On Thu, Jan 8, 2026 at 12:29 PM Martin KaFai Lau [off-list ref] wrote:

On 12/18/25 9:56 AM, Amery Hung wrote:

quoted

To prepare for changing bpf_local_storage_map_bucket::lock to rqspinlock,
convert bpf_selem_unlink_map() to failable. It still always succeeds and
returns 0 for now.

Since some operations updating local storage cannot fail in the middle,
open-code bpf_selem_unlink_map() to take the b->lock before the
operation. There are two such locations:

- bpf_local_storage_alloc()

   The first selem will be unlinked from smap if cmpxchg owner_storage_ptr
   fails, which should not fail. Therefore, hold b->lock when linking
   until allocation complete. Helpers that assume b->lock is held by
   callers are introduced: bpf_selem_link_map_nolock() and
   bpf_selem_unlink_map_nolock().

- bpf_local_storage_update()

   The three step update process: link_map(new_selem),
   link_storage(new_selem), and unlink_map(old_selem) should not fail in
   the middle.

In bpf_selem_unlink(), bpf_selem_unlink_map() and
bpf_selem_unlink_storage() should either all succeed or fail as a whole
instead of failing in the middle. So, return if unlink_map() failed.

In bpf_local_storage_destroy(), since it cannot deadlock with itself or
bpf_local_storage_map_free() who the function might be racing with,
retry if bpf_selem_unlink_map() fails due to rqspinlock returning
errors.

Signed-off-by: Amery Hung <redacted>
---
  kernel/bpf/bpf_local_storage.c | 64 +++++++++++++++++++++++++++++-----
  1 file changed, 55 insertions(+), 9 deletions(-)

diff --git a/kernel/bpf/bpf_local_storage.c b/kernel/bpf/bpf_local_storage.c
index e2fe6c32822b..4e3f227fd634 100644
--- a/kernel/bpf/bpf_local_storage.c
+++ b/kernel/bpf/bpf_local_storage.c

@@ -347,7 +347,7 @@ void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
      hlist_add_head_rcu(&selem->snode, &local_storage->list);
  }

-static void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
+static int bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)

This will end up only be used by bpf_selem_unlink(). It may as well
remove this function and open code in the bpf_selem_unlink(). I think it
may depend on how patch 10 goes and also if it makes sense to remove
bpf_selem_"link"_map and bpf_selem_unlink_map_nolock also, so treat it
as a nit note for now.

Noted

quoted

  {
      struct bpf_local_storage_map *smap;
      struct bpf_local_storage_map_bucket *b;

@@ -355,7 +355,7 @@ static void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)

      if (unlikely(!selem_linked_to_map_lockless(selem)))

In the later patch where both local_storage's and map-bucket's locks
must be acquired, will this check still be needed if there is an earlier
check that ensures the selem is still linked to the local_storage? It
does not matter in terms of perf, but I think it will help code reading
in the future for the common code path (i.e. the code paths other than
bpf_local_storage_destroy and bpf_local_storage_map_free).

Makes sense to remove it. Common code path still follow the unlink
order and do not partial unlink.

quoted

              /* selem has already be unlinked from smap */
-             return;
+             return 0;

      smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
      b = select_bucket(smap, selem);

@@ -363,6 +363,14 @@ static void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
      if (likely(selem_linked_to_map(selem)))
              hlist_del_init_rcu(&selem->map_node);
      raw_spin_unlock_irqrestore(&b->lock, flags);
+
+     return 0;
+}
+
+static void bpf_selem_unlink_map_nolock(struct bpf_local_storage_elem *selem)
+{
+     if (likely(selem_linked_to_map(selem)))

Take this chance to remove the selem_linked_to_map() check.
hlist_del_init_rcu has the same check.

Noted

quoted

+             hlist_del_init_rcu(&selem->map_node);
  }

  void bpf_selem_link_map(struct bpf_local_storage_map *smap,

@@ -376,13 +384,26 @@ void bpf_selem_link_map(struct bpf_local_storage_map *smap,
      raw_spin_unlock_irqrestore(&b->lock, flags);
  }

+static void bpf_selem_link_map_nolock(struct bpf_local_storage_map *smap,
+                                   struct bpf_local_storage_elem *selem,
+                                   struct bpf_local_storage_map_bucket *b)
+{
+     RCU_INIT_POINTER(SDATA(selem)->smap, smap);

Is it needed? bpf_selem_alloc should have init the SDATA(selem)->smap.

Good catch. Forgot to remove it when rebasing. This is redundant as we
already do it in bpf_selem_alloc()

quoted

+     hlist_add_head_rcu(&selem->map_node, &b->list);
+}
+

[ ... ]

quoted

@@ -574,20 +603,37 @@ bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
              goto unlock;
      }

+     b = select_bucket(smap, selem);
+
+     if (old_sdata) {
+             old_b = select_bucket(smap, SELEM(old_sdata));
+             old_b = old_b == b ? NULL : old_b;
+     }
+
+     raw_spin_lock_irqsave(&b->lock, b_flags);
+
+     if (old_b)
+             raw_spin_lock_irqsave(&old_b->lock, old_b_flags);

This will deadlock because of the lock ordering of b and old_b.
Replacing it with res_spin_lock in the later patch can detect it and
break it more gracefully. imo, we should not introduce a known deadlock
logic in the kernel code in the syscall code path and ask the current
user to retry the map_update_elem syscall.

What happened to the patch in the earlier revision that uses the
local_storage (or owner) for select_bucket?

Thanks for reviewing!

I decided to revert it because this introduces the dependency of selem
to local_storage when unlinking. bpf_selem_unlink_lockless() cannot
assume map or local_storage associated with a selem to be alive. In
the case where local_storage is already destroyed, we won't be able to
figure out the bucket if select_bucket() uses local_storage for
hashing.

A middle ground is to use local_storage for hashing, but save the
bucket index in selem so that local_storage pointer won't be needed
later. WDYT?

[ will continue with the rest of the patches a bit later ]

quoted

+
      alloc_selem = NULL;
      /* First, link the new selem to the map */
-     bpf_selem_link_map(smap, selem);
+     bpf_selem_link_map_nolock(smap, selem, b);

      /* Second, link (and publish) the new selem to local_storage */
      bpf_selem_link_storage_nolock(local_storage, selem);

      /* Third, remove old selem, SELEM(old_sdata) */
      if (old_sdata) {
-             bpf_selem_unlink_map(SELEM(old_sdata));
+             bpf_selem_unlink_map_nolock(SELEM(old_sdata));
              bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
                                              &old_selem_free_list);
      }

+     if (old_b)
+             raw_spin_unlock_irqrestore(&old_b->lock, old_b_flags);
+
+     raw_spin_unlock_irqrestore(&b->lock, b_flags);
+
  unlock:
      raw_spin_unlock_irqrestore(&local_storage->lock, flags);
      bpf_selem_free_list(&old_selem_free_list, false);

@@ -679,7 +725,7 @@ void bpf_local_storage_destroy(struct bpf_local_storage *local_storage)
              /* Always unlink from map before unlinking from
               * local_storage.
               */
-             bpf_selem_unlink_map(selem);
+             WARN_ON(bpf_selem_unlink_map(selem));
              /* If local_storage list has only one element, the
               * bpf_selem_unlink_storage_nolock() will return true.
               * Otherwise, it will return false. The current loop iteration