vhost: avoid concurrency when logging dirty pages

This patch aims at fixing a migration performance regression
faced since atomic operation is used to log pages as dirty when
doing live migration.

Instead of setting a single bit by doing an atomic read-modify-write
operation to log a page as dirty, this patch write 0xFF to the
corresponding byte, and so logs 8 page as dirty.

The advantage is that it avoids concurrent atomic operations by
multiple PMD threads, the drawback is that some clean pages are
marked as dirty and so are transferred twice.

Fixes: 897f13a1f726 ("vhost: make page logging atomic")
Cc: stable@dpdk.org
Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Reviewed-by: Jianfeng Tan <jianfeng.tan@intel.com>

diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
index 58aec2e..7ca2a03 100644 (file)
--- a/lib/librte_vhost/vhost.h
+++ b/lib/librte_vhost/vhost.h
@@ -247,18 +247,14 @@ struct virtio_net {
 #define VHOST_LOG_PAGE 4096
 
 /*
- * Atomically set a bit in memory.
+ * Mark all pages belonging to the same dirty log bitmap byte
+ * as dirty. The goal is to avoid concurrency between different
+ * threads doing atomic read-modify-writes on the same byte.
  */
-static __rte_always_inline void
-vhost_set_bit(unsigned int nr, volatile uint8_t *addr)
-{
-       __sync_fetch_and_or_8(addr, (1U << nr));
-}
-
 static __rte_always_inline void
 vhost_log_page(uint8_t *log_base, uint64_t page)
 {
-       vhost_set_bit(page % 8, &log_base[page / 8]);
+       log_base[page / 8] = 0xff;
 }
 
 static __rte_always_inline void