]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/commitdiff
blk-mq: fix race between timeout and freeing request
authorMing Lei <ming.lei@canonical.com>
Sun, 9 Aug 2015 07:41:51 +0000 (03:41 -0400)
committerJens Axboe <axboe@fb.com>
Sat, 15 Aug 2015 15:45:21 +0000 (09:45 -0600)
Inside timeout handler, blk_mq_tag_to_rq() is called
to retrieve the request from one tag. This way is obviously
wrong because the request can be freed any time and some
fiedds of the request can't be trusted, then kernel oops
might be triggered[1].

Currently wrt. blk_mq_tag_to_rq(), the only special case is
that the flush request can share same tag with the request
cloned from, and the two requests can't be active at the same
time, so this patch fixes the above issue by updating tags->rqs[tag]
with the active request(either flush rq or the request cloned
from) of the tag.

Also blk_mq_tag_to_rq() gets much simplified with this patch.

Given blk_mq_tag_to_rq() is mainly for drivers and the caller must
make sure the request can't be freed, so in bt_for_each() this
helper is replaced with tags->rqs[tag].

[1] kernel oops log
[  439.696220] BUG: unable to handle kernel NULL pointer dereference at 0000000000000158^M
[  439.697162] IP: [<ffffffff812d89ba>] blk_mq_tag_to_rq+0x21/0x6e^M
[  439.700653] PGD 7ef765067 PUD 7ef764067 PMD 0 ^M
[  439.700653] Oops: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC ^M
[  439.700653] Dumping ftrace buffer:^M
[  439.700653]    (ftrace buffer empty)^M
[  439.700653] Modules linked in: nbd ipv6 kvm_intel kvm serio_raw^M
[  439.700653] CPU: 6 PID: 2779 Comm: stress-ng-sigfd Not tainted 4.2.0-rc5-next-20150805+ #265^M
[  439.730500] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011^M
[  439.730500] task: ffff880605308000 ti: ffff88060530c000 task.ti: ffff88060530c000^M
[  439.730500] RIP: 0010:[<ffffffff812d89ba>]  [<ffffffff812d89ba>] blk_mq_tag_to_rq+0x21/0x6e^M
[  439.730500] RSP: 0018:ffff880819203da0  EFLAGS: 00010283^M
[  439.730500] RAX: ffff880811b0e000 RBX: ffff8800bb465f00 RCX: 0000000000000002^M
[  439.730500] RDX: 0000000000000000 RSI: 0000000000000202 RDI: 0000000000000000^M
[  439.730500] RBP: ffff880819203db0 R08: 0000000000000002 R09: 0000000000000000^M
[  439.730500] R10: 0000000000000000 R11: 0000000000000000 R12: 0000000000000202^M
[  439.730500] R13: ffff880814104800 R14: 0000000000000002 R15: ffff880811a2ea00^M
[  439.730500] FS:  00007f165b3f5740(0000) GS:ffff880819200000(0000) knlGS:0000000000000000^M
[  439.730500] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b^M
[  439.730500] CR2: 0000000000000158 CR3: 00000007ef766000 CR4: 00000000000006e0^M
[  439.730500] Stack:^M
[  439.730500]  0000000000000008 ffff8808114eed90 ffff880819203e00 ffffffff812dc104^M
[  439.755663]  ffff880819203e40 ffffffff812d9f5e 0000020000000000 ffff8808114eed80^M
[  439.755663] Call Trace:^M
[  439.755663]  <IRQ> ^M
[  439.755663]  [<ffffffff812dc104>] bt_for_each+0x6e/0xc8^M
[  439.755663]  [<ffffffff812d9f5e>] ? blk_mq_rq_timed_out+0x6a/0x6a^M
[  439.755663]  [<ffffffff812d9f5e>] ? blk_mq_rq_timed_out+0x6a/0x6a^M
[  439.755663]  [<ffffffff812dc1b3>] blk_mq_tag_busy_iter+0x55/0x5e^M
[  439.755663]  [<ffffffff812d88b4>] ? blk_mq_bio_to_request+0x38/0x38^M
[  439.755663]  [<ffffffff812d8911>] blk_mq_rq_timer+0x5d/0xd4^M
[  439.755663]  [<ffffffff810a3e10>] call_timer_fn+0xf7/0x284^M
[  439.755663]  [<ffffffff810a3d1e>] ? call_timer_fn+0x5/0x284^M
[  439.755663]  [<ffffffff812d88b4>] ? blk_mq_bio_to_request+0x38/0x38^M
[  439.755663]  [<ffffffff810a46d6>] run_timer_softirq+0x1ce/0x1f8^M
[  439.755663]  [<ffffffff8104c367>] __do_softirq+0x181/0x3a4^M
[  439.755663]  [<ffffffff8104c76e>] irq_exit+0x40/0x94^M
[  439.755663]  [<ffffffff81031482>] smp_apic_timer_interrupt+0x33/0x3e^M
[  439.755663]  [<ffffffff815559a4>] apic_timer_interrupt+0x84/0x90^M
[  439.755663]  <EOI> ^M
[  439.755663]  [<ffffffff81554350>] ? _raw_spin_unlock_irq+0x32/0x4a^M
[  439.755663]  [<ffffffff8106a98b>] finish_task_switch+0xe0/0x163^M
[  439.755663]  [<ffffffff8106a94d>] ? finish_task_switch+0xa2/0x163^M
[  439.755663]  [<ffffffff81550066>] __schedule+0x469/0x6cd^M
[  439.755663]  [<ffffffff8155039b>] schedule+0x82/0x9a^M
[  439.789267]  [<ffffffff8119b28b>] signalfd_read+0x186/0x49a^M
[  439.790911]  [<ffffffff8106d86a>] ? wake_up_q+0x47/0x47^M
[  439.790911]  [<ffffffff811618c2>] __vfs_read+0x28/0x9f^M
[  439.790911]  [<ffffffff8117a289>] ? __fget_light+0x4d/0x74^M
[  439.790911]  [<ffffffff811620a7>] vfs_read+0x7a/0xc6^M
[  439.790911]  [<ffffffff8116292b>] SyS_read+0x49/0x7f^M
[  439.790911]  [<ffffffff81554c17>] entry_SYSCALL_64_fastpath+0x12/0x6f^M
[  439.790911] Code: 48 89 e5 e8 a9 b8 e7 ff 5d c3 0f 1f 44 00 00 55 89
f2 48 89 e5 41 54 41 89 f4 53 48 8b 47 60 48 8b 1c d0 48 8b 7b 30 48 8b
53 38 <48> 8b 87 58 01 00 00 48 85 c0 75 09 48 8b 97 88 0c 00 00 eb 10
^M
[  439.790911] RIP  [<ffffffff812d89ba>] blk_mq_tag_to_rq+0x21/0x6e^M
[  439.790911]  RSP <ffff880819203da0>^M
[  439.790911] CR2: 0000000000000158^M
[  439.790911] ---[ end trace d40af58949325661 ]---^M

Cc: <stable@vger.kernel.org>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
block/blk-flush.c
block/blk-mq-tag.c
block/blk-mq-tag.h
block/blk-mq.c
block/blk.h

index 20badd7b9d1b664f534a791b83435534f94a70ec..9c423e53324a29c1ba1213c6009822b15d75dac7 100644 (file)
@@ -73,6 +73,7 @@
 
 #include "blk.h"
 #include "blk-mq.h"
+#include "blk-mq-tag.h"
 
 /* FLUSH/FUA sequences */
 enum {
@@ -226,7 +227,12 @@ static void flush_end_io(struct request *flush_rq, int error)
        struct blk_flush_queue *fq = blk_get_flush_queue(q, flush_rq->mq_ctx);
 
        if (q->mq_ops) {
+               struct blk_mq_hw_ctx *hctx;
+
+               /* release the tag's ownership to the req cloned from */
                spin_lock_irqsave(&fq->mq_flush_lock, flags);
+               hctx = q->mq_ops->map_queue(q, flush_rq->mq_ctx->cpu);
+               blk_mq_tag_set_rq(hctx, flush_rq->tag, fq->orig_rq);
                flush_rq->tag = -1;
        }
 
@@ -308,11 +314,18 @@ static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
 
        /*
         * Borrow tag from the first request since they can't
-        * be in flight at the same time.
+        * be in flight at the same time. And acquire the tag's
+        * ownership for flush req.
         */
        if (q->mq_ops) {
+               struct blk_mq_hw_ctx *hctx;
+
                flush_rq->mq_ctx = first_rq->mq_ctx;
                flush_rq->tag = first_rq->tag;
+               fq->orig_rq = first_rq;
+
+               hctx = q->mq_ops->map_queue(q, first_rq->mq_ctx->cpu);
+               blk_mq_tag_set_rq(hctx, first_rq->tag, flush_rq);
        }
 
        flush_rq->cmd_type = REQ_TYPE_FS;
index 9b6e28830b823866ec63ccd2f6b138c88be562b5..9115c6d59948addbc445a26ad0f9ccaf4237b137 100644 (file)
@@ -429,7 +429,7 @@ static void bt_for_each(struct blk_mq_hw_ctx *hctx,
                for (bit = find_first_bit(&bm->word, bm->depth);
                     bit < bm->depth;
                     bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
-                       rq = blk_mq_tag_to_rq(hctx->tags, off + bit);
+                       rq = hctx->tags->rqs[off + bit];
                        if (rq->q == hctx->queue)
                                fn(hctx, rq, data, reserved);
                }
@@ -453,7 +453,7 @@ static void bt_tags_for_each(struct blk_mq_tags *tags,
                for (bit = find_first_bit(&bm->word, bm->depth);
                     bit < bm->depth;
                     bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
-                       rq = blk_mq_tag_to_rq(tags, off + bit);
+                       rq = tags->rqs[off + bit];
                        fn(rq, data, reserved);
                }
 
index 75893a34237d2eb2b4d07bd21a70cc35948c172a..9eb2cf4f01cb874706d64af87a01e94e0121f7e4 100644 (file)
@@ -89,4 +89,16 @@ static inline void blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
        __blk_mq_tag_idle(hctx);
 }
 
+/*
+ * This helper should only be used for flush request to share tag
+ * with the request cloned from, and both the two requests can't be
+ * in flight at the same time. The caller has to make sure the tag
+ * can't be freed.
+ */
+static inline void blk_mq_tag_set_rq(struct blk_mq_hw_ctx *hctx,
+               unsigned int tag, struct request *rq)
+{
+       hctx->tags->rqs[tag] = rq;
+}
+
 #endif
index 81edbd95bda87eb5da6e21bfa84253daf9ca0747..f2d67b4047a04d7015c3c2af16871972c3b5a720 100644 (file)
@@ -559,23 +559,9 @@ void blk_mq_abort_requeue_list(struct request_queue *q)
 }
 EXPORT_SYMBOL(blk_mq_abort_requeue_list);
 
-static inline bool is_flush_request(struct request *rq,
-               struct blk_flush_queue *fq, unsigned int tag)
-{
-       return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
-                       fq->flush_rq->tag == tag);
-}
-
 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
 {
-       struct request *rq = tags->rqs[tag];
-       /* mq_ctx of flush rq is always cloned from the corresponding req */
-       struct blk_flush_queue *fq = blk_get_flush_queue(rq->q, rq->mq_ctx);
-
-       if (!is_flush_request(rq, fq, tag))
-               return rq;
-
-       return fq->flush_rq;
+       return tags->rqs[tag];
 }
 EXPORT_SYMBOL(blk_mq_tag_to_rq);
 
index 026d9594142bdedf1d5a0b919390bb54b3f857c1..838188b35a83f03696fba66cbb71728e402403bd 100644 (file)
@@ -22,6 +22,12 @@ struct blk_flush_queue {
        struct list_head        flush_queue[2];
        struct list_head        flush_data_in_flight;
        struct request          *flush_rq;
+
+       /*
+        * flush_rq shares tag with this rq, both can't be active
+        * at the same time
+        */
+       struct request          *orig_rq;
        spinlock_t              mq_flush_lock;
 };