Re: [PATCH RFC 0/2] mm/page_alloc: free_pcppages_bulk safeguard

[Alexei Starovoitov]

On Tue, Aug 22, 2023 at 2:15 PM Chris Li <chrisl@xxxxxxxxxx> wrote:
>
> Hi Kemeng,
>
> On Mon, Aug 21, 2023 at 6:27 PM Kemeng Shi <shikemeng@xxxxxxxxxxxxxxx> wrote:
> > >
> > Agreed. We assume pcp->count is protected by pcp->lock. Instead of make code
> > work in case pcp->count could be changed without lock held, it's more reasonble
> > to modify pcp->count with pcp->lock held in BPF program.
>
> The lock is holded when pcp->count is modified. It is going through
> the kernel page
> allocation API. The issue is nest memory allocation inside spin_lock()
> introduced by BPF.
>
> The execution sequence is like this:
>
>        count = min(pcp->count, count);
>
>         /* Ensure requested pindex is drained first. */
>         pindex = pindex - 1;
>         bpf_injected_spin_lock_irqsave {
>                  alloc_page();
>                  original spin_lock_irqsave(&zone->lock, flags) ;
>         }

bpf doesn't call into alloc_page() or slab alloc or pcpu alloc from
tracing progs.
All memory is preallocated.
Can you reproduce the issue on the latest upstream kernel?