On Wed, Apr 03, 2024 at 10:06:16PM +0000, yueyang.pan@xxxxxxx wrote:
> Dear Matthew,
> I am Yueyang Pan a PhD student from EPFL, and I am currently
> checking the swap code in the kernel. Sorry to bother you because this
> email should go to Mel Gorman. He did not reply to me so I turned to
> you for some help.
Hi Yueyang,
You'd probably have more luck if you cc'd the mailing list. Somebody
other than Mel might have answered you. Added it now.
> 1) I have some questions about `try_to_unmap_flush_dirty`. I
> wonder why this is necessary in shrink_folio_list because in
> the folio_check_references, we have already checked that the PTEs
> pointing to this page does not have any access bit set. The current
> shrink_folio_list then unmaps the page, clears the dirty bit, issues
> the TLB flush if the dirty bit was set previously and then starts
> to write the page to the swap. I wonder why here we cannot take an
> opportunistic approach. My understanding is that if we don’t unmap the
> page and perform flush, when there is a concurrent write to the page,
> both the access bit and dirty bit will be set (because the dirty bit
> is cleared) so we can simply check the access bit again after pageout
> to see whether we can free this page or not.
> I checked the git blame and saw Mel's commit in 2015 where he
> mentioned that it was better to assume a writeable entry exist
> in TLB but I wonder why this can be true if we have already use
> folio_check_references to check the PTE access bit. Does this imply
> even if the folio_check_references gives no reference there can be
> still entries in the TLB?
This is far outside my realm of expertise. I suspect it's
possible that there can be stale entries in the TLB if
ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH.
> 2) I have some questions at the end of shrink_folio_list before the
> ref_folios are spliced again back to folio_list. What if at the same
> time, there is another function trying to free the page or mlock
> the page? Will this page be circulated again in the inactive LRU
> list and being double freed since the page lock was released at the
> list_splice? Because from what I understood, the mlock will take the
> page directly from the inactive/active list the page is in and move
> the page to the mlock list but at this moment the page does not reside
> in either of the list.
I think the answer is that these folios have their LRU flag cleared
throughout shrink_folio_list() so they cannot be mlocked?
static struct lruvec *__mlock_folio(struct folio *folio, struct lruvec *lruvec)
{
/* There is nothing more we can do while it's off LRU */
if (!folio_test_clear_lru(folio))
return lruvec;
