Re: Online checksums verification in the backend

Hi,

On 2020-10-29 11:17:29 -0700, Andres Freund wrote:
> The code does IO while holding the buffer mapping lock. That seems
> *entirely* unacceptable to me. That basically locks 1/128 of shared
> buffers against concurrent mapping changes, while reading data that is
> likely not to be on disk? Seriously?

Also, uh, I don't think the locking of the buffer table provides you
with the full guarantees CheckBuffer() seems to assume:

* Check the state of a buffer without loading it into the shared buffers. To
* avoid torn pages and possible false positives when reading data, a shared
* LWLock is taken on the target buffer pool partition mapping, and we check
* if the page is in shared buffers or not. An I/O lock is taken on the block
* to prevent any concurrent activity from happening.

this doesn't actually prevent all concurrent write IO, unless you hold
an appropriate lock on the relation. There's a few places that use
smgrwrite()/smgrextend() to write out data bypassing shared buffers.

Maybe that isn't a problem for the uses of CheckBuffer() is envisioned
for, but that'd need a pretty detailed explanation as to when it's safe
to use CheckBuffer() for which blocks.

Greetings,

Andres Freund