Re: Non-reproducible AIO failure
Konstantin Knizhnik <knizhnik@garret.ru>
From: Konstantin Knizhnik <knizhnik@garret.ru>
To: pgsql-hackers@lists.postgresql.org
Date: 2025-06-06T11:03:12Z
Lists: pgsql-hackers
Commits
Same data as JSON:
GET /api/v1/messages/:b64id/commits
the thread's linked commits as JSON, with link sources.
API reference →
-
aio: Stop using enum bitfields due to bad code generation
- ce161b194e84 18.0 landed
- 5865150b6d53 19 (unreleased) landed
-
amcheck: Fix posting tree checks in gin_index_check()
- 0cf205e122ae 18.0 cited
-
aio: Add missing memory barrier when waiting for IO handle
- e9a3615a5224 18.0 landed
There is really essential difference in code generated by clang 15
(working) and 16 (not working).
```
pgaio_io_stage(PgAioHandle *ioh, PgAioOp op)
{
...
HOLD_INTERRUPTS();
ioh->op = op;
ioh->result = 0;
pgaio_io_update_state(ioh, PGAIO_HS_DEFINED);
...
}
```
clang15:
```
postgres[0x10035d4ac] <+84>: add x22, x22, #0x5d4 ;
InterruptHoldoffCount
postgres[0x10035d4b0] <+88>: ldr w8, [x22]
postgres[0x10035d4b4] <+92>: add w8, w8, #0x1
postgres[0x10035d4b8] <+96>: str w8, [x22]
*postgres[0x10035d4bc] <+100>: strb w20, [x19, #0x2]
*postgres[0x10035d4c0] <+104>: str wzr, [x19, #0x14]
postgres[0x10035d4c4] <+108>: mov x0, x19
postgres[0x10035d4c8] <+112>: mov w1, #0x2
postgres[0x10035d4cc] <+116>: bl 0x10035cbdc ;
pgaio_io_update_state at aio.c:384
```
clang16:
```
postgres[0x100699e8c] <+200>: add x9, x9, #0xd74 ; InterruptHoldoffCount
postgres[0x100699e90] <+204>: ldr w8, [x9]
postgres[0x100699e94] <+208>: mov w10, #0x1 ; =1
postgres[0x100699e98] <+212>: stur w10, [x29, #-0x18]
postgres[0x100699e9c] <+216>: add w8, w8, #0x1
postgres[0x100699ea0] <+220>: str w8, [x9]
postgres[0x100699ea4] <+224>: ldur w10, [x29, #-0xc]
postgres[0x100699ea8] <+228>: ldur x9, [x29, #-0x8]
postgres[0x100699eac] <+232>: ldrh w8, [x9]
postgres[0x100699eb0] <+236>: and w10, w10, #0xff
postgres[0x100699eb4] <+240>: orr w10, w8, w10, lsl #16
postgres[0x100699eb8] <+244>: lsr w8, w10, #16
*postgres[0x100699ebc] <+248>: strh w10, [x9]
postgres[0x100699ec0] <+252>: strb w8, [x9, #0x2]
*postgres[0x100699ec4] <+256>: ldur x8, [x29, #-0x8]
postgres[0x100699ec8] <+260>: str wzr, [x8, #0x14]
postgres[0x100699ecc] <+264>: ldur x0, [x29, #-0x8]
postgres[0x100699ed0] <+268>: mov w1, #0x2 ; =2
postgres[0x100699ed4] <+272>: bl 0x100698efc ;
pgaio_io_update_state at aio.c:384
```
As you can see pg16 just updates third byte of the structure (`op`
field). It is not affecting 'state' field at all.
While clang16 is also updating `state` field! And looks like it is moved
before memory barrier (in `pgaio_io_update_state` function),