Re: [PATCH v2 1/1] mm/madvise: enhance lazyfreeing with mTHP in madvise_free

From: David Hildenbrand
Date: Thu Mar 07 2024 - 03:32:29 EST


On 07.03.24 09:10, Barry Song wrote:
On Thu, Mar 7, 2024 at 9:00 PM Lance Yang <ioworker0@xxxxxxxxx> wrote:

Hey Barry,

Thanks for taking time to review!

On Thu, Mar 7, 2024 at 3:00 PM Barry Song <21cnbao@xxxxxxxxx> wrote:

On Thu, Mar 7, 2024 at 7:15 PM Lance Yang <ioworker0@xxxxxxxxx> wrote:

[...]
+static inline bool can_mark_large_folio_lazyfree(unsigned long addr,
+ struct folio *folio, pte_t *start_pte)
+{
+ int nr_pages = folio_nr_pages(folio);
+ fpb_t flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
+
+ for (int i = 0; i < nr_pages; i++)
+ if (page_mapcount(folio_page(folio, i)) != 1)

No new direct subpage mapcount users please if avoidable. Also, without holding the page lock this might be racy (did not stare at the bigger code).

+ return false;

we have moved to folio_estimated_sharers though it is not precise, so
we don't do
this check with lots of loops and depending on the subpage's mapcount.

If we don't check the subpage’s mapcount, and there is a cow folio associated
with this folio and the cow folio has smaller size than this folio,
should we still
mark this folio as lazyfree?

I agree, this is true. However, we've somehow accepted the fact that
folio_likely_mapped_shared
can result in false negatives or false positives to balance the
overhead. So I really don't know :-)

Maybe David and Vishal can give some comments here.

Well, I did not accept the fact yet that folio_likely_mapped_shared() can give false negatives :)

So I've been working on a replacement [1], also in the context of removing the subpage mapcounts. But there is still a lot to resolve around that, first step being to introduce the total mapcount.

But independent of that, MADV_FREE is special (compared to MADV_DONTNEED), because we really have to make sure all page table mappings that map the folio will be covered here.

What could work for large folios is making sure that #ptes that map the folio here correspond to the folio_mapcount(). And folio_mapcount() should be called under folio lock, to avoid racing with swapout/migration.

So instead of checking each page_mapcount(), see how many PTEs you could batch, try taking the folio lock, and compare the number of batched PTEs against folio_mapcount(). If they match, we're good. if not, there is some other folio mapping somewhere else (other process, mremap).

folio_likely_mapped_shared() should still be used as an early exit point.

[1] https://lore.kernel.org/all/20231124132626.235350-1-david@xxxxxxxxxx/T/#u

--
Cheers,

David / dhildenb