[PATCH bpf-next v7 0/3] bpf: Implement stack_map_get_build_id_offset_sleepable()

[Ihor Solodrai]

The series introduces stack_map_get_build_id_offset_sleepable(),
fixing a gap with parsing build_id in sleepable context in stackmap.c

In particular, this fixes a deadlock in
stack_map_get_build_id_offset() doing a blocking __kernel_read(),
which happens since commit 777a8560fd29 ("lib/buildid: use
__kernel_read() for sleepable context").

See previous revisions for more details.

---

v6->v7:
  * Addressed feedback from Andrii (mostly patch #2):
    * implement proper CONFIG_PER_VMA_LOCK=n support, following a
      VMA locking pattern similar to one used in PROCMAP_QUERY
    * change the contract of stack_map_lock_vma(): if a non-NULL VMA
      is returned, then a read lock is held
      * remove now unnecessary vma_locked flag
    * and various other nits
  * Add vma_is_anonymous() checks where appropriate (AIs)
v6: https://lore.kernel.org/bpf/20260521225022.2695755-1-ihor.solodrai@xxxxxxxxx/

v5->v6:
  * Misc refactoring (Andrii):
    * add stack_map_build_id_set_valid() helper
    * simplify control flow in stack_map_get_build_id_offset_sleepable()
v5: https://lore.kernel.org/bpf/20260515005244.1333013-1-ihor.solodrai@xxxxxxxxx/

v4->v5:
  * Add comments explaining mmap_read_trylock() (Shakeel)
  * Rebase on bpf-next (Alexei)
v4: https://lore.kernel.org/bpf/20260514184727.1067141-1-ihor.solodrai@xxxxxxxxx/

v3->v4:
  * Change Fixes tag in patch #2 (AI)
  * Nit in caching implementation (Mykyta)
v3: https://lore.kernel.org/bpf/20260512032906.2670326-1-ihor.solodrai@xxxxxxxxx/

v2->v3:
  * Split patch #2 in two: stack_map_get_build_id_offset_sleepable()
    implementation, and then introduce caching
  * Drop taking mmap_lock if CONFIG_PER_VMA_LOCK=n, fall back to raw
    IPs instead
  * Cache vm_{start,end} in addition to prev_file (Mykyta)
v2: https://lore.kernel.org/bpf/20260409010604.1439087-1-ihor.solodrai@xxxxxxxxx/

v1->v2:
  * Addressed feedback from Puranjay:
    * split out a small refactoring patch
    * use mmap_read_trylock()
    * take into account CONFIG_PER_VMA_LOCK
    * replace find_vma() with vma_lookup()
    * cache prev_build_id to avoid re-parsing the same file
  * Snapshot vm_pgoff and vm_start before unlocking (AI)
  * To avoid repetitive unlocking statements, introduce struct
    stack_map_vma_lock to hold relevant lock state info and add an
    unlock helper
v1: https://lore.kernel.org/bpf/20260407223003.720428-1-ihor.solodrai@xxxxxxxxx/

---


Ihor Solodrai (3):
  bpf: Factor out stack_map build ID helpers
  bpf: Avoid faultable build ID reads under mm locks
  bpf: Cache build IDs in sleepable stackmap path

 kernel/bpf/stackmap.c | 215 ++++++++++++++++++++++++++++++++++++++----
 1 file changed, 198 insertions(+), 17 deletions(-)

-- 
2.54.0