Re: [PATCH bpf v2] bpf: Fix deadlock when freeing cgroup storage

[Martin KaFai Lau <martin.lau@linux.dev>]

On 1/27/25 2:15 PM, Alexei Starovoitov wrote:
> On Sun, Jan 26, 2025 at 1:31 AM Abel Wu <wuyun.abel@bytedance.com> wrote:
>>
>> On 1/25/25 4:20 AM, Martin KaFai Lau Wrote:
>>> On 12/20/24 10:10 PM, Abel Wu wrote:
>>>> The following commit
>>>> bc235cdb423a ("bpf: Prevent deadlock from recursive bpf_task_storage_[get|delete]")
>>>> first introduced deadlock prevention for fentry/fexit programs attaching
>>>> on bpf_task_storage helpers. That commit also employed the logic in map
>>>> free path in its v6 version.
>>>>
>>>> Later bpf_cgrp_storage was first introduced in
>>>> c4bcfb38a95e ("bpf: Implement cgroup storage available to non-cgroup-attached bpf progs")
>>>> which faces the same issue as bpf_task_storage, instead of its busy
>>>> counter, NULL was passed to bpf_local_storage_map_free() which opened
>>>> a window to cause deadlock:
>>>>
>>>>      <TASK>
>>>>          (acquiring local_storage->lock)
>>>>      _raw_spin_lock_irqsave+0x3d/0x50
>>>>      bpf_local_storage_update+0xd1/0x460
>>>>      bpf_cgrp_storage_get+0x109/0x130
>>>>      bpf_prog_a4d4a370ba857314_cgrp_ptr+0x139/0x170
>>>>      ? __bpf_prog_enter_recur+0x16/0x80
>>>>      bpf_trampoline_6442485186+0x43/0xa4
>>>>      cgroup_storage_ptr+0x9/0x20
>>>>          (holding local_storage->lock)
>>>>      bpf_selem_unlink_storage_nolock.constprop.0+0x135/0x160
>>>>      bpf_selem_unlink_storage+0x6f/0x110
>>>>      bpf_local_storage_map_free+0xa2/0x110
>>>>      bpf_map_free_deferred+0x5b/0x90
>>>>      process_one_work+0x17c/0x390
>>>>      worker_thread+0x251/0x360
>>>>      kthread+0xd2/0x100
>>>>      ret_from_fork+0x34/0x50
>>>>      ret_from_fork_asm+0x1a/0x30
>>>>      </TASK>
>>>>
>>>> Progs:
>>>>    - A: SEC("fentry/cgroup_storage_ptr")
>>>
>>> The v1 thread has suggested using notrace in a few functions. I didn't see any counterarguments that wouldn't be sufficient.
>>>
>>> imo, that should be a better option instead of having more unnecessary failures in all other normal use cases which will not be interested in tracing cgroup_storage_ptr().
> 
> Martin,
> 
> task_storage_map_free() is doing this busy inc/dec already,
> in that sense doing the same in cgroup_storage_map_free() fits.

sgtm. Agree to be consistent with the task_storage_map_free.

would be nice if the busy inc/dec usage can be revisited after the rqspinlock work.

Acked-by: Martin KaFai Lau <martin.lau@kernel.org>