Skip to content

vfio/pci: Fix up breakage against split irqchip and INTx

Peter Xu qemu-devel 4 versions – Applied

Status from this message: Queued, thanks.

Get series with b4
b4 am [email protected]

Versions newest first · linked by author and title

  1. v4 Latest5 patches 3 repliesApplied
  2. v35 patches 9 repliesSuperseded
  3. v25 patches 11 repliesSuperseded
Earlier versions · 1
VersionPostedPatchesRepliesStatus
v15 patches23 repliesSuperseded

Changes in v4 from the cover letter

- pick r-b and a-b for Alex without patch 4
- only kick resamplefd for level triggered irq (as 3.1 change on patch
4) [Alex]
- fix mingw build error with below squashed into patch 4:

Patches in v4 Applied

#Subject and filesSizeRepliesApplied
1/5vfio/pci: Disable INTx fast path if using split irqchiphw/vfio/pci.c+12 -00 replies
2/5vfio/pci: Use kvm_irqchip_add_irqfd_notifier_gsi() for irqfdshw/vfio/pci.c+15 -220 replies
3/5KVM: Pass EventNotifier into kvm_irqchip_assign_irqfdaccel/kvm/kvm-all.c+10 -60 replies
4/5KVM: Kick resamplefd for split kernel irqchipaccel/kvm/kvm-all.c accel/kvm/trace-events hw/intc/ioapic.c +1 more+101 -20 replies
5/5Revert "vfio/pci: Disable INTx fast path if using split irqchip"hw/vfio/pci.c+0 -120 replies

Reviews trailers per patch

4 of 5 patches have a Reviewed-by or Acked-by.

PatchEric AugerAlex WilliamsonCornelia Huck
1/5T-b R-b R-b A-b
2/5R-b R-b A-b R-b
3/5R-b R-b
4/5
5/5R-b R-b A-b

Trailers given in replies are filled; those carried in the patches from earlier versions are outlined with dashes. Scroll the table sideways for all 3 reviewers.

Read the v4 thread

Cover letter [PATCH v4 0/5] vfio/pci: Fix up breakage against split irqchip and INTx

v4:
- pick r-b and a-b for Alex without patch 4
- only kick resamplefd for level triggered irq (as 3.1 change on patch
4) [Alex]
- fix mingw build error with below squashed into patch 4:

----------------------------------------

hw/intc/ioapic.c +2 -0
index 2ae96e10be..b9ec570c03 100644
@@ -241,6 +241,7 @@ void ioapic_eoi_broadcast(int vector)
                 continue;
             }
 
+#ifdef CONFIG_KVM
             /*
              * When IOAPIC is in the userspace while APIC is still in
              * the kernel (i.e., split irqchip), we have a trick to
@@ -257,6 +258,7 @@ void ioapic_eoi_broadcast(int vector)
              * emulated devices that are using/sharing the same IRQ.
              */
             kvm_resample_fd_notify(n);
+#endif
 
             if (!(entry & IOAPIC_LVT_REMOTE_IRR)) {
                 continue;

----------------------------------------

v3:
- collect r-bs for Eric
- unconditionally call kvm_resample_fd_notify(), change comment [Alex]
- remove the split irqchip check in kvm_resample_fd_notify(), then let it return nothing [Alex]
- test against shared irq to make sure it won't break

v2:
- pick tags
- don't register resamplefd with KVM kernel when the userspace resamplefd path is enabled (should enable fast path on new kernels)
- fix resamplefd mem leak
- fix commit message of patch 4 [Eric]
- let kvm_resample_fd_notify() return a boolean, skip ioapic check if returned true
- more comments here and there in the code to state the fact that userspace ioapic irr & remote-irr are bypassed [Paolo]

VFIO INTx is not working with split irqchip. On new kernels KVM_IRQFD will directly fail with resamplefd attached so QEMU will automatically fallback to the INTx slow path. However on old kernels it's still broken.

Only until recently I noticed that this could also break PXE boot for assigned NICs [1]. My wild guess is that the PXE ROM will be mostly using INTx as well, which means we can't bypass that even if we enables MSI for the guest kernel.

This series tries to first fix this issue function-wise, then speed up for the INTx again with resamplefd (mostly following the ideas proposed by Paolo one year ago [2]). My TCP_RR test shows that:

- Before this series: this is broken, no number to show

- After patch 1 (enable slow path): get 63% perf comparing to full kernel irqchip

- After whole series (enable fast path partly, irq injection will be the same as fast path, however userspace needs to intercept for EOI broadcast to resamplefd, though should still be faster than the MMIO trick for intx eoi): get 93% perf comparing to full kernel irqchip, which is a 46% performance boost

I think we can consider to apply patch 1 even sooner than the rest of the series to unbreak intx+split first.

The whole test matrix for reference:

  |----------+---------+-------------------+--------------+--------------------|
  | IRQ type | irqchip | TCP_STREAM (Gbps) | TCP_RR (pps) | note               |
  |----------+---------+-------------------+--------------+--------------------|
  | msi      | on      |              9.39 |        17567 |                    |
  | nomsi    | on      |              9.29 |        14056 |                    |
  | msi      | split   |              9.36 |        17330 |                    |
  | nomsi    | split   |                 / |            / | currently broken   |
  | nomsi    | split   |              8.98 |         8977 | after patch 1      |
  | nomsi    | split   |              9.21 |        13142 | after whole series |
  |----------+---------+-------------------+--------------+--------------------|

Any review comment is welcomed. Thanks,

[1] https://bugzilla.redhat.com/show_bug.cgi?id=1786404
[2] https://patchwork.kernel.org/patch/10738541/#22609933

Peter Xu (5):
vfio/pci: Disable INTx fast path if using split irqchip vfio/pci: Use kvm_irqchip_add_irqfd_notifier_gsi() for irqfds
KVM: Pass EventNotifier into kvm_irqchip_assign_irqfd
KVM: Kick resamplefd for split kernel irqchip
Revert "vfio/pci: Disable INTx fast path if using split irqchip"

 accel/kvm/kvm-all.c    | 95 ++++++++++++++++++++++++++++++++++++++----
 accel/kvm/trace-events |  1 +
 hw/intc/ioapic.c       | 19 +++++++++
 hw/vfio/pci.c          | 37 +++++++---------
 include/sysemu/kvm.h   |  4 ++
 5 files changed, 126 insertions(+), 30 deletions(-)
Signature · 1 line
2.24.1