Re: LogwrtResult contended spinlock

On Sat, Apr 6, 2024 at 6:55 AM Alvaro Herrera <alvherre(at)alvh(dot)no-ip(dot)org> wrote:
> Pushed 0001.

Could that be related to the 3 failures on parula that look like this?

TRAP: failed Assert("node->next == 0 && node->prev == 0"), File:
"../../../../src/include/storage/proclist.h", Line: 63, PID: 29119
2024-04-05 16:16:26.812 UTC [29114:15] pg_regress/drop_operator LOG:
statement: DROP OPERATOR <|(bigint, bigint);
postgres: postgres regression [local] CREATE
ROLE(ExceptionalCondition+0x4c)[0x9b3fdc]
postgres: postgres regression [local] CREATE ROLE[0x8529e4]
postgres: postgres regression [local] CREATE
ROLE(LWLockWaitForVar+0xec)[0x8538fc]
postgres: postgres regression [local] CREATE ROLE[0x54c7d4]
postgres: postgres regression [local] CREATE ROLE(XLogFlush+0xf0)[0x552600]
postgres: postgres regression [local] CREATE ROLE[0x54a9b0]
postgres: postgres regression [local] CREATE ROLE[0x54bbdc]

Hmm, the comments for LWLockWaitForVar say:

* Be aware that LWLockConflictsWithVar() does not include a memory barrier,
* hence the caller of this function may want to rely on an explicit barrier or
* an implied barrier via spinlock or LWLock to avoid memory ordering issues.

But that seems to be more likely to make LWLockWaitForVar suffer data
races (ie hang), not break assertions about LWLock sanity, so I don't
know what's going on there. I happened to have a shell on a Graviton
box, but I couldn't reproduce it after a while...