Re: [PATCH] irqbypass: fix error handle when irq_bypass_register_producer() return fails

[Alex Williamson]

On Fri, 9 Oct 2020 12:30:04 +0800
gchen chen <gchen.guomin@xxxxxxxxx> wrote:

> Alex Williamson <alex.williamson@xxxxxxxxxx> 于2020年9月30日周三 下午10:09写道：
> >
> >
> > Please version your postings so we know which one to consider as the
> > current proposal.
> >
> > On Wed, 30 Sep 2020 20:54:39 +0800
> > guomin_chen@xxxxxxxx wrote:
> >  
> > > From: guomin chen <guomin_chen@xxxxxxxx>
> > >
> > > When the producer object registration fails,In the future, due to
> > > incorrect matching when unregistering, list_del(&producer->node)
> > > may still be called, then trigger a BUG:
> > >
> > >     vfio-pci 0000:db:00.0: irq bypass producer (token 0000000060c8cda5) registration fails: -16
> > >     vfio-pci 0000:db:00.0: irq bypass producer (token 0000000060c8cda5) registration fails: -16
> > >     vfio-pci 0000:db:00.0: irq bypass producer (token 0000000060c8cda5) registration fails: -16
> > >     ...
> > >     list_del corruption, ffff8f7fb8ba0828->next is LIST_POISON1 (dead000000000100)
> > >     ------------[ cut here ]------------
> > >     kernel BUG at lib/list_debug.c:47!
> > >     invalid opcode: 0000 [#1] SMP NOPTI
> > >     CPU: 29 PID: 3914 Comm: qemu-kvm Kdump: loaded Tainted: G      E
> > >     -------- - -4.18.0-193.6.3.el8.x86_64 #1
> > >     Hardware name: Lenovo ThinkSystem SR650 -[7X06CTO1WW]-/-[7X06CTO1WW]-,
> > >     BIOS -[IVE636Z-2.13]- 07/18/2019
> > >     RIP: 0010:__list_del_entry_valid.cold.1+0x12/0x4c
> > >     Code: ce ff 0f 0b 48 89 c1 4c 89 c6 48 c7 c7 40 85 4d 88 e8 8c bc
> > >           ce ff 0f 0b 48 89 fe 48 89 c2 48 c7 c7 d0 85 4d 88 e8 78 bc
> > >           ce ff <0f> 0b 48 c7 c7 80 86 4d 88 e8 6a bc ce ff 0f 0b 48
> > >           89 f2 48 89 fe
> > >     RSP: 0018:ffffaa9d60197d20 EFLAGS: 00010246
> > >     RAX: 000000000000004e RBX: ffff8f7fb8ba0828 RCX: 0000000000000000
> > >     RDX: 0000000000000000 RSI: ffff8f7fbf4d6a08 RDI: ffff8f7fbf4d6a08
> > >     RBP: 0000000000000000 R08: 000000000000084b R09: 000000000000005d
> > >     R10: 0000000000000000 R11: ffffaa9d60197bd0 R12: ffff8f4fbe863000
> > >     R13: 00000000000000c2 R14: 0000000000000000 R15: 0000000000000000
> > >     FS:  00007f7cb97fa700(0000) GS:ffff8f7fbf4c0000(0000)
> > >     knlGS:0000000000000000
> > >     CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > >     CR2: 00007fcf31da4000 CR3: 0000005f6d404001 CR4: 00000000007626e0
> > >     DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > >     DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > >     PKRU: 55555554
> > >     Call Trace:
> > >         irq_bypass_unregister_producer+0x9b/0xf0 [irqbypass]
> > >         vfio_msi_set_vector_signal+0x8c/0x290 [vfio_pci]
> > >         ? load_fixmap_gdt+0x22/0x30
> > >         vfio_msi_set_block+0x6e/0xd0 [vfio_pci]
> > >         vfio_pci_ioctl+0x218/0xbe0 [vfio_pci]
> > >         ? kvm_vcpu_ioctl+0xf2/0x5f0 [kvm]
> > >         do_vfs_ioctl+0xa4/0x630
> > >         ? syscall_trace_enter+0x1d3/0x2c0
> > >         ksys_ioctl+0x60/0x90
> > >         __x64_sys_ioctl+0x16/0x20
> > >         do_syscall_64+0x5b/0x1a0
> > >         entry_SYSCALL_64_after_hwframe+0x65/0xca
> > >
> > > Cc: Alex Williamson <alex.williamson@xxxxxxxxxx>
> > > Cc: Cornelia Huck <cohuck@xxxxxxxxxx>
> > > Cc: Jiang Yi <giangyi@xxxxxxxxxx>
> > > Cc: Marc Zyngier <maz@xxxxxxxxxx>
> > > Cc: Peter Xu <peterx@xxxxxxxxxx>
> > > Cc: Eric Auger <eric.auger@xxxxxxxxxx>
> > > Cc: "Michael S. Tsirkin" <mst@xxxxxxxxxx>
> > > Cc: Jason Wang <jasowang@xxxxxxxxxx>
> > > Cc: kvm@xxxxxxxxxxxxxxx
> > > Cc: linux-kernel@xxxxxxxxxxxxxxx
> > > Signed-off-by: guomin chen <guomin_chen@xxxxxxxx>
> > > ---
> > >  drivers/vfio/pci/vfio_pci_intrs.c | 13 +++++++++++--
> > >  drivers/vhost/vdpa.c              |  7 +++++++
> > >  2 files changed, 18 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/vfio/pci/vfio_pci_intrs.c b/drivers/vfio/pci/vfio_pci_intrs.c
> > > index 1d9fb25..c371943 100644
> > > --- a/drivers/vfio/pci/vfio_pci_intrs.c
> > > +++ b/drivers/vfio/pci/vfio_pci_intrs.c
> > > @@ -352,12 +352,21 @@ static int vfio_msi_set_vector_signal(struct vfio_pci_device *vdev,
> > >       vdev->ctx[vector].producer.token = trigger;
> > >       vdev->ctx[vector].producer.irq = irq;
> > >       ret = irq_bypass_register_producer(&vdev->ctx[vector].producer);
> > > -     if (unlikely(ret))
> > > +     if (unlikely(ret)) {
> > >               dev_info(&pdev->dev,
> > >               "irq bypass producer (token %p) registration fails: %d\n",
> > >               vdev->ctx[vector].producer.token, ret);
> > >
> > > -     vdev->ctx[vector].trigger = trigger;
> > > +             kfree(vdev->ctx[vector].name);
> > > +             eventfd_ctx_put(trigger);
> > > +
> > > +             cmd = vfio_pci_memory_lock_and_enable(vdev);
> > > +             free_irq(irq, trigger);
> > > +             vfio_pci_memory_unlock_and_restore(vdev, cmd);
> > > +
> > > +             vdev->ctx[vector].trigger = NULL;
> > > +     } else
> > > +             vdev->ctx[vector].trigger = trigger;
> > >
> > >       return 0;
> > >  }  
> >
> > Once again, the irq bypass registration cannot cause the vector setup
> > to fail, either by returning an error code or failing to configure the
> > vector while returning success.  It's my assertion that we simply need
> > to set the producer.token to NULL on failure such that unregistering
> > the producer will not generate a match, as you've done below.  The
> > vector still works even if this registration fails.
> >  
> Yes,  the irq bypass registration cannot cause the vector setup to fail.
> But if I simply set producer.token to NULL when fails, instead of
> cleaning up vector, it will trigger the following BUG:
> 
> vfio_ecap_init: 0000:db:00.0 hiding ecap 0x1e@0x310
> vfio-pci 0000:db:00.0: irq bypass producer (token 000000004409229f)
> registration fails: -16
> ------------[ cut here ]------------
> kernel BUG at drivers/pci/msi.c:352!
> invalid opcode: 0000 [#1] SMP NOPTI
> CPU: 55 PID: 9389 Comm: qemu-kvm Kdump: loaded Tainted: G
> E    --------- -  - 4.18.0-193.irqb.r1.el8.x86_64 #1
> Hardware name: Lenovo ThinkSystem SR650 -[7X06CTO1WW]-/-[7X06CTO1WW]-,
>   BIOS -[IVE636Z-2.13]- 07/18/2019
> RIP: 0010:free_msi_irqs+0x180/0x1b0
> Code: 14 85 c0 0f 84 d5 fe ff ff 31 ed eb 0f 83 c5 01 39 6b 14 0f 86
>       c5 fe ff ff 8b 7b 10 01 ef e8 d7 4a c9 ff 48 83 78 70 00 74 e3
>   <0f> 0b 49 8d b5 b0 00 00 00 e8 e2 e3 c9 ff e9 c7 fe ff ff 48
>   8b 7b
> RSP: 0018:ffffaeca4f4bfcd8 EFLAGS: 00010286
> RAX: ffff8bec77441600 RBX: ffff8bbcdb637e40 RCX: 0000000000000000
> RDX: 0000000000000000 RSI: 00000000000001ab RDI: ffffffff8ea5b2a0
> RBP: 0000000000000000 R08: ffff8bec7e746828 R09: ffff8bec7e7466a8
> R10: 0000000000000000 R11: 0000000000000000 R12: ffff8bbcde921308
> R13: ffff8bbcde921000 R14: 000000000000000b R15: 0000000000000021
> FS:  00007fd18d7fa700(0000) GS:ffff8bec7f6c0000(0000) knlGS:0000000000000000
> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 00007f83650024a0 CR3: 000000476e70c001 CR4: 00000000007626e0
> DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> PKRU: 55555554
> Call Trace:
>  pci_disable_msix+0xf3/0x120
>  pci_free_irq_vectors+0xe/0x20
>  vfio_msi_disable+0x89/0xd0 [vfio_pci]
>  vfio_pci_set_msi_trigger+0x229/0x2d0 [vfio_pci]
>  vfio_pci_ioctl+0x24f/0xdb0 [vfio_pci]
>  ? pollwake+0x74/0x90
>  ? wake_up_q+0x70/0x70
>  do_vfs_ioctl+0xa4/0x630
>  ? __alloc_fd+0x33/0x140
>  ? syscall_trace_enter+0x1d3/0x2c0
>  ksys_ioctl+0x60/0x90
>  __x64_sys_ioctl+0x16/0x20
>  do_syscall_64+0x5b/0x1a0
>  entry_SYSCALL_64_after_hwframe+0x65/0xca

Please post the patch that triggers this, I'm not yet convinced we're
speaking of the same solution.  The user ioctl cannot fail due to the
failure to setup a bypass accelerator, nor can the ioctl return success
without configuring all of the user requested vectors, which is what I
understand the v2 patch above to do.  We simply want to configure the
failed producer such that when we unregister it at user request, we
avoid creating a bogus match.  It's not apparent to me why doing that
would cause any changes to the setup or teardown of the MSI vector in
PCI code.  Thanks,

Alex