[PATCH] Avoid pallocs in async.c's SignalBackends critical section - Mailing list pgsql-hackers

From Joel Jacobson
Subject [PATCH] Avoid pallocs in async.c's SignalBackends critical section
Date
Msg-id 5798d6fa-7fdc-49d4-9b4e-57767cd40198@app.fastmail.com
Whole thread Raw
Responses Re: [PATCH] Avoid pallocs in async.c's SignalBackends critical section
List pgsql-hackers
Hi hackers,

This patch addresses this comment in async.c's SignalBackends:

    * XXX in principle these pallocs could fail, which would be bad.
    * Maybe preallocate the arrays?  They're not that large, though.

This is unsafe, since AtCommit_Notify effectively runs in a critical
section, so an OOM there would PANIC ("AbortTransaction while in COMMIT
state"), as we can no longer abort safely.

This patch fixes this by adding two static arrays, notifySignalPids and
notifySignalProcs, allocated lazily in TopMemoryContext by
initSignalArrays. PreCommit_Notify now calls initSignalArrays while it's
still safe to ERROR, ensuring the arrays exist before entering the
commit path.

SignalBackends is updated to use these preallocated arrays instead of
allocating temporary ones.

This removes the risk of palloc in a critical section and eliminates
repeated palloc/pfree cycles.

/Joel
Attachment

pgsql-hackers by date:

Previous
From: Tomas Vondra
Date:
Subject: Re: should we have a fast-path planning for OLTP starjoins?
Next
From: Fujii Masao
Date:
Subject: Re: Exit walsender before confirming remote flush in logical replication