On 2024/3/27 23:10, Zi Yan wrote:
On 27 Mar 2024, at 10:54, Matthew Wilcox wrote:
On Wed, Mar 27, 2024 at 10:45:42AM -0400, Zi Yan wrote:
for (pfn = start_pfn; pfn < end_pfn; pfn++) {
- struct folio *folio;
+ struct page *page, *head;
You could get rid of head too. It is only used to calculate next pfn,
so pfn = folio_to_pfn(folio) + folio_nr_pages(folio) - 1 would work.
And the PageHuge(page) and PageTransHuge(page) can be simplified, since
their pfn calculations are the same. Something like:
if (folio_test_large(folio)) {
pfn = folio_to_pfn(folio) + folio_nr_pages(folio) - 1;
if (folio_test_hugetlb(folio)) {
isolate_hugetlb(folio, &source);
continue;
}
}
How much of this is safe without a refcount on the folio?
folio_to_pfn() should be fine, isolate_hugetlb() checks the folio
under hugetlb_lock, but folio_nr_pages() might return a bogus number
that makes pfn go beyond end_pfn and ends for loop early. The code
below increases the refcount, so it might be better to move this
part of the code after refcount is increased.
The PageHWPoison() is per-page check, for hugetlb, it will directly
try to isolate and ignore the PageHWPoison check, I'm not sure about
moveing PageHWPoison ahead, we need to take the i_mmap_lock_write and
add TTU_RMAP_LOCKED for for hugetlb pages in shared mappings when
try_to_unmap(), but now hugetlb pages won't unmap if there is posion
page, if the get_page_unless_zero() is moved ahead, the logical need be
changed a lot, this minimize changes aim to remove isolate_lru/movable_page,
so could we keep it simple, but as your suggested, we could do more
optimization about do_migrate_range() in the next.
Thanks.
--
Best Regards,
Yan, Zi
