[PATCH v2 0/2] Follow-up on async memtx index cleanup

[Vladimir Davydov]

Turned out that after commit 2a1482f33fa0 ("memtx: free tuples
asynchronously when primary index is dropped"), we can't merge
1.10 to 2.0 without breaking tests, because one of the tests
(namely sql-tap/gh-3083-ephemeral-unref-tuples) executes a lot
of SQL SELECT statements without yielding. Those statements
in turn create a lot of ephemeral spaces that consume all
available memory since memtx garbage collection doesn't have
a chance to clean up. To fix that, this patch set improves
the garbage collection procedure: now it is invoked not only
by a background fiber, but also on demand, i.e. when any of
memtx functions fails to allocate memory.

https://github.com/tarantool/tarantool/issues/3408
https://github.com/tarantool/tarantool/commits/memtx-run-gc-on-demand

Changes in v2:
 - Patches 1-6 pushed with minor changes. Patches 7-8 => Patches 1-2.
 - Add gc task vtab to separate task destruction from gc callback.
 - Return gc task state in an argument.

v1: https://www.freelists.org/post/tarantool-patches/PATCH-08-Followup-on-async-memtx-index-cleanup

Vladimir Davydov (2):
  memtx: rework background garbage collection procedure
  memtx: run garbage collection on demand

 src/box/index.cc       |  7 ++++++-
 src/box/memtx_engine.c | 54 +++++++++++++++++++++++++++++++++++++++++++-------
 src/box/memtx_engine.h | 17 +++++++++++++---
 src/box/memtx_hash.c   | 35 ++++++++++++++++++++++----------
 src/box/memtx_hash.h   |  1 +
 src/box/memtx_tree.c   | 38 ++++++++++++++++++++++++-----------
 src/box/memtx_tree.h   |  1 +
 7 files changed, 119 insertions(+), 34 deletions(-)

-- 
2.11.0