arm64: avoid flushing icache multiple times on contiguous HugeTLB
authorMuchun Song <songmuchun@bytedance.com>
Wed, 2 Mar 2022 08:46:23 +0000 (16:46 +0800)
committerWill Deacon <will@kernel.org>
Mon, 7 Mar 2022 21:42:34 +0000 (21:42 +0000)
When a contiguous HugeTLB page is mapped, set_pte_at() will be called
CONT_PTES/CONT_PMDS times.  Therefore, __sync_icache_dcache() will
flush cache multiple times if the page is executable (to ensure
the I-D cache coherency).  However, the first flushing cache already
covers subsequent cache flush operations.  So only flusing cache
for the head page if it is a HugeTLB page to avoid redundant cache
flushing.  In the next patch, it is also depends on this change
since the tail vmemmap pages of HugeTLB is mapped with read-only
meanning only head page struct can be modified.

Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
Link: https://lore.kernel.org/r/20220302084624.33340-1-songmuchun@bytedance.com
Signed-off-by: Will Deacon <will@kernel.org>
arch/arm64/mm/flush.c

index 2aaf950..a06c6ac 100644 (file)
@@ -52,6 +52,13 @@ void __sync_icache_dcache(pte_t pte)
 {
        struct page *page = pte_page(pte);
 
+       /*
+        * HugeTLB pages are always fully mapped, so only setting head page's
+        * PG_dcache_clean flag is enough.
+        */
+       if (PageHuge(page))
+               page = compound_head(page);
+
        if (!test_bit(PG_dcache_clean, &page->flags)) {
                sync_icache_aliases((unsigned long)page_address(page),
                                    (unsigned long)page_address(page) +