Re: could not extend file "base/5/3501" with FileFallocate(): Interrupted system call

Hi,

On 2023-04-24 10:53:35 +0200, Christoph Berg wrote:
> I'm often seeing PG16 builds erroring out in the pgbench tests:

Interesting!

> I don't think the disk is full since it's always hitting that same
> spot, on some of the builds:

Yea, the EINTR pretty clearly indicates that it's not really out-of-space.

> https://pgdgbuild.dus.dg-i.net/job/postgresql-16-binaries-snapshot/833/
>
> This is overlayfs with tmpfs (upper)/ext4 (lower). Manually running
> that test works though, and the FS seems to support posix_fallocate:

I guess it requires a bunch of memory (?) pressure for this to happen
(triggering blocking during fallocate, opening the window for a signal to
arrive), which likely only happens when running things concurrently.

We obviously can add a retry loop to FileFallocate(), similar to what's
already present e.g. in FileRead(). But I wonder if we shouldn't go a bit
further, and do it for all the fd.c routines where it's remotely plausible
EINTR could be returned? It's a bit silly to add EINTR retries one-by-one to
the functions.

The following are documented to potentially return EINTR, without fd.c having
code to retry:

- FileWriteback() / pg_flush_data()
- FileSync() / pg_fsync()
- FileFallocate()
- FileTruncate()

With the first two there's the added complication that it's not entirely
obvious whether it'd be better to handle this in File* or pg_*. I'd argue the
latter is a bit more sensible?

Greetings,

Andres Freund