Thread

pg18: leaks io-uring FDs during restart_after_crash

Justin Pryzby <pryzby@telsasoft.com> — 2025-10-17T20:29:53Z

If a backend crashes/killed due to OOM, then the FDs opened in the
postmaster aren't closed, so after a few crashes you can end up with a
multiple as many FDs opened.  Which can then lead to queries failing
with "too many opened files".

$ postgres -c io_method=io_uring &
$ lsof -p 3732231 2>/dev/null |grep -c io_uring
284
$ kill -9 3732232
$ lsof -p 3732231 2>/dev/null |grep -c io_uring
568

I'm not sure, but that might be related to these other errors.

error_severity         | ERROR
sql_state_code         | XX000
message                | could not read blocks 122..126 in file "base/246942/132857974": Operation canceled
query                  | SELECT DISTINCT attrelid FROM pg_attribute WHERE attacl IS NOT NULL
application_name       | pg_dump

-- 
Justin