Skip to contentSkip to conversation

aio: fix race in ring buffer page lookup introduced by page migration support

PATCH In review Benjamin LaHaise 2 participants 2 messages Sep 9, 2013 – Sep 10, 2013

Subsystems (2)
lore
Share & tools
lore

Benjamin LaHaise

author maintainer
#0 lore raw

PATCH 1/1 aio: fix race in ring buffer page lookup introduced by page migration support

Hi Al, Gu,

I've added this patch to my tree at git://git.kvack.org/~bcrl/aio-next.git
to fix the get_user_pages() issue introduced by Gu's changes in the page migration patch. Thanks Al for spotting this.

-ben

commit d6c355c7dabcd753a75bc77d150d36328a355267

Author: Benjamin LaHaise <[email protected]>
Date:   Mon Sep 9 11:57:59 2013 -0400

aio: fix race in ring buffer page lookup introduced by page migration support

Prior to the introduction of page migration support in "fs/aio: Add support to aio ring pages migration" / 36bc08cc01709b4a9bb563b35aa530241ddc63e3, mapping of the ring buffer pages was done via get_user_pages() while retaining mmap_sem held for write. This avoided possible races with userland racing an munmap() or mremap(). The page migration patch, however, switched to using mm_populate() to prime the page mapping. mm_populate() cannot be called with mmap_sem held.

Instead of dropping the mmap_sem, revert to the old behaviour and simply drop the use of mm_populate() since get_user_pages() will cause the pages to get mapped anyways. Thanks to Al Viro for spotting this issue.

fs/aio.c +12 -3
index 6e26755..f4a27af 100644
@@ -307,16 +307,25 @@ static int aio_setup_ring(struct kioctx *ctx)
 		aio_free_ring(ctx);
 		return -EAGAIN;
 	}
-	up_write(&mm->mmap_sem);
-
-	mm_populate(ctx->mmap_base, populate);
 
 	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
+
+	/* We must do this while still holding mmap_sem for write, as we
+	 * need to be protected against userspace attempting to mremap()
+	 * or munmap() the ring buffer.
+	 */
 	ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
 				       1, 0, ctx->ring_pages, NULL);
+
+	/* Dropping the reference here is safe as the page cache will hold
+	 * onto the pages for us.  It is also required so that page migration
+	 * can unmap the pages and get the right reference count.
+	 */
 	for (i = 0; i < ctx->nr_pages; i++)
 		put_page(ctx->ring_pages[i]);
 
+	up_write(&mm->mmap_sem);
+
 	if (unlikely(ctx->nr_pages != nr_pages)) {
 		aio_free_ring(ctx);
 		return -EAGAIN;
Signature · 1 line
"Thought is the essence of where you are now."
#1 lore raw

Hi Ben, Al,

Quoted 5 lines · Benjamin LaHaise

Hi Al, Gu,

I've added this patch to my tree at git://git.kvack.org/~bcrl/aio-next.git
to fix the get_user_pages() issue introduced by Gu's changes in the page migration patch. Thanks Al for spotting this.

Thanks very much for spotting and fixing this issue.

Best regards,
Gu

Quoted 56 lines

-ben

commit d6c355c7dabcd753a75bc77d150d36328a355267

Author: Benjamin LaHaise <[email protected]>
Date:   Mon Sep 9 11:57:59 2013 -0400

aio: fix race in ring buffer page lookup introduced by page migration support

Prior to the introduction of page migration support in "fs/aio: Add support to aio ring pages migration" / 36bc08cc01709b4a9bb563b35aa530241ddc63e3, mapping of the ring buffer pages was done via get_user_pages() while retaining mmap_sem held for write. This avoided possible races with userland racing an munmap() or mremap(). The page migration patch, however, switched to using mm_populate() to prime the page mapping. mm_populate() cannot be called with mmap_sem held.

Instead of dropping the mmap_sem, revert to the old behaviour and simply drop the use of mm_populate() since get_user_pages() will cause the pages to get mapped anyways. Thanks to Al Viro for spotting this issue.

fs/aio.c +12 -3
index 6e26755..f4a27af 100644
@@ -307,16 +307,25 @@ static int aio_setup_ring(struct kioctx *ctx)
 		aio_free_ring(ctx);
 		return -EAGAIN;
 	}
-	up_write(&mm->mmap_sem);
-
-	mm_populate(ctx->mmap_base, populate);
 
 	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
+
+	/* We must do this while still holding mmap_sem for write, as we
+	 * need to be protected against userspace attempting to mremap()
+	 * or munmap() the ring buffer.
+	 */
 	ctx->nr_pages = get_user_pages(current, mm, ctx->mmap_base, nr_pages,
 				       1, 0, ctx->ring_pages, NULL);
+
+	/* Dropping the reference here is safe as the page cache will hold
+	 * onto the pages for us.  It is also required so that page migration
+	 * can unmap the pages and get the right reference count.
+	 */
 	for (i = 0; i < ctx->nr_pages; i++)
 		put_page(ctx->ring_pages[i]);
 
+	up_write(&mm->mmap_sem);
+
 	if (unlikely(ctx->nr_pages != nr_pages)) {
 		aio_free_ring(ctx);
 		return -EAGAIN;