[PATCH v3] kernel/watch_queue: Make pipe NULL while clearing watch_queue

[Siddh Raman Pant]

On Mon, 01 Aug 2022 21:46:42 +0530  Dipanjan Das <mail.dipanjan.das at gmail.com> wrote:
> Are you referring to the reproducer attached to our original report?
> https://lore.kernel.org/all/CANX2M5bHye2ZEEhEV6PUj1kYL2KdWYeJtgXw8KZRzwrNpLYz+A@mail.gmail.com/
 
Yes, I meant the reproducer you gave.

I suspect I must have missed CONFIG_WATCH_QUEUE=y while setting the kernel
up, extremely sorry for it.

I now tried 5.10.y with it (using a modification of syzkaller's dashboard
config I had been using[1]), and I'm getting a __post_watch_notification()
crash (which is a related crash, as the fix[2][3] for that causes the
reproducer to not reproduce the post_one_notification crash on mainline),
but not the post_one_notification() crash you had reported. It seems if I
apply my patch, it doesn't trigger this related crash, so these bugs are
seem to be very related maybe due to racing? I haven't looked at that yet.

I then tried on v5.10.131 since that was the exact version you had
reproduced on, and it reproduces the post_one_notification() error
successfully. Applying 353f7988dd84 causes __post_watch_notification()
crash, and then applying this v3 patch does not trigger the issue, but
the patch to fix __post_watch_notification() crash is [2], which does
not really address the issue of post_one_notification() crash which
is due to the dangling reference to a freed pipe.

Can you please try reproducer at your end?

Thanks,
Siddh

[1] https://gist.github.com/siddhpant/06c7d64ca83273f0fd6604e4677f7c0d
[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=e64ab2dbd882933b65cd82ff6235d705ad65dbb6
[3] https://lore.kernel.org/linux-mm/18259769e5e.52eb2082293078.3991591702430862151@siddh.me/