Hi,
On 2025-02-11 16:18:37 -0500, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > And when using something like io_uring for AIO, it'd allow to
> > max_files_per_process in addition to the files requires for the io_uring
> > instances.
>
> Not following? Surely we'd not be configuring that so early in
> postmaster start?
The issue is that, with io_uring, we need to create one FD for each possible
child process, so that one backend can wait for completions for IO issued by
another backend [1]. Those io_uring instances need to be created in
postmaster, so they're visible to each backend. Obviously that helps to much
more quickly run into an unadjusted soft RLIMIT_NOFILE, particularly if
max_connections is set to a higher value.
In the current version of the AIO patchset, the creation of those io_uring
instances does happen as part of an shmem init callback, as the io uring
creation also sets up queues visible in shmem. And shmem init callbacks are
currently happening *before* postmaster's set_max_safe_fds() call:
/*
* Set up shared memory and semaphores.
*
* Note: if using SysV shmem and/or semas, each postmaster startup will
* normally choose the same IPC keys. This helps ensure that we will
* clean up dead IPC objects if the postmaster crashes and is restarted.
*/
CreateSharedMemoryAndSemaphores();
/*
* Estimate number of openable files. This must happen after setting up
* semaphores, because on some platforms semaphores count as open files.
*/
set_max_safe_fds();
So the issue would actually be that we're currently doing set_max_safe_fds()
too late, not too early :/
Greetings,
Andres Freund
[1] Initially I tried to avoid that, by sharing a smaller number of io_uring
instances across backends. Making that work was a fair bit of code *and*
was considerably slower, due to now needing a lock around submission of
IOs. Moving to one io_uring instance per backend fairly dramatically
simplified the code while also speeding it up.
