OSDN Git Service

page_pool: help compiler remove code in case CONFIG_NUMA=n
authorJesper Dangaard Brouer <brouer@redhat.com>
Fri, 27 Dec 2019 17:13:24 +0000 (18:13 +0100)
committerDavid S. Miller <davem@davemloft.net>
Thu, 2 Jan 2020 23:37:52 +0000 (15:37 -0800)
When kernel is compiled without NUMA support, then page_pool NUMA
config setting (pool->p.nid) doesn't make any practical sense. The
compiler cannot see that it can remove the code paths.

This patch avoids reading pool->p.nid setting in case of !CONFIG_NUMA,
in allocation and numa check code, which helps compiler to see the
optimisation potential. It leaves update code intact to keep API the
same.

 $ ./scripts/bloat-o-meter net/core/page_pool.o-numa-enabled \
                           net/core/page_pool.o-numa-disabled
 add/remove: 0/0 grow/shrink: 0/3 up/down: 0/-113 (-113)
 Function                                     old     new   delta
 page_pool_create                             401     398      -3
 __page_pool_alloc_pages_slow                 439     426     -13
 page_pool_refill_alloc_cache                 425     328     -97
 Total: Before=3611, After=3498, chg -3.13%

Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/core/page_pool.c

index 515bfb9..9b7cbe3 100644 (file)
@@ -113,7 +113,12 @@ static struct page *page_pool_refill_alloc_cache(struct page_pool *pool,
        /* Softirq guarantee CPU and thus NUMA node is stable. This,
         * assumes CPU refilling driver RX-ring will also run RX-NAPI.
         */
+#ifdef CONFIG_NUMA
        pref_nid = (pool->p.nid == NUMA_NO_NODE) ? numa_mem_id() : pool->p.nid;
+#else
+       /* Ignore pool->p.nid setting if !CONFIG_NUMA, helps compiler */
+       pref_nid = numa_mem_id(); /* will be zero like page_to_nid() */
+#endif
 
        /* Slower-path: Get pages from locked ring queue */
        spin_lock(&r->consumer_lock);
@@ -200,7 +205,11 @@ static struct page *__page_pool_alloc_pages_slow(struct page_pool *pool,
         */
 
        /* Cache was empty, do real allocation */
+#ifdef CONFIG_NUMA
        page = alloc_pages_node(pool->p.nid, gfp, pool->p.order);
+#else
+       page = alloc_pages(gfp, pool->p.order);
+#endif
        if (!page)
                return NULL;