[dm-devel] [PATCH 6/9] dm-crypt: avoid deadlock in mempools
Mikulas Patocka
mpatocka at redhat.com
Sat Apr 5 18:07:47 UTC 2014
This patch fixes a theoretical deadlock introduced in the previous patch.
The function crypt_alloc_buffer may be called concurrently. If we allocate
from the mempool concurrently, there is a possibility of deadlock.
For example, if we have mempool of 256 pages, two processes, each wanting 256,
pages allocate from the mempool concurrently, it may deadlock in a situation
where both processes have allocated 128 pages and the mempool is exhausted.
In order to avoid this scenarios, we allocate the pages under a mutex.
In order to not degrade performance with excessive locking, we try
non-blocking allocations without a mutex first and if it fails, we fallback
to a blocking allocation with a mutex.
Signed-off-by: Mikulas Patocka <mpatocka at redhat.com>
---
drivers/md/dm-crypt.c | 41 ++++++++++++++++++++++++++++++++++++-----
1 file changed, 36 insertions(+), 5 deletions(-)
Index: linux-3.14/drivers/md/dm-crypt.c
===================================================================
--- linux-3.14.orig/drivers/md/dm-crypt.c 2014-04-04 20:59:46.000000000 +0200
+++ linux-3.14/drivers/md/dm-crypt.c 2014-04-04 21:04:40.000000000 +0200
@@ -124,6 +124,7 @@ struct crypt_config {
mempool_t *req_pool;
mempool_t *page_pool;
struct bio_set *bs;
+ struct mutex bio_alloc_lock;
struct workqueue_struct *io_queue;
struct workqueue_struct *crypt_queue;
@@ -954,27 +955,51 @@ static void crypt_free_buffer_pages(stru
/*
* Generate a new unfragmented bio with the given size
* This should never violate the device limitations
+ *
+ * This function may be called concurrently. If we allocate from the mempool
+ * concurrently, there is a possibility of deadlock. For example, if we have
+ * mempool of 256 pages, two processes, each wanting 256, pages allocate from
+ * the mempool concurrently, it may deadlock in a situation where both processes
+ * have allocated 128 pages and the mempool is exhausted.
+ *
+ * In order to avoid this scenarios, we allocate the pages under a mutex.
+ *
+ * In order to not degrade performance with excessive locking, we try
+ * non-blocking allocations without a mutex first and if it fails, we fallback
+ * to a blocking allocation with a mutex.
*/
static struct bio *crypt_alloc_buffer(struct dm_crypt_io *io, unsigned size)
{
struct crypt_config *cc = io->cc;
struct bio *clone;
unsigned int nr_iovecs = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
- gfp_t gfp_mask = GFP_NOIO | __GFP_HIGHMEM;
- unsigned i, len;
+ gfp_t gfp_mask = GFP_NOWAIT | __GFP_HIGHMEM;
+ unsigned i, len, remaining_size;
struct page *page;
struct bio_vec *bvec;
+retry:
+ if (unlikely(gfp_mask & __GFP_WAIT))
+ mutex_lock(&cc->bio_alloc_lock);
+
clone = bio_alloc_bioset(GFP_NOIO, nr_iovecs, cc->bs);
if (!clone)
- return NULL;
+ goto return_clone;
clone_init(io, clone);
+ remaining_size = size;
+
for (i = 0; i < nr_iovecs; i++) {
page = mempool_alloc(cc->page_pool, gfp_mask);
+ if (!page) {
+ crypt_free_buffer_pages(cc, clone);
+ bio_put(clone);
+ gfp_mask |= __GFP_WAIT;
+ goto retry;
+ }
- len = (size > PAGE_SIZE) ? PAGE_SIZE : size;
+ len = (remaining_size > PAGE_SIZE) ? PAGE_SIZE : remaining_size;
bvec = &clone->bi_io_vec[clone->bi_vcnt++];
bvec->bv_page = page;
@@ -983,9 +1008,13 @@ static struct bio *crypt_alloc_buffer(st
clone->bi_iter.bi_size += len;
- size -= len;
+ remaining_size -= len;
}
+return_clone:
+ if (unlikely(gfp_mask & __GFP_WAIT))
+ mutex_unlock(&cc->bio_alloc_lock);
+
return clone;
}
@@ -1671,6 +1700,8 @@ static int crypt_ctr(struct dm_target *t
goto bad;
}
+ mutex_init(&cc->bio_alloc_lock);
+
ret = -EINVAL;
if (sscanf(argv[2], "%llu%c", &tmpll, &dummy) != 1) {
ti->error = "Invalid iv_offset sector";
More information about the dm-devel
mailing list