Avoid cache line miss dereferencing struct page if we can.
page_copy_sane() mostly deals with order-0 pages.
Extra cache line miss is visible on TCP recvmsg() calls dealing
with GRO packets (typically 45 page frags are attached to one skb).
Bringing the 45 struct pages into cpu cache while copying the data
is not free, since the freeing of the skb (and associated
page frags put_page()) can happen after cache lines have been evicted.
Signed-off-by: Eric Dumazet <[email protected]>
Cc: Al Viro <[email protected]>
---
lib/iov_iter.c | 17 +++++++++++++++--
1 file changed, 15 insertions(+), 2 deletions(-)
diff --git a/lib/iov_iter.c b/lib/iov_iter.c
index be4bd627caf060cd89aa41ac88208946da568035..4d6b19c1b1294e1c30f6bbb7137e98cca5121f13 100644
--- a/lib/iov_iter.c
+++ b/lib/iov_iter.c
@@ -861,8 +861,21 @@ EXPORT_SYMBOL(_copy_from_iter_full_nocache);
static inline bool page_copy_sane(struct page *page, size_t offset, size_t n)
{
- struct page *head = compound_head(page);
- size_t v = n + offset + page_address(page) - page_address(head);
+ struct page *head;
+ size_t v = n + offset;
+
+ /*
+ * The general case needs to access the page order in order
+ * to compute the page size.
+ * However, we mostly deal with order-0 pages and thus can
+ * avoid a possible cache line miss for requests that fit all
+ * page orders.
+ */
+ if (n <= v && v <= PAGE_SIZE)
+ return true;
+
+ head = compound_head(page);
+ v += (page - head) << PAGE_SHIFT;
if (likely(n <= v && v <= (PAGE_SIZE << compound_order(head))))
return true;
--
2.21.0.rc2.261.ga7da99ff1b-goog
On Tue, Feb 26, 2019 at 10:42:39AM -0800, Eric Dumazet wrote:
> Avoid cache line miss dereferencing struct page if we can.
>
> page_copy_sane() mostly deals with order-0 pages.
>
> Extra cache line miss is visible on TCP recvmsg() calls dealing
> with GRO packets (typically 45 page frags are attached to one skb).
>
> Bringing the 45 struct pages into cpu cache while copying the data
> is not free, since the freeing of the skb (and associated
> page frags put_page()) can happen after cache lines have been evicted.
Applied.