On Mon, Oct 31, 2022 at 8:43 AM Nadav Amit <nadav.amit@xxxxxxxxx> wrote:
>
> On Oct 30, 2022, at 10:00 PM, Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> > So the current ordering rules are basically that we need to do
> > set_page_dirty() *and* we need to flush the TLB's before dropping the
> > page table lock. That's what gets us serialized with "mkclean”.
>
> I understand. I am still not sure whether ordering the set_page_dirty() and
> dropping the mapcount reference cannot suffice for the reclaim logic not to
> free the buffers if the page is dirtied.
Ahh, ok.
> According to the code, shrink_page_list() first checks for folio_mapped()
> and then for folio_test_dirty() to check whether pageout() is necessary.
> IIUC, the buffers are not dropped up to this point and set_page_dirty()
> would always set the page-struct dirty bit.
>
> IOW: In shrink_page_list(), when we decide on whether to pageout(), we
> should see whether the page is dirty (give or take smp_rmb()).
>
> But this is an optimization and I do not know all the cases in which buffers
> might be dropped. My intuition says that they cannot be dropped while
> mapcount != 0, but I need to further explore it.
Yes, the above sounds like one fairly good way out of the whole forced
TLB flushing for dirty pages, while still keeping the filesystem code
happy.
But at this point it's an independent issue.
And I really would like any fix like that to also fix the whole issue
with GUP too, which seems related.
Linus
