null_blk: mem garbage on NUMA systems during init
authorMatias Bjorling <m@bjorling.me>
Tue, 10 Dec 2013 15:50:38 +0000 (16:50 +0100)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sun, 15 Dec 2013 20:17:16 +0000 (12:17 -0800)
For NUMA systems, initializing the blk-mq layer and using per node hctx.
We initialize submit queues to 1, while blk-mq nr_hw_queues is
initialized to the number of NUMA nodes.

This makes the null_init_hctx function overwrite memory outside of what
it allocated.  In my case it lead to writing garbage into struct
request_queue's mq_map.

Signed-off-by: Matias Bjorling <m@bjorling.me>
Cc: Jens Axboe <axboe@kernel.dk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/block/null_blk.c

index ea192ec029c45bf7354d8184b44e1eb460cdbf77..f370fc13aea5d7d8c2a2c78fb40722da0cc27407 100644 (file)
@@ -495,23 +495,23 @@ static int null_add_dev(void)
 
        spin_lock_init(&nullb->lock);
 
+       if (queue_mode == NULL_Q_MQ && use_per_node_hctx)
+               submit_queues = nr_online_nodes;
+
        if (setup_queues(nullb))
                goto err;
 
        if (queue_mode == NULL_Q_MQ) {
                null_mq_reg.numa_node = home_node;
                null_mq_reg.queue_depth = hw_queue_depth;
+               null_mq_reg.nr_hw_queues = submit_queues;
 
                if (use_per_node_hctx) {
                        null_mq_reg.ops->alloc_hctx = null_alloc_hctx;
                        null_mq_reg.ops->free_hctx = null_free_hctx;
-
-                       null_mq_reg.nr_hw_queues = nr_online_nodes;
                } else {
                        null_mq_reg.ops->alloc_hctx = blk_mq_alloc_single_hw_queue;
                        null_mq_reg.ops->free_hctx = blk_mq_free_single_hw_queue;
-
-                       null_mq_reg.nr_hw_queues = submit_queues;
                }
 
                nullb->q = blk_mq_init_queue(&null_mq_reg, nullb);