Hi,
I’ve been trying to review this patch, and it struck me that we’re
currently grabbing the content lock exclusively just to flip a header bit:
if (!(buf_state & BM_DIRTY))
{
LWLockAcquire(BufferDescriptorGetContentLock(desc), LW_EXCLUSIVE);
MarkBufferDirty(buf);
LWLockRelease(BufferDescriptorGetContentLock(desc));
}
Since our sole goal here is to simulate a buffer’s dirty state for
testing/debugging, I wonder—could we instead:
1. Acquire the shared content lock (which already blocks eviction),
2. Call MarkBufferDirtyHint() to flip the BM_DIRTY bit under the header
spinlock, and
3. Release the shared lock?
This seems to satisfy Assert(LWLockHeldByMe(...)) inside the hint routine
and would:
- Prevent exclusive‐lock contention when sweeping many buffers,
- Avoid full‐page WAL writes (we only need a hint record, if any)
Would love to hear if this makes sense or or am I overlooking something
here. Thanks for any feedback!
Cheers,
Xuneng
Nazir Bilal Yavuz <byavu...@gmail.com> 于2025年4月11日周五 17:14写道：
> Hi,
>
> There is another thread [1] to add both pg_buffercache_evict_[relation
> | all] and pg_buffercache_mark_dirty[_all] functions to the
> pg_buffercache. I decided to create another thread as
> pg_buffercache_evict_[relation | all] functions are committed but
> pg_buffercache_mark_dirty[_all] functions still need review.
>
> pg_buffercache_mark_dirty(): This function takes a buffer id as an
> argument and tries to mark this buffer as dirty. Returns true on
> success.
> pg_buffercache_mark_dirty_all(): This is very similar to the
> pg_buffercache_mark_dirty() function. The difference is
> pg_buffercache_mark_dirty_all() does not take an argument. Instead it
> just loops over the shared buffers and tries to mark all of them as
> dirty. It returns the number of buffers marked as dirty.
>
> Since that patch is targeted for the PG 19, pg_buffercache is bumped to
> v1.7.
>
> Latest version is attached and people who already reviewed the patches are
> CCed.
>
> [1]
> postgr.es/m/CAN55FZ0h_YoSqqutxV6DES1RW8ig6wcA8CR9rJk358YRMxZFmw%40mail.gmail.com
>
> --
> Regards,
> Nazir Bilal Yavuz
> Microsoft
>
