Re: Known and unfixed active data loss bug in MM + XFS with large folios since Dec 2021 (any kernel from 6.1 upwards)
From: Matthew Wilcox
Date: Mon Sep 30 2024 - 16:57:17 EST
On Mon, Sep 30, 2024 at 01:12:37PM -0700, Linus Torvalds wrote:
> It's basically been that way forever. The code has changed many times,
> but we've basically always had that "wait on bit will wait not until
> the next wakeup, but until it actually sees the bit being clear".
>
> And by "always" I mean "going back at least to before the git tree". I
> didn't search further. It's not new.
>
> The only reason I pointed at that (relatively recent) commit from 2021
> is that when we rewrote the page bit waiting logic (for some unrelated
> horrendous scalability issues with tens of thousands of pages on wait
> queues), the rewritten code _tried_ to not do it, and instead go "we
> were woken up by a bit clear op, so now we've waited enough".
>
> And that then caused problems as explained in that commit c2407cf7d22d
> ("mm: make wait_on_page_writeback() wait for multiple pending
> writebacks") because the wakeups aren't atomic wrt the actual bit
> setting/clearing/testing.
Could we break out if folio->mapping has changed? Clearly if it has,
we're no longer waiting for the folio we thought we were waiting for,
but for a folio which now belongs to a different file.
maybe this:
+void __folio_wait_writeback(struct address_space *mapping, struct folio *folio)
+{
+ while (folio_test_writeback(folio) && folio->mapping == mapping) {
+ trace_folio_wait_writeback(folio, mapping);
+ folio_wait_bit(folio, PG_writeback);
+ }
+}
[...]
void folio_wait_writeback(struct folio *folio)
{
- while (folio_test_writeback(folio)) {
- trace_folio_wait_writeback(folio, folio_mapping(folio));
- folio_wait_bit(folio, PG_writeback);
- }
+ __folio_wait_writeback(folio->mapping, folio);
}