mm: add support for __GFP_ZERO flag to dma_pool_alloc()
authorSean O. Stalley <[email protected]>
Tue, 8 Sep 2015 22:02:24 +0000 (15:02 -0700)
committerLinus Torvalds <[email protected]>
Tue, 8 Sep 2015 22:35:28 +0000 (15:35 -0700)
Currently a call to dma_pool_alloc() with a ___GFP_ZERO flag returns a
non-zeroed memory region.

This patchset adds support for the __GFP_ZERO flag to dma_pool_alloc(),
adds 2 wrapper functions for allocing zeroed memory from a pool, and
provides a coccinelle script for finding & replacing instances of
dma_pool_alloc() followed by memset(0) with a single dma_pool_zalloc()
call.

There was some concern that this always calls memset() to zero, instead
of passing __GFP_ZERO into the page allocator.
[https://lkml.org/lkml/2015/7/15/881]

I ran a test on my system to get an idea of how often dma_pool_alloc()
calls into pool_alloc_page().

After Boot: [   30.119863] alloc_calls:541, page_allocs:7
After an hour: [ 3600.951031] alloc_calls:9566, page_allocs:12
After copying 1GB file onto a USB drive:
[ 4260.657148] alloc_calls:17225, page_allocs:12

It doesn't look like dma_pool_alloc() calls down to the page allocator
very often (at least on my system).

This patch (of 4):

Currently the __GFP_ZERO flag is ignored by dma_pool_alloc().
Make dma_pool_alloc() zero the memory if this flag is set.

Signed-off-by: Sean O. Stalley <[email protected]>
Acked-by: David Rientjes <[email protected]>
Cc: Vinod Koul <[email protected]>
Cc: Bjorn Helgaas <[email protected]>
Cc: Gilles Muller <[email protected]>
Cc: Nicolas Palix <[email protected]>
Cc: Michal Marek <[email protected]>
Cc: Sebastian Andrzej Siewior <[email protected]>
Cc: Jonathan Corbet <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
mm/dmapool.c

index 4b657099111f8ab2f92f2c5c870c7a2b33e4231b..71a8998cd03a6b8b0d2dbfe36a24ce70766047e8 100644 (file)
@@ -337,7 +337,7 @@ void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
        /* pool_alloc_page() might sleep, so temporarily drop &pool->lock */
        spin_unlock_irqrestore(&pool->lock, flags);
 
-       page = pool_alloc_page(pool, mem_flags);
+       page = pool_alloc_page(pool, mem_flags & (~__GFP_ZERO));
        if (!page)
                return NULL;
 
@@ -375,9 +375,14 @@ void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
                        break;
                }
        }
-       memset(retval, POOL_POISON_ALLOCATED, pool->size);
+       if (!(mem_flags & __GFP_ZERO))
+               memset(retval, POOL_POISON_ALLOCATED, pool->size);
 #endif
        spin_unlock_irqrestore(&pool->lock, flags);
+
+       if (mem_flags & __GFP_ZERO)
+               memset(retval, 0, pool->size);
+
        return retval;
 }
 EXPORT_SYMBOL(dma_pool_alloc);