KVM: Dirty ring fixes and cleanups
Sean Christopherson
reviewerv2 cover KVM: Dirty ring fixes and cleanups
Fix issues with dirty ring harvesting where KVM doesn't bound the processing of entries in any way, which allows userspace to keep KVM in a tight loop indefinitely.
E.g.
struct kvm_dirty_gfn *dirty_gfns = vcpu_map_dirty_ring(vcpu);
if (fork()) {
int r;
for (;;) {
r = kvm_vm_reset_dirty_ring(vcpu->vm);
if (r)
printf("RESET %d dirty ring entries\n", r);
}
} else {
int i;
for (i = 0; i < test_dirty_ring_count; i++) {
dirty_gfns[i].slot = TEST_MEM_SLOT_INDEX;
dirty_gfns[i].offset = (i * 64) % host_num_pages;
}
for (;;) {
for (i = 0; i < test_dirty_ring_count; i++)
WRITE_ONCE(dirty_gfns[i].flags, KVM_DIRTY_GFN_F_RESET);
}
}Patches 1-3 address that class of bugs. Patches 4 and 5 are cleanups.
v2: Expand on comments in dirty ring harvesting code. [Yan]
v1: https://lore.kernel.org/all/[email protected]
Sean Christopherson (5):
KVM: Bound the number of dirty ring entries in a single reset at
INT_MAX
KVM: Bail from the dirty ring reset flow if a signal is pending
KVM: Conditionally reschedule when resetting the dirty ring
KVM: Check for empty mask of harvested dirty ring entries in caller
KVM: Use mask of harvested dirty ring entries to coalesce dirty ring
resets
include/linux/kvm_dirty_ring.h | 8 ++- virt/kvm/dirty_ring.c | 103 ++++++++++++++++++++++----------- virt/kvm/kvm_main.c | 9 ++- 3 files changed, 81 insertions(+), 39 deletions(-)
base-commit: d1b88431c59e94799aff0e31ce1467af2a86b0cf
Signature · 1 line
2.49.0.1015.ga840276032-goog
Sean Christopherson
reviewerPATCH v2 1/5 KVM: Bound the number of dirty ring entries in a single reset at INT_MAX
Cap the number of ring entries that are reset in a single ioctl to INT_MAX to ensure userspace isn't confused by a wrap into negative space, and so that, in a truly pathological scenario, KVM doesn't miss a TLB flush due to the count wrapping to zero. While the size of the ring is fixed at 0x10000 entries and KVM (currently) supports at most 4096, userspace is allowed to harvest entries from the ring while the reset is in-progress, i.e. it's possible for the ring to always have harvested entries.
Opportunistically return an actual error code from the helper so that a future fix to handle pending signals can gracefully return -EINTR.
--- include/linux/kvm_dirty_ring.h | 8 +++++--- virt/kvm/dirty_ring.c | 10 +++++----- virt/kvm/kvm_main.c | 9 ++++++--- 3 files changed, 16 insertions(+), 11 deletions(-)
include/linux/kvm_dirty_ring.h +5 -3
@@ -49,9 +49,10 @@ static inline int kvm_dirty_ring_alloc(struct kvm *kvm, struct kvm_dirty_ring *r }static inline int kvm_dirty_ring_reset (struct kvm *kvm, -struct kvm_dirty_ring *ring) +struct kvm_dirty_ring *ring, +int *nr_entries_reset) { -return 0 ; +return -ENOENT ; } static inline void kvm_dirty_ring_push(struct kvm_vcpu *vcpu,
@@ -82,7 +83,8 @@ int kvm_dirty_ring_alloc(struct kvm *kvm, struct kvm_dirty_ring *ring,* called with kvm->slots_lock held, returns the number of * processed pages. */ -int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring); +int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring, +int *nr_entries_reset);/* * returns =0: successfully pushed
virt/kvm/dirty_ring.c +5 -5
@@ -105,19 +105,19 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)return smp_load_acquire (&gfn->flags) &KVM_DIRTY_GFN_F_RESET ; } -int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring) +int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring, +int *nr_entries_reset) {u32 cur_slot, next_slot;u64 cur_offset, next_offset;unsigned long mask; -int count =0 ;struct kvm_dirty_gfn *entry;bool first_round = true;/* This is only needed to make compilers happy */ cur_slot = cur_offset = mask =0 ; -while (true) { +while (likely ((*nr_entries_reset) <INT_MAX )) { entry = &ring->dirty_gfns[ring->reset_index & (ring->size -1 )];if (!kvm_dirty_gfn_harvested (entry))
@@ -130,7 +130,7 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring)kvm_dirty_gfn_set_invalid (entry); ring->reset_index++; - count++; + (*nr_entries_reset)++;/* * Try to coalesce the reset operations when the guest is * scanning pages in the same slot.
@@ -167,7 +167,7 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring)trace_kvm_dirty_ring_reset (ring); -return count; +return 0 ; }void kvm_dirty_ring_push (struct kvm_vcpu *vcpu,u32 slot,u64 offset)
virt/kvm/kvm_main.c +6 -3
@@ -4903,15 +4903,18 @@ static int kvm_vm_ioctl_reset_dirty_pages(struct kvm *kvm) {unsigned long i;struct kvm_vcpu *vcpu; -int cleared =0 ; +int cleared =0 , r;if (!kvm->dirty_ring_size)return -EINVAL ;mutex_lock (&kvm->slots_lock); -kvm_for_each_vcpu (i, vcpu, kvm) - cleared +=kvm_dirty_ring_reset (vcpu->kvm, &vcpu->dirty_ring); +kvm_for_each_vcpu (i, vcpu, kvm) { + r =kvm_dirty_ring_reset (vcpu->kvm, &vcpu->dirty_ring, &cleared); +if (r) +break ; + }mutex_unlock (&kvm->slots_lock);
Signature · 1 line
2.49.0.1015.ga840276032-goog
Re: [PATCH v2 1/5] KVM: Bound the number of dirty ring entries in a single reset at INT_MAX
Quoted 40 lines · Sean Christopherson
Cap the number of ring entries that are reset in a single ioctl to
INT_MAXto ensure userspace isn't confused by a wrap into negative space, and so that, in a truly pathological scenario, KVM doesn't miss a TLB flush due to the count wrapping to zero. While the size of the ring is fixed at0x10000entries and KVM (currently) supports at most 4096, userspace is allowed to harvest entries from the ring while the reset is in-progress, i.e. it's possible for the ring to always have harvested entries.Opportunistically return an actual error code from the helper so that a future fix to handle pending signals can gracefully return
-EINTR.Cc Peter Xu <[email protected]>Cc Yan Zhao <[email protected]>Cc Maxim Levitsky <[email protected]>Fixes fb04a1eddb1a ("KVM: X86: Implement ring-based dirty memory tracking")Signed-off-by Sean Christopherson <[email protected]>--- include/linux/kvm_dirty_ring.h | 8 +++++--- virt/kvm/dirty_ring.c | 10 +++++----- virt/kvm/kvm_main.c | 9 ++++++--- 3 files changed, 16 insertions(+), 11 deletions(-)include/linux/kvm_dirty_ring.h +3
-2@@ -49,9 +49,10 @@ static inline int kvm_dirty_ring_alloc(struct kvm *kvm, struct kvm_dirty_ring *r }static inline int kvm_dirty_ring_reset (struct kvm *kvm, -struct kvm_dirty_ring *ring) +struct kvm_dirty_ring *ring, +int *nr_entries_reset) { -return 0 ; +return -ENOENT ; } static inline void kvm_dirty_ring_push(struct kvm_vcpu *vcpu,
include/linux/kvm_dirty_ring.h int kvm_dirty_ring_alloc(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 85
Quoted 3 lines
include/linux/kvm_dirty_ring.h +0
-0@@ -82,7 +83,8 @@ int kvm_dirty_ring_alloc(struct kvm *kvm, struct kvm_dirty_ring *ring,* called with kvm->slots_lock held, returns the number of * processed pages. */
The comment should be updated as well, since the return value is not the number of processed pages now.
Quoted 4 lines
include/linux/kvm_dirty_ring.h +2
-1-int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring); +int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring, +int *nr_entries_reset);
[...]
Sean Christopherson
reviewerPATCH v2 2/5 KVM: Bail from the dirty ring reset flow if a signal is pending
In reply to Sean ChristophersonAbort a dirty ring reset if the current task has a pending signal, as the hard limit of INT_MAX entries doesn't ensure KVM will respond to a signal in a timely fashion.
--- virt/kvm/dirty_ring.c | 3 +++ 1 file changed, 3 insertions(+)
virt/kvm/dirty_ring.c +3 -0
@@ -118,6 +118,9 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, cur_slot = cur_offset = mask =0 ;while (likely ((*nr_entries_reset) <INT_MAX )) { +if (signal_pending (current)) +return -EINTR ; + entry = &ring->dirty_gfns[ring->reset_index & (ring->size -1 )];if (!kvm_dirty_gfn_harvested (entry))
Signature · 1 line
2.49.0.1015.ga840276032-goog
Sean Christopherson
reviewerPATCH v2 3/5 KVM: Conditionally reschedule when resetting the dirty ring
In reply to Sean ChristophersonWhen resetting a dirty ring, conditionally reschedule on each iteration after the first. The recently introduced hard limit mitigates the issue of an endless reset, but isn't sufficient to completely prevent RCU stalls, soft lockups, etc., nor is the hard limit intended to guard against such badness.
Note! Take care to check for reschedule even in the "continue" paths, as a pathological scenario (or malicious userspace) could dirty the same gfn over and over, i.e. always hit the continue path.
rcu: INFO: rcu_sched self-detected stall on CPU rcu: 4-....: (5249 ticks this GP) idle=51e4/1/0x4000000000000000 softirq=309/309 fqs=2563 rcu: (t=5250 jiffies g=-319 q=608 ncpus=24) CPU: 4 UID: 1000 PID: 1067 Comm: dirty_log_test Tainted: G L 6.13.0-rc3-17fa7a24ea1e-HEAD-vm #814 Tainted: [L]=SOFTLOCKUP Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 RIP: 0010:kvm_arch_mmu_enable_log_dirty_pt_masked+0x26/0x200 [kvm] Call Trace: <TASK> kvm_reset_dirty_gfn.part.0+0xb4/0xe0 [kvm] kvm_dirty_ring_reset+0x58/0x220 [kvm] kvm_vm_ioctl+0x10eb/0x15d0 [kvm] __x64_sys_ioctl+0x8b/0xb0 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x4b/0x53 </TASK> Tainted: [L]=SOFTLOCKUP Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 RIP: 0010:kvm_arch_mmu_enable_log_dirty_pt_masked+0x17/0x200 [kvm] Call Trace: <TASK> kvm_reset_dirty_gfn.part.0+0xb4/0xe0 [kvm] kvm_dirty_ring_reset+0x58/0x220 [kvm] kvm_vm_ioctl+0x10eb/0x15d0 [kvm] __x64_sys_ioctl+0x8b/0xb0 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x4b/0x53 </TASK>
--- virt/kvm/dirty_ring.c | 10 ++++++++++ 1 file changed, 10 insertions(+)
virt/kvm/dirty_ring.c +10 -0
@@ -134,6 +134,16 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++; (*nr_entries_reset)++; + +/* +* While the size of each ring is fixed, it's possible for the +* ring to be constantly re-dirtied/harvested while the reset +* is in-progress (the hard limit exists only to guard against +* wrapping the count into negative space). +*/ +if (!first_round) +cond_resched (); +/* * Try to coalesce the reset operations when the guest is * scanning pages in the same slot.
Signature · 1 line
2.49.0.1015.ga840276032-goog
Re: [PATCH v2 3/5] KVM: Conditionally reschedule when resetting the dirty ring
Quoted 53 lines · Sean Christopherson
When resetting a dirty ring, conditionally reschedule on each iteration after the first. The recently introduced hard limit mitigates the issue of an endless reset, but isn't sufficient to completely prevent RCU stalls, soft lockups, etc., nor is the hard limit intended to guard against such badness.
Note! Take care to check for reschedule even in the "continue" paths, as a pathological scenario (or malicious userspace) could dirty the same gfn over and over, i.e. always hit the continue path.
rcu: INFO: rcu_sched self-detected stall on CPU rcu: 4-....: (5249 ticks this GP) idle=51e4/1/0x4000000000000000 softirq=309/309 fqs=2563 rcu: (t=5250 jiffies g=-319 q=608 ncpus=24) CPU: 4 UID: 1000 PID: 1067 Comm: dirty_log_test Tainted: G L 6.13.0-rc3-17fa7a24ea1e-HEAD-vm #814 Tainted: [L]=SOFTLOCKUP Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 RIP: 0010:kvm_arch_mmu_enable_log_dirty_pt_masked+0x26/0x200 [kvm] Call Trace: <TASK> kvm_reset_dirty_gfn.part.0+0xb4/0xe0 [kvm] kvm_dirty_ring_reset+0x58/0x220 [kvm] kvm_vm_ioctl+0x10eb/0x15d0 [kvm] __x64_sys_ioctl+0x8b/0xb0 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x4b/0x53 </TASK> Tainted: [L]=SOFTLOCKUP Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 RIP: 0010:kvm_arch_mmu_enable_log_dirty_pt_masked+0x17/0x200 [kvm] Call Trace: <TASK> kvm_reset_dirty_gfn.part.0+0xb4/0xe0 [kvm] kvm_dirty_ring_reset+0x58/0x220 [kvm] kvm_vm_ioctl+0x10eb/0x15d0 [kvm] __x64_sys_ioctl+0x8b/0xb0 do_syscall_64+0x5b/0x160 entry_SYSCALL_64_after_hwframe+0x4b/0x53 </TASK>Fixes fb04a1eddb1a ("KVM: X86: Implement ring-based dirty memory tracking")Signed-off-by Sean Christopherson <[email protected]>--- virt/kvm/dirty_ring.c | 10 ++++++++++ 1 file changed, 10 insertions(+)virt/kvm/dirty_ring.c +0
-0@@ -134,6 +134,16 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++;
virt/kvm/dirty_ring.c int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 145
Quoted 10 lines
virt/kvm/dirty_ring.c +9
-0(*nr_entries_reset)++; + +/* +* While the size of each ring is fixed, it's possible for the +* ring to be constantly re-dirtied/harvested while the reset +* is in-progress (the hard limit exists only to guard against +* wrapping the count into negative space). +*/ +if (!first_round) +cond_resched ();
Should we be dropping slots_lock here?
It seems like we need to be holding slots_lock to call kvm_reset_dirty_gfn(), but that's it. Userspace can already change the memslots after enabling the dirty ring, so entry->slot can already be stale, so dropping slots_lock for the cond_resched() seems harmless (and better than not dropping it).
Sean Christopherson
reviewerRe: [PATCH v2 3/5] KVM: Conditionally reschedule when resetting the dirty ring
Quoted 11 lines · James Houghton
Quoted 10 lines · Sean Christopherson
--- virt/kvm/dirty_ring.c | 10 ++++++++++ 1 file changed, 10 insertions(+)virt/kvm/dirty_ring.c +0
-0@@ -134,6 +134,16 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,
Quoted 13 lines
virt/kvm/dirty_ring.c int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 145
Quoted 11 lines
virt/kvm/dirty_ring.c +9
-0ring->reset_index++; (*nr_entries_reset)++; + +/* +* While the size of each ring is fixed, it's possible for the +* ring to be constantly re-dirtied/harvested while the reset +* is in-progress (the hard limit exists only to guard against +* wrapping the count into negative space). +*/ +if (!first_round) +cond_resched ();Should we be dropping
slots_lockhere?
Could we? Yes. Should we? Eh. I don't see any value in doing so, because in practice, it's extremely unlikely anything will be waiting on slots_lock.
kvm_vm_ioctl_reset_dirty_pages() operates on all vCPUs, i.e. there won't be competing calls to reset other rings. A well-behaved userspace won't be modifying memslots or dirty logs, and won't be toggling nx_huge_pages.
That leaves kvm_vm_ioctl_set_mem_attributes(), kvm_inhibit_apic_access_page(), kvm_assign_ioeventfd(), snp_launch_update(), and coalesced IO/MMIO (un)registration. Except for snp_launch_update(), those are all brutally slow paths, e.g. require SRCU synchronization and/or zapping of SPTEs. And snp_launch_update() is probably fairly slow too.
And dropping slots_lock only makes any sense for non-preemptible kernels, because preemptible kernels include an equivalent check in KVM_MMU_UNLOCK().
It's also possible that dropping slots_lock in this case could be a net negative. I don't think it's likely, but I don't think it's any more or less likely that droppings slots_lock is a net positive. Without performance data to guide us, it'd be little more than a guess, and I really, really don't want to set a precedence of dropping a mutex on cond_resched() without a very strong reason for doing so.
Quoted 4 lines
It seems like we need to be holding
slots_lockto callkvm_reset_dirty_gfn(), but that's it. Userspace can already change the memslots after enabling the dirty ring, soentry->slotcan already be stale, so droppingslots_lockfor thecond_resched()seems harmless (and better than not dropping it).
Re: [PATCH v2 3/5] KVM: Conditionally reschedule when resetting the dirty ring
Quoted 34 lines · Sean Christopherson
Quoted 11 lines · James Houghton
Quoted 10 lines · Sean Christopherson
--- virt/kvm/dirty_ring.c | 10 ++++++++++ 1 file changed, 10 insertions(+)virt/kvm/dirty_ring.c +0
-0@@ -134,6 +134,16 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,Quoted 13 lines
virt/kvm/dirty_ring.c int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 145
Quoted 11 lines
virt/kvm/dirty_ring.c +9
-0ring->reset_index++; (*nr_entries_reset)++; + +/* +* While the size of each ring is fixed, it's possible for the +* ring to be constantly re-dirtied/harvested while the reset +* is in-progress (the hard limit exists only to guard against +* wrapping the count into negative space). +*/ +if (!first_round) +cond_resched ();Should we be dropping
slots_lockhere?Could we? Yes. Should we? Eh. I don't see any value in doing so, because in practice, it's extremely unlikely anything will be waiting on
slots_lock.
kvm_vm_ioctl_reset_dirty_pages()operates on all vCPUs, i.e. there won't be competing calls to reset other rings. A well-behaved userspace won't be modifying memslots or dirty logs, and won't be togglingnx_huge_pages.
Quoted 5 lines
That leaves
kvm_vm_ioctl_set_mem_attributes(),kvm_inhibit_apic_access_page(),kvm_assign_ioeventfd(),snp_launch_update(), and coalesced IO/MMIO (un)registration. Except forsnp_launch_update(), those are all brutally slow paths, e.g. require SRCU synchronization and/or zapping of SPTEs. Andsnp_launch_update()is probably fairly slow too.
Okay, that makes sense.
Quoted 2 lines
And dropping
slots_lockonly makes any sense for non-preemptible kernels, because preemptible kernels include an equivalent check inKVM_MMU_UNLOCK().
I'm not really sure what "equivalent check" you mean, sorry. For preemptible kernels, we could reschedule at any time, so dropping the slots_lock on a cond_resched() wouldn't do much, yeah. Hopefully that's partially what you mean.
Quoted 6 lines
It's also possible that dropping
slots_lockin this case could be a net negative. I don't think it's likely, but I don't think it's any more or less likely that droppingsslots_lockis a net positive. Without performance data to guide us, it'd be little more than a guess, and I really, really don't want to set a precedence of dropping a mutex oncond_resched()without a very strong reason for doing so.
Fair enough.
Also, while we're at it, could you add a lockdep_assert_held(&kvm->slots_lock) to this function? :) Not necessarily in this patch.
Sean Christopherson
reviewerRe: [PATCH v2 3/5] KVM: Conditionally reschedule when resetting the dirty ring
Quoted 34 lines · James Houghton
Quoted 33 lines · Sean Christopherson
Quoted 11 lines · James Houghton
Quoted 10 lines · Sean Christopherson
--- virt/kvm/dirty_ring.c | 10 ++++++++++ 1 file changed, 10 insertions(+)virt/kvm/dirty_ring.c +0
-0@@ -134,6 +134,16 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,Quoted 13 lines
virt/kvm/dirty_ring.c int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 145
Quoted 11 lines
virt/kvm/dirty_ring.c +9
-0ring->reset_index++; (*nr_entries_reset)++; + +/* +* While the size of each ring is fixed, it's possible for the +* ring to be constantly re-dirtied/harvested while the reset +* is in-progress (the hard limit exists only to guard against +* wrapping the count into negative space). +*/ +if (!first_round) +cond_resched ();Should we be dropping
slots_lockhere?Could we? Yes. Should we? Eh. I don't see any value in doing so, because in practice, it's extremely unlikely anything will be waiting on
slots_lock.
kvm_vm_ioctl_reset_dirty_pages()operates on all vCPUs, i.e. there won't be competing calls to reset other rings. A well-behaved userspace won't be modifying memslots or dirty logs, and won't be togglingnx_huge_pages.
Quoted 7 lines
Quoted 5 lines
That leaves
kvm_vm_ioctl_set_mem_attributes(),kvm_inhibit_apic_access_page(),kvm_assign_ioeventfd(),snp_launch_update(), and coalesced IO/MMIO (un)registration. Except forsnp_launch_update(), those are all brutally slow paths, e.g. require SRCU synchronization and/or zapping of SPTEs. Andsnp_launch_update()is probably fairly slow too.Okay, that makes sense.
As discussed offlist, dropping slots_lock would also be functionally problematic, as concurrent calls to KVM_RESET_DIRTY_RINGS could get interwoven, which could result in one of the calls returning to userspace without actually completing the reset, i.e. if a different task has reaped entries but not yet called kvm_reset_dirty_gfn().
Quoted 7 lines
Quoted 2 lines
And dropping
slots_lockonly makes any sense for non-preemptible kernels, because preemptible kernels include an equivalent check inKVM_MMU_UNLOCK().I'm not really sure what "equivalent check" you mean, sorry. For preemptible kernels, we could reschedule at any time, so dropping the
slots_lockon acond_resched()wouldn't do much, yeah. Hopefully that's partially what you mean.
Ya, that's essentially what I mean. What I was referencing with KVM_MMU_UNLOCK() is the explicit check for NEED_RESCHED that happens when the preempt count hits '0' on preemptible kernels.
Quoted 7 lines
Quoted 6 lines
It's also possible that dropping
slots_lockin this case could be a net negative. I don't think it's likely, but I don't think it's any more or less likely that droppingsslots_lockis a net positive. Without performance data to guide us, it'd be little more than a guess, and I really, really don't want to set a precedence of dropping a mutex oncond_resched()without a very strong reason for doing so.
Quoted 5 lines
Fair enough.
Also, while we're at it, could you add a
lockdep_assert_held(&kvm->slots_lock)to this function? :) Not necessarily in this patch.
Heh, yep, my mind jumped to that as well. I'll tack on a patch to add a lockdep assertion, along with a comment explaining what all it protects.
Sean Christopherson
reviewerPATCH v2 4/5 KVM: Check for empty mask of harvested dirty ring entries in caller
In reply to Sean ChristophersonWhen resetting a dirty ring, explicitly check that there is work to be done before calling kvm_reset_dirty_gfn(), e.g. if no harvested entries are found and/or on the loop's first iteration, and delete the extremely misleading comment "This is only needed to make compilers happy". KVM absolutely relies on mask to be zero-initialized, i.e. the comment is an outright lie. Furthermore, the compiler is right to complain that KVM is calling a function with uninitialized data, as there are no guarantees the implementation details of kvm_reset_dirty_gfn() will be visible to kvm_dirty_ring_reset().
While the flaw could be fixed by simply deleting (or rewording) the comment, and duplicating the check is unfortunate, checking mask in the caller will allow for additional cleanups.
Opportunisticaly drop the zero-initialization of cur_slot and cur_offset. If a bug were introduced where either the slot or offset was consumed before mask is set to a non-zero value, then it is highly desirable for the compiler (or some other sanitizer) to yell.
--- virt/kvm/dirty_ring.c | 44 ++++++++++++++++++++++++++++++++++--------- 1 file changed, 35 insertions(+), 9 deletions(-)
virt/kvm/dirty_ring.c +35 -9
@@ -55,9 +55,6 @@ static void kvm_reset_dirty_gfn(struct kvm *kvm, u32 slot, u64 offset, u64 mask)struct kvm_memory_slot *memslot;int as_id, id; -if (!mask) -return ; - as_id = slot >>16 ; id = (u16 )slot;
@@ -108,15 +105,24 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring,int *nr_entries_reset) { +/* +* To minimize mmu_lock contention, batch resets for harvested entries +* whose gfns are in the same slot, and are within N frame numbers of +* each other, where N is the number of bits in an unsigned long. For +* simplicity, process the current set of entries when the next entry +* can't be included in the batch. +* +* Track the current batch slot, the gfn offset into the slot for the +* batch, and the bitmask of gfns that need to be reset (relative to +* offset). Note, the offset may be adjusted backwards, e.g. so that +* a sequence of gfns X, X-1, ... X-N can be batched. +*/ u32 cur_slot, next_slot;u64 cur_offset, next_offset; -unsigned long mask; +unsigned long mask =0 ;struct kvm_dirty_gfn *entry;bool first_round = true; -/* This is only needed to make compilers happy */ - cur_slot = cur_offset = mask =0 ; -while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))return -EINTR ;
@@ -164,14 +170,34 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,continue ; } } -kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + +/* +* Reset the slot for all the harvested entries that have been +* gathered, but not yet fully processed. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + +/* +* The current slot was reset or this is the first harvested +* entry, (re)initialize the metadata. +*/ cur_slot = next_slot; cur_offset = next_offset; mask =1 ; first_round = false; } -kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); +/* +* Perform a final reset if there are harvested entries that haven't +* been processed, which is guaranteed if at least one harvested was +* found. The loop only performs a reset when the "next" entry can't +* be batched with "current" the entry(s), and that reset processes the +* _current_ entry(s), i.e. the last harvested entry, a.k.a. next, will +* always be left pending. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask);/* * The request KVM_REQ_DIRTY_RING_SOFT_FULL will be cleared
Signature · 1 line
2.49.0.1015.ga840276032-goog
Re: [PATCH v2 4/5] KVM: Check for empty mask of harvested dirty ring entries in caller
Quoted 10 lines · Sean Christopherson
When resetting a dirty ring, explicitly check that there is work to be done before calling
kvm_reset_dirty_gfn(), e.g. if no harvested entries are found and/or on the loop's first iteration, and delete the extremely misleading comment "This is only needed to make compilers happy". KVM absolutely relies on mask to be zero-initialized, i.e. the comment is an outright lie. Furthermore, the compiler is right to complain that KVM is calling a function with uninitialized data, as there are no guarantees the implementation details ofkvm_reset_dirty_gfn()will be visible tokvm_dirty_ring_reset().
Quoted 5 lines
While the flaw could be fixed by simply deleting (or rewording) the comment, and duplicating the check is unfortunate, checking mask in the caller will allow for additional cleanups.
Opportunisticaly drop the zero-initialization of
cur_slotandcur_offset.
Opportunisticaly -> Opportunistically
Quoted 12 lines
If a bug were introduced where either the slot or offset was consumed before mask is set to a non-zero value, then it is highly desirable for the compiler (or some other sanitizer) to yell.
Cc Peter Xu <[email protected]>Cc Yan Zhao <[email protected]>Cc Maxim Levitsky <[email protected]>Signed-off-by Sean Christopherson <[email protected]>--- virt/kvm/dirty_ring.c | 44 ++++++++++++++++++++++++++++++++++--------- 1 file changed, 35 insertions(+), 9 deletions(-)
[...]
virt/kvm/dirty_ring.c int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, · line 195
Quoted 7 lines
virt/kvm/dirty_ring.c +5
-1-kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); +/* +* Perform a final reset if there are harvested entries that haven't +* been processed, which is guaranteed if at least one harvested was +* found. The loop only performs a reset when the "next" entry can't +* be batched with "current" the entry(s), and that reset processes the
"current" the entry(s) -> the "current" entry(s) ?
Quoted 8 lines
virt/kvm/dirty_ring.c +5
-0+* _current_ entry(s), i.e. the last harvested entry, a.k.a. next, will +* always be left pending. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask);/* * The request KVM_REQ_DIRTY_RING_SOFT_FULL will be cleared
Re: [PATCH v2 4/5] KVM: Check for empty mask of harvested dirty ring entries in caller
In reply to Sean ChristophersonQuoted 19 lines · Sean Christopherson
When resetting a dirty ring, explicitly check that there is work to be done before calling
kvm_reset_dirty_gfn(), e.g. if no harvested entries are found and/or on the loop's first iteration, and delete the extremely misleading comment "This is only needed to make compilers happy". KVM absolutely relies on mask to be zero-initialized, i.e. the comment is an outright lie. Furthermore, the compiler is right to complain that KVM is calling a function with uninitialized data, as there are no guarantees the implementation details ofkvm_reset_dirty_gfn()will be visible tokvm_dirty_ring_reset().While the flaw could be fixed by simply deleting (or rewording) the comment, and duplicating the check is unfortunate, checking mask in the caller will allow for additional cleanups.
Opportunisticaly drop the zero-initialization of
cur_slotandcur_offset. If a bug were introduced where either the slot or offset was consumed before mask is set to a non-zero value, then it is highly desirable for the compiler (or some other sanitizer) to yell.
Quoted 4 lines
Cc Peter Xu <[email protected]>Cc Yan Zhao <[email protected]>Cc Maxim Levitsky <[email protected]>Signed-off-by Sean Christopherson <[email protected]>
LGTM
Quoted 84 lines
--- virt/kvm/dirty_ring.c | 44 ++++++++++++++++++++++++++++++++++--------- 1 file changed, 35 insertions(+), 9 deletions(-)virt/kvm/dirty_ring.c +35
-9@@ -55,9 +55,6 @@ static void kvm_reset_dirty_gfn(struct kvm *kvm, u32 slot, u64 offset, u64 mask)struct kvm_memory_slot *memslot;int as_id, id; -if (!mask) -return ; - as_id = slot >>16 ; id = (u16 )slot;@@ -108,15 +105,24 @@ static inline bool kvm_dirty_gfn_harvested(struct kvm_dirty_gfn *gfn)int kvm_dirty_ring_reset (struct kvm *kvm,struct kvm_dirty_ring *ring,int *nr_entries_reset) { +/* +* To minimize mmu_lock contention, batch resets for harvested entries +* whose gfns are in the same slot, and are within N frame numbers of +* each other, where N is the number of bits in an unsigned long. For +* simplicity, process the current set of entries when the next entry +* can't be included in the batch. +* +* Track the current batch slot, the gfn offset into the slot for the +* batch, and the bitmask of gfns that need to be reset (relative to +* offset). Note, the offset may be adjusted backwards, e.g. so that +* a sequence of gfns X, X-1, ... X-N can be batched. +*/ u32 cur_slot, next_slot;u64 cur_offset, next_offset; -unsigned long mask; +unsigned long mask =0 ;struct kvm_dirty_gfn *entry;bool first_round = true; -/* This is only needed to make compilers happy */ - cur_slot = cur_offset = mask =0 ; -while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))return -EINTR ;@@ -164,14 +170,34 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,continue ; } } -kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + +/* +* Reset the slot for all the harvested entries that have been +* gathered, but not yet fully processed. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + +/* +* The current slot was reset or this is the first harvested +* entry, (re)initialize the metadata. +*/ cur_slot = next_slot; cur_offset = next_offset; mask =1 ; first_round = false; } -kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); +/* +* Perform a final reset if there are harvested entries that haven't +* been processed, which is guaranteed if at least one harvested was +* found. The loop only performs a reset when the "next" entry can't +* be batched with "current" the entry(s), and that reset processes the +* _current_ entry(s), i.e. the last harvested entry, a.k.a. next, will +* always be left pending. +*/ +if (mask) +kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask);/* * The request KVM_REQ_DIRTY_RING_SOFT_FULL will be cleared
Sean Christopherson
reviewerPATCH v2 5/5 KVM: Use mask of harvested dirty ring entries to coalesce dirty ring resets
In reply to Sean ChristophersonUse "mask" instead of a dedicated boolean to track whether or not there is at least one to-be-reset entry for the current slot+offset. In the body of the loop, mask is zero only on the first iteration, i.e. !mask is equivalent to first_round.
Opportunstically combine the adjacent "if (mask)" statements into a single if-statement.
No function change intended.
--- virt/kvm/dirty_ring.c | 60 +++++++++++++++++++++---------------------- 1 file changed, 29 insertions(+), 31 deletions(-)
virt/kvm/dirty_ring.c +29 -31
@@ -121,7 +121,6 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,u64 cur_offset, next_offset;unsigned long mask =0 ;struct kvm_dirty_gfn *entry; -bool first_round = true;while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))
@@ -141,42 +140,42 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++; (*nr_entries_reset)++; -/* -* While the size of each ring is fixed, it's possible for the -* ring to be constantly re-dirtied/harvested while the reset -* is in-progress (the hard limit exists only to guard against -* wrapping the count into negative space). -*/ -if (!first_round) +if (mask) { +/* +* While the size of each ring is fixed, it's possible +* for the ring to be constantly re-dirtied/harvested +* while the reset is in-progress (the hard limit exists +* only to guard against the count becoming negative). +*/ cond_resched (); -/* -* Try to coalesce the reset operations when the guest is -* scanning pages in the same slot. -*/ -if (!first_round && next_slot == cur_slot) { -s64 delta = next_offset - cur_offset; +/* +* Try to coalesce the reset operations when the guest +* is scanning pages in the same slot. +*/ +if (next_slot == cur_slot) { +s64 delta = next_offset - cur_offset; -if (delta >=0 && delta <BITS_PER_LONG ) { - mask |=1ull << delta; -continue ; - } +if (delta >=0 && delta <BITS_PER_LONG ) { + mask |=1ull << delta; +continue ; + } -/* Backwards visit, careful about overflows! */ -if (delta > -BITS_PER_LONG && delta <0 && - (mask << -delta >> -delta) == mask) { - cur_offset = next_offset; - mask = (mask << -delta) |1 ; -continue ; +/* Backwards visit, careful about overflows! */ +if (delta > -BITS_PER_LONG && delta <0 && + (mask << -delta >> -delta) == mask) { + cur_offset = next_offset; + mask = (mask << -delta) |1 ; +continue ; + } } - } -/* -* Reset the slot for all the harvested entries that have been -* gathered, but not yet fully processed. -*/ -if (mask) +/* +* Reset the slot for all the harvested entries that +* have been gathered, but not yet fully processed. +*/ kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + }/* * The current slot was reset or this is the first harvested
@@ -185,7 +184,6 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, cur_slot = next_slot; cur_offset = next_offset; mask =1 ; - first_round = false; }/*
Signature · 1 line
2.49.0.1015.ga840276032-goog
Re: [PATCH v2 5/5] KVM: Use mask of harvested dirty ring entries to coalesce dirty ring resets
Quoted 10 lines · Sean Christopherson
Use "mask" instead of a dedicated boolean to track whether or not there is at least one to-be-reset entry for the current slot+offset. In the body of the loop, mask is zero only on the first iteration, i.e. !mask is equivalent to
first_round.Opportunstically combine the adjacent "if (mask)" statements into a single if-statement.
No function change intended.
nit: "Opportunistically" and "functional" :)
Quoted 5 lines
Cc Peter Xu <[email protected]>Cc Yan Zhao <[email protected]>Cc Maxim Levitsky <[email protected]>Signed-off-by Sean Christopherson <[email protected]>
Thanks, Sean! This logic is much easier to read now.
Feel free to add to the entire series if you'd like:
Re: [PATCH v2 5/5] KVM: Use mask of harvested dirty ring entries to coalesce dirty ring resets
In reply to Sean ChristophersonQuoted 8 lines · Sean Christopherson
Use "mask" instead of a dedicated boolean to track whether or not there is at least one to-be-reset entry for the current slot+offset. In the body of the loop, mask is zero only on the first iteration, i.e. !mask is equivalent to
first_round.Opportunstically combine the adjacent "if (mask)" statements into a single if-statement.
Quoted 6 lines
No function change intended.
Cc Peter Xu <[email protected]>Cc Yan Zhao <[email protected]>Cc Maxim Levitsky <[email protected]>Signed-off-by Sean Christopherson <[email protected]>
Thanks!
Quoted 96 lines
--- virt/kvm/dirty_ring.c | 60 +++++++++++++++++++++---------------------- 1 file changed, 29 insertions(+), 31 deletions(-)virt/kvm/dirty_ring.c +29
-31@@ -121,7 +121,6 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,u64 cur_offset, next_offset;unsigned long mask =0 ;struct kvm_dirty_gfn *entry; -bool first_round = true;while (likely ((*nr_entries_reset) <INT_MAX )) {if (signal_pending (current))@@ -141,42 +140,42 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, ring->reset_index++; (*nr_entries_reset)++; -/* -* While the size of each ring is fixed, it's possible for the -* ring to be constantly re-dirtied/harvested while the reset -* is in-progress (the hard limit exists only to guard against -* wrapping the count into negative space). -*/ -if (!first_round) +if (mask) { +/* +* While the size of each ring is fixed, it's possible +* for the ring to be constantly re-dirtied/harvested +* while the reset is in-progress (the hard limit exists +* only to guard against the count becoming negative). +*/ cond_resched (); -/* -* Try to coalesce the reset operations when the guest is -* scanning pages in the same slot. -*/ -if (!first_round && next_slot == cur_slot) { -s64 delta = next_offset - cur_offset; +/* +* Try to coalesce the reset operations when the guest +* is scanning pages in the same slot. +*/ +if (next_slot == cur_slot) { +s64 delta = next_offset - cur_offset; -if (delta >=0 && delta <BITS_PER_LONG ) { - mask |=1ull << delta; -continue ; - } +if (delta >=0 && delta <BITS_PER_LONG ) { + mask |=1ull << delta; +continue ; + } -/* Backwards visit, careful about overflows! */ -if (delta > -BITS_PER_LONG && delta <0 && - (mask << -delta >> -delta) == mask) { - cur_offset = next_offset; - mask = (mask << -delta) |1 ; -continue ; +/* Backwards visit, careful about overflows! */ +if (delta > -BITS_PER_LONG && delta <0 && + (mask << -delta >> -delta) == mask) { + cur_offset = next_offset; + mask = (mask << -delta) |1 ; +continue ; + } } - } -/* -* Reset the slot for all the harvested entries that have been -* gathered, but not yet fully processed. -*/ -if (mask) +/* +* Reset the slot for all the harvested entries that +* have been gathered, but not yet fully processed. +*/ kvm_reset_dirty_gfn (kvm, cur_slot, cur_offset, mask); + }/* * The current slot was reset or this is the first harvested @@ -185,7 +184,6 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring, cur_slot = next_slot; cur_offset = next_offset; mask =1 ; - first_round = false; }/*