As discussed on the thread "Spinlocks and compiler/memory barriers",
now that we've made the spinlock primitives function as compiler
barriers (we think), it should be possible to remove volatile
qualifiers from many places in the source code. The attached patch
does this in lwlock.c. If the changes in commit
0709b7ee72e4bc71ad07b7120acd117265ab51d0 (and follow-on commits) are
correct and complete, applying this shouldn't break anything, while
possibly giving the compiler room to optimize things better than it
does today.
However, demonstrating the necessity of that commit for these changes
seems to be non-trivial. I tried applying this patch and reverting
commits 5b26278822c69dd76ef89fd50ecc7cdba9c3f035,
b4c28d1b92c81941e4fc124884e51a7c110316bf, and
0709b7ee72e4bc71ad07b7120acd117265ab51d0 on a PPC64 POWER8 box with a
whopping 192 hardware threads (thanks, IBM!). I then ran the
regression tests repeatedly, and I ran several long pgbench runs with
as many as 350 concurrent clients. No failures.
So I'm posting this patch in the hope that others can help. The
relevant tests are:
1. If you apply this patch to master and run tests of whatever kind
strikes your fancy, does anything break under high concurrency? If it
does, then the above commits weren't enough to make this safe on your
platform.
2. If you apply this patch to master, revert the commits mentioned
above, and again run tests, does anything now break? If it does (but
the first tests were OK), then that shows that those commits did
something useful on your platform.
The change to make spinlocks act as compiler barriers provides the
foundation for, hopefully, a cleaner code base and easier future work
on scalabilty and performance projects, so help is much appreciated.
Thanks,
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company