On Tue, Feb 18, 2025 at 7:36 AM Vlastimil Babka <vbabka@xxxxxxx> wrote:
>
> On 2/13/25 04:35, Alexei Starovoitov wrote:
> > From: Alexei Starovoitov <ast@xxxxxxxxxx>
> >
> > Use try_alloc_pages() and free_pages_nolock() for BPF needs
> > when context doesn't allow using normal alloc_pages.
> > This is a prerequisite for further work.
> >
> > Signed-off-by: Alexei Starovoitov <ast@xxxxxxxxxx>
> > ---
> > include/linux/bpf.h | 2 +-
> > kernel/bpf/arena.c | 5 ++---
> > kernel/bpf/syscall.c | 23 ++++++++++++++++++++---
> > 3 files changed, 23 insertions(+), 7 deletions(-)
> >
> > diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> > index f3f50e29d639..e1838a341817 100644
> > --- a/include/linux/bpf.h
> > +++ b/include/linux/bpf.h
> > @@ -2348,7 +2348,7 @@ int generic_map_delete_batch(struct bpf_map *map,
> > struct bpf_map *bpf_map_get_curr_or_next(u32 *id);
> > struct bpf_prog *bpf_prog_get_curr_or_next(u32 *id);
> >
> > -int bpf_map_alloc_pages(const struct bpf_map *map, gfp_t gfp, int nid,
> > +int bpf_map_alloc_pages(const struct bpf_map *map, int nid,
> > unsigned long nr_pages, struct page **page_array);
> > #ifdef CONFIG_MEMCG
> > void *bpf_map_kmalloc_node(const struct bpf_map *map, size_t size, gfp_t flags,
> > diff --git a/kernel/bpf/arena.c b/kernel/bpf/arena.c
> > index 0975d7f22544..8ecc62e6b1a2 100644
> > --- a/kernel/bpf/arena.c
> > +++ b/kernel/bpf/arena.c
> > @@ -287,7 +287,7 @@ static vm_fault_t arena_vm_fault(struct vm_fault *vmf)
> > return VM_FAULT_SIGSEGV;
> >
> > /* Account into memcg of the process that created bpf_arena */
> > - ret = bpf_map_alloc_pages(map, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE, 1, &page);
> > + ret = bpf_map_alloc_pages(map, NUMA_NO_NODE, 1, &page);
> > if (ret) {
> > range_tree_set(&arena->rt, vmf->pgoff, 1);
> > return VM_FAULT_SIGSEGV;
> > @@ -465,8 +465,7 @@ static long arena_alloc_pages(struct bpf_arena *arena, long uaddr, long page_cnt
> > if (ret)
> > goto out_free_pages;
> >
> > - ret = bpf_map_alloc_pages(&arena->map, GFP_KERNEL | __GFP_ZERO,
> > - node_id, page_cnt, pages);
> > + ret = bpf_map_alloc_pages(&arena->map, node_id, page_cnt, pages);
> > if (ret)
> > goto out;
> >
> > diff --git a/kernel/bpf/syscall.c b/kernel/bpf/syscall.c
> > index c420edbfb7c8..a7af8d0185d0 100644
> > --- a/kernel/bpf/syscall.c
> > +++ b/kernel/bpf/syscall.c
> > @@ -569,7 +569,24 @@ static void bpf_map_release_memcg(struct bpf_map *map)
> > }
> > #endif
> >
> > -int bpf_map_alloc_pages(const struct bpf_map *map, gfp_t gfp, int nid,
> > +static bool can_alloc_pages(void)
> > +{
> > + return preempt_count() == 0 && !irqs_disabled() &&
> > + !IS_ENABLED(CONFIG_PREEMPT_RT);
> > +}
> > +
>
> I see this is new since v6 and wasn't yet discussed (or I missed it?)
It was in v1:
https://lore.kernel.org/bpf/20241116014854.55141-1-alexei.starovoitov@xxxxxxxxx/
See Peter's comments.
In this version I open coded preemptible(), since it's more accurate
and disabled the detection on PREEMPT_RT.
> I wonder how reliable these preempt/irq_disabled checks are for correctness
> purposes, e.g. we don't have CONFIG_PREEMPT_COUNT enabled always?
I believe the above doesn't produce false positives.
It's not exhaustive and might change as we learn more and tune it.
Hence I moved it to be bpf specific to iterate quickly instead of
being in inux/gfp.h and also considering Sebastian's comment
that normal kernel code should better know the calling context.
> As longs
> as the callers of bpf_map_alloc_pages() know the context and pass gfp
> accordingly, can't we use i.e. gfpflags_allow_blocking() to determine if
> try_alloc_pages() should be used or not?
bpf infra has a very coarse knowledge of the context.
There are two categories: sleepable or not.
In sleepable GFP_KERNEL is allowed, but it's very narrow and
represents a tiny slice of use cases compared to
non-sleepable. The try_alloc_pages() is for the latter.
netconsole has a similar problem/challenge.
It doesn't know the context where it will be called.
Currently it's just doing GFP_ATOMIC and praying.
This is something to fix eventually when slab is taught about
gfpflags_allow_blocking.
