir_data;
-		list_add(&irqfd->vcpu_list, &svm->ir_list);
-		return 0;
+		ret = avic_pi_add_irte(irqfd, pi_data.ir_data, vcpu);
+		if (WARN_ON_ONCE(ret))
+			irq_set_vcpu_affinity(host_irq, NULL);
+
+		return ret;
  	}
  	return irq_set_vcpu_affinity(host_irq, NULL);
  }
drivers/iommu/amd/iommu.c +7 -17
index 5d45795c367a..855c6309900c 100644
@@ -3970,7 +3970,6 @@ EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode);
 
  static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *info)
  {
-	int ret;
  	struct amd_iommu_pi_data *pi_data = info;
  	struct amd_ir_data *ir_data = data->chip_data;
  	struct irq_2_irte *irte_info = &ir_data->irq_2_irte;
@@ -3993,25 +3992,16 @@ static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *info)
 
  	ir_data->cfg = irqd_cfg(data);
 
-	if (pi_data) {
-		pi_data->ir_data = ir_data;
+	if (!pi_data)
+		return amd_iommu_deactivate_guest_mode(ir_data);
 
-		ir_data->ga_root_ptr = (pi_data->vapic_addr >> 12);
-		ir_data->ga_vector = pi_data->vector;
-		ir_data->ga_tag = pi_data->ga_tag;
-		if (pi_data->is_guest_mode)
-			ret = amd_iommu_activate_guest_mode(ir_data, pi_data->cpu,
-							    pi_data->ga_log_intr);
-		else
-			ret = amd_iommu_deactivate_guest_mode(ir_data);
-	} else {
-		ret = amd_iommu_deactivate_guest_mode(ir_data);
-	}
-
-	return ret;
+	pi_data->ir_data = ir_data;
+	ir_data->ga_root_ptr = (pi_data->vapic_addr >> 12);
+	ir_data->ga_vector = pi_data->vector;
+	ir_data->ga_tag = pi_data->ga_tag;
+	return 0;
  }
 
-
  static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu,
  			       struct amd_ir_data *ir_data,
  			       struct irq_2_irte *irte_info,

base-commit: 9448598b22c50c8a5bb77a9103e2d49f134c9578 --

author maintainer
#79 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

Quoted 15 lines · Paolo Bonzini
Quoted 10 lines · Sean Christopherson
Quoted 4 lines · Paolo Bonzini

notably, __setup_irq() calls wake_up_process outside desc->lock. Therefore I'd like so much to treat it as a kernel/irq/ bug; and the simplest (perhaps too simple...) fix is to drop the wake_up_process(). The only cost is extra latency on the next interrupt after an affinity change.

Alternatively, what if we rework the KVM<=>IOMMU exchange to decouple updating the IRTE from binding the metadata to the vCPU? KVM already has the necessary exports to do "out-of-band" updates due to the AVIC architecture requiring IRTE updates on scheduling changes.

In fact this was actually my first idea, exactly because it makes svm->ir_list_lock a leaf lock!

Quoted 7 lines

I threw it away because it makes amd_ir_set_vcpu_affinity() weird, passing back the ir_data but not really doing anything else. Basically its role becomes little more than violate abstractions, which seemed wrong. On the other hand, drivers/iommu is already very much tied to the KVM vendor modules (in particular avic.c already calls amd_iommu_{,de}activate_guest_mode), so who am I to judge what the IOMMU driver does.

Yeah, I 100% agree the whole thing is a bit gross, but practically speaking it's just not feasible to properly abstract the interaction, because in reality the IOMMU implementation is tightly coupled to the CPU implementation. E.g. passing in the address of a PID isn't going to work well with an AMD IOMMU, and passing in the address of a vCPI isn't going to work well with an Intel IOMMU.

And FWIW, the lack of true abstraction isn't limited to x86. ARM's GIGv4 passes around "struct its_cmd_info" and PPC uses "struct kvmppc_xive_irq_state".

I mean, we could do what GICv4 does and use irq_set_vcpu_affinity() to pass different commands to the IOMMU, e.g. by formalizing "enum avic_vcpu_action" between KVM and IOMMU so that avic_update_iommu_vcpu_affinity() wouldn't need to _directly_ call AMD IOMMU code. But IMO that would be a net negative because in practice all it would do is make it harder to understand what's going on. And it would more directly create this potential deadlock.

Huh. Which begs the question of whether or not ARM is also affected by this deadlock, without the extra hop through svm->ir_list_lock:

(a) irq_set_thread_affinity() triggers the scheduler via wake_up_process(),

      while irq_desc->lock is taken

(b) the scheduler calls into KVM with rq_lock taken, and KVM uses

      irq_set_vcpu_affinity() via vgic_v4_{load,put}()

If ARM is affected, then maybe fixing this in irq_set_thread_affinity() is indeed better than fudging around the deadlock in avic.c.

#80 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

In reply to Paolo Bonzini
Quoted 82 lines · Paolo Bonzini
Quoted 30 lines · Ankit Soni
   ======================================================
   WARNING: possible circular locking dependency detected
   6.19.0-rc2 #20 Tainted: G            E
   ------------------------------------------------------
   CPU 58/KVM/28597 is trying to acquire lock:
     ff12c47d4b1f34c0 (&irq_desc_lock_class){-.-.}-{2:2}, at: __irq_get_desc_lock+0x58/0xa0

     but task is already holding lock:
     ff12c49b28552110 (&svm->ir_list_lock){....}-{2:2}, at: avic_pi_update_irte+0x147/0x270 [kvm_amd]

     which lock already depends on the new lock.

   Chain exists of:
     &irq_desc_lock_class --> &rq->__lock --> &svm->ir_list_lock

   Possible unsafe locking scenario:

         CPU0                            CPU1
         ----                            ----
    lock(&svm->ir_list_lock);
                                       lock(&rq->__lock);
                                       lock(&svm->ir_list_lock);
    lock(&irq_desc_lock_class);

         *** DEADLOCK ***

So lockdep sees:

   &irq_desc_lock_class -> &rq->__lock -> &svm->ir_list_lock
Quoted 7 lines

while avic_pi_update_irte() currently holds svm->ir_list_lock and then takes irq_desc_lock via irq_set_vcpu_affinity(), which creates the potential inversion.

- Is this lockdep warning expected/benign in this code path, or does it indicate a real potential deadlock between svm->ir_list_lock and irq_desc_lock with AVIC + irq_bypass + VFIO?

I'd treat it as a potential (if unlikely) deadlock:

(a) irq_set_thread_affinity triggers the scheduler via wake_up_process, while irq_desc->lock is taken

(b) the scheduler calls into KVM with rq_lock taken, and KVM uses ir_list_lock within __avic_vcpu_load/__avic_vcpu_put

(c) KVM wants to block scheduling for a while and uses ir_list_lock for that purpose, but then takes irq_set_vcpu_affinity takes irq_desc->lock.

I don't think there's an alternative choice of lock for (c); and there's no easy way to pull the irq_desc->lock out of the IRQ subsystem--in fact the stickiness of the situation comes from rq->rq_lock and irq_desc->lock being both internal and not leaf.

Of the three, the most sketchy is (a); notably, __setup_irq() calls wake_up_process outside desc->lock. Therefore I'd like so much to treat it as a kernel/irq/ bug; and the simplest (perhaps too simple...) fix is to drop the wake_up_process(). The only cost is extra latency on the next interrupt after an affinity change.

kernel/irq/manage.c +2 -6
index 8b1b4c8a4f54..fc135bd079a4 100644
@@ -189,14 +189,10 @@ static void irq_set_thread_affinity(struct irq_desc *desc)
 	struct irqaction *action;
 	for_each_action_of_desc(desc, action) {
-		if (action->thread) {
+		if (action->thread)
 			set_bit(IRQTF_AFFINITY, &action->thread_flags);
-			wake_up_process(action->thread);
-		}
-		if (action->secondary && action->secondary->thread) {
+		if (action->secondary && action->secondary->thread)
 			set_bit(IRQTF_AFFINITY, &action->secondary->thread_flags);
-			wake_up_process(action->secondary->thread);
-		}
 	}
 }

Marc, what do you think?

Quoted 2 lines

Paolo

Hi Paolo,
With the above patch I’m still seeing the same circular dependency warning. However, and with Sean’s patch I’m not seeing any warnings.

[  335.128640] ======================================================
[  335.128650] WARNING: possible circular locking dependency detected
[  335.128660] 6.19.0-rc2-dirty #21 Tainted: G            E
[  335.128668] ------------------------------------------------------
[  335.128675] CPU 96/KVM/28699 is trying to acquire lock:
[  335.128682] ff453c6b59bb18c0 (&irq_desc_lock_class){-.-.}-{2:2}, at: __irq_get_desc_lock+0x58/0xa0
[  335.128704]
               but task is already holding lock:
[  335.128712] ff453c8b0946e600 (&svm->ir_list_lock){....}-{2:2}, at: avic_pi_update_irte+0x147/0x270 [kvm_amd]
[  335.128732]
               which lock already depends on the new lock.
[  335.128742]
               the existing dependency chain (in reverse order) is:
[  335.128751]
               -> #4 (&svm->ir_list_lock){....}-{2:2}:
[  335.128760]        _raw_spin_lock_irqsave+0x4e/0xb0
[  335.128772]        __avic_vcpu_put+0x7a/0x150 [kvm_amd]
[  335.128783]        avic_vcpu_put+0x50/0x70 [kvm_amd]
[  335.128791]        svm_vcpu_put+0x38/0x70 [kvm_amd]
[  335.128800]        kvm_arch_vcpu_put+0x21b/0x330 [kvm]
[  335.128854]        kvm_sched_out+0x62/0x90 [kvm]
[  335.128893]        __schedule+0x8d3/0x1d10
[  335.128901]        __cond_resched+0x5c/0x80
[  335.128909]        __kmalloc_cache_noprof+0x3d7/0x730
[  335.128920]        kvm_hv_vcpu_init+0x48/0x260 [kvm]
[  335.128957]        kvm_hv_set_msr_common+0x5b/0x12a0 [kvm]
[  335.128988]        kvm_set_msr_common+0x468/0x1310 [kvm]
[  335.129019]        svm_set_msr+0x645/0x730 [kvm_amd]
[  335.129028]        __kvm_set_msr+0xa3/0x2f0 [kvm]
[  335.129066]        kvm_set_msr_ignored_check+0x23/0x1b0 [kvm]
[  335.129096]        do_set_msr+0x76/0xd0 [kvm]
[  335.129126]        msr_io+0xbe/0x1c0 [kvm]
[  335.129152]        kvm_arch_vcpu_ioctl+0x700/0x2090 [kvm]
[  335.129181]        kvm_vcpu_ioctl+0x632/0xc60 [kvm]
[  335.129215]        __x64_sys_ioctl+0xa5/0x100
[  335.129224]        x64_sys_call+0x1243/0x26b0
[  335.129234]        do_syscall_64+0x93/0x1470
[  335.129242]        entry_SYSCALL_64_after_hwframe+0x76/0x7e
[  335.129251]
               -> #3 (&rq->__lock){-.-.}-{2:2}:
[  335.129259]        _raw_spin_lock_nested+0x32/0x80
[  335.129267]        raw_spin_rq_lock_nested+0x22/0xa0
[  335.129276]        task_rq_lock+0x5f/0x150
[  335.129392]        cgroup_move_task+0x46/0x110
[  335.129494]        css_set_move_task+0xe1/0x240
[  335.129614]        cgroup_post_fork+0x98/0x2d0
[  335.129704]        copy_process+0x1ea8/0x2330
[  335.129795]        kernel_clone+0xa7/0x440
[  335.129883]        user_mode_thread+0x63/0x90
[  335.129970]        rest_init+0x28/0x200
[  335.130056]        start_kernel+0xae0/0xcd0
[  335.130144]        x86_64_start_reservations+0x18/0x30
[  335.130230]        x86_64_start_kernel+0xfd/0x150
[  335.130312]        common_startup_64+0x13e/0x141
[  335.130396]
               -> #2 (&p->pi_lock){-.-.}-{2:2}:
[  335.130554]        _raw_spin_lock_irqsave+0x4e/0xb0
[  335.130631]        try_to_wake_up+0x59/0xaa0
[  335.130709]        wake_up_process+0x15/0x30
[  335.130785]        create_worker+0x154/0x250
[  335.130862]        workqueue_init+0x414/0x760
[  335.130938]        kernel_init_freeable+0x24f/0x630
[  335.131013]        kernel_init+0x1b/0x200
[  335.131086]        ret_from_fork+0x344/0x3a0
[  335.131161]        ret_from_fork_asm+0x1a/0x30
[  335.131235]
               -> #1 (&pool->lock){-.-.}-{2:2}:
[  335.131375]        _raw_spin_lock+0x34/0x80
[  335.131448]        __queue_work+0xf4/0x740
[  335.131523]        queue_work_on+0x70/0xd0
[  335.131595]        irq_set_affinity_locked+0x13b/0x250
[  335.131669]        __irq_apply_affinity_hint+0xf8/0x120
[  335.131742]        mlx5_irq_alloc+0x28e/0x4e0 [mlx5_core]
[  335.131884]        mlx5_irq_request+0x125/0x140 [mlx5_core]
[  335.131997]        mlx5_irq_request_vector+0xb4/0x110 [mlx5_core]
[  335.132107]        comp_irq_request_pci+0x68/0xf0 [mlx5_core]
[  335.132219]        mlx5_comp_eqn_get+0x127/0x850 [mlx5_core]
[  335.132327]        mlx5e_create_cq+0x58/0x260 [mlx5_core]
[  335.132452]        mlx5e_open_drop_rq+0x11c/0x220 [mlx5_core]
[  335.132573]        mlx5e_init_nic_rx+0x2c/0x270 [mlx5_core]
[  335.132687]        mlx5e_attach_netdev+0xed/0x340 [mlx5_core]
[  335.132794]        _mlx5e_resume+0x6a/0xd0 [mlx5_core]
[  335.132900]        mlx5e_probe+0x5e3/0xa50 [mlx5_core]
[  335.133010]        auxiliary_bus_probe+0x45/0x90
[  335.133086]        really_probe+0xf1/0x410
[  335.133160]        __driver_probe_device+0x8c/0x190
[  335.133233]        driver_probe_device+0x24/0xd0
[  335.133305]        __device_attach_driver+0xcd/0x170
[  335.133377]        bus_for_each_drv+0x99/0x100
[  335.133450]        __device_attach+0xba/0x1f0
[  335.133523]        device_initial_probe+0x4e/0x50
[  335.133595]        bus_probe_device+0x3c/0xa0
[  335.133668]        device_add+0x6af/0x8a0
[  335.133740]        __auxiliary_device_add+0x43/0xc0
[  335.133813]        add_adev+0xd3/0x160 [mlx5_core]
[  335.133931]        mlx5_rescan_drivers_locked+0x1ee/0x340 [mlx5_core]
[  335.134045]        mlx5_register_device+0x37/0xb0 [mlx5_core]
[  335.134156]        mlx5_init_one_devl_locked+0x43e/0x710 [mlx5_core]
[  335.134267]        probe_one+0x35b/0x530 [mlx5_core]
[  335.134382]        local_pci_probe+0x47/0xb0
[  335.134460]        work_for_cpu_fn+0x1a/0x30
[  335.134538]        process_one_work+0x22b/0x6f0
[  335.134613]        worker_thread+0x1c6/0x3b0
[  335.134687]        kthread+0x110/0x230
[  335.134764]        ret_from_fork+0x344/0x3a0
[  335.134838]        ret_from_fork_asm+0x1a/0x30
[  335.134913]
               -> #0 (&irq_desc_lock_class){-.-.}-{2:2}:
[  335.135059]        __lock_acquire+0x1595/0x2640
[  335.135135]        lock_acquire+0xc4/0x2c0
[  335.135210]        _raw_spin_lock_irqsave+0x4e/0xb0
[  335.135286]        __irq_get_desc_lock+0x58/0xa0
[  335.135362]        irq_set_vcpu_affinity+0x4a/0x100
[  335.135438]        avic_pi_update_irte+0x170/0x270 [kvm_amd]
[  335.135521]        kvm_pi_update_irte+0xea/0x220 [kvm]
[  335.135633]        kvm_arch_irq_bypass_add_producer+0x9b/0xb0 [kvm]
[  335.135737]        __connect+0x5f/0x100 [irqbypass]
[  335.135815]        irq_bypass_register_producer+0xe4/0xb90 [irqbypass]
[  335.135895]        vfio_msi_set_vector_signal+0x1b0/0x330 [vfio_pci_core]
[  335.135979]        vfio_msi_set_block+0x5a/0xd0 [vfio_pci_core]
[  335.136060]        vfio_pci_set_msi_trigger+0x19e/0x260 [vfio_pci_core]
[  335.136142]        vfio_pci_set_irqs_ioctl+0x46/0x140 [vfio_pci_core]
[  335.136224]        vfio_pci_core_ioctl+0x6ea/0xc20 [vfio_pci_core]
[  335.136306]        vfio_device_fops_unl_ioctl+0xb1/0x9d0 [vfio]
[  335.136390]        __x64_sys_ioctl+0xa5/0x100
[  335.136470]        x64_sys_call+0x1243/0x26b0
[  335.136553]        do_syscall_64+0x93/0x1470
[  335.136632]        entry_SYSCALL_64_after_hwframe+0x76/0x7e
[  335.136712]
               other info that might help us debug this:
[  335.136942] Chain exists of:
                 &irq_desc_lock_class --> &rq->__lock --> &svm->ir_list_lock
[  335.137178]  Possible unsafe locking scenario:
[  335.137340]        CPU0                    CPU1
[  335.137420]        ----                    ----
[  335.137502]   lock(&svm->ir_list_lock);
[  335.137582]                                lock(&rq->__lock);
[  335.137664]                                lock(&svm->ir_list_lock);
[  335.137746]   lock(&irq_desc_lock_class);
[  335.137826]
                *** DEADLOCK ***
[  335.138057] 4 locks held by CPU 96/KVM/28699:
[  335.138136]  #0: ff453c6c28bae800 (&vdev->igate){+.+.}-{4:4}, at: vfio_pci_core_ioctl+0x6d2/0xc20 [vfio_pci_core]
[  335.138223]  #1: ffffffffc0faf110 (lock#10){+.+.}-{4:4}, at: irq_bypass_register_producer+0x31/0xb90 [irqbypass]
[  335.138310]  #2: ff453c8b11179478 (&kvm->irqfds.lock){....}-{3:3}, at: kvm_arch_irq_bypass_add_producer+0x2d/0xb0 [kvm]
[  335.138423]  #3: ff453c8b0946e600 (&svm->ir_list_lock){....}-{2:2}, at: avic_pi_update_irte+0x147/0x270 [kvm_amd]

-Ankit

maintainer
#81 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

In reply to Paolo Bonzini
Quoted 44 lines · Paolo Bonzini
Quoted 30 lines · Ankit Soni
   ======================================================
   WARNING: possible circular locking dependency detected
   6.19.0-rc2 #20 Tainted: G            E
   ------------------------------------------------------
   CPU 58/KVM/28597 is trying to acquire lock:
     ff12c47d4b1f34c0 (&irq_desc_lock_class){-.-.}-{2:2}, at: __irq_get_desc_lock+0x58/0xa0

     but task is already holding lock:
     ff12c49b28552110 (&svm->ir_list_lock){....}-{2:2}, at: avic_pi_update_irte+0x147/0x270 [kvm_amd]

     which lock already depends on the new lock.

   Chain exists of:
     &irq_desc_lock_class --> &rq->__lock --> &svm->ir_list_lock

   Possible unsafe locking scenario:

         CPU0                            CPU1
         ----                            ----
    lock(&svm->ir_list_lock);
                                       lock(&rq->__lock);
                                       lock(&svm->ir_list_lock);
    lock(&irq_desc_lock_class);

         *** DEADLOCK ***

So lockdep sees:

   &irq_desc_lock_class -> &rq->__lock -> &svm->ir_list_lock
Quoted 7 lines

while avic_pi_update_irte() currently holds svm->ir_list_lock and then takes irq_desc_lock via irq_set_vcpu_affinity(), which creates the potential inversion.

- Is this lockdep warning expected/benign in this code path, or does it indicate a real potential deadlock between svm->ir_list_lock and irq_desc_lock with AVIC + irq_bypass + VFIO?

I'd treat it as a potential (if unlikely) deadlock:

(a) irq_set_thread_affinity triggers the scheduler via wake_up_process, while irq_desc->lock is taken

Quoted 8 lines

(b) the scheduler calls into KVM with rq_lock taken, and KVM uses ir_list_lock within __avic_vcpu_load/__avic_vcpu_put

(c) KVM wants to block scheduling for a while and uses ir_list_lock for that purpose, but then takes irq_set_vcpu_affinity takes irq_desc->lock.

I don't think there's an alternative choice of lock for (c); and there's no easy way to pull the irq_desc->lock out of the IRQ subsystem--in fact

Don't even think about that.

Quoted 6 lines

the stickiness of the situation comes from rq->rq_lock and irq_desc->lock being both internal and not leaf.

Of the three, the most sketchy is (a); notably, __setup_irq() calls wake_up_process outside desc->lock. Therefore I'd like so much to treat it as a kernel/irq/ bug; and the simplest (perhaps too simple...) fix is

It's not more sketchy than VIRT assuming that it can do what it wants under rq->lock. :)

Quoted 2 lines

to drop the wake_up_process(). The only cost is extra latency on the next interrupt after an affinity change.

The real problematic cost is that in an isolation scenario the wakeup happens at the next interrupt which might be far into the isolated phase. That's why the wakeup is there. See:

c99303a2d2a2 ("genirq: Wake interrupt threads immediately when changing affinity")

Obviously you did not even bother to look that up otherwise you would have CC'ed Crystal and asked her to take a look...

Thanks,

tglx

maintainer
#82 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

Quoted 7 lines · Thomas Gleixner
Quoted 3 lines · Paolo Bonzini

Of the three, the most sketchy is (a); notably, __setup_irq() calls wake_up_process outside desc->lock. Therefore I'd like so much to treat it as a kernel/irq/ bug; and the simplest (perhaps too simple...) fix is

It's not more sketchy than VIRT assuming that it can do what it wants under rq->lock. :)

And just for the record, that's not the only place in the irq core which has that lock chain.

irq_set_affinity_locked() // invoked with desc::lock held
if (desc->affinity_notify)
schedule_work() // Ends up taking rq::lock

and that's the case since cd7eab44e994 ("genirq: Add IRQ affinity notifiers"), which was added 15 years ago.

Are you still claiming that this is a kernel/irq bug?

Thanks,

tglx

maintainer
#83 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

maintainer
#84 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

In reply to Thomas Gleixner

Sorry, not sure how the previous email ended up encrypted.

Quoted 12 lines · Thomas Gleixner
Quoted 7 lines · Thomas Gleixner
Quoted 3 lines · Paolo Bonzini

Of the three, the most sketchy is (a); notably, __setup_irq() calls wake_up_process outside desc->lock. Therefore I'd like so much to treat it as a kernel/irq/ bug; and the simplest (perhaps too simple...) fix is

It's not more sketchy than VIRT assuming that it can do what it wants under rq->lock. 🙂

And just for the record, that's not the only place in the irq core which has that lock chain.

Quoted 8 lines

irq_set_affinity_locked() // invoked with desc::lock held
if (desc->affinity_notify)
schedule_work() // Ends up taking rq::lock

and that's the case since cd7eab44e994 ("genirq: Add IRQ affinity notifiers"), which was added 15 years ago.

Are you still claiming that this is a kernel/irq bug?

Not really, I did say I'd like to treat it as a kernel/irq bug... but certainly didn't have hopes high enough to "claim" that. I do think that it's ugly to have locks that are internal, non-leaf and held around callbacks; but people smarter than me have thought about it and you can't call it a bug anyway.

For x86/AMD we have a way to fix it, so that part is not a problem.

For the call(*) to irq_set_affinity() in arch/arm64/kvm/'s vgic_v4_load() I think it can be solved as well. kvm_make_request(KVM_REQ_RELOAD_GICv4) will delay vgic_v4_load() to a safe spot, so just cache the previous smp_processor_id() and, if it is different, do the kvm_make_request() and return instead of calling irq_set_affinity().

vgic_v3_load() is the only place that calls it from the preempt notifier, so this behavior can be tied to a "bool delay_set_affinity" argument to vgic_v4_load() or placed in a different function.

Marc/Oliver, does that sound doable?

Paolo

(*) kvm_sched_in() preempt notifier -> kvm_arch_vcpu_load() ->

    kvm_vgic_load() -> vgic_v3_load() -> vgic_v4_load()
maintainer
#85 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

Quoted 40 lines · Paolo Bonzini

Sorry, not sure how the previous email ended up encrypted.

Quoted 12 lines · Thomas Gleixner
Quoted 7 lines · Thomas Gleixner
Quoted 3 lines · Paolo Bonzini

Of the three, the most sketchy is (a); notably, __setup_irq() calls wake_up_process outside desc->lock. Therefore I'd like so much to treat it as a kernel/irq/ bug; and the simplest (perhaps too simple...) fix is

It's not more sketchy than VIRT assuming that it can do what it wants under rq->lock. 🙂

And just for the record, that's not the only place in the irq core which has that lock chain.

Quoted 8 lines

irq_set_affinity_locked() // invoked with desc::lock held
if (desc->affinity_notify)
schedule_work() // Ends up taking rq::lock

and that's the case since cd7eab44e994 ("genirq: Add IRQ affinity notifiers"), which was added 15 years ago.

Are you still claiming that this is a kernel/irq bug?

Not really, I did say I'd like to treat it as a kernel/irq bug... but certainly didn't have hopes high enough to "claim" that. I do think that it's ugly to have locks that are internal, non-leaf and held around callbacks; but people smarter than me have thought about it and you can't call it a bug anyway.

For x86/AMD we have a way to fix it, so that part is not a problem.

For the call(*) to irq_set_affinity() in arch/arm64/kvm/'s vgic_v4_load() I think it can be solved as well. kvm_make_request(KVM_REQ_RELOAD_GICv4) will delay vgic_v4_load() to a safe spot, so just cache the previous smp_processor_id() and, if it is different, do the kvm_make_request() and return instead of calling irq_set_affinity().

Quoted 5 lines

vgic_v3_load() is the only place that calls it from the preempt notifier, so this behavior can be tied to a "bool delay_set_affinity" argument to vgic_v4_load() or placed in a different function.

Marc/Oliver, does that sound doable?

Potentially. But there are a few gotchas that may need surgery beyond KVM itself, all the way down to the ITS code that abstract the differences between v4.0 and v4.1.

I'll have a look over the weekend.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
maintainer
#86 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

In reply to Paolo Bonzini
Quoted 8 lines · Paolo Bonzini
Quoted 1 line · Thomas Gleixner

Are you still claiming that this is a kernel/irq bug?

Not really, I did say I'd like to treat it as a kernel/irq bug... but certainly didn't have hopes high enough to "claim" that. I do think that it's ugly to have locks that are internal, non-leaf and held around callbacks; but people smarter than me have thought about it and you can't call it a bug anyway.

Deep core code has a tendency to be ugly. But if it makes your life easier, then these wakeups can be delayed via an irq_work to be outside of the lock. That needs some life-time issues to be addressed, but should be doable.

Thanks,

tglx

maintainer
#87 lore raw

Re: possible deadlock due to irq_set_thread_affinity() calling into the scheduler (was Re: [PATCH v3 38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU)

Quoted 11 lines · Thomas Gleixner
Quoted 8 lines · Paolo Bonzini
Quoted 1 line · Thomas Gleixner

Are you still claiming that this is a kernel/irq bug?

Not really, I did say I'd like to treat it as a kernel/irq bug... but certainly didn't have hopes high enough to "claim" that. I do think that it's ugly to have locks that are internal, non-leaf and held around callbacks; but people smarter than me have thought about it and you can't call it a bug anyway.

Quoted 4 lines

Deep core code has a tendency to be ugly. But if it makes your life easier, then these wakeups can be delayed via an irq_work to be outside of the lock. That needs some life-time issues to be addressed, but should be doable.

Thanks for the suggestion---hopefully it's not needed at all and we can delay taking the lock in KVM.

Paolo

author maintainer
#88 lore raw

PATCH v3 39/62 iommu/amd: Document which IRTE fields amd_iommu_update_ga() can modify

In reply to Sean Christopherson

Add a comment to amd_iommu_update_ga() to document what fields it can safely modify without issuing an invalidation of the IRTE, and to explain its role in keeping GA IRTEs up-to-date.

Per page 93 of the IOMMU spec dated Feb 2025:

When virtual interrupts are enabled by setting MMIO Offset 0018h[GAEn] and IRTE[GuestMode=1], IRTE[IsRun], IRTE[Destination], and if present IRTE[GATag], are not cached by the IOMMU. Modifications to these fields do not require an invalidation of the Interrupt Remapping Table.

---
 drivers/iommu/amd/iommu.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
drivers/iommu/amd/iommu.c +12 -0
index 36749efcc781..5adc932b947e 100644
@@ -3986,6 +3986,18 @@ int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
 	return 0;
 }
 
+/*
+ * Update the pCPU information for an IRTE that is configured to post IRQs to
+ * a vCPU, without issuing an IOMMU invalidation for the IRTE.
+ *
+ * This API is intended to be used when a vCPU is scheduled in/out (or stops
+ * running for any reason), to do a fast update of IsRun and (conditionally)
+ * Destination.
+ *
+ * Per the IOMMU spec, the Destination, IsRun, and GATag fields are not cached
+ * and thus don't require an invalidation to ensure the IOMMU consumes fresh
+ * information.
+ */
 int amd_iommu_update_ga(int cpu, bool is_run, void *data)
 {
 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#89 lore raw

PATCH v3 40/62 iommu/amd: KVM: SVM: Infer IsRun from validity of pCPU destination

In reply to Sean Christopherson

Infer whether or not a vCPU should be marked running from the validity of the pCPU on which it is running. amd_iommu_update_ga() already skips the IRTE update if the pCPU is invalid, i.e. passing %true for is_run with an invalid pCPU would be a blatant and egregrious KVM bug.

---
 arch/x86/kvm/svm/avic.c   | 11 +++++------
 drivers/iommu/amd/iommu.c | 14 +++++++++-----
 include/linux/amd-iommu.h |  6 ++----
 3 files changed, 16 insertions(+), 15 deletions(-)
arch/x86/kvm/svm/avic.c +5 -6
index 4747fb09aca4..c79648d96752 100644
@@ -832,7 +832,7 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		entry = svm->avic_physical_id_entry;
 		if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
 			amd_iommu_update_ga(entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK,
-					    true, pi_data.ir_data);
+					    pi_data.ir_data);
 
 		irqfd->irq_bypass_data = pi_data.ir_data;
 		list_add(&irqfd->vcpu_list, &svm->ir_list);
@@ -841,8 +841,7 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 	return irq_set_vcpu_affinity(host_irq, NULL);
 }
 
-static inline int
-avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu, bool r)
+static inline int avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu)
 {
 	int ret = 0;
 	struct vcpu_svm *svm = to_svm(vcpu);
@@ -861,7 +860,7 @@ avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu, bool r)
 		return 0;
 
 	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list) {
-		ret = amd_iommu_update_ga(cpu, r, irqfd->irq_bypass_data);
+		ret = amd_iommu_update_ga(cpu, irqfd->irq_bypass_data);
 		if (ret)
 			return ret;
 	}
@@ -923,7 +922,7 @@ void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 
 	WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
 
-	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id, true);
+	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id);
 
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
 }
@@ -963,7 +962,7 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 	 */
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
 
-	avic_update_iommu_vcpu_affinity(vcpu, -1, 0);
+	avic_update_iommu_vcpu_affinity(vcpu, -1);
 
 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
 	svm->avic_physical_id_entry = entry;
drivers/iommu/amd/iommu.c +9 -5
index 5adc932b947e..bb804bbc916b 100644
@@ -3990,15 +3990,17 @@ int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
  * Update the pCPU information for an IRTE that is configured to post IRQs to
  * a vCPU, without issuing an IOMMU invalidation for the IRTE.
  *
- * This API is intended to be used when a vCPU is scheduled in/out (or stops
- * running for any reason), to do a fast update of IsRun and (conditionally)
- * Destination.
+ * If the vCPU is associated with a pCPU (@cpu >= 0), configure the Destination
+ * with the pCPU's APIC ID and set IsRun, else clear IsRun.  I.e. treat vCPUs
+ * that are associated with a pCPU as running.  This API is intended to be used
+ * when a vCPU is scheduled in/out (or stops running for any reason), to do a
+ * fast update of IsRun and (conditionally) Destination.
  *
  * Per the IOMMU spec, the Destination, IsRun, and GATag fields are not cached
  * and thus don't require an invalidation to ensure the IOMMU consumes fresh
  * information.
  */
-int amd_iommu_update_ga(int cpu, bool is_run, void *data)
+int amd_iommu_update_ga(int cpu, void *data)
 {
 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
@@ -4015,8 +4017,10 @@ int amd_iommu_update_ga(int cpu, bool is_run, void *data)
 					APICID_TO_IRTE_DEST_LO(cpu);
 		entry->hi.fields.destination =
 					APICID_TO_IRTE_DEST_HI(cpu);
+		entry->lo.fields_vapic.is_run = true;
+	} else {
+		entry->lo.fields_vapic.is_run = false;
 	}
-	entry->lo.fields_vapic.is_run = is_run;
 
 	return __modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
 				ir_data->irq_2_irte.index, entry);
include/linux/amd-iommu.h +2 -4
index 99b4fa9a0296..fe0e16ffe0e5 100644
@@ -30,8 +30,7 @@ static inline void amd_iommu_detect(void) { }
 /* IOMMU AVIC Function */
 extern int amd_iommu_register_ga_log_notifier(int (*notifier)(u32));
 
-extern int
-amd_iommu_update_ga(int cpu, bool is_run, void *data);
+extern int amd_iommu_update_ga(int cpu, void *data);
 
 extern int amd_iommu_activate_guest_mode(void *data);
 extern int amd_iommu_deactivate_guest_mode(void *data);
@@ -44,8 +43,7 @@ amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
 	return 0;
 }
 
-static inline int
-amd_iommu_update_ga(int cpu, bool is_run, void *data)
+static inline int amd_iommu_update_ga(int cpu, void *data)
 {
 	return 0;
 }
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#90 lore raw

PATCH v3 41/62 iommu/amd: Factor out helper for manipulating IRTE GA/CPU info

In reply to Sean Christopherson

Split the guts of amd_iommu_update_ga() to a dedicated helper so that the logic can be shared with flows that put the IRTE into posted mode.

Opportunistically move amd_iommu_update_ga() and its new helper above amd_iommu_activate_guest_mode() so that it's all co-located.

---
 drivers/iommu/amd/iommu.c | 87 +++++++++++++++++++++------------------
 1 file changed, 46 insertions(+), 41 deletions(-)
drivers/iommu/amd/iommu.c +46 -41
index bb804bbc916b..15718b7b8bd4 100644
@@ -3804,6 +3804,52 @@ static const struct irq_domain_ops amd_ir_domain_ops = {
 	.deactivate = irq_remapping_deactivate,
 };
 
+static void __amd_iommu_update_ga(struct irte_ga *entry, int cpu)
+{
+	if (cpu >= 0) {
+		entry->lo.fields_vapic.destination =
+					APICID_TO_IRTE_DEST_LO(cpu);
+		entry->hi.fields.destination =
+					APICID_TO_IRTE_DEST_HI(cpu);
+		entry->lo.fields_vapic.is_run = true;
+	} else {
+		entry->lo.fields_vapic.is_run = false;
+	}
+}
+
+/*
+ * Update the pCPU information for an IRTE that is configured to post IRQs to
+ * a vCPU, without issuing an IOMMU invalidation for the IRTE.
+ *
+ * If the vCPU is associated with a pCPU (@cpu >= 0), configure the Destination
+ * with the pCPU's APIC ID and set IsRun, else clear IsRun.  I.e. treat vCPUs
+ * that are associated with a pCPU as running.  This API is intended to be used
+ * when a vCPU is scheduled in/out (or stops running for any reason), to do a
+ * fast update of IsRun and (conditionally) Destination.
+ *
+ * Per the IOMMU spec, the Destination, IsRun, and GATag fields are not cached
+ * and thus don't require an invalidation to ensure the IOMMU consumes fresh
+ * information.
+ */
+int amd_iommu_update_ga(int cpu, void *data)
+{
+	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
+	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
+
+	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
+	    !entry || !entry->lo.fields_vapic.guest_mode)
+		return 0;
+
+	if (!ir_data->iommu)
+		return -ENODEV;
+
+	__amd_iommu_update_ga(entry, cpu);
+
+	return __modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
+				ir_data->irq_2_irte.index, entry);
+}
+EXPORT_SYMBOL(amd_iommu_update_ga);
+
 int amd_iommu_activate_guest_mode(void *data)
 {
 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
@@ -3985,45 +4031,4 @@ int amd_iommu_create_irq_domain(struct amd_iommu *iommu)
 
 	return 0;
 }
-
-/*
- * Update the pCPU information for an IRTE that is configured to post IRQs to
- * a vCPU, without issuing an IOMMU invalidation for the IRTE.
- *
- * If the vCPU is associated with a pCPU (@cpu >= 0), configure the Destination
- * with the pCPU's APIC ID and set IsRun, else clear IsRun.  I.e. treat vCPUs
- * that are associated with a pCPU as running.  This API is intended to be used
- * when a vCPU is scheduled in/out (or stops running for any reason), to do a
- * fast update of IsRun and (conditionally) Destination.
- *
- * Per the IOMMU spec, the Destination, IsRun, and GATag fields are not cached
- * and thus don't require an invalidation to ensure the IOMMU consumes fresh
- * information.
- */
-int amd_iommu_update_ga(int cpu, void *data)
-{
-	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
-	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
-
-	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
-	    !entry || !entry->lo.fields_vapic.guest_mode)
-		return 0;
-
-	if (!ir_data->iommu)
-		return -ENODEV;
-
-	if (cpu >= 0) {
-		entry->lo.fields_vapic.destination =
-					APICID_TO_IRTE_DEST_LO(cpu);
-		entry->hi.fields.destination =
-					APICID_TO_IRTE_DEST_HI(cpu);
-		entry->lo.fields_vapic.is_run = true;
-	} else {
-		entry->lo.fields_vapic.is_run = false;
-	}
-
-	return __modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
-				ir_data->irq_2_irte.index, entry);
-}
-EXPORT_SYMBOL(amd_iommu_update_ga);
 #endif
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#91 lore raw

PATCH v3 42/62 iommu/amd: KVM: SVM: Set pCPU info in IRTE when setting vCPU affinity

In reply to Sean Christopherson

Now that setting vCPU affinity is guarded with ir_list_lock, i.e. now that avic_physical_id_entry can be safely accessed, set the pCPU info straight-away when setting vCPU affinity. Putting the IRTE into posted mode, and then immediately updating the IRTE a second time if the target vCPU is running is wasteful and confusing.

This also fixes a flaw where a posted IRQ that arrives between putting the IRTE into guest_mode and setting the correct destination could cause the IOMMU to ring the doorbell on the wrong pCPU.

---
 arch/x86/include/asm/irq_remapping.h |  1 +
 arch/x86/kvm/svm/avic.c              | 26 ++++++++++++++------------
 drivers/iommu/amd/iommu.c            |  6 ++++--
 include/linux/amd-iommu.h            |  4 ++--
 4 files changed, 21 insertions(+), 16 deletions(-)
arch/x86/include/asm/irq_remapping.h +1 -0
index 2dbc9cb61c2f..4c75a17632f6 100644
@@ -35,6 +35,7 @@ struct amd_iommu_pi_data {
 	u64 vapic_addr;		/* Physical address of the vCPU's vAPIC. */
 	u32 ga_tag;
 	u32 vector;		/* Guest vector of the interrupt */
+	int cpu;
 	bool is_guest_mode;
 	void *ir_data;
 };
arch/x86/kvm/svm/avic.c +14 -12
index c79648d96752..16557328aa58 100644
@@ -726,6 +726,7 @@ void avic_apicv_post_state_restore(struct kvm_vcpu *vcpu)
 
 static int avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
 {
+	int apic_id = kvm_cpu_get_apicid(vcpu->cpu);
 	int ret = 0;
 	unsigned long flags;
 	struct vcpu_svm *svm = to_svm(vcpu);
@@ -745,7 +746,7 @@ static int avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
 
 	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list) {
 		if (activate)
-			ret = amd_iommu_activate_guest_mode(irqfd->irq_bypass_data);
+			ret = amd_iommu_activate_guest_mode(irqfd->irq_bypass_data, apic_id);
 		else
 			ret = amd_iommu_deactivate_guest_mode(irqfd->irq_bypass_data);
 		if (ret)
@@ -809,6 +810,18 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		 */
 		guard(spinlock_irqsave)(&svm->ir_list_lock);
 
+		/*
+		 * Update the target pCPU for IOMMU doorbells if the vCPU is
+		 * running.  If the vCPU is NOT running, i.e. is blocking or
+		 * scheduled out, KVM will update the pCPU info when the vCPU
+		 * is awakened and/or scheduled in.  See also avic_vcpu_load().
+		 */
+		entry = svm->avic_physical_id_entry;
+		if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
+			pi_data.cpu = entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
+		else
+			pi_data.cpu = -1;
+
 		ret = irq_set_vcpu_affinity(host_irq, &pi_data);
 		if (ret)
 			return ret;
@@ -823,17 +836,6 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			return -EIO;
 		}
 
-		/*
-		 * Update the target pCPU for IOMMU doorbells if the vCPU is
-		 * running.  If the vCPU is NOT running, i.e. is blocking or
-		 * scheduled out, KVM will update the pCPU info when the vCPU
-		 * is awakened and/or scheduled in.  See also avic_vcpu_load().
-		 */
-		entry = svm->avic_physical_id_entry;
-		if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
-			amd_iommu_update_ga(entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK,
-					    pi_data.ir_data);
-
 		irqfd->irq_bypass_data = pi_data.ir_data;
 		list_add(&irqfd->vcpu_list, &svm->ir_list);
 		return 0;
drivers/iommu/amd/iommu.c +4 -2
index 15718b7b8bd4..718bd9604f71 100644
@@ -3850,7 +3850,7 @@ int amd_iommu_update_ga(int cpu, void *data)
 }
 EXPORT_SYMBOL(amd_iommu_update_ga);
 
-int amd_iommu_activate_guest_mode(void *data)
+int amd_iommu_activate_guest_mode(void *data, int cpu)
 {
 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
@@ -3871,6 +3871,8 @@ int amd_iommu_activate_guest_mode(void *data)
 	entry->hi.fields.vector            = ir_data->ga_vector;
 	entry->lo.fields_vapic.ga_tag      = ir_data->ga_tag;
 
+	__amd_iommu_update_ga(entry, cpu);
+
 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
 			      ir_data->irq_2_irte.index, entry);
 }
@@ -3937,7 +3939,7 @@ static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *info)
 		ir_data->ga_root_ptr = (pi_data->vapic_addr >> 12);
 		ir_data->ga_vector = pi_data->vector;
 		ir_data->ga_tag = pi_data->ga_tag;
-		ret = amd_iommu_activate_guest_mode(ir_data);
+		ret = amd_iommu_activate_guest_mode(ir_data, pi_data->cpu);
 	} else {
 		ret = amd_iommu_deactivate_guest_mode(ir_data);
 	}
include/linux/amd-iommu.h +2 -2
index fe0e16ffe0e5..c9f2df0c4596 100644
@@ -32,7 +32,7 @@ extern int amd_iommu_register_ga_log_notifier(int (*notifier)(u32));
 
 extern int amd_iommu_update_ga(int cpu, void *data);
 
-extern int amd_iommu_activate_guest_mode(void *data);
+extern int amd_iommu_activate_guest_mode(void *data, int cpu);
 extern int amd_iommu_deactivate_guest_mode(void *data);
 
 #else /* defined(CONFIG_AMD_IOMMU) && defined(CONFIG_IRQ_REMAP) */
@@ -48,7 +48,7 @@ static inline int amd_iommu_update_ga(int cpu, void *data)
 	return 0;
 }
 
-static inline int amd_iommu_activate_guest_mode(void *data)
+static inline int amd_iommu_activate_guest_mode(void *data, int cpu)
 {
 	return 0;
 }
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#92 lore raw

PATCH v3 43/62 iommu/amd: KVM: SVM: Add IRTE metadata to affined vCPU's list if AVIC is inhibited

In reply to Sean Christopherson

If an IRQ can be posted to a vCPU, but AVIC is currently inhibited on the vCPU, go through the dance of "affining" the IRTE to the vCPU, but leave the actual IRTE in remapped mode. KVM already handles the case where AVIC is inhibited => uninhibited with posted IRQs (see avic_set_pi_irte_mode()), but doesn't handle the scenario where a postable IRQ comes along while AVIC is inhibited.

---
 arch/x86/kvm/svm/avic.c   | 16 ++++++----------
 drivers/iommu/amd/iommu.c |  5 ++++-
 2 files changed, 10 insertions(+), 11 deletions(-)
arch/x86/kvm/svm/avic.c +6 -10
index 16557328aa58..2e3a8fda0355 100644
@@ -780,21 +780,17 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 	 */
 	svm_ir_list_del(irqfd);
 
-	/**
-	 * Here, we setup with legacy mode in the following cases:
-	 * 1. When cannot target interrupt to a specific vcpu.
-	 * 2. Unsetting posted interrupt.
-	 * 3. APIC virtualization is disabled for the vcpu.
-	 * 4. IRQ has incompatible delivery mode (SMI, INIT, etc)
-	 */
-	if (vcpu && kvm_vcpu_apicv_active(vcpu)) {
+	if (vcpu) {
 		/*
-		 * Try to enable guest_mode in IRTE.
+		 * Try to enable guest_mode in IRTE, unless AVIC is inhibited,
+		 * in which case configure the IRTE for legacy mode, but track
+		 * the IRTE metadata so that it can be converted to guest mode
+		 * if AVIC is enabled/uninhibited in the future.
 		 */
 		struct amd_iommu_pi_data pi_data = {
 			.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
 					     vcpu->vcpu_id),
-			.is_guest_mode = true,
+			.is_guest_mode = kvm_vcpu_apicv_active(vcpu),
 			.vapic_addr = avic_get_backing_page_address(to_svm(vcpu)),
 			.vector = vector,
 		};
drivers/iommu/amd/iommu.c +4 -1
index 718bd9604f71..becef69a306d 100644
@@ -3939,7 +3939,10 @@ static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *info)
 		ir_data->ga_root_ptr = (pi_data->vapic_addr >> 12);
 		ir_data->ga_vector = pi_data->vector;
 		ir_data->ga_tag = pi_data->ga_tag;
-		ret = amd_iommu_activate_guest_mode(ir_data, pi_data->cpu);
+		if (pi_data->is_guest_mode)
+			ret = amd_iommu_activate_guest_mode(ir_data, pi_data->cpu);
+		else
+			ret = amd_iommu_deactivate_guest_mode(ir_data);
 	} else {
 		ret = amd_iommu_deactivate_guest_mode(ir_data);
 	}
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#93 lore raw

PATCH v3 44/62 KVM: SVM: Don't check for assigned device(s) when updating affinity

In reply to Sean Christopherson

Don't bother checking if a VM has an assigned device when updating AVIC vCPU affinity, querying ir_list is just as cheap and nothing prevents racing with changes in device assignment.

---
 arch/x86/kvm/svm/avic.c | 3 ---
 1 file changed, 3 deletions(-)
arch/x86/kvm/svm/avic.c +0 -3
index 2e3a8fda0355..dadd982b03c0 100644
@@ -847,9 +847,6 @@ static inline int avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu
 
 	lockdep_assert_held(&svm->ir_list_lock);
 
-	if (!kvm_arch_has_assigned_device(vcpu->kvm))
-		return 0;
-
 	/*
 	 * Here, we go through the per-vcpu ir_list to update all existing
 	 * interrupt remapping table entry targeting this vcpu.
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#94 lore raw

PATCH v3 45/62 KVM: SVM: Don't check for assigned device(s) when activating AVIC

In reply to Sean Christopherson

Don't short-circuit IRTE updating when (de)activating AVIC based on the VM having assigned devices, as nothing prevents AVIC (de)activation from racing with device (de)assignment. And from a performance perspective, bailing early when there is no assigned device doesn't add much, as ir_list_lock will never be contended if there's no assigned device.

---
 arch/x86/kvm/svm/avic.c | 3 ---
 1 file changed, 3 deletions(-)
arch/x86/kvm/svm/avic.c +0 -3
index dadd982b03c0..ab7fb8950cc0 100644
@@ -732,9 +732,6 @@ static int avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
 	struct vcpu_svm *svm = to_svm(vcpu);
 	struct kvm_kernel_irqfd *irqfd;
 
-	if (!kvm_arch_has_assigned_device(vcpu->kvm))
-		return 0;
-
 	/*
 	 * Here, we go through the per-vcpu ir_list to update all existing
 	 * interrupt remapping table entry targeting this vcpu.
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#95 lore raw

PATCH v3 46/62 KVM: SVM: WARN if (de)activating guest mode in IOMMU fails

In reply to Sean Christopherson

WARN if (de)activating "guest mode" for an IRTE entry fails as modifying an IRTE should only fail if KVM is buggy, e.g. has stale metadata.

---
 arch/x86/kvm/svm/avic.c | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)
arch/x86/kvm/svm/avic.c +5 -7
index ab7fb8950cc0..6048cd90e731 100644
@@ -724,10 +724,9 @@ void avic_apicv_post_state_restore(struct kvm_vcpu *vcpu)
 	avic_handle_ldr_update(vcpu);
 }
 
-static int avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
+static void avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
 {
 	int apic_id = kvm_cpu_get_apicid(vcpu->cpu);
-	int ret = 0;
 	unsigned long flags;
 	struct vcpu_svm *svm = to_svm(vcpu);
 	struct kvm_kernel_irqfd *irqfd;
@@ -742,16 +741,15 @@ static int avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
 		goto out;
 
 	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list) {
+		void *data = irqfd->irq_bypass_data;
+
 		if (activate)
-			ret = amd_iommu_activate_guest_mode(irqfd->irq_bypass_data, apic_id);
+			WARN_ON_ONCE(amd_iommu_activate_guest_mode(data, apic_id));
 		else
-			ret = amd_iommu_deactivate_guest_mode(irqfd->irq_bypass_data);
-		if (ret)
-			break;
+			WARN_ON_ONCE(amd_iommu_deactivate_guest_mode(data));
 	}
 out:
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
-	return ret;
 }
 
 static void svm_ir_list_del(struct kvm_kernel_irqfd *irqfd)
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#96 lore raw

PATCH v3 47/62 KVM: SVM: Process all IRTEs on affinity change even if one update fails

In reply to Sean Christopherson

When updating IRTE GA fields, keep processing all other IRTEs if an update fails, as not updating later entries risks making a bad situation worse.

---
 arch/x86/kvm/svm/avic.c | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)
arch/x86/kvm/svm/avic.c +3 -5
index 6048cd90e731..24e07f075646 100644
@@ -849,12 +849,10 @@ static inline int avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu
 	if (list_empty(&svm->ir_list))
 		return 0;
 
-	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list) {
+	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list)
 		ret = amd_iommu_update_ga(cpu, irqfd->irq_bypass_data);
-		if (ret)
-			return ret;
-	}
-	return 0;
+
+	return ret;
 }
 
 void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#97 lore raw

PATCH v3 48/62 KVM: SVM: WARN if updating IRTE GA fields in IOMMU fails

In reply to Sean Christopherson

WARN if updating GA information for an IRTE entry fails as modifying an IRTE should only fail if KVM is buggy, e.g. has stale metadata, and because returning an error that is always ignored is pointless.

---
 arch/x86/kvm/svm/avic.c | 9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)
arch/x86/kvm/svm/avic.c +3 -6
index 24e07f075646..d1f7b35c1b02 100644
@@ -834,9 +834,8 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 	return irq_set_vcpu_affinity(host_irq, NULL);
 }
 
-static inline int avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu)
+static void avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu)
 {
-	int ret = 0;
 	struct vcpu_svm *svm = to_svm(vcpu);
 	struct kvm_kernel_irqfd *irqfd;
 
@@ -847,12 +846,10 @@ static inline int avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu
 	 * interrupt remapping table entry targeting this vcpu.
 	 */
 	if (list_empty(&svm->ir_list))
-		return 0;
+		return;
 
 	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list)
-		ret = amd_iommu_update_ga(cpu, irqfd->irq_bypass_data);
-
-	return ret;
+		WARN_ON_ONCE(amd_iommu_update_ga(cpu, irqfd->irq_bypass_data));
 }
 
 void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#98 lore raw

PATCH v3 49/62 KVM: x86: Drop superfluous "has assigned device" check in kvm_pi_update_irte()

In reply to Sean Christopherson

Don't bother checking if the VM has an assigned device when updating IRTE entries. kvm_arch_irq_bypass_add_producer() explicitly increments the assigned device count, kvm_arch_irq_bypass_del_producer() explicitly decrements the count before invoking kvm_pi_update_irte(), and kvm_irq_routing_update() only updates IRTE entries if there's an active IRQ bypass producer.

---
 arch/x86/kvm/irq.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
arch/x86/kvm/irq.c +1 -3
index 4119c1e880e7..45cb9f1ee618 100644
@@ -517,9 +517,7 @@ static int kvm_pi_update_irte(struct kvm_kernel_irqfd *irqfd,
 	struct kvm_lapic_irq irq;
 	int r;
 
-	if (!irqchip_in_kernel(kvm) ||
-	    !kvm_arch_has_irq_bypass() ||
-	    !kvm_arch_has_assigned_device(kvm))
+	if (!irqchip_in_kernel(kvm) || !kvm_arch_has_irq_bypass())
 		return 0;
 
 	if (entry && entry->type == KVM_IRQ_ROUTING_MSI) {
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#99 lore raw

PATCH v3 50/62 KVM: x86: WARN if IRQ bypass isn't supported in kvm_pi_update_irte()

In reply to Sean Christopherson

WARN if kvm_pi_update_irte() is reached without IRQ bypass support, as the code is only reachable if the VM already has an IRQ bypass producer (see kvm_irq_routing_update()), or from kvm_arch_irq_bypass_{add,del}_producer(), which, stating the obvious, are called if and only if KVM enables its IRQ bypass hooks.

---
 arch/x86/kvm/irq.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kvm/irq.c +1 -1
index 45cb9f1ee618..9c03a67fedbe 100644
@@ -517,7 +517,7 @@ static int kvm_pi_update_irte(struct kvm_kernel_irqfd *irqfd,
 	struct kvm_lapic_irq irq;
 	int r;
 
-	if (!irqchip_in_kernel(kvm) || !kvm_arch_has_irq_bypass())
+	if (!irqchip_in_kernel(kvm) || WARN_ON_ONCE(!kvm_arch_has_irq_bypass()))
 		return 0;
 
 	if (entry && entry->type == KVM_IRQ_ROUTING_MSI) {
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#100 lore raw

PATCH v3 51/62 KVM: x86: WARN if IRQ bypass routing is updated without in-kernel local APIC

In reply to Sean Christopherson

Yell if kvm_pi_update_irte() is reached without an in-kernel local APIC, as kvm_arch_irqfd_allowed() should prevent attaching an irqfd and thus any and all postable IRQs to an APIC-less VM.

---
 arch/x86/kvm/irq.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
arch/x86/kvm/irq.c +2 -2
index 9c03a67fedbe..02efa7162252 100644
@@ -517,8 +517,8 @@ static int kvm_pi_update_irte(struct kvm_kernel_irqfd *irqfd,
 	struct kvm_lapic_irq irq;
 	int r;
 
-	if (!irqchip_in_kernel(kvm) || WARN_ON_ONCE(!kvm_arch_has_irq_bypass()))
-		return 0;
+	if (WARN_ON_ONCE(!irqchip_in_kernel(kvm) || !kvm_arch_has_irq_bypass()))
+		return -EINVAL;
 
 	if (entry && entry->type == KVM_IRQ_ROUTING_MSI) {
 		kvm_set_msi_irq(kvm, entry, &irq);
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#101 lore raw

PATCH v3 52/62 KVM: SVM: WARN if ir_list is non-empty at vCPU free

In reply to Sean Christopherson

Now that AVIC IRTE tracking is in a mostly sane state, WARN if a vCPU is freed with ir_list entries, i.e. if KVM leaves a dangling IRTE.

Initialize the per-vCPU interrupt remapping list and its lock even if AVIC is disabled so that the WARN doesn't hit false positives (and so that KVM doesn't need to call into AVIC code for a simple sanity check).

---
 arch/x86/kvm/svm/avic.c | 5 +++--
 arch/x86/kvm/svm/svm.c  | 2 ++
 2 files changed, 5 insertions(+), 2 deletions(-)
arch/x86/kvm/svm/avic.c +3 -2
index d1f7b35c1b02..c55cbb0610b4 100644
@@ -704,6 +704,9 @@ int avic_init_vcpu(struct vcpu_svm *svm)
 	int ret;
 	struct kvm_vcpu *vcpu = &svm->vcpu;
 
+	INIT_LIST_HEAD(&svm->ir_list);
+	spin_lock_init(&svm->ir_list_lock);
+
 	if (!enable_apicv || !irqchip_in_kernel(vcpu->kvm))
 		return 0;
 
@@ -711,8 +714,6 @@ int avic_init_vcpu(struct vcpu_svm *svm)
 	if (ret)
 		return ret;
 
-	INIT_LIST_HEAD(&svm->ir_list);
-	spin_lock_init(&svm->ir_list_lock);
 	svm->dfr_reg = APIC_DFR_FLAT;
 
 	return ret;
arch/x86/kvm/svm/svm.c +2 -0
index 56d11f7b4bef..2cd991062acb 100644
@@ -1497,6 +1497,8 @@ static void svm_vcpu_free(struct kvm_vcpu *vcpu)
 {
 	struct vcpu_svm *svm = to_svm(vcpu);
 
+	WARN_ON_ONCE(!list_empty(&svm->ir_list));
+
 	svm_leave_nested(vcpu);
 	svm_free_nested(svm);
 
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#102 lore raw

PATCH v3 53/62 KVM: x86: Decouple device assignment from IRQ bypass

In reply to Sean Christopherson

Use a dedicated counter to track the number of IRQs that can utilize IRQ bypass instead of piggybacking the assigned device count. As evidenced by commit 2edd9cb79fb3 ("kvm: detect assigned device via irqbypass manager"), it's possible for a device to be able to post IRQs to a vCPU without said device being assigned to a VM.

Leave the calls to kvm_arch_{start,end}_assignment() alone for the moment to avoid regressing the MMIO stale data mitigation. KVM is abusing the assigned device count when applying mmio_stale_data_clear, and it's not at all clear if vDPA devices rely on this behavior. This will hopefully be cleaned up in the future, as the number of assigned devices is a terrible heuristic for detecting if a VM has access to host MMIO.

---
 arch/x86/include/asm/kvm-x86-ops.h |  2 +-
 arch/x86/include/asm/kvm_host.h    |  4 +++-
 arch/x86/kvm/irq.c                 |  9 ++++++++-
 arch/x86/kvm/vmx/main.c            |  2 +-
 arch/x86/kvm/vmx/posted_intr.c     | 16 ++++++++++------
 arch/x86/kvm/vmx/posted_intr.h     |  2 +-
 arch/x86/kvm/x86.c                 |  3 +--
 7 files changed, 25 insertions(+), 13 deletions(-)
arch/x86/include/asm/kvm-x86-ops.h +1 -1
index 8d50e3e0a19b..8897f509860c 100644
@@ -112,7 +112,7 @@ KVM_X86_OP_OPTIONAL(update_cpu_dirty_logging)
 KVM_X86_OP_OPTIONAL(vcpu_blocking)
 KVM_X86_OP_OPTIONAL(vcpu_unblocking)
 KVM_X86_OP_OPTIONAL(pi_update_irte)
-KVM_X86_OP_OPTIONAL(pi_start_assignment)
+KVM_X86_OP_OPTIONAL(pi_start_bypass)
 KVM_X86_OP_OPTIONAL(apicv_pre_state_restore)
 KVM_X86_OP_OPTIONAL(apicv_post_state_restore)
 KVM_X86_OP_OPTIONAL_RET0(dy_apicv_has_pending_interrupt)
arch/x86/include/asm/kvm_host.h +3 -1
index c722adfedd96..01edcefbd937 100644
@@ -1382,6 +1382,8 @@ struct kvm_arch {
 	atomic_t noncoherent_dma_count;
 #define __KVM_HAVE_ARCH_ASSIGNED_DEVICE
 	atomic_t assigned_device_count;
+	unsigned long nr_possible_bypass_irqs;
+
 #ifdef CONFIG_KVM_IOAPIC
 	struct kvm_pic *vpic;
 	struct kvm_ioapic *vioapic;
@@ -1855,7 +1857,7 @@ struct kvm_x86_ops {
 	int (*pi_update_irte)(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 			      unsigned int host_irq, uint32_t guest_irq,
 			      struct kvm_vcpu *vcpu, u32 vector);
-	void (*pi_start_assignment)(struct kvm *kvm);
+	void (*pi_start_bypass)(struct kvm *kvm);
 	void (*apicv_pre_state_restore)(struct kvm_vcpu *vcpu);
 	void (*apicv_post_state_restore)(struct kvm_vcpu *vcpu);
 	bool (*dy_apicv_has_pending_interrupt)(struct kvm_vcpu *vcpu);
arch/x86/kvm/irq.c +8 -1
index 02efa7162252..3e4338c6a712 100644
@@ -570,10 +570,15 @@ int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
 	spin_lock_irq(&kvm->irqfds.lock);
 	irqfd->producer = prod;
 
+	if (!kvm->arch.nr_possible_bypass_irqs++)
+		kvm_x86_call(pi_start_bypass)(kvm);
+
 	if (irqfd->irq_entry.type == KVM_IRQ_ROUTING_MSI) {
 		ret = kvm_pi_update_irte(irqfd, &irqfd->irq_entry);
-		if (ret)
+		if (ret) {
+			kvm->arch.nr_possible_bypass_irqs--;
 			kvm_arch_end_assignment(irqfd->kvm);
+		}
 	}
 	spin_unlock_irq(&kvm->irqfds.lock);
 
@@ -606,6 +611,8 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
 	}
 	irqfd->producer = NULL;
 
+	kvm->arch.nr_possible_bypass_irqs--;
+
 	spin_unlock_irq(&kvm->irqfds.lock);
 
 
arch/x86/kvm/vmx/main.c +1 -1
index d1e02e567b57..a986fc45145e 100644
@@ -1014,7 +1014,7 @@ struct kvm_x86_ops vt_x86_ops __initdata = {
 	.nested_ops = &vmx_nested_ops,
 
 	.pi_update_irte = vmx_pi_update_irte,
-	.pi_start_assignment = vmx_pi_start_assignment,
+	.pi_start_bypass = vmx_pi_start_bypass,
 
 #ifdef CONFIG_X86_64
 	.set_hv_timer = vt_op(set_hv_timer),
arch/x86/kvm/vmx/posted_intr.c +10 -6
index 3a23c30f73cb..5671d59a6b6d 100644
@@ -146,8 +146,13 @@ void vmx_vcpu_pi_load(struct kvm_vcpu *vcpu, int cpu)
 
 static bool vmx_can_use_vtd_pi(struct kvm *kvm)
 {
+	/*
+	 * Note, reading the number of possible bypass IRQs can race with a
+	 * bypass IRQ being attached to the VM.  vmx_pi_start_bypass() ensures
+	 * blockng vCPUs will see an elevated count or get KVM_REQ_UNBLOCK.
+	 */
 	return irqchip_in_kernel(kvm) && kvm_arch_has_irq_bypass() &&
-	       kvm_arch_has_assigned_device(kvm);
+	       READ_ONCE(kvm->arch.nr_possible_bypass_irqs);
 }
 
 /*
@@ -285,12 +290,11 @@ bool pi_has_pending_interrupt(struct kvm_vcpu *vcpu)
 
 
 /*
- * Bail out of the block loop if the VM has an assigned
- * device, but the blocking vCPU didn't reconfigure the
- * PI.NV to the wakeup vector, i.e. the assigned device
- * came along after the initial check in vmx_vcpu_pi_put().
+ * Kick all vCPUs when the first possible bypass IRQ is attached to a VM, as
+ * blocking vCPUs may scheduled out without reconfiguring PID.NV to the wakeup
+ * vector, i.e. if the bypass IRQ came along after vmx_vcpu_pi_put().
  */
-void vmx_pi_start_assignment(struct kvm *kvm)
+void vmx_pi_start_bypass(struct kvm *kvm)
 {
 	if (!kvm_arch_has_irq_bypass())
 		return;
arch/x86/kvm/vmx/posted_intr.h +1 -1
index 94ed66ea6249..a4af39948cf0 100644
@@ -17,7 +17,7 @@ bool pi_has_pending_interrupt(struct kvm_vcpu *vcpu);
 int vmx_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		       unsigned int host_irq, uint32_t guest_irq,
 		       struct kvm_vcpu *vcpu, u32 vector);
-void vmx_pi_start_assignment(struct kvm *kvm);
+void vmx_pi_start_bypass(struct kvm *kvm);
 
 static inline int pi_find_highest_vector(struct pi_desc *pi_desc)
 {
arch/x86/kvm/x86.c +1 -2
index d3571c484790..59391cb56674 100644
@@ -13440,8 +13440,7 @@ bool kvm_arch_can_dequeue_async_page_present(struct kvm_vcpu *vcpu)
 
 void kvm_arch_start_assignment(struct kvm *kvm)
 {
-	if (atomic_inc_return(&kvm->arch.assigned_device_count) == 1)
-		kvm_x86_call(pi_start_assignment)(kvm);
+	atomic_inc(&kvm->arch.assigned_device_count);
 }
 EXPORT_SYMBOL_GPL(kvm_arch_start_assignment);
 
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#103 lore raw

PATCH v3 54/62 KVM: VMX: WARN if VT-d Posted IRQs aren't possible when starting IRQ bypass

In reply to Sean Christopherson

WARN if KVM attempts to "start" IRQ bypass when VT-d Posted IRQs are disabled, to make it obvious that the logic is a sanity check, and so that a bug related to nr_possible_bypass_irqs is more like to cause noisy failures, e.g. so that KVM doesn't silently fail to wake blocking vCPUs.

---
 arch/x86/kvm/vmx/posted_intr.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kvm/vmx/posted_intr.c +1 -1
index 5671d59a6b6d..4a6d9a17da23 100644
@@ -296,7 +296,7 @@ bool pi_has_pending_interrupt(struct kvm_vcpu *vcpu)
  */
 void vmx_pi_start_bypass(struct kvm *kvm)
 {
-	if (!kvm_arch_has_irq_bypass())
+	if (WARN_ON_ONCE(!vmx_can_use_vtd_pi(kvm)))
 		return;
 
 	kvm_make_all_cpus_request(kvm, KVM_REQ_UNBLOCK);
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#104 lore raw

PATCH v3 55/62 KVM: SVM: Use vcpu_idx, not vcpu_id, for GA log tag/metadata

In reply to Sean Christopherson

Use a vCPU's index, not its ID, for the GA log tag/metadata that's used to find and kick vCPUs when a device posted interrupt serves as a wake event. Lookups on a vCPU index are O(fast) (not sure what xa_load() actually provides), whereas a vCPU ID lookup is O(n) if a vCPU's ID doesn't match its index.

Unlike the Physical APIC Table, which is accessed by hardware when virtualizing IPIs, hardware doesn't consume the GA tag, i.e. KVM _must_ use APIC IDs to fill the Physical APIC Table, but KVM has free rein over the format/meaning of the GA tag.

---
 arch/x86/kvm/svm/avic.c | 37 ++++++++++++++++++++-----------------
 1 file changed, 20 insertions(+), 17 deletions(-)
arch/x86/kvm/svm/avic.c +20 -17
index c55cbb0610b4..bb74705d6cfd 100644
@@ -29,36 +29,39 @@
 #include "svm.h"
 
 /*
- * Encode the arbitrary VM ID and the vCPU's default APIC ID, i.e the vCPU ID,
- * into the GATag so that KVM can retrieve the correct vCPU from a GALog entry
- * if an interrupt can't be delivered, e.g. because the vCPU isn't running.
+ * Encode the arbitrary VM ID and the vCPU's _index_ into the GATag so that
+ * KVM can retrieve the correct vCPU from a GALog entry if an interrupt can't
+ * be delivered, e.g. because the vCPU isn't running.  Use the vCPU's index
+ * instead of its ID (a.k.a. its default APIC ID), as KVM is guaranteed a fast
+ * lookup on the index, where as vCPUs whose index doesn't match their ID need
+ * to walk the entire xarray of vCPUs in the worst case scenario.
  *
- * For the vCPU ID, use however many bits are currently allowed for the max
+ * For the vCPU index, use however many bits are currently allowed for the max
  * guest physical APIC ID (limited by the size of the physical ID table), and
  * use whatever bits remain to assign arbitrary AVIC IDs to VMs.  Note, the
  * size of the GATag is defined by hardware (32 bits), but is an opaque value
  * as far as hardware is concerned.
  */
-#define AVIC_VCPU_ID_MASK		AVIC_PHYSICAL_MAX_INDEX_MASK
+#define AVIC_VCPU_IDX_MASK		AVIC_PHYSICAL_MAX_INDEX_MASK
 
 #define AVIC_VM_ID_SHIFT		HWEIGHT32(AVIC_PHYSICAL_MAX_INDEX_MASK)
 #define AVIC_VM_ID_MASK			(GENMASK(31, AVIC_VM_ID_SHIFT) >> AVIC_VM_ID_SHIFT)
 
 #define AVIC_GATAG_TO_VMID(x)		((x >> AVIC_VM_ID_SHIFT) & AVIC_VM_ID_MASK)
-#define AVIC_GATAG_TO_VCPUID(x)		(x & AVIC_VCPU_ID_MASK)
+#define AVIC_GATAG_TO_VCPUIDX(x)	(x & AVIC_VCPU_IDX_MASK)
 
-#define __AVIC_GATAG(vm_id, vcpu_id)	((((vm_id) & AVIC_VM_ID_MASK) << AVIC_VM_ID_SHIFT) | \
-					 ((vcpu_id) & AVIC_VCPU_ID_MASK))
-#define AVIC_GATAG(vm_id, vcpu_id)					\
+#define __AVIC_GATAG(vm_id, vcpu_idx)	((((vm_id) & AVIC_VM_ID_MASK) << AVIC_VM_ID_SHIFT) | \
+					 ((vcpu_idx) & AVIC_VCPU_IDX_MASK))
+#define AVIC_GATAG(vm_id, vcpu_idx)					\
 ({									\
-	u32 ga_tag = __AVIC_GATAG(vm_id, vcpu_id);			\
+	u32 ga_tag = __AVIC_GATAG(vm_id, vcpu_idx);			\
 									\
-	WARN_ON_ONCE(AVIC_GATAG_TO_VCPUID(ga_tag) != (vcpu_id));	\
+	WARN_ON_ONCE(AVIC_GATAG_TO_VCPUIDX(ga_tag) != (vcpu_idx));	\
 	WARN_ON_ONCE(AVIC_GATAG_TO_VMID(ga_tag) != (vm_id));		\
 	ga_tag;								\
 })
 
-static_assert(__AVIC_GATAG(AVIC_VM_ID_MASK, AVIC_VCPU_ID_MASK) == -1u);
+static_assert(__AVIC_GATAG(AVIC_VM_ID_MASK, AVIC_VCPU_IDX_MASK) == -1u);
 
 static bool force_avic;
 module_param_unsafe(force_avic, bool, 0444);
@@ -139,16 +142,16 @@ int avic_ga_log_notifier(u32 ga_tag)
 	struct kvm_svm *kvm_svm;
 	struct kvm_vcpu *vcpu = NULL;
 	u32 vm_id = AVIC_GATAG_TO_VMID(ga_tag);
-	u32 vcpu_id = AVIC_GATAG_TO_VCPUID(ga_tag);
+	u32 vcpu_idx = AVIC_GATAG_TO_VCPUIDX(ga_tag);
 
-	pr_debug("SVM: %s: vm_id=%#x, vcpu_id=%#x\n", __func__, vm_id, vcpu_id);
-	trace_kvm_avic_ga_log(vm_id, vcpu_id);
+	pr_debug("SVM: %s: vm_id=%#x, vcpu_idx=%#x\n", __func__, vm_id, vcpu_idx);
+	trace_kvm_avic_ga_log(vm_id, vcpu_idx);
 
 	spin_lock_irqsave(&svm_vm_data_hash_lock, flags);
 	hash_for_each_possible(svm_vm_data_hash, kvm_svm, hnode, vm_id) {
 		if (kvm_svm->avic_vm_id != vm_id)
 			continue;
-		vcpu = kvm_get_vcpu_by_id(&kvm_svm->kvm, vcpu_id);
+		vcpu = kvm_get_vcpu(&kvm_svm->kvm, vcpu_idx);
 		break;
 	}
 	spin_unlock_irqrestore(&svm_vm_data_hash_lock, flags);
@@ -785,7 +788,7 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		 */
 		struct amd_iommu_pi_data pi_data = {
 			.ga_tag = AVIC_GATAG(to_kvm_svm(kvm)->avic_vm_id,
-					     vcpu->vcpu_id),
+					     vcpu->vcpu_idx),
 			.is_guest_mode = kvm_vcpu_apicv_active(vcpu),
 			.vapic_addr = avic_get_backing_page_address(to_svm(vcpu)),
 			.vector = vector,
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#105 lore raw

PATCH v3 56/62 iommu/amd: WARN if KVM calls GA IRTE helpers without virtual APIC support

In reply to Sean Christopherson

WARN if KVM attempts to update IRTE entries when virtual APIC isn't fully supported, as KVM should guard all such calls on IRQ posting being enabled.

---
 drivers/iommu/amd/iommu.c | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)
drivers/iommu/amd/iommu.c +12 -5
index becef69a306d..926dcdfe08c8 100644
@@ -3836,8 +3836,10 @@ int amd_iommu_update_ga(int cpu, void *data)
 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
 
-	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
-	    !entry || !entry->lo.fields_vapic.guest_mode)
+	if (WARN_ON_ONCE(!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)))
+		return -EINVAL;
+
+	if (!entry || !entry->lo.fields_vapic.guest_mode)
 		return 0;
 
 	if (!ir_data->iommu)
@@ -3856,7 +3858,10 @@ int amd_iommu_activate_guest_mode(void *data, int cpu)
 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
 	u64 valid;
 
-	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) || !entry)
+	if (WARN_ON_ONCE(!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)))
+		return -EINVAL;
+
+	if (!entry)
 		return 0;
 
 	valid = entry->lo.fields_vapic.valid;
@@ -3885,8 +3890,10 @@ int amd_iommu_deactivate_guest_mode(void *data)
 	struct irq_cfg *cfg = ir_data->cfg;
 	u64 valid;
 
-	if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) ||
-	    !entry || !entry->lo.fields_vapic.guest_mode)
+	if (WARN_ON_ONCE(!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)))
+		return -EINVAL;
+
+	if (!entry || !entry->lo.fields_vapic.guest_mode)
 		return 0;
 
 	valid = entry->lo.fields_remap.valid;
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#106 lore raw

PATCH v3 57/62 KVM: SVM: Fold avic_set_pi_irte_mode() into its sole caller

In reply to Sean Christopherson

Fold avic_set_pi_irte_mode() into avic_refresh_apicv_exec_ctrl() in anticipation of moving the __avic_vcpu_{load,put}() calls into the critical section, and because having a one-off helper with a name that's easily confused with avic_pi_update_irte() is unnecessary.

No functional change intended.

---
 arch/x86/kvm/svm/avic.c | 52 ++++++++++++++++++-----------------------
 1 file changed, 23 insertions(+), 29 deletions(-)
arch/x86/kvm/svm/avic.c +23 -29
index bb74705d6cfd..9ddec6f3ad41 100644
@@ -728,34 +728,6 @@ void avic_apicv_post_state_restore(struct kvm_vcpu *vcpu)
 	avic_handle_ldr_update(vcpu);
 }
 
-static void avic_set_pi_irte_mode(struct kvm_vcpu *vcpu, bool activate)
-{
-	int apic_id = kvm_cpu_get_apicid(vcpu->cpu);
-	unsigned long flags;
-	struct vcpu_svm *svm = to_svm(vcpu);
-	struct kvm_kernel_irqfd *irqfd;
-
-	/*
-	 * Here, we go through the per-vcpu ir_list to update all existing
-	 * interrupt remapping table entry targeting this vcpu.
-	 */
-	spin_lock_irqsave(&svm->ir_list_lock, flags);
-
-	if (list_empty(&svm->ir_list))
-		goto out;
-
-	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list) {
-		void *data = irqfd->irq_bypass_data;
-
-		if (activate)
-			WARN_ON_ONCE(amd_iommu_activate_guest_mode(data, apic_id));
-		else
-			WARN_ON_ONCE(amd_iommu_deactivate_guest_mode(data));
-	}
-out:
-	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
-}
-
 static void svm_ir_list_del(struct kvm_kernel_irqfd *irqfd)
 {
 	struct kvm_vcpu *vcpu = irqfd->irq_bypass_vcpu;
@@ -990,6 +962,10 @@ void avic_refresh_virtual_apic_mode(struct kvm_vcpu *vcpu)
 void avic_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu)
 {
 	bool activated = kvm_vcpu_apicv_active(vcpu);
+	int apic_id = kvm_cpu_get_apicid(vcpu->cpu);
+	struct vcpu_svm *svm = to_svm(vcpu);
+	struct kvm_kernel_irqfd *irqfd;
+	unsigned long flags;
 
 	if (!enable_apicv)
 		return;
@@ -1001,7 +977,25 @@ void avic_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu)
 	else
 		avic_vcpu_put(vcpu);
 
-	avic_set_pi_irte_mode(vcpu, activated);
+	/*
+	 * Here, we go through the per-vcpu ir_list to update all existing
+	 * interrupt remapping table entry targeting this vcpu.
+	 */
+	spin_lock_irqsave(&svm->ir_list_lock, flags);
+
+	if (list_empty(&svm->ir_list))
+		goto out;
+
+	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list) {
+		void *data = irqfd->irq_bypass_data;
+
+		if (activated)
+			WARN_ON_ONCE(amd_iommu_activate_guest_mode(data, apic_id));
+		else
+			WARN_ON_ONCE(amd_iommu_deactivate_guest_mode(data));
+	}
+out:
+	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
 }
 
 void avic_vcpu_blocking(struct kvm_vcpu *vcpu)
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#107 lore raw

PATCH v3 58/62 KVM: SVM: Don't check vCPU's blocking status when toggling AVIC on/off

In reply to Sean Christopherson

Don't query a vCPU's blocking status when toggling AVIC on/off; barring KVM bugs, the vCPU can't be blocking when refreshing AVIC controls. And if there are KVM bugs, ensuring the vCPU and its associated IRTEs are in the correct state is desirable, i.e. well worth any overhead in a buggy scenario.

Isolating the "real" load/put flows will allow moving the IOMMU IRTE (de)activation logic from avic_refresh_apicv_exec_ctrl() to avic_update_iommu_vcpu_affinity(), i.e. will allow updating the vCPU's physical ID entry and its IRTEs in a common path, under a single critical section of ir_list_lock.

---
 arch/x86/kvm/svm/avic.c | 65 +++++++++++++++++++++++------------------
 1 file changed, 37 insertions(+), 28 deletions(-)
arch/x86/kvm/svm/avic.c +37 -28
index 9ddec6f3ad41..1e6e5d1f6b4e 100644
@@ -828,7 +828,7 @@ static void avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu)
 		WARN_ON_ONCE(amd_iommu_update_ga(cpu, irqfd->irq_bypass_data));
 }
 
-void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
+static void __avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 {
 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	int h_physical_id = kvm_cpu_get_apicid(cpu);
@@ -844,16 +844,6 @@ void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	if (WARN_ON_ONCE(vcpu->vcpu_id * sizeof(entry) >= PAGE_SIZE))
 		return;
 
-	/*
-	 * No need to update anything if the vCPU is blocking, i.e. if the vCPU
-	 * is being scheduled in after being preempted.  The CPU entries in the
-	 * Physical APIC table and IRTE are consumed iff IsRun{ning} is '1'.
-	 * If the vCPU was migrated, its new CPU value will be stuffed when the
-	 * vCPU unblocks.
-	 */
-	if (kvm_vcpu_is_blocking(vcpu))
-		return;
-
 	/*
 	 * Grab the per-vCPU interrupt remapping lock even if the VM doesn't
 	 * _currently_ have assigned devices, as that can change.  Holding
@@ -888,31 +878,33 @@ void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
 }
 
-void avic_vcpu_put(struct kvm_vcpu *vcpu)
+void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
+{
+	/*
+	 * No need to update anything if the vCPU is blocking, i.e. if the vCPU
+	 * is being scheduled in after being preempted.  The CPU entries in the
+	 * Physical APIC table and IRTE are consumed iff IsRun{ning} is '1'.
+	 * If the vCPU was migrated, its new CPU value will be stuffed when the
+	 * vCPU unblocks.
+	 */
+	if (kvm_vcpu_is_blocking(vcpu))
+		return;
+
+	__avic_vcpu_load(vcpu, cpu);
+}
+
+static void __avic_vcpu_put(struct kvm_vcpu *vcpu)
 {
 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	struct vcpu_svm *svm = to_svm(vcpu);
 	unsigned long flags;
-	u64 entry;
+	u64 entry = svm->avic_physical_id_entry;
 
 	lockdep_assert_preemption_disabled();
 
 	if (WARN_ON_ONCE(vcpu->vcpu_id * sizeof(entry) >= PAGE_SIZE))
 		return;
 
-	/*
-	 * Note, reading the Physical ID entry outside of ir_list_lock is safe
-	 * as only the pCPU that has loaded (or is loading) the vCPU is allowed
-	 * to modify the entry, and preemption is disabled.  I.e. the vCPU
-	 * can't be scheduled out and thus avic_vcpu_{put,load}() can't run
-	 * recursively.
-	 */
-	entry = svm->avic_physical_id_entry;
-
-	/* Nothing to do if IsRunning == '0' due to vCPU blocking. */
-	if (!(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK))
-		return;
-
 	/*
 	 * Take and hold the per-vCPU interrupt remapping lock while updating
 	 * the Physical ID entry even though the lock doesn't protect against
@@ -932,7 +924,24 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 		WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
 
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
+}
 
+void avic_vcpu_put(struct kvm_vcpu *vcpu)
+{
+	/*
+	 * Note, reading the Physical ID entry outside of ir_list_lock is safe
+	 * as only the pCPU that has loaded (or is loading) the vCPU is allowed
+	 * to modify the entry, and preemption is disabled.  I.e. the vCPU
+	 * can't be scheduled out and thus avic_vcpu_{put,load}() can't run
+	 * recursively.
+	 */
+	u64 entry = to_svm(vcpu)->avic_physical_id_entry;
+
+	/* Nothing to do if IsRunning == '0' due to vCPU blocking. */
+	if (!(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK))
+		return;
+
+	__avic_vcpu_put(vcpu);
 }
 
 void avic_refresh_virtual_apic_mode(struct kvm_vcpu *vcpu)
@@ -973,9 +982,9 @@ void avic_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu)
 	avic_refresh_virtual_apic_mode(vcpu);
 
 	if (activated)
-		avic_vcpu_load(vcpu, vcpu->cpu);
+		__avic_vcpu_load(vcpu, vcpu->cpu);
 	else
-		avic_vcpu_put(vcpu);
+		__avic_vcpu_put(vcpu);
 
 	/*
 	 * Here, we go through the per-vcpu ir_list to update all existing
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#108 lore raw

PATCH v3 59/62 KVM: SVM: Consolidate IRTE update when toggling AVIC on/off

In reply to Sean Christopherson

Fold the IRTE modification logic in avic_refresh_apicv_exec_ctrl() into __avic_vcpu_{load,put}(), and add a param to the helpers to communicate whether or not AVIC is being toggled, i.e. if IRTE needs a "full" update, or just a quick update to set the CPU and IsRun.

---
 arch/x86/kvm/svm/avic.c | 85 ++++++++++++++++++++++-------------------
 1 file changed, 46 insertions(+), 39 deletions(-)
arch/x86/kvm/svm/avic.c +46 -39
index 1e6e5d1f6b4e..2e47559a4134 100644
@@ -810,7 +810,28 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 	return irq_set_vcpu_affinity(host_irq, NULL);
 }
 
-static void avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu)
+enum avic_vcpu_action {
+	/*
+	 * There is no need to differentiate between activate and deactivate,
+	 * as KVM only refreshes AVIC state when the vCPU is scheduled in and
+	 * isn't blocking, i.e. the pCPU must always be (in)valid when AVIC is
+	 * being (de)activated.
+	 */
+	AVIC_TOGGLE_ON_OFF	= BIT(0),
+	AVIC_ACTIVATE		= AVIC_TOGGLE_ON_OFF,
+	AVIC_DEACTIVATE		= AVIC_TOGGLE_ON_OFF,
+
+	/*
+	 * No unique action is required to deal with a vCPU that stops/starts
+	 * running, as IRTEs are configured to generate GALog interrupts at all
+	 * times.
+	 */
+	AVIC_START_RUNNING	= 0,
+	AVIC_STOP_RUNNING	= 0,
+};
+
+static void avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu,
+					    enum avic_vcpu_action action)
 {
 	struct vcpu_svm *svm = to_svm(vcpu);
 	struct kvm_kernel_irqfd *irqfd;
@@ -824,11 +845,20 @@ static void avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu)
 	if (list_empty(&svm->ir_list))
 		return;
 
-	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list)
-		WARN_ON_ONCE(amd_iommu_update_ga(cpu, irqfd->irq_bypass_data));
+	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list) {
+		void *data = irqfd->irq_bypass_data;
+
+		if (!(action & AVIC_TOGGLE_ON_OFF))
+			WARN_ON_ONCE(amd_iommu_update_ga(cpu, data));
+		else if (cpu >= 0)
+			WARN_ON_ONCE(amd_iommu_activate_guest_mode(data, cpu));
+		else
+			WARN_ON_ONCE(amd_iommu_deactivate_guest_mode(data));
+	}
 }
 
-static void __avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
+static void __avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu,
+			     enum avic_vcpu_action action)
 {
 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	int h_physical_id = kvm_cpu_get_apicid(cpu);
@@ -873,7 +903,7 @@ static void __avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 
 	WRITE_ONCE(kvm_svm->avic_physical_id_table[vcpu->vcpu_id], entry);
 
-	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id);
+	avic_update_iommu_vcpu_affinity(vcpu, h_physical_id, action);
 
 	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
 }
@@ -890,10 +920,10 @@ void avic_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	if (kvm_vcpu_is_blocking(vcpu))
 		return;
 
-	__avic_vcpu_load(vcpu, cpu);
+	__avic_vcpu_load(vcpu, cpu, AVIC_START_RUNNING);
 }
 
-static void __avic_vcpu_put(struct kvm_vcpu *vcpu)
+static void __avic_vcpu_put(struct kvm_vcpu *vcpu, enum avic_vcpu_action action)
 {
 	struct kvm_svm *kvm_svm = to_kvm_svm(vcpu->kvm);
 	struct vcpu_svm *svm = to_svm(vcpu);
@@ -915,7 +945,7 @@ static void __avic_vcpu_put(struct kvm_vcpu *vcpu)
 	 */
 	spin_lock_irqsave(&svm->ir_list_lock, flags);
 
-	avic_update_iommu_vcpu_affinity(vcpu, -1);
+	avic_update_iommu_vcpu_affinity(vcpu, -1, action);
 
 	entry &= ~AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK;
 	svm->avic_physical_id_entry = entry;
@@ -941,7 +971,7 @@ void avic_vcpu_put(struct kvm_vcpu *vcpu)
 	if (!(entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK))
 		return;
 
-	__avic_vcpu_put(vcpu);
+	__avic_vcpu_put(vcpu, AVIC_STOP_RUNNING);
 }
 
 void avic_refresh_virtual_apic_mode(struct kvm_vcpu *vcpu)
@@ -970,41 +1000,18 @@ void avic_refresh_virtual_apic_mode(struct kvm_vcpu *vcpu)
 
 void avic_refresh_apicv_exec_ctrl(struct kvm_vcpu *vcpu)
 {
-	bool activated = kvm_vcpu_apicv_active(vcpu);
-	int apic_id = kvm_cpu_get_apicid(vcpu->cpu);
-	struct vcpu_svm *svm = to_svm(vcpu);
-	struct kvm_kernel_irqfd *irqfd;
-	unsigned long flags;
-
 	if (!enable_apicv)
 		return;
 
+	/* APICv should only be toggled on/off while the vCPU is running. */
+	WARN_ON_ONCE(kvm_vcpu_is_blocking(vcpu));
+
 	avic_refresh_virtual_apic_mode(vcpu);
 
-	if (activated)
-		__avic_vcpu_load(vcpu, vcpu->cpu);
+	if (kvm_vcpu_apicv_active(vcpu))
+		__avic_vcpu_load(vcpu, vcpu->cpu, AVIC_ACTIVATE);
 	else
-		__avic_vcpu_put(vcpu);
-
-	/*
-	 * Here, we go through the per-vcpu ir_list to update all existing
-	 * interrupt remapping table entry targeting this vcpu.
-	 */
-	spin_lock_irqsave(&svm->ir_list_lock, flags);
-
-	if (list_empty(&svm->ir_list))
-		goto out;
-
-	list_for_each_entry(irqfd, &svm->ir_list, vcpu_list) {
-		void *data = irqfd->irq_bypass_data;
-
-		if (activated)
-			WARN_ON_ONCE(amd_iommu_activate_guest_mode(data, apic_id));
-		else
-			WARN_ON_ONCE(amd_iommu_deactivate_guest_mode(data));
-	}
-out:
-	spin_unlock_irqrestore(&svm->ir_list_lock, flags);
+		__avic_vcpu_put(vcpu, AVIC_DEACTIVATE);
 }
 
 void avic_vcpu_blocking(struct kvm_vcpu *vcpu)
@@ -1030,7 +1037,7 @@ void avic_vcpu_blocking(struct kvm_vcpu *vcpu)
 	 * CPU and cause noisy neighbor problems if the VM is sending interrupts
 	 * to the vCPU while it's scheduled out.
 	 */
-	avic_vcpu_put(vcpu);
+	__avic_vcpu_put(vcpu, AVIC_STOP_RUNNING);
 }
 
 void avic_vcpu_unblocking(struct kvm_vcpu *vcpu)
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#109 lore raw

PATCH v3 60/62 iommu/amd: KVM: SVM: Allow KVM to control need for GA log interrupts

In reply to Sean Christopherson

Add plumbing to the AMD IOMMU driver to allow KVM to control whether or not an IRTE is configured to generate GA log interrupts. KVM only needs a notification if the target vCPU is blocking, so the vCPU can be awakened. If a vCPU is preempted or exits to userspace, KVM clears is_run, but will set the vCPU back to running when userspace does KVM_RUN and/or the vCPU task is scheduled back in, i.e. KVM doesn't need a notification.

Unconditionally pass "true" in all KVM paths to isolate the IOMMU changes from the KVM changes insofar as possible.

Opportunistically swap the ordering of parameters for amd_iommu_update_ga() so that the match amd_iommu_activate_guest_mode().

Note, as of this writing, the AMD IOMMU manual doesn't list GALogIntr as a non-cached field, but per AMD hardware architects, it's not cached and can be safely updated without an invalidation.

---
 arch/x86/include/asm/irq_remapping.h |  1 +
 arch/x86/kvm/svm/avic.c              | 10 ++++++----
 drivers/iommu/amd/iommu.c            | 28 +++++++++++++++++-----------
 include/linux/amd-iommu.h            |  9 ++++-----
 4 files changed, 28 insertions(+), 20 deletions(-)
arch/x86/include/asm/irq_remapping.h +1 -0
index 4c75a17632f6..5a0d42464d44 100644
@@ -36,6 +36,7 @@ struct amd_iommu_pi_data {
 	u32 ga_tag;
 	u32 vector;		/* Guest vector of the interrupt */
 	int cpu;
+	bool ga_log_intr;
 	bool is_guest_mode;
 	void *ir_data;
 };
arch/x86/kvm/svm/avic.c +6 -4
index 2e47559a4134..e61ecc3514ea 100644
@@ -784,10 +784,12 @@ int avic_pi_update_irte(struct kvm_kernel_irqfd *irqfd, struct kvm *kvm,
 		 * is awakened and/or scheduled in.  See also avic_vcpu_load().
 		 */
 		entry = svm->avic_physical_id_entry;
-		if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK)
+		if (entry & AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK) {
 			pi_data.cpu = entry & AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK;
-		else
+		} else {
 			pi_data.cpu = -1;
+			pi_data.ga_log_intr = true;
+		}
 
 		ret = irq_set_vcpu_affinity(host_irq, &pi_data);
 		if (ret)
@@ -849,9 +851,9 @@ static void avic_update_iommu_vcpu_affinity(struct kvm_vcpu *vcpu, int cpu,
 		void *data = irqfd->irq_bypass_data;
 
 		if (!(action & AVIC_TOGGLE_ON_OFF))
-			WARN_ON_ONCE(amd_iommu_update_ga(cpu, data));
+			WARN_ON_ONCE(amd_iommu_update_ga(data, cpu, true));
 		else if (cpu >= 0)
-			WARN_ON_ONCE(amd_iommu_activate_guest_mode(data, cpu));
+			WARN_ON_ONCE(amd_iommu_activate_guest_mode(data, cpu, true));
 		else
 			WARN_ON_ONCE(amd_iommu_deactivate_guest_mode(data));
 	}
drivers/iommu/amd/iommu.c +17 -11
index 926dcdfe08c8..e79f583da36b 100644
@@ -3804,7 +3804,8 @@ static const struct irq_domain_ops amd_ir_domain_ops = {
 	.deactivate = irq_remapping_deactivate,
 };
 
-static void __amd_iommu_update_ga(struct irte_ga *entry, int cpu)
+static void __amd_iommu_update_ga(struct irte_ga *entry, int cpu,
+				  bool ga_log_intr)
 {
 	if (cpu >= 0) {
 		entry->lo.fields_vapic.destination =
@@ -3812,8 +3813,10 @@ static void __amd_iommu_update_ga(struct irte_ga *entry, int cpu)
 		entry->hi.fields.destination =
 					APICID_TO_IRTE_DEST_HI(cpu);
 		entry->lo.fields_vapic.is_run = true;
+		entry->lo.fields_vapic.ga_log_intr = false;
 	} else {
 		entry->lo.fields_vapic.is_run = false;
+		entry->lo.fields_vapic.ga_log_intr = ga_log_intr;
 	}
 }
 
@@ -3822,16 +3825,19 @@ static void __amd_iommu_update_ga(struct irte_ga *entry, int cpu)
  * a vCPU, without issuing an IOMMU invalidation for the IRTE.
  *
  * If the vCPU is associated with a pCPU (@cpu >= 0), configure the Destination
- * with the pCPU's APIC ID and set IsRun, else clear IsRun.  I.e. treat vCPUs
- * that are associated with a pCPU as running.  This API is intended to be used
- * when a vCPU is scheduled in/out (or stops running for any reason), to do a
- * fast update of IsRun and (conditionally) Destination.
+ * with the pCPU's APIC ID, set IsRun, and clear GALogIntr.  If the vCPU isn't
+ * associated with a pCPU (@cpu < 0), clear IsRun and set/clear GALogIntr based
+ * on input from the caller (e.g. KVM only requests GALogIntr when the vCPU is
+ * blocking and requires a notification wake event).  I.e. treat vCPUs that are
+ * associated with a pCPU as running.  This API is intended to be used when a
+ * vCPU is scheduled in/out (or stops running for any reason), to do a fast
+ * update of IsRun, GALogIntr, and (conditionally) Destination.
  *
  * Per the IOMMU spec, the Destination, IsRun, and GATag fields are not cached
  * and thus don't require an invalidation to ensure the IOMMU consumes fresh
  * information.
  */
-int amd_iommu_update_ga(int cpu, void *data)
+int amd_iommu_update_ga(void *data, int cpu, bool ga_log_intr)
 {
 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
@@ -3845,14 +3851,14 @@ int amd_iommu_update_ga(int cpu, void *data)
 	if (!ir_data->iommu)
 		return -ENODEV;
 
-	__amd_iommu_update_ga(entry, cpu);
+	__amd_iommu_update_ga(entry, cpu, ga_log_intr);
 
 	return __modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
 				ir_data->irq_2_irte.index, entry);
 }
 EXPORT_SYMBOL(amd_iommu_update_ga);
 
-int amd_iommu_activate_guest_mode(void *data, int cpu)
+int amd_iommu_activate_guest_mode(void *data, int cpu, bool ga_log_intr)
 {
 	struct amd_ir_data *ir_data = (struct amd_ir_data *)data;
 	struct irte_ga *entry = (struct irte_ga *) ir_data->entry;
@@ -3871,12 +3877,11 @@ int amd_iommu_activate_guest_mode(void *data, int cpu)
 
 	entry->lo.fields_vapic.valid       = valid;
 	entry->lo.fields_vapic.guest_mode  = 1;
-	entry->lo.fields_vapic.ga_log_intr = 1;
 	entry->hi.fields.ga_root_ptr       = ir_data->ga_root_ptr;
 	entry->hi.fields.vector            = ir_data->ga_vector;
 	entry->lo.fields_vapic.ga_tag      = ir_data->ga_tag;
 
-	__amd_iommu_update_ga(entry, cpu);
+	__amd_iommu_update_ga(entry, cpu, ga_log_intr);
 
 	return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid,
 			      ir_data->irq_2_irte.index, entry);
@@ -3947,7 +3952,8 @@ static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *info)
 		ir_data->ga_vector = pi_data->vector;
 		ir_data->ga_tag = pi_data->ga_tag;
 		if (pi_data->is_guest_mode)
-			ret = amd_iommu_activate_guest_mode(ir_data, pi_data->cpu);
+			ret = amd_iommu_activate_guest_mode(ir_data, pi_data->cpu,
+							    pi_data->ga_log_intr);
 		else
 			ret = amd_iommu_deactivate_guest_mode(ir_data);
 	} else {
include/linux/amd-iommu.h +4 -5
index c9f2df0c4596..8cced632ecd0 100644
@@ -30,9 +30,8 @@ static inline void amd_iommu_detect(void) { }
 /* IOMMU AVIC Function */
 extern int amd_iommu_register_ga_log_notifier(int (*notifier)(u32));
 
-extern int amd_iommu_update_ga(int cpu, void *data);
-
-extern int amd_iommu_activate_guest_mode(void *data, int cpu);
+extern int amd_iommu_update_ga(void *data, int cpu, bool ga_log_intr);
+extern int amd_iommu_activate_guest_mode(void *data, int cpu, bool ga_log_intr);
 extern int amd_iommu_deactivate_guest_mode(void *data);
 
 #else /* defined(CONFIG_AMD_IOMMU) && defined(CONFIG_IRQ_REMAP) */
@@ -43,12 +42,12 @@ amd_iommu_register_ga_log_notifier(int (*notifier)(u32))
 	return 0;
 }
 
-static inline int amd_iommu_update_ga(int cpu, void *data)
+static inline int amd_iommu_update_ga(void *data, int cpu, bool ga_log_intr)
 {
 	return 0;
 }
 
-static inline int amd_iommu_activate_guest_mode(void *data, int cpu)
+static inline int amd_iommu_activate_guest_mode(void *data, int cpu, bool ga_log_intr)
 {
 	return 0;
 }
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#110 lore raw

PATCH v3 61/62 KVM: SVM: Generate GA log IRQs only if the associated vCPUs is blocking

In reply to Sean Christopherson

Configure IRTEs to GA log interrupts for device posted IRQs that hit non-running vCPUs if and only if the target vCPU is blocking, i.e. actually needs a wake event. If the vCPU has exited to userspace or was preempted, generating GA log entries and interrupts is wasteful and unnecessary, as the vCPU will be re-loaded and/or scheduled back in irrespective of the GA log notification (avic_ga_log_notifier() is just a fancy wrapper for kvm_vcpu_wake_up()).

Use a should-be-zero bit in the vCPU's Physical APIC ID Table Entry to track whether or not the vCPU's associated IRTEs are configured to generate GA logs, but only set the synthetic bit in KVM's "cache", i.e. never set the should-be-zero bit in tables that are used by hardware. Use a synthetic bit instead of a dedicated boolean to minimize the odds of messing up the locking, i.e. so that all the existing rules that apply to avic_physical_id_entry for IS_RUNNING are reused verbatim for GA_LOG_INTR.

Note, because KVM (by design) "puts" AVIC state in a "pre-blocking" phase, using kvm_vcpu_is_blocking() to track the need for notifications isn't a viable option.

---
 arch/x86/include/asm/svm.h |  7 +++++
 arch/x86/kvm/svm/avic.c    | 63 ++++++++++++++++++++++++++++++--------
 2 files changed, 58 insertions(+), 12 deletions(-)
arch/x86/include/asm/svm.h +7 -0
index 36f67c69ea66..ffc27f676243 100644
@@ -252,6 +252,13 @@ struct __attribute__ ((__packed__)) vmcb_control_area {
 #define AVIC_LOGICAL_ID_ENTRY_VALID_BIT			31
 #define AVIC_LOGICAL_ID_ENTRY_VALID_MASK		(1 << 31)
 
+/*
+ * GA_LOG_INTR is a synthetic flag that's never propagated to hardware-visible
+ * tables.  GA_LOG_INTR is set if the vCPU needs device posted IRQs to generate
+ * GA log interrupts to wake the vCPU (because it's blocking or about to block).
+ */
+#define AVIC_PHYSICAL_ID_ENTRY_GA_LOG_INTR		BIT_ULL(61)
+
 #define AVIC_PHYSICAL_ID_ENTRY_HOST_PHYSICAL_ID_MASK	GENMASK_ULL(11, 0)
 #define AVIC_PHYSICAL_ID_ENTRY_BACKING_PAGE_MASK	GENMASK_ULL(51, 12)
 #define AVIC_PHYSICAL_ID_ENTRY_IS_RUNNING_MASK		(1ULL << 62)
arch/x86/kvm/svm/avic.c +51 -12

120 lines not on this page · raw message

Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer
#111 lore raw

PATCH v3 62/62 KVM: x86: Rename kvm_set_msi_irq() => kvm_msi_to_lapic_irq()

In reply to Sean Christopherson

Rename kvm_set_msi_irq() to kvm_msi_to_lapic_irq() to better capture what it actually does, e.g. it's _really_ easy to conflate kvm_set_msi_irq() with kvm_set_msi().

Opportunistically delete the public declaration and export, as they are no longer used/needed.

No functional change intended.

---
 arch/x86/include/asm/kvm_host.h |  3 ---
 arch/x86/kvm/irq.c              | 14 +++++++-------
 2 files changed, 7 insertions(+), 10 deletions(-)
arch/x86/include/asm/kvm_host.h +0 -3
index 01edcefbd937..af08b1acba04 100644
@@ -2381,9 +2381,6 @@ bool kvm_vcpu_is_bsp(struct kvm_vcpu *vcpu);
 bool kvm_intr_is_single_vcpu(struct kvm *kvm, struct kvm_lapic_irq *irq,
 			     struct kvm_vcpu **dest_vcpu);
 
-void kvm_set_msi_irq(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
-		     struct kvm_lapic_irq *irq);
-
 static inline bool kvm_irq_is_postable(struct kvm_lapic_irq *irq)
 {
 	/* We can only post Fixed and LowPrio IRQs */
arch/x86/kvm/irq.c +7 -7
index 3e4338c6a712..83c91845389c 100644
@@ -252,8 +252,9 @@ int kvm_irq_delivery_to_apic(struct kvm *kvm, struct kvm_lapic *src,
 	return r;
 }
 
-void kvm_set_msi_irq(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
-		     struct kvm_lapic_irq *irq)
+static void kvm_msi_to_lapic_irq(struct kvm *kvm,
+				 struct kvm_kernel_irq_routing_entry *e,
+				 struct kvm_lapic_irq *irq)
 {
 	struct msi_msg msg = { .address_lo = e->msi.address_lo,
 			       .address_hi = e->msi.address_hi,
@@ -271,7 +272,6 @@ void kvm_set_msi_irq(struct kvm *kvm, struct kvm_kernel_irq_routing_entry *e,
 	irq->level = 1;
 	irq->shorthand = APIC_DEST_NOSHORT;
 }
-EXPORT_SYMBOL_GPL(kvm_set_msi_irq);
 
 static inline bool kvm_msi_route_invalid(struct kvm *kvm,
 		struct kvm_kernel_irq_routing_entry *e)
@@ -290,7 +290,7 @@ int kvm_set_msi(struct kvm_kernel_irq_routing_entry *e,
 	if (!level)
 		return -1;
 
-	kvm_set_msi_irq(kvm, e, &irq);
+	kvm_msi_to_lapic_irq(kvm, e, &irq);
 
 	return kvm_irq_delivery_to_apic(kvm, NULL, &irq, NULL);
 }
@@ -313,7 +313,7 @@ int kvm_arch_set_irq_inatomic(struct kvm_kernel_irq_routing_entry *e,
 		if (kvm_msi_route_invalid(kvm, e))
 			return -EINVAL;
 
-		kvm_set_msi_irq(kvm, e, &irq);
+		kvm_msi_to_lapic_irq(kvm, e, &irq);
 
 		if (kvm_irq_delivery_to_apic_fast(kvm, NULL, &irq, &r, NULL))
 			return r;
@@ -486,7 +486,7 @@ void kvm_scan_ioapic_routes(struct kvm_vcpu *vcpu,
 			if (entry->type != KVM_IRQ_ROUTING_MSI)
 				continue;
 
-			kvm_set_msi_irq(vcpu->kvm, entry, &irq);
+			kvm_msi_to_lapic_irq(vcpu->kvm, entry, &irq);
 
 			if (!irq.trig_mode)
 				continue;
@@ -521,7 +521,7 @@ static int kvm_pi_update_irte(struct kvm_kernel_irqfd *irqfd,
 		return -EINVAL;
 
 	if (entry && entry->type == KVM_IRQ_ROUTING_MSI) {
-		kvm_set_msi_irq(kvm, entry, &irq);
+		kvm_msi_to_lapic_irq(kvm, entry, &irq);
 
 		/*
 		 * Force remapped mode if hardware doesn't support posting the
Signature · 1 line
2.50.0.rc1.591.g9c95f17f64-goog
author maintainer Applied
#112 lore raw
In reply to Sean Christopherson
Quoted 8 lines · Sean Christopherson

TL;DR: Overhaul device posted interrupts in KVM and IOMMU, and AVIC in
general.

This applies on the series to add CONFIG_KVM_IOAPIC (and to kill irq_comm.c):

https://lore.kernel.org/all/[email protected]

[...]

Applied to kvm-x86 irqs (except for patch 1, which is already in Linus' tree), with minor massaging (comments and changelogs) to address feedback.

Note, in the branch, the arm64 vLPI is split quite far from the rest of the series. This is by design so that I can provide a stable branch/tag for arm64 if a conflict or dependency crops up, without the arm64 patch inheriting all of the dependencies and prep work for the series at large.

[01/62] KVM: arm64: Explicitly treat routing entry type changes as changes
(already in Linus' tree)
[02/62] KVM: arm64: WARN if unmapping a vLPI fails in any path

        https://github.com/kvm-x86/linux/commit/cd4178d19420
[03/62] KVM: Pass new routing entries and irqfd when updating IRTEs
        https://github.com/kvm-x86/linux/commit/cb210737675e
[04/62] KVM: SVM: Track per-vCPU IRTEs using kvm_kernel_irqfd structure
        https://github.com/kvm-x86/linux/commit/05c5e23657e1
[05/62] KVM: SVM: Delete IRTE link from previous vCPU before setting new IRTE
        https://github.com/kvm-x86/linux/commit/0a917e9d4b70
[06/62] iommu/amd: KVM: SVM: Delete now-unused cached/previous GA tag fields
        https://github.com/kvm-x86/linux/commit/1da19c5ce053
[07/62] KVM: SVM: Delete IRTE link from previous vCPU irrespective of new routing
        https://github.com/kvm-x86/linux/commit/a0ca34bb1aad
[08/62] KVM: SVM: Drop pointless masking of default APIC base when setting V_APIC_BAR
        https://github.com/kvm-x86/linux/commit/430579577892
[09/62] KVM: SVM: Drop pointless masking of kernel page pa's with AVIC HPA masks
        https://github.com/kvm-x86/linux/commit/2e002ddc8966
[10/62] KVM: SVM: Add helper to deduplicate code for getting AVIC backing page
        https://github.com/kvm-x86/linux/commit/3338c639da15
[11/62] KVM: SVM: Drop vcpu_svm's pointless avic_backing_page field
        https://github.com/kvm-x86/linux/commit/d8527f133c0a
[12/62] KVM: SVM: Inhibit AVIC if ID is too big instead of rejecting vCPU creation
        https://github.com/kvm-x86/linux/commit/1aa6e256e46f
[13/62] KVM: SVM: Drop redundant check in AVIC code on ID during vCPU creation
        https://github.com/kvm-x86/linux/commit/c24ed209c474
[14/62] KVM: SVM: Track AVIC tables as natively sized pointers, not "struct pages"
        https://github.com/kvm-x86/linux/commit/26baab4eea4c
[15/62] KVM: SVM: Drop superfluous "cache" of AVIC Physical ID entry pointer
        https://github.com/kvm-x86/linux/commit/d29433336a7b
[16/62] KVM: VMX: Move enable_ipiv knob to common x86
        https://github.com/kvm-x86/linux/commit/bafddc70001d
[17/62] KVM: SVM: Add enable_ipiv param, never set IsRunning if disabled
        https://github.com/kvm-x86/linux/commit/d921665e01ba
[18/62] KVM: SVM: Disable (x2)AVIC IPI virtualization if CPU has erratum #1235
        https://github.com/kvm-x86/linux/commit/8de4a1c8164e
[19/62] KVM: VMX: Suppress PI notifications whenever the vCPU is put
        https://github.com/kvm-x86/linux/commit/6737557442e5
[20/62] KVM: SVM: Add a comment to explain why avic_vcpu_blocking() ignores IRQ blocking
        https://github.com/kvm-x86/linux/commit/52d826c9e54c
[21/62] iommu/amd: KVM: SVM: Use pi_desc_addr to derive ga_root_ptr
        https://github.com/kvm-x86/linux/commit/c4cdbaf9d81c
[22/62] iommu/amd: KVM: SVM: Pass NULL @vcpu_info to indicate "not guest mode"
        https://github.com/kvm-x86/linux/commit/95d50ebe6df8
[23/62] KVM: SVM: Stop walking list of routing table entries when updating IRTE
        https://github.com/kvm-x86/linux/commit/1e663ed23992
[24/62] KVM: VMX: Stop walking list of routing table entries when updating IRTE
        https://github.com/kvm-x86/linux/commit/23ca102e6fb2
[25/62] KVM: SVM: Extract SVM specific code out of get_pi_vcpu_info()
        https://github.com/kvm-x86/linux/commit/0a64c447f6f8
[26/62] KVM: x86: Move IRQ routing/delivery APIs from x86.c => irq.c
        https://github.com/kvm-x86/linux/commit/f5369619f7f8
[27/62] KVM: x86: Nullify irqfd->producer after updating IRTEs
        https://github.com/kvm-x86/linux/commit/9517aedecd0e
[28/62] KVM: x86: Dedup AVIC vs. PI code for identifying target vCPU
        https://github.com/kvm-x86/linux/commit/cf04ec393ed0
[29/62] KVM: x86: Move posted interrupt tracepoint to common code
        https://github.com/kvm-x86/linux/commit/c5af31698d71
[30/62] KVM: SVM: Clean up return handling in avic_pi_update_irte()
        https://github.com/kvm-x86/linux/commit/803928483669
[31/62] iommu: KVM: Split "struct vcpu_data" into separate AMD vs. Intel structs
        https://github.com/kvm-x86/linux/commit/53527ea1b702
[32/62] KVM: Don't WARN if updating IRQ bypass route fails
        https://github.com/kvm-x86/linux/commit/b33252b9d172
[33/62] KVM: Fold kvm_arch_irqfd_route_changed() into kvm_arch_update_irqfd_routing()
        https://github.com/kvm-x86/linux/commit/77bb184ab880
[34/62] KVM: x86: Track irq_bypass_vcpu in common x86 code
        https://github.com/kvm-x86/linux/commit/511754bc548b
[35/62] KVM: x86: Skip IOMMU IRTE updates if there's no old or new vCPU being targeted
        https://github.com/kvm-x86/linux/commit/dc6adb13046a
[36/62] KVM: x86: Don't update IRTE entries when old and new routes were !MSI
        https://github.com/kvm-x86/linux/commit/cc8b13105eac
[37/62] KVM: SVM: Revert IRTE to legacy mode if IOMMU doesn't provide IR metadata
        https://github.com/kvm-x86/linux/commit/71d6b3b8e69d
[38/62] KVM: SVM: Take and hold ir_list_lock across IRTE updates in IOMMU
        https://github.com/kvm-x86/linux/commit/c3d591c91f9c
[39/62] iommu/amd: Document which IRTE fields amd_iommu_update_ga() can modify
        https://github.com/kvm-x86/linux/commit/3be405e89f3d
[40/62] iommu/amd: KVM: SVM: Infer IsRun from validity of pCPU destination
        https://github.com/kvm-x86/linux/commit/08d9ccdd1a5c
[41/62] iommu/amd: Factor out helper for manipulating IRTE GA/CPU info
        https://github.com/kvm-x86/linux/commit/0b2b541fa3cd
[42/62] iommu/amd: KVM: SVM: Set pCPU info in IRTE when setting vCPU affinity
        https://github.com/kvm-x86/linux/commit/f965255dc503
[43/62] iommu/amd: KVM: SVM: Add IRTE metadata to affined vCPU's list if AVIC is inhibited
        https://github.com/kvm-x86/linux/commit/6df262f915ab
[44/62] KVM: SVM: Don't check for assigned device(s) when updating affinity
        https://github.com/kvm-x86/linux/commit/f5998661ff73
[45/62] KVM: SVM: Don't check for assigned device(s) when activating AVIC
        https://github.com/kvm-x86/linux/commit/fe0213923dd9
[46/62] KVM: SVM: WARN if (de)activating guest mode in IOMMU fails
        https://github.com/kvm-x86/linux/commit/16562766f171
[47/62] KVM: SVM: Process all IRTEs on affinity change even if one update fails
        https://github.com/kvm-x86/linux/commit/48f79c6c86b3
[48/62] KVM: SVM: WARN if updating IRTE GA fields in IOMMU fails
        https://github.com/kvm-x86/linux/commit/cd86240fea26
[49/62] KVM: x86: Drop superfluous "has assigned device" check in kvm_pi_update_irte()
        https://github.com/kvm-x86/linux/commit/04c4ca0ae479
[50/62] KVM: x86: WARN if IRQ bypass isn't supported in kvm_pi_update_irte()
        https://github.com/kvm-x86/linux/commit/d1bccaa1793d
[51/62] KVM: x86: WARN if IRQ bypass routing is updated without in-kernel local APIC
        https://github.com/kvm-x86/linux/commit/25ef059e8bc5
[52/62] KVM: SVM: WARN if ir_list is non-empty at vCPU free
        https://github.com/kvm-x86/linux/commit/99836eb9c5dc
[53/62] KVM: x86: Decouple device assignment from IRQ bypass
        https://github.com/kvm-x86/linux/commit/77e1b8332d1d
[54/62] KVM: VMX: WARN if VT-d Posted IRQs aren't possible when starting IRQ bypass
        https://github.com/kvm-x86/linux/commit/ce9d54f41be0
[55/62] KVM: SVM: Use vcpu_idx, not vcpu_id, for GA log tag/metadata
        https://github.com/kvm-x86/linux/commit/11a60455d4c9
[56/62] iommu/amd: WARN if KVM calls GA IRTE helpers without virtual APIC support
        https://github.com/kvm-x86/linux/commit/a23480fe21de
[57/62] KVM: SVM: Fold avic_set_pi_irte_mode() into its sole caller
        https://github.com/kvm-x86/linux/commit/f2bc961d383b
[58/62] KVM: SVM: Don't check vCPU's blocking status when toggling AVIC on/off
        https://github.com/kvm-x86/linux/commit/6eab2340f339
[59/62] KVM: SVM: Consolidate IRTE update when toggling AVIC on/off
        https://github.com/kvm-x86/linux/commit/5f3d06b1648e
[60/62] iommu/amd: KVM: SVM: Allow KVM to control need for GA log interrupts
        https://github.com/kvm-x86/linux/commit/b9e53f9ff4a8
[61/62] KVM: SVM: Generate GA log IRQs only if the associated vCPUs is blocking
        https://github.com/kvm-x86/linux/commit/b03500f03ea0
[62/62] KVM: x86: Rename kvm_set_msi_irq() => kvm_msi_to_lapic_irq()
        https://github.com/kvm-x86/linux/commit/6f343724837b

--
https://github.com/kvm-x86/kvm-unit-tests/tree/next