Commit fe48d9eb authored Feb 05, 2014 by Mitchel Humpherys

ion: invalidate buffers before zero'ing



Currently, we invalidate the cache for buffers being zero'd *after* we
zero them. However, this is racey since a dirty cache line from an
earlier mapping could be written out to memory between the zero'ing and
the invalidation. Fix this by invalidating the cache *before* zero'ing
the buffer.

Before this change, I was seeing corruption with:

    $ cat <<EOF | ./memory_prof -e -i -
    simple_alloc,1,ION_SYSTEM_HEAP_ID,ION_FLAG_CACHED,0x100000,1MB
    simple_basic_sanity,1
    simple_profile,1
    simple_free,1
    EOF

After this change, no corruption.

Change-Id: I1f2603604efc9be4ebf9112b06728e0aeca60ea0
Signed-off-by: Mitchel Humpherys <mitchelh@codeaurora.org>

parent 4db175e5

drivers/staging/android/ion/ion_heap.c

+5 −2

Original line number	Diff line number	Diff line
		@@ -146,9 +146,11 @@ int ion_heap_pages_zero(struct page **pages, int num_pages)
		if (!ptr)
		return -ENOMEM;

		memset(ptr, 0, npages_to_vmap * PAGE_SIZE);
		/*
		* invalidate the cache to pick up the zeroing
		* We have to invalidate the cache here because there
		* might be dirty lines to these physical pages (which
		* we don't care about) that could get written out at
		* any moment.
		*/
		for (k = 0; k < npages_to_vmap; k++) {
		void *p = kmap_atomic(pages[i + k]);
		@@ -156,6 +158,7 @@ int ion_heap_pages_zero(struct page **pages, int num_pages)
		dmac_inv_range(p, p + PAGE_SIZE);
		kunmap_atomic(p);
		}
		memset(ptr, 0, npages_to_vmap * PAGE_SIZE);
		vunmap(ptr);
		}