Hi,

On Tue, 31 Mar 2026 at 13:19, Christoph Berg <[email protected]> wrote:
>
> One of the AIO commits around this one (2026-03-28 00:00Z) broke the
> postgresql-19 builds on apt.pg.o. Only the Debian unstable (sid,
> assertions enabled) builds are working, everything else including the
> nearly identical testing (forky) is broken (where assertions are off):

Thank you for the report! I can reproduce this when I disable
assertions. The problem is that, in StartReadBuffersImpl() in
bufmgr.c:

```
#ifdef USE_ASSERT_CHECKING

                /*
                 * Initialize enough of ReadBuffersOperation to make
                 * CheckReadBuffersOperation() work. Outside of assertions
                 * that's not necessary when no IO is issued.
                 */
                operation->buffers = buffers;
                operation->blocknum = blockNum;
                operation->nblocks = 1;
                operation->nblocks_done = 1;
                CheckReadBuffersOperation(operation, true);
#endif
```

if (found) and if (i == 0) then we set operation->buffers and
operation->nblocks if the assertions are enabled but AIO tests expect
to read these values. So, read_buffers() in AIO tests read incorrect
values [1] when the assertions are disabled. Moving operation->buffers
and operation->nblocks outside of #ifdef USE_ASSERT_CHECKING like in
the attached fixes the problem.

[1]
read_buffers(PG_FUNCTION_ARGS)
{
    ...
    for (int nio = 0; nio < nios; nio++)
    {
        ReadBuffersOperation *operation = &operations[nio];
        int            nblocks_this_io = operation->nblocks;
        Datum        values[6] = {0};
        bool        nulls[6] = {0};
        ArrayType  *buffers_arr;

        /* convert buffer array to datum array */
        for (int i = 0; i < nblocks_this_io; i++)
        {
            Buffer        buf = operation->buffers[i];
    ...
    }


-- 
Regards,
Nazir Bilal Yavuz
Microsoft
From 56218fc850cdaccaeb87767c2ca8977ee2f26697 Mon Sep 17 00:00:00 2001
From: Nazir Bilal Yavuz <[email protected]>
Date: Tue, 31 Mar 2026 15:00:40 +0300
Subject: [PATCH] Set the variables that callers might read in
 StartReadBuffersImpl()

When the first requested block is already in the buffer pool, StartRead-
BuffersImpl() returned early after setting *nblocks = 1 without starting
any I/O. In that path, operation->nblocks and operation->buffers were
only initialised inside #ifdef USE_ASSERT_CHECKING, leaving them as zero
/ NULL in non-assert builds.

Callers might want to inspect these values like in the AIO tests. Fix
that problem by unconditionally setting operation->nblocks and
operation->buffers in the early-return hit path.
---
 src/backend/storage/buffer/bufmgr.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index cd21ae3fc36..e6f99ad7668 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -1457,6 +1457,10 @@ StartReadBuffersImpl(ReadBuffersOperation *operation,
 			{
 				*nblocks = 1;
 
+				/* Set the variables that callers might read */
+				operation->buffers = buffers;
+				operation->nblocks = 1;
+
 #ifdef USE_ASSERT_CHECKING
 
 				/*
@@ -1464,9 +1468,7 @@ StartReadBuffersImpl(ReadBuffersOperation *operation,
 				 * CheckReadBuffersOperation() work. Outside of assertions
 				 * that's not necessary when no IO is issued.
 				 */
-				operation->buffers = buffers;
 				operation->blocknum = blockNum;
-				operation->nblocks = 1;
 				operation->nblocks_done = 1;
 				CheckReadBuffersOperation(operation, true);
 #endif
-- 
2.47.3

Reply via email to