On Thu, 26 Jul 2018 14:58:59 +0200, Dmitry Vyukov wrote:

On Thu, Jul 26, 2018 at 2:50 PM, Takashi Iwai tiwai@suse.de wrote:

...

On Thu, 26 Jul 2018 07:53:26 +0200, Dae R. Jeong wrote:

...

Reporting the crash: BUG: soft lockup in snd_virmidi_output_trigger

This crash has been found in v4.18-rc3 using RaceFuzzer (a modified version of Syzkaller), which we describe more at the end of this report.

Note that this bug is previously reported by Syzkaller a few month ago. (https://syzkaller.appspot.com/bug?id=642c927972318775e0ea1b4779ccd8624ce9e6f...) Nonetheless, the crash still can be detected. We guess that the crash has not fixed yet. We report the crash log and the attached reproducer with a hope that they are helpful to diagnose the crash and to fix a bug.

C Reproducer: https://kiwi.cs.purdue.edu/static/race-fuzzer/repro-dab082.c

Kernel config: https://kiwi.cs.purdue.edu/static/race-fuzzer/config-dab082

(Please disregard all code related to "should_hypercall" in the C repro, as this is only for our debugging purposes using our own hypervisor.)

Crash log:

watchdog: BUG: soft lockup - CPU#1 stuck for 22s! [syz-executor0:24261] Modules linked in: irq event stamp: 6692 hardirqs last enabled at (6691): [<ffffffff84eb1f12>] __raw_spin_unlock_irqrestore include/linux/spinlock_api_smp.h:160 [inline] hardirqs last enabled at (6691): [<ffffffff84eb1f12>] _raw_spin_unlock_irqrestore+0x62/0x90 kernel/locking/spinlock.c:184 hardirqs last disabled at (6692): [<ffffffff85000905>] interrupt_entry+0xb5/0xf0 arch/x86/entry/entry_64.S:625 softirqs last enabled at (1468): [<ffffffff852005ff>] __do_softirq+0x5ff/0x7f4 kernel/softirq.c:314 softirqs last disabled at (1455): [<ffffffff8130de56>] invoke_softirq kernel/softirq.c:368 [inline] softirqs last disabled at (1455): [<ffffffff8130de56>] irq_exit+0x126/0x130 kernel/softirq.c:408 CPU: 1 PID: 24261 Comm: syz-executor0 Not tainted 4.18.0-rc3 #1 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.8.2-0-g33fbe13 by qemu-project.org 04/01/2014 RIP: 0010:arch_local_irq_restore arch/x86/include/asm/paravirt.h:783 [inline] RIP: 0010:__raw_spin_unlock_irqrestore include/linux/spinlock_api_smp.h:160 [inline] RIP: 0010:_raw_spin_unlock_irqrestore+0x7d/0x90 kernel/locking/spinlock.c:184 Code: 0d c8 ce 16 7b 5b 41 5c 5d c3 e8 3e 0f 52 fc 48 c7 c7 68 b0 71 86 e8 62 4c 8d fc 48 83 3d 42 91 86 01 00 74 0e 48 89 df 57 9d <0f> 1f 44 00 00 eb cd 0f 0b 0f 0b 0f 1f 84 00 00 00 00 00 55 48 89 RSP: 0018:ffff8801b8e4f7e8 EFLAGS: 00000282 ORIG_RAX: ffffffffffffff13 RAX: 0000000000000000 RBX: 0000000000000282 RCX: ffffffff84eb1f1e RDX: dffffc0000000000 RSI: dffffc0000000000 RDI: 0000000000000282 RBP: ffff8801b8e4f7f8 R08: ffffed002573a06a R09: 0000000000000000 R10: 0000000000000000 R11: 0000000000000000 R12: ffff88012b9d0348 R13: ffff8801e6417580 R14: ffff8801b8e4f868 R15: ffff8801e64175a8 FS: 00007f3447a49700(0000) GS:ffff8801f6d00000(0000) knlGS:0000000000000000 CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 CR2: 00007f3447a48db8 CR3: 00000001ee58d000 CR4: 00000000000006e0 Call Trace: spin_unlock_irqrestore include/linux/spinlock.h:365 [inline] snd_virmidi_output_trigger+0x37d/0x420 sound/core/seq/seq_virmidi.c:205 snd_rawmidi_output_trigger sound/core/rawmidi.c:150 [inline] snd_rawmidi_kernel_write1+0x383/0x410 sound/core/rawmidi.c:1288 snd_rawmidi_write+0x249/0xa50 sound/core/rawmidi.c:1338

This seems to be a long loop in snd_virmidi_output_trigger processed in a spinlock. The below patch puts a cond_resched() for avoiding the possible stalls, and it seems working.

Let me know if this works for you.

thanks,

Takashi

-- 8< -- From: Takashi Iwai tiwai@suse.de Subject: [PATCH] ALSA: virmidi: Fix too long output trigger loop

The virmidi output trigger tries to parse the all available bytes and process sequencer events as much as possible. In a normal situation, this is supposed to be relatively short, but a program may give a huge buffer and it'll take a long time in a single spin lock, which may eventually lead to a soft lockup.

This patch simply adds a workaround, a cond_resched() call in the loop if applicable. A better solution would be to move the event processor into a work, but let's put a duct-tape quickly at first.

Reported-by: Dae R. Jeong threeearcat@gmail.com Cc: stable@vger.kernel.org Signed-off-by: Takashi Iwai tiwai@suse.de

---

sound/core/seq/seq_virmidi.c | 6 ++++++ 1 file changed, 6 insertions(+)

diff --git a/sound/core/seq/seq_virmidi.c b/sound/core/seq/seq_virmidi.c index 289ae6bb81d9..65549d41533e 100644 --- a/sound/core/seq/seq_virmidi.c +++ b/sound/core/seq/seq_virmidi.c @@ -163,6 +163,7 @@ static void snd_virmidi_output_trigger(struct snd_rawmidi_substream *substream, int count, res; unsigned char buf[32], *pbuf; unsigned long flags;

•   bool check_resched = !in_atomic();
  
    if (up) {
            vmidi->trigger = 1;
  

@@ -200,6 +201,11 @@ static void snd_virmidi_output_trigger(struct snd_rawmidi_substream *substream, vmidi->event.type = SNDRV_SEQ_EVENT_NONE; } }

•                   if (!check_resched)
  

•                           continue;
  

•                   spin_unlock_irqrestore(&substream->runtime->lock, flags);
  

•                   cond_resched();
  

•                   spin_lock_irqsave(&substream->runtime->lock, flags);
            }
    out:
            spin_unlock_irqrestore(&substream->runtime->lock, flags);
  

Just in case: is it a killable loop? It seems that user can make kernel stuck there almost infinitely.

Well, the original buffer size in rawmidi is limited, so this loop should be limited, too. I haven't seen the unkillable task through tests.

thanks,

Takashi