blktrace: fix debugfs use after free

Its been a fun ride, but all patch series come to an end. My hope is that this is it. The simplification of the fix is considerable now, with only a few lines of code and with no data structure changes.

We were only creating the debugfs_dir upon initialization only if you had CONFIG_BLK_DEBUG_FS for for make_request block drivers (multiqueue). That's where the UAF bug could happen. Folks liked the idea of open coding the debugfs initialization even if CONFIG_BLK_DEBUG_FS was disabled, given that debugfs code will simply ignore that code if debugfs is disabled, but to make the fix easier to backport, that shift is done now in another patch. Likewise, although we were only creating the debugfs_dir only for make_request block drivers (multiqueue), the same new additional patch also creates the debugfs_dir for request-based block drivers. That *begged* us to just rename the mutex to clarify its for the debugfs_dir, blktrace then just becomes its biggest user.

The only patches changed here is the last one from the last series, which actually fixed the UAF oops, and that one is now split in 3 patches, which makes a secondary fix much clearer.

I've waited a while to post these, to let 0-day give me its blessings, both for Linus' tree and linux-next. No issues have been found. I've also taken time to run blktests prior and after this series and I have found no regressions. In fact, I think I should just extend blktests with the break-blktrace tests, I'll do that later.

Luis Chamberlain (8):
block: add docs for gendisk / request_queue refcount helpers
block: clarify context for refcount increment helpers
block: revert back to synchronous request_queue removal
blktrace: annotate required lock on do_blk_trace_setup()
loop: be paranoid on exit and prevent new additions / removals
blktrace: fix debugfs use after free
blktrace: ensure our debugfs dir exists
block: create the request_queue debugfs_dir on registration

 block/blk-core.c        | 31 +++++++++++++----
 block/blk-mq-debugfs.c  |  5 ---
 block/blk-sysfs.c       | 52 ++++++++++++++++------------
 block/blk.h             |  2 --
 block/genhd.c           | 73 ++++++++++++++++++++++++++++++++++++++-
 drivers/block/loop.c    |  4 +++
 include/linux/blkdev.h  |  7 ++--
 kernel/trace/blktrace.c | 76 ++++++++++++++++++++++++-----------------
 8 files changed, 179 insertions(+), 71 deletions(-)

2.26.2

This adds documentation for the gendisk / request_queue refcount helpers.

---
 block/blk-core.c | 13 +++++++++++++
 block/genhd.c    | 50 +++++++++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 62 insertions(+), 1 deletion(-)

index 62a4904db921..a0760aac110a 100644
@@ -321,6 +321,13 @@ void blk_clear_pm_only(struct request_queue *q)
 }
 EXPORT_SYMBOL_GPL(blk_clear_pm_only);
 
+/**
+ * blk_put_queue - decrement the request_queue refcount
+ * @q: the request_queue structure to decrement the refcount for
+ *
+ * Decrements the refcount of the request_queue kobject. When this reaches 0
+ * we'll have blk_release_queue() called.
+ */
 void blk_put_queue(struct request_queue *q)
 {
 	kobject_put(&q->kobj);

@@ -598,6 +605,12 @@ struct request_queue *blk_alloc_queue(make_request_fn make_request, int node_id)
 }
 EXPORT_SYMBOL(blk_alloc_queue);
 
+/**
+ * blk_get_queue - increment the request_queue refcount
+ * @q: the request_queue structure to increment the refcount for
+ *
+ * Increment the refcount of the request_queue kobject.
+ */
 bool blk_get_queue(struct request_queue *q)
 {
 	if (likely(!blk_queue_dying(q))) {

2.26.2

Let us clarify the context under which the helpers to increment the refcount for the gendisk and request_queue can be called under. We make this explicit on the places where we may sleep with might_sleep().

We don't address the decrement context yet, as that needs some extra work and fixes, but will be addressed in the next patch.

---
 block/blk-core.c | 2 ++
 block/genhd.c    | 6 ++++++
 2 files changed, 8 insertions(+)

index a0760aac110a..14c09daf55f3 100644
@@ -610,6 +610,8 @@ EXPORT_SYMBOL(blk_alloc_queue);
  * @q: the request_queue structure to increment the refcount for
  *
  * Increment the refcount of the request_queue kobject.
+ *
+ * Context: Any context.
  */
 bool blk_get_queue(struct request_queue *q)
 {

index f741613d731f..1be86b1f43ec 100644
@@ -985,11 +985,15 @@ static ssize_t disk_badblocks_store(struct device *dev,
  *
  * This function gets the structure containing partitioning
  * information for the given device @devt.
+ *
+ * Context: can sleep
  */
 struct gendisk *get_gendisk(dev_t devt, int *partno)
 {
 	struct gendisk *disk = NULL;
 
+	might_sleep();
+
 	if (MAJOR(devt) != BLOCK_EXT_MAJOR) {
 		struct kobject *kobj;
 

@@ -1764,6 +1768,8 @@ EXPORT_SYMBOL(__alloc_disk_node);
  *
  * This increments the refcount for the struct gendisk, and the gendisk's
  * fops module owner.
+ *
+ * Context: Any context.
  */
 struct kobject *get_disk_and_module(struct gendisk *disk)
 {

2.26.2

Commit dc9edc44de6c ("block: Fix a blk_exit_rl() regression") merged on v4.12 moved the work behind blk_release_queue() into a workqueue after a splat floated around which indicated some work on blk_release_queue() could sleep in blk_exit_rl(). This splat would be possible when a driver called blk_put_queue() or blk_cleanup_queue() (which calls blk_put_queue() as its final call) from an atomic context.

blk_put_queue() decrements the refcount for the request_queue kobject, and upon reaching 0 blk_release_queue() is called. Although blk_exit_rl() is now removed through commit db6d99523560 ("block: remove request_list code") on v5.0, we reserve the right to be able to sleep within blk_release_queue() context.

The last reference for the request_queue must not be called from atomic context. *When* the last reference to the request_queue reaches 0 varies, and so let's take the opportunity to document when that is expected to happen and also document the context of the related calls as best as possible so we can avoid future issues, and with the hopes that the synchronous request_queue removal sticks.

We revert back to synchronous request_queue removal because asynchronous removal creates a regression with expected userspace interaction with several drivers. An example is when removing the loopback driver, one uses ioctls from userspace to do so, but upon return and if successful, one expects the device to be removed. Likewise if one races to add another device the new one may not be added as it is still being removed. This was expected behavior before and it now fails as the device is still present and busy still. Moving to asynchronous request_queue removal could have broken many scripts which relied on the removal to have been completed if there was no error. Document this expectation as well so that this doesn't regress userspace again.

Using asynchronous request_queue removal however has helped us find other bugs. In the future we can test what could break with this arrangement by enabling CONFIG_DEBUG_KOBJECT_RELEASE.

While at it, update the docs with the context expectations for the request_queue / gendisk refcount decrement, and make these expectations explicit by using might_sleep().

---
 block/blk-core.c       |  8 ++++++++
 block/blk-sysfs.c      | 43 +++++++++++++++++++++---------------------
 block/genhd.c          | 17 +++++++++++++++++
 include/linux/blkdev.h |  2 --
 4 files changed, 47 insertions(+), 23 deletions(-)

index 14c09daf55f3..a5126c0be777 100644
@@ -327,6 +327,9 @@ EXPORT_SYMBOL_GPL(blk_clear_pm_only);
  *
  * Decrements the refcount of the request_queue kobject. When this reaches 0
  * we'll have blk_release_queue() called.
+ *
+ * Context: Any context, but the last reference must not be dropped from
+ *          atomic context.
  */
 void blk_put_queue(struct request_queue *q)
 {

@@ -359,9 +362,14 @@ EXPORT_SYMBOL_GPL(blk_set_queue_dying);
  *
  * Mark @q DYING, drain all pending requests, mark @q DEAD, destroy and
  * put it.  All future requests will be failed immediately with -ENODEV.
+ *
+ * Context: can sleep
  */
 void blk_cleanup_queue(struct request_queue *q)
 {
+	/* cannot be called from atomic context */
+	might_sleep();
+
 	WARN_ON_ONCE(blk_queue_registered(q));
 
 	/* mark @q DYING, no new request or merges will be allowed afterwards */

index 1be86b1f43ec..60ae4e1b4d38 100644
@@ -889,12 +889,19 @@ static void invalidate_partition(struct gendisk *disk, int partno)
  * The final removal of the struct gendisk happens when its refcount reaches 0
  * with put_disk(), which should be called after del_gendisk(), if
  * __device_add_disk() was used.
+ *
+ * Drivers exist which depend on the release of the gendisk to be synchronous,
+ * it should not be deferred.
+ *
+ * Context: can sleep
  */
 void del_gendisk(struct gendisk *disk)
 {
 	struct disk_part_iter piter;
 	struct hd_struct *part;
 
+	might_sleep();
+
 	blk_integrity_del(disk);
 	disk_del_events(disk);
 

@@ -1548,11 +1555,15 @@ int disk_expand_part_tbl(struct gendisk *disk, int partno)
  * drivers we also call blk_put_queue() for them, and we expect the
  * request_queue refcount to reach 0 at this point, and so the request_queue
  * will also be freed prior to the disk.
+ *
+ * Context: can sleep
  */
 static void disk_release(struct device *dev)
 {
 	struct gendisk *disk = dev_to_disk(dev);
 
+	might_sleep();
+
 	blk_free_devt(dev->devt);
 	disk_release_events(disk);
 	kfree(disk->random);

@@ -1797,6 +1808,9 @@ EXPORT_SYMBOL(get_disk_and_module);
  *
  * This decrements the refcount for the struct gendisk. When this reaches 0
  * we'll have disk_release() called.
+ *
+ * Context: Any context, but the last reference must not be dropped from
+ *          atomic context.
  */
 void put_disk(struct gendisk *disk)
 {

@@ -1811,6 +1825,9 @@ EXPORT_SYMBOL(put_disk);
  *
  * This is a counterpart of get_disk_and_module() and thus also of
  * get_gendisk().
+ *
+ * Context: Any context, but the last reference must not be dropped from
+ *          atomic context.
  */
 void put_disk_and_module(struct gendisk *disk)
 {

index 6e067dca94cf..780d6fa94746 100644
@@ -584,8 +584,6 @@ struct request_queue {
 
 	size_t			cmd_size;
 
-	struct work_struct	release_work;
-
 #define BLK_MAX_WRITE_HINTS	5
 	u64			write_hints[BLK_MAX_WRITE_HINTS];
 };

2.26.2

Ensure it is clear which lock is required on do_blk_trace_setup().

---
 kernel/trace/blktrace.c | 2 ++
 1 file changed, 2 insertions(+)

index 7f60029bdaff..7ff2ea5cd05e 100644
@@ -483,6 +483,8 @@ static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
 	struct dentry *dir = NULL;
 	int ret;
 
+	lockdep_assert_held(&q->blk_trace_mutex);
+
 	if (!buts->buf_size || !buts->buf_nr)
 		return -EINVAL;
 

2.26.2

Be pedantic on removal as well and hold the mutex. This should prevent uses of addition while we exit.

---
 drivers/block/loop.c | 4 ++++
 1 file changed, 4 insertions(+)

index c33bbbfd1bd9..d55e1b52f076 100644
@@ -2402,6 +2402,8 @@ static void __exit loop_exit(void)
 
 	range = max_loop ? max_loop << part_shift : 1UL << MINORBITS;
 
+	mutex_lock(&loop_ctl_mutex);
+
 	idr_for_each(&loop_index_idr, &loop_exit_cb, NULL);
 	idr_destroy(&loop_index_idr);
 

@@ -2409,6 +2411,8 @@ static void __exit loop_exit(void)
 	unregister_blkdev(LOOP_MAJOR, "loop");
 
 	misc_deregister(&loop_misc);
+
+	mutex_unlock(&loop_ctl_mutex);
 }
 
 module_init(loop_init);

2.26.2

That's not the only way to add or remove the loop module though.

You may add/remove it manually. And again, as mentioned in the commit log, I couldn't trigger a race myself, however this seemed the more pedantic and careful strategy we can take.

Note: this will bring you sanity if you try to figure out *why* we still get:

[235530.144343] debugfs: Directory 'loop0' with parent 'block' already present!
[235530.149477] blktrace: debugfs_dir not present for loop0 so skipping
[235530.232328] debugfs: Directory 'loop0' with parent 'block' already present!
[235530.238962] blktrace: debugfs_dir not present for loop0 so skipping

If you run run_0004.sh from break-blktrace [0]. Even with all my patches merged we still run into this. And so the bug lies within the block layer or on the driver. I haven't been able to find the issue yet.

[0] https://github.com/mcgrof/break-blktrace

Luis

Thanks Luis for having shared this information. If I can find the time I will have a look into this myself.

Bart.

Let's keep track of it:

https://bugzilla.kernel.org/show_bug.cgi?id=208301

Luis

On 22/06/2020 14:27, Luis Chamberlain wrote:
[...]> If you run run_0004.sh from break-blktrace [0]. Even with all my patches

Would it be a good idea to merge this into blktests? Maybe start a blktrace section for it, which could host other regression test for blktrace.

Thoughts?

Absolutely! That is the goal as well.

Luis

On commit 6ac93117ab00 ("blktrace: use existing disk debugfs directory") merged on v4.12 Omar fixed the original blktrace code for request-based drivers (multiqueue). This however left in place a possible crash, if you happen to abuse blktrace while racing to remove / add a device.

We used to use asynchronous removal of the request_queue, and with that the issue was easier to reproduce. Now that we have reverted to synchronous removal of the request_queue, the issue is still possible to reproduce, its however just a bit more difficult.

We essentially run two instances of break-blktrace which add/remove a loop device, and setup a blktrace and just never tear the blktrace down. We do this twice in parallel. This is easily reproduced with the script run_0004.sh from break-blktrace [0].

We can end up with two types of panics each reflecting where we race, one a failed blktrace setup:

[  252.426751] debugfs: Directory 'loop0' with parent 'block' already present!
[  252.432265] BUG: kernel NULL pointer dereference, address: 00000000000000a0
[  252.436592] #PF: supervisor write access in kernel mode
[  252.439822] #PF: error_code(0x0002) - not-present page
[  252.442967] PGD 0 P4D 0
[  252.444656] Oops: 0002 [#1] SMP NOPTI
[  252.446972] CPU: 10 PID: 1153 Comm: break-blktrace Tainted: G            E     5.7.0-rc2-next-20200420+ #164
[  252.452673] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1 04/01/2014
[  252.456343] RIP: 0010:down_write+0x15/0x40
[  252.458146] Code: eb ca e8 ae 22 8d ff cc cc cc cc cc cc cc cc cc cc cc cc
               cc cc 0f 1f 44 00 00 55 48 89 fd e8 52 db ff ff 31 c0 ba 01 00
               00 00 <f0> 48 0f b1 55 00 75 0f 48 8b 04 25 c0 8b 01 00 48 89
               45 08 5d
[  252.463638] RSP: 0018:ffffa626415abcc8 EFLAGS: 00010246
[  252.464950] RAX: 0000000000000000 RBX: ffff958c25f0f5c0 RCX: ffffff8100000000
[  252.466727] RDX: 0000000000000001 RSI: ffffff8100000000 RDI: 00000000000000a0
[  252.468482] RBP: 00000000000000a0 R08: 0000000000000000 R09: 0000000000000001
[  252.470014] R10: 0000000000000000 R11: ffff958d1f9227ff R12: 0000000000000000
[  252.471473] R13: ffff958c25ea5380 R14: ffffffff8cce15f1 R15: 00000000000000a0
[  252.473346] FS:  00007f2e69dee540(0000) GS:ffff958c2fc80000(0000) knlGS:0000000000000000
[  252.475225] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  252.476267] CR2: 00000000000000a0 CR3: 0000000427d10004 CR4: 0000000000360ee0
[  252.477526] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[  252.478776] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[  252.479866] Call Trace:
[  252.480322]  simple_recursive_removal+0x4e/0x2e0
[  252.481078]  ? debugfs_remove+0x60/0x60
[  252.481725]  ? relay_destroy_buf+0x77/0xb0
[  252.482662]  debugfs_remove+0x40/0x60
[  252.483518]  blk_remove_buf_file_callback+0x5/0x10
[  252.484328]  relay_close_buf+0x2e/0x60
[  252.484930]  relay_open+0x1ce/0x2c0
[  252.485520]  do_blk_trace_setup+0x14f/0x2b0
[  252.486187]  __blk_trace_setup+0x54/0xb0
[  252.486803]  blk_trace_ioctl+0x90/0x140
[  252.487423]  ? do_sys_openat2+0x1ab/0x2d0
[  252.488053]  blkdev_ioctl+0x4d/0x260
[  252.488636]  block_ioctl+0x39/0x40
[  252.489139]  ksys_ioctl+0x87/0xc0
[  252.489675]  __x64_sys_ioctl+0x16/0x20
[  252.490380]  do_syscall_64+0x52/0x180
[  252.491032]  entry_SYSCALL_64_after_hwframe+0x44/0xa9

And the other on the device removal:

[  128.528940] debugfs: Directory 'loop0' with parent 'block' already present!
[  128.615325] BUG: kernel NULL pointer dereference, address: 00000000000000a0
[  128.619537] #PF: supervisor write access in kernel mode
[  128.622700] #PF: error_code(0x0002) - not-present page
[  128.625842] PGD 0 P4D 0
[  128.627585] Oops: 0002 [#1] SMP NOPTI
[  128.629871] CPU: 12 PID: 544 Comm: break-blktrace Tainted: G            E     5.7.0-rc2-next-20200420+ #164
[  128.635595] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1 04/01/2014
[  128.640471] RIP: 0010:down_write+0x15/0x40
[  128.643041] Code: eb ca e8 ae 22 8d ff cc cc cc cc cc cc cc cc cc cc cc cc
               cc cc 0f 1f 44 00 00 55 48 89 fd e8 52 db ff ff 31 c0 ba 01 00
               00 00 <f0> 48 0f b1 55 00 75 0f 65 48 8b 04 25 c0 8b 01 00 48 89
               45 08 5d
[  128.650180] RSP: 0018:ffffa9c3c05ebd78 EFLAGS: 00010246
[  128.651820] RAX: 0000000000000000 RBX: ffff8ae9a6370240 RCX: ffffff8100000000
[  128.653942] RDX: 0000000000000001 RSI: ffffff8100000000 RDI: 00000000000000a0
[  128.655720] RBP: 00000000000000a0 R08: 0000000000000002 R09: ffff8ae9afd2d3d0
[  128.657400] R10: 0000000000000056 R11: 0000000000000000 R12: 0000000000000000
[  128.659099] R13: 0000000000000000 R14: 0000000000000003 R15: 00000000000000a0
[  128.660500] FS:  00007febfd995540(0000) GS:ffff8ae9afd00000(0000) knlGS:0000000000000000
[  128.662204] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  128.663426] CR2: 00000000000000a0 CR3: 0000000420042003 CR4: 0000000000360ee0
[  128.664776] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[  128.666022] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[  128.667282] Call Trace:
[  128.667801]  simple_recursive_removal+0x4e/0x2e0
[  128.668663]  ? debugfs_remove+0x60/0x60
[  128.669368]  debugfs_remove+0x40/0x60
[  128.669985]  blk_trace_free+0xd/0x50
[  128.670593]  __blk_trace_remove+0x27/0x40
[  128.671274]  blk_trace_shutdown+0x30/0x40
[  128.671935]  blk_release_queue+0x95/0xf0
[  128.672589]  kobject_put+0xa5/0x1b0
[  128.673188]  disk_release+0xa2/0xc0
[  128.673786]  device_release+0x28/0x80
[  128.674376]  kobject_put+0xa5/0x1b0
[  128.674915]  loop_remove+0x39/0x50 [loop]
[  128.675511]  loop_control_ioctl+0x113/0x130 [loop]
[  128.676199]  ksys_ioctl+0x87/0xc0
[  128.676708]  __x64_sys_ioctl+0x16/0x20
[  128.677274]  do_syscall_64+0x52/0x180
[  128.677823]  entry_SYSCALL_64_after_hwframe+0x44/0xa9

The common theme here is:

debugfs: Directory 'loop0' with parent 'block' already present

This crash happens because of how blktrace uses the debugfs directory where it places its files. Upon init we always create the same directory which would be needed by blktrace but we only do this for make_request drivers (multiqueue) block drivers. When you race a removal of these devices with a blktrace setup you end up in a situation where the make_request recursive debugfs removal will sweep away the blktrace files and then later blktrace will also try to remove individual dentries which are already NULL. The inverse is also possible and hence the two types of use after frees.

We don't create the block debugfs directory on init for these types of block devices:

* request-based block driver block devices
* every possible partition
* scsi-generic

And so, this race should in theory only be possible with make_request drivers.

We can fix the UAF by simply re-using the debugfs directory for make_request drivers (multiqueue) and only creating the ephemeral directory for the other type of block devices. The new clarifications on relying on the q->blk_trace_mutex *and* also checking for q->blk_trace *prior* to processing a blktrace ensures the debugfs directories are only created if no possible directory name clashes are possible.

This goes tested with:

o nvme partitions
o ISCSI with tgt, and blktracing against scsi-generic with:
o block
o tape
o cdrom
o media changer
o blktests

This patch is part of the work which disputes the severity of CVE-2019-19770 which shows this issue is not a core debugfs issue, but a misuse of debugfs within blktace.

---
 kernel/trace/blktrace.c | 18 ++++++++++++------
 1 file changed, 12 insertions(+), 6 deletions(-)

index 7ff2ea5cd05e..e6e2d25fdbd6 100644
@@ -524,10 +524,18 @@ static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
 	if (!bt->msg_data)
 		goto err;
 
-	ret = -ENOENT;
-
-	dir = debugfs_lookup(buts->name, blk_debugfs_root);
-	if (!dir)
+#ifdef CONFIG_BLK_DEBUG_FS
+	/*
+	 * When tracing whole make_request drivers (multiqueue) block devices,
+	 * reuse the existing debugfs directory created by the block layer on
+	 * init. For request-based block devices, all partitions block devices,
+	 * and scsi-generic block devices we create a temporary new debugfs
+	 * directory that will be removed once the trace ends.
+	 */
+	if (queue_is_mq(q) && bdev && bdev == bdev->bd_contains)
+		dir = q->debugfs_dir;
+	else
+#endif
 		bt->dir = dir = debugfs_create_dir(buts->name, blk_debugfs_root);
 
 	bt->dev = dev;

@@ -565,8 +573,6 @@ static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
 
 	ret = 0;
 err:
-	if (dir && !bt->dir)
-		dput(dir);
 	if (ret)
 		blk_trace_free(bt);
 	return ret;

2.26.2

Still not a fan of the wall of text in the commit log, but the changes look good:

       ^^^^
       got?

                                                  ^^^^^^^^^^^^^^^^^^^^^

Yes, the later.

Great question, the answer is no. The reason is that we first use the mutex and then we check for q->blk_trace. If you hold the lock *and* you have checked for q->blk_trace and its NULL, you are sure you should not have a duplicate.

Its why the commit log mentioned:

The new clarifications on relying on the q->blk_trace_mutex *and* also checking for q->blk_trace *prior* to processing a blktrace ensures the debugfs directories are only created if no possible directory name clashes are possible.

These clarifications were prompted through discussions with Jan Kara on the patches he posted which you CC'd me on. I agreed with his patch *but* I suggested it would hold true only if check for the q->blk_trace first, and this is why my patch titled "blktrace: break out of blktrace setup on concurrent calls" got merged prior to Jan Kara's "blktrace: Avoid sparse warnings when assigning q->blk_trace".

If there is an error, bt->dir will be removed still, as I never modified the removal of bt->dir in this patch.

Luis

We make an assumption that a debugfs directory exists, but since this can fail ensure it exists before allowing blktrace setup to complete. Otherwise we end up stuffing blktrace files on the debugfs root directory. In the worst case scenario this *in theory* can create an eventual panic *iff* in the future a similarly named file is created prior on the debugfs root directory. This theoretical crash can happen due to a recursive removal followed by a specific dentry removal.

This doesn't fix any known crash, however I have seen the files go into the main debugfs root directory in cases where the debugfs directory was not created due to other internal bugs with blktrace now fixed.

blktrace is also completely useless without this directory, so this ensures to userspace we only setup blktrace if the kernel can stuff files where they are supposed to go into.

debugfs directory creations typically aren't checked for, and we have maintainers doing sweep removals of these checks, but since we need this check to ensure proper userspace blktrace functionality we make sure to annotate the justification for the check.

---
 kernel/trace/blktrace.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

index e6e2d25fdbd6..098780ec018f 100644
@@ -538,6 +538,18 @@ static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
 #endif
 		bt->dir = dir = debugfs_create_dir(buts->name, blk_debugfs_root);
 
+	/*
+	 * As blktrace relies on debugfs for its interface the debugfs directory
+	 * is required, contrary to the usual mantra of not checking for debugfs
+	 * files or directories.
+	 */
+	if (IS_ERR_OR_NULL(dir)) {
+		pr_warn("debugfs_dir not present for %s so skipping\n",
+			buts->name);
+		ret = -ENOENT;
+		goto err;
+	}
+
 	bt->dev = dev;
 	atomic_set(&bt->dropped, 0);
 	INIT_LIST_HEAD(&bt->running_list);

2.26.2

Looks good,

We were only creating the request_queue debugfs_dir only for make_request block drivers (multiqueue), but never for request-based block drivers. We did this as we were only creating non-blktrace additional debugfs files on that directory for make_request drivers. However, since blktrace *always* creates that directory anyway, we special-case the use of that directory on blktrace. Other than this being an eye-sore, this exposes request-based block drivers to the same debugfs fragile race that used to exist with make_request block drivers where if we start adding files onto that directory we can later run a race with a double removal of dentries on the directory if we don't deal with this carefully on blktrace.

Instead, just simplify things by always creating the request_queue debugfs_dir on request_queue registration. Rename the mutex also to reflect the fact that this is used outside of the blktrace context.

---
 block/blk-core.c        |  8 +-----
 block/blk-mq-debugfs.c  |  5 ----
 block/blk-sysfs.c       |  9 +++++++
 block/blk.h             |  2 --
 include/linux/blkdev.h  |  5 ++--
 kernel/trace/blktrace.c | 58 ++++++++++++++++++-----------------------
 6 files changed, 39 insertions(+), 48 deletions(-)

index a5126c0be777..72b102a389a5 100644
@@ -51,9 +51,7 @@
 #include "blk-pm.h"
 #include "blk-rq-qos.h"
 
-#ifdef CONFIG_DEBUG_FS
 struct dentry *blk_debugfs_root;
-#endif
 
 EXPORT_TRACEPOINT_SYMBOL_GPL(block_bio_remap);
 EXPORT_TRACEPOINT_SYMBOL_GPL(block_rq_remap);

@@ -555,9 +553,7 @@ struct request_queue *__blk_alloc_queue(int node_id)
 
 	kobject_init(&q->kobj, &blk_queue_ktype);
 
-#ifdef CONFIG_BLK_DEV_IO_TRACE
-	mutex_init(&q->blk_trace_mutex);
-#endif
+	mutex_init(&q->debugfs_mutex);
 	mutex_init(&q->sysfs_lock);
 	mutex_init(&q->sysfs_dir_lock);
 	spin_lock_init(&q->queue_lock);

@@ -1937,9 +1933,7 @@ int __init blk_dev_init(void)
 	blk_requestq_cachep = kmem_cache_create("request_queue",
 			sizeof(struct request_queue), 0, SLAB_PANIC, NULL);
 
-#ifdef CONFIG_DEBUG_FS
 	blk_debugfs_root = debugfs_create_dir("block", NULL);
-#endif
 
 	return 0;
 }

index 15df3a36e9fa..a2800bc56fb4 100644
@@ -824,9 +824,6 @@ void blk_mq_debugfs_register(struct request_queue *q)
 	struct blk_mq_hw_ctx *hctx;
 	int i;
 
-	q->debugfs_dir = debugfs_create_dir(kobject_name(q->kobj.parent),
-					    blk_debugfs_root);
-
 	debugfs_create_files(q->debugfs_dir, q, blk_mq_debugfs_queue_attrs);
 
 	/*

@@ -857,9 +854,7 @@ void blk_mq_debugfs_register(struct request_queue *q)
 
 void blk_mq_debugfs_unregister(struct request_queue *q)
 {
-	debugfs_remove_recursive(q->debugfs_dir);
 	q->sched_debugfs_dir = NULL;
-	q->debugfs_dir = NULL;
 }
 
 static void blk_mq_debugfs_register_ctx(struct blk_mq_hw_ctx *hctx,

index 561624d4cc4e..be67952e7be2 100644
@@ -11,6 +11,7 @@
 #include <linux/blktrace_api.h>
 #include <linux/blk-mq.h>
 #include <linux/blk-cgroup.h>
+#include <linux/debugfs.h>
 
 #include "blk.h"
 #include "blk-mq.h"

@@ -917,6 +918,9 @@ static void blk_release_queue(struct kobject *kobj)
 		blk_mq_release(q);
 
 	blk_trace_shutdown(q);
+	mutex_lock(&q->debugfs_mutex);
+	debugfs_remove_recursive(q->debugfs_dir);
+	mutex_unlock(&q->debugfs_mutex);
 
 	if (queue_is_mq(q))
 		blk_mq_debugfs_unregister(q);

@@ -989,6 +993,11 @@ int blk_register_queue(struct gendisk *disk)
 		goto unlock;
 	}
 
+	mutex_lock(&q->debugfs_mutex);
+	q->debugfs_dir = debugfs_create_dir(kobject_name(q->kobj.parent),
+					    blk_debugfs_root);
+	mutex_unlock(&q->debugfs_mutex);
+
 	if (queue_is_mq(q)) {
 		__blk_mq_register_dev(dev, q);
 		blk_mq_debugfs_register(q);

index b5d1f0fc6547..499308c6ab3b 100644
@@ -14,9 +14,7 @@
 /* Max future timer expiry for timeouts */
 #define BLK_MAX_TIMEOUT		(5 * HZ)
 
-#ifdef CONFIG_DEBUG_FS
 extern struct dentry *blk_debugfs_root;
-#endif
 
 struct blk_flush_queue {
 	unsigned int		flush_pending_idx:1;

index 780d6fa94746..e70e03d2cd8c 100644
@@ -528,9 +528,9 @@ struct request_queue {
 	unsigned int		sg_timeout;
 	unsigned int		sg_reserved_size;
 	int			node;
+	struct mutex		debugfs_mutex;
 #ifdef CONFIG_BLK_DEV_IO_TRACE
 	struct blk_trace __rcu	*blk_trace;
-	struct mutex		blk_trace_mutex;
 #endif
 	/*
 	 * for flush operations

@@ -574,8 +574,9 @@ struct request_queue {
 	struct list_head	tag_set_list;
 	struct bio_set		bio_split;
 
-#ifdef CONFIG_BLK_DEBUG_FS
 	struct dentry		*debugfs_dir;
+
+#ifdef CONFIG_BLK_DEBUG_FS
 	struct dentry		*sched_debugfs_dir;
 	struct dentry		*rqos_debugfs_dir;
 #endif

2.26.2

Looks good,

There are two changes in this patch: a bug fix and a rename of a mutex. I don't like it to see two changes in a single patch.

Additionally, is the new mutex name really better than the old name? The proper way to use mutexes is to use mutexes to protect data instead of code. Where is the documentation that mentions which member variable(s) of which data structures are protected by the mutex formerly called blk_trace_mutex? Since the new name makes it even less clear which data is protected by this mutex, is the new name really better than the old name?

Thanks,

Bart.

I thought about doing the split first, and I did it at first, but then I could hear Christoph yelling at me for it. So I merged the two together. Although it makes it more difficult for review, the changes do go together.

Kind of late to split this as its already merged now.

It does not exist, and that is the point. The debugfs_dir use after free showed us *when* that UAF can happen, and so care must be taken if we are to use the mutex to protect the debugfs_dir but also re-use the same directory for other block core shenanigans.

I thought the new name makes it crystal clear what is being protected. I can however add a comment to explain that the q->debugfs_mutex protects the q->debugfs_dir if it is created, otherwise it protects the ephemeral debugfs_dir directory which would otherwise be created in lieue of q->debugfs_dir, however the patch still lies under <debugfs_root>/block/.

Let me know if you think that will help.

Luis

During the past weeks I have been more busy than usual. I will try to make sure that in the future I have the time to read all comments on the previous versions of a patch series before replying to the latest version of a patch series.

My concern is that q->debugfs_mutex would evolve the same way as q->sysfs_lock: at the time of introduction the role of a mutex is very clear but over time the number of use cases grows to a point where it is no longer possible to recognize the original purpose. I think there are two possible approaches: either a comment is added now that explains the role of q->debugfs_mutex or someone who has followed this conversation yells when someone tries to use q->debugfs_mutex for another purpose than what it was intended for.

Thanks,

Bart.

Applied for 5.9, thanks.

Jens Axboe