[Qemu-devel] [PATCH v6 00/15] tb hash improvements

["Emilio G. Cota" <cota@braap.org>]

v5: https://lists.gnu.org/archive/html/qemu-devel/2016-05/msg02366.html

v6 applies cleanly on top of tcg-next (8b1fe3f4 "cpu-exec:
Clean up 'interrupt_request' reloading", tagged "pull-tcg-20160512").

Changes from v5, mostly from Sergey's review:

- processor.h: use #ifdef #elif throughout the file

- tb_hash_func: use uint32 for 'flags' param

- tb_hash_func5: do 'foo >> 32' instead of 'foo >> 31 >> 1', since foo
  is a u64.

- thread.h:
  * qemu_spin_locked: remove acquire semantics; simply use atomic_read().
  * qemu_spin_trylock: return bool instead of 0 or -EBUSY; this saves
    a branch.
  * qemu_spin:
    + use __sync_lock_test_and_set and __sync_lock_release; drop
      the patches touching atomic.h.
    + add unlikely() hint to "while (test_and_set)"; this gives a small
      speedup under no contention.

- qht:
  * merge the parallel-writes patch into the QHT patch.
    [Richard: I dropped your reviewed-by since the patch changed
     quite a bit.]
  * drop unneeded #includes from qht.h
  * document qht.h using kerneldoc.
  * use unsigned int for storing the seqlock version.
  * fix a couple of typos in the comments at the top of qht.c.
  * explain better the "no duplicated pointer" policy: while trying to
    insert an already-existing hash-pointer pair is OK (insert will
    just return false), it's not OK to insert different hash-pointer
    pairs that share the same pointer value, but not the hashes.
  * Add comment about lookups having to be done in an RCU read-critical
    section.
  * remove map->stale; simply check ht->map before and after acquiring
    a bucket lock.
  * only use atomic_read/set on bucket pointers, not hashes. Reading
    partially-updated hashes is OK, since we'll retry anyway thanks
    to the seqlock. Add a comment regarding this at the top of struct
    qht_bucket.
  * s/b->n/b->n_buckets/
  * define qht_debug_assert, enabled #ifdef QHT_DEBUG. Use it instead of
    assert(), except in one case (slow path) where g_assert_cmpuint is
    convenient.
  * use a mutex for ht->lock instead of a spinlock. This makes the resize
    code simpler, since holding ht->lock for a bit of time is OK now;
    other threads won't be busy-waiting. Document that ht->lock needs
    to be grabbed before b->lock.
  * use atomic_rcu_read/set instead of open-coding them.
  * qht_remove: only clear out b->hashes[] and b->pointers[] if they belong
                to what was the last entry in the chain.
  * qht_remove: add debug assert against inserting a NULL pointer.

Thanks,

		Emilio