[PATCH v4 0/3] mempool: fix races between list walk and create/free

[Robert Sanford <rsanford2@gmail.com>]

From: Robert Sanford <rsanford@akamai.com>

A thread walking the mempool list with rte_mempool_walk(),
rte_mempool_lookup() or rte_mempool_list_dump() can crash while other
threads create and free mempools. We hit this in production, in a
process that walks the mempools while other threads create and free
them.

There are two races:

- rte_mempool_free() unlinks and frees the list entry without the
  mempool list lock that the walkers hold, so a walker can reach a
  freed entry or mempool. Patch 1 takes the lock in rte_mempool_free(),
  as rte_mempool_create_empty() already does. A rte_mempool_walk()
  callback can therefore no longer free a mempool; this is documented.

- A mempool is on the list before its store exists, but
  rte_mempool_dump() queries the mempool driver, which reads the store.
  So rte_mempool_list_dump() and the telemetry /mempool/info handler
  crash on such a mempool. Patch 2 adds rte_mempool_is_created(),
  makes them report the mempool without querying the driver until it
  returns true, and documents that rte_mempool_walk() callbacks must
  check it before querying the driver.

Patch 3 adds a test to mempool_autotest. On x86 with 4 or 8 lcores,
it crashed in all 20 runs without patch 1, without patch 2, or
without both, and passed in all 20 runs with both patches.

v4:
- Patch 3, as suggested by Morten Brørup: the dump goes to "NUL" instead
  of "/dev/null" on Windows. Acked-by added.
- Rebased on main.
- Patches 1 and 2: unchanged.

v3:
- Patch 2, as suggested by Morten Brørup: the helper that tests
  RTE_MEMPOOL_F_POOL_CREATED is now the public function
  rte_mempool_is_created(), and the rte_mempool_walk() documentation
  refers to it instead of the flag. Release note added. Acked-by added.
- Patch 3: the walk callback uses rte_mempool_is_created().
- Patch 1: unchanged.

v2:
- Patch 2 reworked as suggested by Morten Brørup: rte_mempool_walk()
  no longer skips mempools that are not populated yet. Instead,
  rte_mempool_dump() and the telemetry callback report them without
  querying the mempool driver, and the rte_mempool_walk() documentation
  asks callbacks to check RTE_MEMPOOL_F_POOL_CREATED. The flag is now
  set and tested with release/acquire atomics instead of plain accesses
  with fences. New title.
- Patch 3: the walk callback dumps every mempool, and checks the flag
  before querying the driver. The test waits 100 us instead of 10 us
  before populating, so that it still finds unpopulated mempools
  reliably.
- Patch 1: unchanged, Reviewed-by added.

Robert Sanford (3):
  mempool: fix race between list walk and free
  mempool: fix dump of unpopulated pool
  test/mempool: add list walk race test

 app/test/test_mempool.c                | 150 +++++++++++++++++++++++++
 doc/guides/rel_notes/release_26_11.rst |   7 ++
 lib/mempool/rte_mempool.c              |  45 +++++++-
 lib/mempool/rte_mempool.h              |  29 +++++
 4 files changed, 227 insertions(+), 4 deletions(-)

-- 
2.34.1