Re: [PATCH v1 01/14] mm: introduce bpf struct ops for OOM handling

[Roman Gushchin <roman.gushchin@linux.dev>]

Kumar Kartikeya Dwivedi <memxor@gmail.com> writes:

> On Thu, 21 Aug 2025 at 02:25, Roman Gushchin <roman.gushchin@linux.dev> wrote:
>>
>> Kumar Kartikeya Dwivedi <memxor@gmail.com> writes:
>>
>> > On Mon, 18 Aug 2025 at 19:01, Roman Gushchin <roman.gushchin@linux.dev> wrote:
>> >>
>> >> Introduce a bpf struct ops for implementing custom OOM handling policies.
>> >>
>> >> The struct ops provides the bpf_handle_out_of_memory() callback,
>> >> which expected to return 1 if it was able to free some memory and 0
>> >> otherwise.
>> >>
>> >> In the latter case it's guaranteed that the in-kernel OOM killer will
>> >> be invoked. Otherwise the kernel also checks the bpf_memory_freed
>> >> field of the oom_control structure, which is expected to be set by
>> >> kfuncs suitable for releasing memory. It's a safety mechanism which
>> >> prevents a bpf program to claim forward progress without actually
>> >> releasing memory. The callback program is sleepable to enable using
>> >> iterators, e.g. cgroup iterators.
>> >>
>> >> The callback receives struct oom_control as an argument, so it can
>> >> easily filter out OOM's it doesn't want to handle, e.g. global vs
>> >> memcg OOM's.
>> >>
>> >> The callback is executed just before the kernel victim task selection
>> >> algorithm, so all heuristics and sysctls like panic on oom,
>> >> sysctl_oom_kill_allocating_task and sysctl_oom_kill_allocating_task
>> >> are respected.
>> >>
>> >> The struct ops also has the name field, which allows to define a
>> >> custom name for the implemented policy. It's printed in the OOM report
>> >> in the oom_policy=<policy> format. "default" is printed if bpf is not
>> >> used or policy name is not specified.
>> >>
>> >> [  112.696676] test_progs invoked oom-killer: gfp_mask=0xcc0(GFP_KERNEL), order=0, oom_score_adj=0
>> >>                oom_policy=bpf_test_policy
>> >> [  112.698160] CPU: 1 UID: 0 PID: 660 Comm: test_progs Not tainted 6.16.0-00015-gf09eb0d6badc #102 PREEMPT(full)
>> >> [  112.698165] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.17.0-5.fc42 04/01/2014
>> >> [  112.698167] Call Trace:
>> >> [  112.698177]  <TASK>
>> >> [  112.698182]  dump_stack_lvl+0x4d/0x70
>> >> [  112.698192]  dump_header+0x59/0x1c6
>> >> [  112.698199]  oom_kill_process.cold+0x8/0xef
>> >> [  112.698206]  bpf_oom_kill_process+0x59/0xb0
>> >> [  112.698216]  bpf_prog_7ecad0f36a167fd7_test_out_of_memory+0x2be/0x313
>> >> [  112.698229]  bpf__bpf_oom_ops_handle_out_of_memory+0x47/0xaf
>> >> [  112.698236]  ? srso_alias_return_thunk+0x5/0xfbef5
>> >> [  112.698240]  bpf_handle_oom+0x11a/0x1e0
>> >> [  112.698250]  out_of_memory+0xab/0x5c0
>> >> [  112.698258]  mem_cgroup_out_of_memory+0xbc/0x110
>> >> [  112.698274]  try_charge_memcg+0x4b5/0x7e0
>> >> [  112.698288]  charge_memcg+0x2f/0xc0
>> >> [  112.698293]  __mem_cgroup_charge+0x30/0xc0
>> >> [  112.698299]  do_anonymous_page+0x40f/0xa50
>> >> [  112.698311]  __handle_mm_fault+0xbba/0x1140
>> >> [  112.698317]  ? srso_alias_return_thunk+0x5/0xfbef5
>> >> [  112.698335]  handle_mm_fault+0xe6/0x370
>> >> [  112.698343]  do_user_addr_fault+0x211/0x6a0
>> >> [  112.698354]  exc_page_fault+0x75/0x1d0
>> >> [  112.698363]  asm_exc_page_fault+0x26/0x30
>> >> [  112.698366] RIP: 0033:0x7fa97236db00
>> >>
>> >> It's possible to load multiple bpf struct programs. In the case of
>> >> oom, they will be executed one by one in the same order they been
>> >> loaded until one of them returns 1 and bpf_memory_freed is set to 1
>> >> - an indication that the memory was freed. This allows to have
>> >> multiple bpf programs to focus on different types of OOM's - e.g.
>> >> one program can only handle memcg OOM's in one memory cgroup.
>> >> But the filtering is done in bpf - so it's fully flexible.
>> >
>> > I think a natural question here is ordering. Is this ability to have
>> > multiple OOM programs critical right now?
>>
>> Good question. Initially I had only supported a single bpf policy.
>> But then I realized that likely people would want to have different
>> policies handling different parts of the cgroup tree.
>> E.g. a global policy and several policies handling OOMs only
>> in some memory cgroups.
>> So having just a single policy is likely a no go.
>
> If the ordering is more to facilitate scoping, would it then be better
> to support attaching the policy to specific memcg/cgroup?

Well, it has some advantages and disadvantages. First, it will require
way more infrastructure on the memcg side. Second, the interface is not
super clear: we don't want to have a struct ops per cgroup, I guess.
And in many case a single policy for all memcgs is just fine, so asking
the user to attach it to all memcgs is just adding a toil and creating
all kinds of races.
So I see your point, but I'm not yet convinced, to be honest.

Thanks!