We saw panic: page ... already unswappable on a downstream codebase. The page was wired. Another thread in physio() -> vm_fault_quick_hold_pages() -> vm_page_wire_mapped() had just wired it. Can we simply relax the assertion? Or should we test for a wired page (with appropriate synchronization)? This is trivially reproduced by adding vm_page_wire(mc[0]) at the top of vm_pageout_flush(), which seems to be a fair repro due to the lack of synchronization.
...due to the lack of synchronization with vm_fault_quick_hold_pages(), that is.
Yes, I agree that the assertion is incorrect and should simply be removed. It has been racy since we removed the use of the page hash lock to synchronize wiring of pages. We should assert that the page's object is locked, though. Just for some background on this queue, I added it to try and work around a problem on swapless systems where the page daemon would spend inordinate amounts of CPU in low memory situations trying to swap out anonymous pages, only to fail over and over. We can't simply not scan queues since we might have a large number of dirty filesystem pages which *can* be paged out.
Thanks as always, Mark. I'll test that and get back to you.
https://reviews.freebsd.org/D37320
A commit in branch main references this bug: URL: https://cgit.FreeBSD.org/src/commit/?id=cfbf1da0deede6b82ac197471b6cd439706a2dea commit cfbf1da0deede6b82ac197471b6cd439706a2dea Author: Anton Rang <rang@acm.org> AuthorDate: 2022-11-09 20:13:01 +0000 Commit: Eric van Gyzen <vangyzen@FreeBSD.org> CommitDate: 2022-11-09 20:28:03 +0000 vm_page_unswappable: remove wrong assertion markj says: ...the assertion is incorrect and should simply be removed. It has been racy since we removed the use of the page hash lock to synchronize wiring of pages. PR: 267621 Reviewed by: markj, Anton Rang <rang@acm.org> MFC after: 1 week Sponsored by: Dell Inc. Differential Revision: https://reviews.freebsd.org/D37320 sys/vm/vm_page.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
^Triage: FreeBSD 12 is now out of support. While this PR is open, do some QA to it.