Re: [PATCH bpf v3 3/7] bpf: Add sleepable arena page allocation path

["Emil Tsalapatis" <emil@etsalapatis.com>]

On Fri Sep 25, 2026 at 9:25 PM UTC, Alexei Starovoitov wrote:
> On Fri, Sep 25, 2026 at 08:39 PM Emil Tsalapatis <emil@etsalapatis.com> wrote:
>> Since allocation now happens before range-tree arbitration, concurrent
>> callers can allocate pages for competing requests. Bound the aggregate
>> number of pages in this pre-arbitration window to the configured maximum
>> size of the arena.
>
> The pages are allocated with __GFP_ACCOUNT. memcg bounds them already.
> Let's drop nr_pages_inflight. Seems like unnecessary complication.
>

The nr_pages_in_flight is there for the edge case where multiple concurrent
allocations in the arena inflate memcg usage close to the limit, then a non-arena
related allocation fails/triggers OOM. If we're fine with that happy to remove
it since it's a bit of a tacked-on hack.

I'll replace it with __GFP_RETRY_MAYFAIL on alloc_pages() so that the allocation
doesn't accidentally trigger the OOM killer. AFAICT it can do so otherwise.

>> +out_unmap:
>> +	if (sleepable)
>> +		flush_vmap_cache(kern_vm_start + uaddr32, mapped << PAGE_SHIFT);
>>  	range_tree_set(&arena->rt, pgoff + mapped, page_cnt - mapped);
>>  	raw_res_spin_unlock_irqrestore(&arena->spinlock, flags);
>> -	if (mapped) {
>> -		flush_vmap_cache(kern_vm_start + uaddr32, mapped << PAGE_SHIFT);
>> +	if (mapped || sleepable) {
>> +		if (!sleepable)
>> +			flush_vmap_cache(kern_vm_start + uaddr32, mapped << PAGE_SHIFT);
>>  		arena_free_pages(arena, uaddr32, mapped, sleepable);
>>  	}
>
> v2 leftovers? It doesn't depend on sleepable ?

Yes, I will remove that part.

>
> pw-bot: cr