[PATCH] vgaarb: fix signal handling in vga_get()

[Kirill A. Shutemov]

On Mon, Dec 14, 2015 at 11:20:00AM +0100, David Herrmann wrote:
> Hi
> 
> On Mon, Dec 14, 2015 at 9:19 AM, Kirill A. Shutemov
> <kirill at shutemov.name> wrote:
> > On Thu, Dec 10, 2015 at 11:28:58AM +0100, David Herrmann wrote:
> >> Hi
> >>
> >> On Mon, Nov 30, 2015 at 3:17 AM, Kirill A. Shutemov
> >> <kirill at shutemov.name> wrote:
> >> > There are few defects in vga_get() related to signal hadning:
> >> >
> >> >   - we shouldn't check for pending signals for TASK_UNINTERRUPTIBLE
> >> >     case;
> >> >
> >> >   - if we found pending signal we must remove ourself from wait queue
> >> >     and change task state back to running;
> >> >
> >> >   - -ERESTARTSYS is more appropriate, I guess.
> >> >
> >> > Signed-off-by: Kirill A. Shutemov <kirill at shutemov.name>
> >> > ---
> >> >
> >> > Alex, I try to get KVM with VGA passthrough working properly. I have i915
> >> > (HD 4600) on the host and GTX 580 for the guest. The guest GPU is not
> >> > capabale of EFI, so I have to use x-vga=on. It's kinda work with your
> >> > patch for i915.enable_hd_vgaarb=1. But guest refuse to initialize the GPU
> >> > after KVM was not shut down correctly, resulting in host crash like this:
> >> >
> >> > BUG: unable to handle kernel paging request at ffff880870187ed8
> >> > IP: [<ffff880870187ed8>] 0xffff880870187ed8
> >> > PGD 2129067 PUD 80000008400001e3
> >> > Oops: 0011 [#1] PREEMPT SMP
> >> > Modules linked in: iwlmvm iwlwifi
> >> > CPU: 6 PID: 3983 Comm: qemu-system-x86 Not tainted 4.3.0-gentoo #6
> >> > Hardware name: Gigabyte Technology Co., Ltd. Z87X-UD7 TH/Z87X-UD7 TH-CF, BIOS F5a 06/12/2014
> >> > task: ffff88087a910000 ti: ffff8808632c0000 task.ti: ffff8808632c0000
> >> > RIP: 0010:[<ffff880870187ed8>]  [<ffff880870187ed8>] 0xffff880870187ed8
> >> > RSP: 0018:ffff8808632c3d08  EFLAGS: 00010006
> >> > RAX: ffff880870187db0 RBX: 0000000070187f58 RCX: 0000000000000000
> >> > RDX: 0000000000000000 RSI: 0000000000000003 RDI: ffff880870187db0
> >> > RBP: ffff8808632c3d48 R08: 0000000000000000 R09: 0000000000000000
> >> > R10: 00000000000103c0 R11: 0000000000000293 R12: ffffffff81ea03c8
> >> > R13: ffffffff8104c7cb R14: 0000000000000000 R15: 0000000000000003
> >> > FS:  00007f984f9b2700(0000) GS:ffff88089f380000(0000) knlGS:0000000000000000
> >> > CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> >> > CR2: ffff880870187ed8 CR3: 00000008645f8000 CR4: 00000000001426e0
> >> > Stack:
> >> >  ffffffff810cc83d 00000000632c3d28 0000000000000000 ffffffff81ea03c0
> >> >  0000000000000046 0000000000000003 0000000000000000 0000000000000000
> >> >  ffff8808632c3d80 ffffffff810cca44 ffff88087af63800 0000000000000286
> >> > Call Trace:
> >> >  [<ffffffff810cc83d>] ? __wake_up_common+0x4d/0x80
> >> >  [<ffffffff810cca44>] __wake_up+0x34/0x50
> >> >  [<ffffffff815d99e3>] __vga_put+0x73/0xd0
> >> >  [<ffffffff815d9db4>] vga_put+0x54/0x80
> >> >  [<ffffffff8169d042>] vfio_pci_vga_rw+0x1d2/0x220
> >> >  [<ffffffff8169a7f3>] vfio_pci_rw+0x33/0x60
> >> >  [<ffffffff8169abf7>] vfio_pci_write+0x17/0x20
> >> >  [<ffffffff816966a6>] vfio_device_fops_write+0x26/0x30
> >> >  [<ffffffff811a4b23>] __vfs_write+0x23/0xe0
> >> >  [<ffffffff811a4a53>] ? __vfs_read+0x23/0xd0
> >> >  [<ffffffff811b6e35>] ? do_vfs_ioctl+0x2b5/0x490
> >> >  [<ffffffff811a5194>] vfs_write+0xa4/0x190
> >> >  [<ffffffff811a5fa6>] SyS_pwrite64+0x66/0xa0
> >> >  [<ffffffff819a17d7>] entry_SYSCALL_64_fastpath+0x12/0x6a
> >> > Code: 88 ff ff e0 7e 18 70 08 88 ff ff 00 8c 57 76 08 88 ff ff 20 7f 18 70 08 88 ff ff 08 7f 18 70 08 88 ff ff 94 51 1a 81 ff ff ff ff <09> 00 00 00 00 00 00 00 01 8c 57 76 08 88 ff ff 00 8c 57 76 08
> >> > RIP  [<ffff880870187ed8>] 0xffff880870187ed8
> >> >  RSP <ffff8808632c3d08>
> >> > CR2: ffff880870187ed8
> >> >
> >> > The patch fixes the crash, but doesn't help with getting GPU in guest
> >> > working again.
> >> >
> >> > Any ideas?
> >> >
> >> > ---
> >> >  drivers/gpu/vga/vgaarb.c | 6 ++++--
> >> >  1 file changed, 4 insertions(+), 2 deletions(-)
> >> >
> >> > diff --git a/drivers/gpu/vga/vgaarb.c b/drivers/gpu/vga/vgaarb.c
> >> > index 3166e4bc4eb6..9abcaa53bd25 100644
> >> > --- a/drivers/gpu/vga/vgaarb.c
> >> > +++ b/drivers/gpu/vga/vgaarb.c
> >> > @@ -395,8 +395,10 @@ int vga_get(struct pci_dev *pdev, unsigned int rsrc, int interruptible)
> >> >                 set_current_state(interruptible ?
> >> >                                   TASK_INTERRUPTIBLE :
> >> >                                   TASK_UNINTERRUPTIBLE);
> >> > -               if (signal_pending(current)) {
> >> > -                       rc = -EINTR;
> >> > +               if (interruptible && signal_pending(current)) {
> >> > +                       __set_current_state(TASK_RUNNING);
> >> > +                       remove_wait_queue(&vga_wait_queue, &wait);
> >> > +                       rc = -ERESTARTSYS;
> >> >                         break;
> >>
> >> All 3 points are valid, and the patch looks good to me:
> >>
> >> Reviewed-by: David Herrmann <dh.herrmann at gmail.com>
> >>
> >> However, there seems to be a race between vga_lock and putting the
> >> thread asleep.
> >
> > I'm not sure I understand the race you're talking about.
> > Could you elaborate?
> 
> Normal wake-up race. You have to check the condition you wait for
> *after* linking your wait-queue and setting your thread-state. In the
> current code, if the wake-up happens between spin_unlock_irqrestore()
> and add_wait_queue(), you will never get woken up.

Ah, I see. Your patch looks good to me to fix the issue.

-- 
 Kirill A. Shutemov