This series is a continuation of the conversion of the i_mmap_mutex to
rwsem, following what we have for the anon memory counterpart. With
Hugh's feedback from the first iteration.
Ultimately, the most obvious paths that require exclusive ownership of the
lock is when we modify the VMA interval tree, via
vma_interval_tree_insert() and vma_interval_tree_remove() families. Cases
such as unmapping, where the ptes content is changed but the tree remains
untouched should make it safe to share the i_mmap_rwsem.
As such, the code of course is straightforward, however the devil is very
much in the details. While its been tested on a number of workloads
without anything exploding, I would not be surprised if there are some
less documented/known assumptions about the lock that could suffer from
these changes. Or maybe I'm just missing something, but either way I
believe its at the point where it could use more eyes and hopefully some
time in linux-next.
Because the lock type conversion is the heart of this patchset,
its worth noting a few comparisons between mutex vs rwsem (xadd):
(i) Same size, no extra footprint.
(ii) Both have CONFIG_XXX_SPIN_ON_OWNER capabilities for
exclusive lock ownership.
(iii) Both can be slightly unfair wrt exclusive ownership, with
writer lock stealing properties, not necessarily respecting
FIFO order for granting the lock when contended.
(iv) Mutexes can be slightly faster than rwsems when
the lock is non-contended.
(v) Both suck at performance for debug (slowpaths), which
shouldn't matter anyway.
Sharing the lock is obviously beneficial, and sem writer ownership is
close enough to mutexes. The biggest winner of these changes is
migration.
As for concrete numbers, the following performance results are for a
4-socket 60-core IvyBridge-EX with 130Gb of RAM.
Both alltests and disk (xfs+ramdisk) workloads of aim7 suite do quite well
with this set, with a steady ~60% throughput (jpm) increase for alltests
and up to ~30% for disk for high amounts of concurrency. Lower counts of
workload users (< 100) does not show much difference at all, so at least
no regressions.
In addition, a 67.5% increase in successfully migrated NUMA pages, thus
improving node locality.
The patch layout is simple but designed for bisection (in case reversion
is needed if the changes break upstream) and easier review:
o Patches 1-4 convert the i_mmap lock from mutex to rwsem.
o Patches 5-10 share the lock in specific paths, each patch
details the rationale behind why it should be safe.
This patchset has been tested with: postgres 9.4 (with brand new hugetlb
support), hugetlbfs test suite (all tests pass, in fact more tests pass
with these changes than with an upstream kernel), ltp, aim7 benchmarks,
memcached and iozone with the -B option for mmap'ing. *Untested* paths
are nommu, memory-failure, uprobes and xip.
This patch (of 8):
Various parts of the kernel acquire and release this mutex, so add
i_mmap_lock_write() and immap_unlock_write() helper functions that will
encapsulate this logic. The next patch will make use of these.
projects / mirror_ubuntu-disco-kernel.git / commit