The TLB flushing will cost quite some CPU cycles during the folio
migration in some situations. For example, when migrate a folio of a
process with multiple active threads that run on multiple CPUs. After
batching the _unmap and _move in migrate_pages(), the TLB flushing can
be batched easily with the existing TLB flush batching mechanism.
This patch implements that.
We use the following test case to test the patch.
On a 2-socket Intel server,
- Run pmbench memory accessing benchmark
- Run `migratepages` to migrate pages of pmbench between node 0 and
node 1 back and forth.
With the patch, the TLB flushing IPI reduces 99.1% during the test and
the number of pages migrated successfully per second increases 291.7%.
Haoxin helped to test the patchset on an ARM64 server with 128 cores,
2 NUMA nodes. Test results show that the page migration performance
increases up to 78%.
NOTE: TLB flushing is batched only for normal folios, not for THP
folios. Because the overhead of TLB flushing for THP folios is much
lower than that for normal folios (about 1/512 on x86 platform).
Signed-off-by: "Huang, Ying" <ying.huang@xxxxxxxxx>
Tested-by: Xin Hao <xhao@xxxxxxxxxxxxxxxxx>
Reviewed-by: Zi Yan <ziy@xxxxxxxxxx>
Reviewed-by: Xin Hao <xhao@xxxxxxxxxxxxxxxxx>
Cc: Yang Shi <shy828301@xxxxxxxxx>
Cc: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
Cc: Oscar Salvador <osalvador@xxxxxxx>
Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx>
Cc: Bharata B Rao <bharata@xxxxxxx>
Cc: Alistair Popple <apopple@xxxxxxxxxx>
Cc: Minchan Kim <minchan@xxxxxxxxxx>
Cc: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
Cc: Hyeonggon Yoo <42.hyeyoo@xxxxxxxxx>
---
mm/migrate.c | 5 ++++-
mm/rmap.c | 20 +++++++++++++++++---
2 files changed, 21 insertions(+), 4 deletions(-)
diff --git a/mm/migrate.c b/mm/migrate.c
index 00713ccb6643..2fa420e4f68c 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -1248,7 +1248,7 @@ static int migrate_folio_unmap(new_page_t get_new_page, free_page_t put_new_page
/* Establish migration ptes */
VM_BUG_ON_FOLIO(folio_test_anon(src) &&
!folio_test_ksm(src) && !anon_vma, src);
- try_to_migrate(src, 0);
+ try_to_migrate(src, TTU_BATCH_FLUSH);
page_was_mapped = 1;
}
@@ -1806,6 +1806,9 @@ static int migrate_pages_batch(struct list_head *from, new_page_t get_new_page,
stats->nr_thp_failed += thp_retry;
stats->nr_failed_pages += nr_retry_pages;
move:
+ /* Flush TLBs for all unmapped folios */
+ try_to_unmap_flush();
+
retry = 1;
for (pass = 0;
pass < NR_MAX_MIGRATE_PAGES_RETRY && (retry || large_retry);
diff --git a/mm/rmap.c b/mm/rmap.c
index 8287f2cc327d..15ae24585fc4 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -1952,7 +1952,21 @@ static bool try_to_migrate_one(struct folio *folio, struct vm_area_struct *vma,
} else {
flush_cache_page(vma, address, pte_pfn(*pvmw.pte));
/* Nuke the page table entry. */
- pteval = ptep_clear_flush(vma, address, pvmw.pte);
+ if (should_defer_flush(mm, flags)) {
+ /*
+ * We clear the PTE but do not flush so potentially
+ * a remote CPU could still be writing to the folio.
+ * If the entry was previously clean then the
+ * architecture must guarantee that a clear->dirty
+ * transition on a cached TLB entry is written through
+ * and traps if the PTE is unmapped.
+ */
+ pteval = ptep_get_and_clear(mm, address, pvmw.pte);
+
+ set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
+ } else {
+ pteval = ptep_clear_flush(vma, address, pvmw.pte);
+ }
}
/* Set the dirty flag on the folio now the pte is gone. */
@@ -2124,10 +2138,10 @@ void try_to_migrate(struct folio *folio, enum ttu_flags flags)
/*
* Migration always ignores mlock and only supports TTU_RMAP_LOCKED and
- * TTU_SPLIT_HUGE_PMD and TTU_SYNC flags.
+ * TTU_SPLIT_HUGE_PMD, TTU_SYNC, and TTU_BATCH_FLUSH flags.
*/
if (WARN_ON_ONCE(flags & ~(TTU_RMAP_LOCKED | TTU_SPLIT_HUGE_PMD |
- TTU_SYNC)))
+ TTU_SYNC | TTU_BATCH_FLUSH)))
return;
if (folio_is_zone_device(folio) &&
--
2.35.1
