[PATCH v6 0/8] block: fix blktrace debugfs use after free

[PATCH v6 0/8] block: fix blktrace debugfs use after free

[Luis Chamberlain]

Here is v6 of the blktrace fixes which address the debugfs use after
free. I've followed the strategy suggested by Christoph of open coding
the solution in place, and extended it with the required work for
partitions and scsi-generic. Jan's blktrace sparse fix ended up
depending on one of my patch, "blktrace: break out of blktrace setup
on concurrent calls", and so he has sent those for inclusion prior to my
series. This series would have to be applied after those two patches
from Jan are merged then.

Since the patch "blktrace: fix debugfs use after free" ends up being the
only one modified lately, I've moved that patch to be the last one in
the series now.

You can find these on my git tree branch 20200608-blktrace-fixes based
on linux-next 20200608 [0].

Hopefully this is it.

[0] https://git.kernel.org/pub/scm/linux/kernel/git/mcgrof/linux-next.git/log/?h=20200608-blktrace-fixes

Luis Chamberlain (6):
  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

 block/blk-core.c             | 27 ++++++++++--
 block/blk-mq-debugfs.c       |  5 ---
 block/blk-sysfs.c            | 83 +++++++++++++++++++++++++++---------
 block/blk.h                  |  2 -
 block/genhd.c                | 73 ++++++++++++++++++++++++++++++-
 block/partitions/core.c      |  3 ++
 drivers/block/loop.c         |  4 ++
 drivers/scsi/sg.c            |  3 ++
 include/linux/blkdev.h       |  6 +--
 include/linux/blktrace_api.h |  1 -
 include/linux/genhd.h        |  1 +
 kernel/trace/blktrace.c      | 55 ++++++++++++++++++------
 12 files changed, 214 insertions(+), 49 deletions(-)

-- 
2.26.2

[PATCH v6 1/6] block: add docs for gendisk / request_queue refcount helpers

[Luis Chamberlain]

This adds documentation for the gendisk / request_queue refcount
helpers.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
---
 block/blk-core.c | 13 +++++++++++++
 block/genhd.c    | 50 +++++++++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 62 insertions(+), 1 deletion(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 62a4904db921..a0760aac110a 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -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))) {
diff --git a/block/genhd.c b/block/genhd.c
index 1a7659327664..f741613d731f 100644
--- a/block/genhd.c
+++ b/block/genhd.c
@@ -876,6 +876,20 @@ static void invalidate_partition(struct gendisk *disk, int partno)
 	bdput(bdev);
 }
 
+/**
+ * del_gendisk - remove the gendisk
+ * @disk: the struct gendisk to remove
+ *
+ * Removes the gendisk and all its associated resources. This deletes the
+ * partitions associated with the gendisk, and unregisters the associated
+ * request_queue.
+ *
+ * This is the counter to the respective __device_add_disk() call.
+ *
+ * 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.
+ */
 void del_gendisk(struct gendisk *disk)
 {
 	struct disk_part_iter piter;
@@ -1514,6 +1528,23 @@ int disk_expand_part_tbl(struct gendisk *disk, int partno)
 	return 0;
 }
 
+/**
+ * disk_release - releases all allocated resources of the gendisk
+ * @dev: the device representing this disk
+ *
+ * This function releases all allocated resources of the gendisk.
+ *
+ * The struct gendisk refcount is incremented with get_gendisk() or
+ * get_disk_and_module(), and its refcount is decremented with
+ * put_disk_and_module() or put_disk(). Once the refcount reaches 0 this
+ * function is called.
+ *
+ * Drivers which used __device_add_disk() have a gendisk with a request_queue
+ * assigned. Since the request_queue sits on top of the gendisk for these
+ * 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.
+ */
 static void disk_release(struct device *dev)
 {
 	struct gendisk *disk = dev_to_disk(dev);
@@ -1727,6 +1758,13 @@ struct gendisk *__alloc_disk_node(int minors, int node_id)
 }
 EXPORT_SYMBOL(__alloc_disk_node);
 
+/**
+ * get_disk_and_module - increments the gendisk and gendisk fops module refcount
+ * @disk: the struct gendisk to to increment the refcount for
+ *
+ * This increments the refcount for the struct gendisk, and the gendisk's
+ * fops module owner.
+ */
 struct kobject *get_disk_and_module(struct gendisk *disk)
 {
 	struct module *owner;
@@ -1747,6 +1785,13 @@ struct kobject *get_disk_and_module(struct gendisk *disk)
 }
 EXPORT_SYMBOL(get_disk_and_module);
 
+/**
+ * put_disk - decrements the gendisk refcount
+ * @disk: the struct gendisk to to decrement the refcount for
+ *
+ * This decrements the refcount for the struct gendisk. When this reaches 0
+ * we'll have disk_release() called.
+ */
 void put_disk(struct gendisk *disk)
 {
 	if (disk)
@@ -1754,7 +1799,10 @@ void put_disk(struct gendisk *disk)
 }
 EXPORT_SYMBOL(put_disk);
 
-/*
+/**
+ * put_disk_and_module - decrements the module and gendisk refcount
+ * @disk: the struct gendisk to to decrement the refcount for
+ *
  * This is a counterpart of get_disk_and_module() and thus also of
  * get_gendisk().
  */
-- 
2.26.2

[PATCH v6 2/6] block: clarify context for refcount increment helpers

[Luis Chamberlain]

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.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
---
 block/blk-core.c | 2 ++
 block/genhd.c    | 6 ++++++
 2 files changed, 8 insertions(+)

diff --git a/block/blk-core.c b/block/blk-core.c
index a0760aac110a..14c09daf55f3 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -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)
 {
diff --git a/block/genhd.c b/block/genhd.c
index f741613d731f..1be86b1f43ec 100644
--- a/block/genhd.c
+++ b/block/genhd.c
@@ -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

[PATCH v6 3/6] block: revert back to synchronous request_queue removal

[Luis Chamberlain]

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 db6d9952356 ("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().

Cc: Bart Van Assche <bvanassche@acm.org>
Cc: Omar Sandoval <osandov@fb.com>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Nicolai Stange <nstange@suse.de>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: yu kuai <yukuai3@huawei.com>
Suggested-by: Nicolai Stange <nstange@suse.de>
Fixes: dc9edc44de6c ("block: Fix a blk_exit_rl() regression")
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
---
 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(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 14c09daf55f3..a5126c0be777 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -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 */
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 02643e149d5e..561624d4cc4e 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -873,22 +873,32 @@ static void blk_exit_queue(struct request_queue *q)
 	bdi_put(q->backing_dev_info);
 }
 
-
 /**
- * __blk_release_queue - release a request queue
- * @work: pointer to the release_work member of the request queue to be released
+ * blk_release_queue - releases all allocated resources of the request_queue
+ * @kobj: pointer to a kobject, whose container is a request_queue
+ *
+ * This function releases all allocated resources of the request queue.
+ *
+ * The struct request_queue refcount is incremented with blk_get_queue() and
+ * decremented with blk_put_queue(). Once the refcount reaches 0 this function
+ * is called.
+ *
+ * For drivers that have a request_queue on a gendisk and added with
+ * __device_add_disk() the refcount to request_queue will reach 0 with
+ * the last put_disk() called by the driver. For drivers which don't use
+ * __device_add_disk() this happens with blk_cleanup_queue().
  *
- * Description:
- *     This function is called when a block device is being unregistered. The
- *     process of releasing a request queue starts with blk_cleanup_queue, which
- *     set the appropriate flags and then calls blk_put_queue, that decrements
- *     the reference counter of the request queue. Once the reference counter
- *     of the request queue reaches zero, blk_release_queue is called to release
- *     all allocated resources of the request queue.
+ * Drivers exist which depend on the release of the request_queue to be
+ * synchronous, it should not be deferred.
+ *
+ * Context: can sleep
  */
-static void __blk_release_queue(struct work_struct *work)
+static void blk_release_queue(struct kobject *kobj)
 {
-	struct request_queue *q = container_of(work, typeof(*q), release_work);
+	struct request_queue *q =
+		container_of(kobj, struct request_queue, kobj);
+
+	might_sleep();
 
 	if (test_bit(QUEUE_FLAG_POLL_STATS, &q->queue_flags))
 		blk_stat_remove_callback(q, q->poll_cb);
@@ -917,15 +927,6 @@ static void __blk_release_queue(struct work_struct *work)
 	call_rcu(&q->rcu_head, blk_free_queue_rcu);
 }
 
-static void blk_release_queue(struct kobject *kobj)
-{
-	struct request_queue *q =
-		container_of(kobj, struct request_queue, kobj);
-
-	INIT_WORK(&q->release_work, __blk_release_queue);
-	schedule_work(&q->release_work);
-}
-
 static const struct sysfs_ops queue_sysfs_ops = {
 	.show	= queue_attr_show,
 	.store	= queue_attr_store,
diff --git a/block/genhd.c b/block/genhd.c
index 1be86b1f43ec..60ae4e1b4d38 100644
--- a/block/genhd.c
+++ b/block/genhd.c
@@ -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)
 {
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 27887bf36d50..2462b78c1013 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -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

Re: [PATCH v6 3/6] block: revert back to synchronous request_queue removal

[Bart Van Assche]

On 2020-06-08 10:01, Luis Chamberlain wrote:
> + * Drivers exist which depend on the release of the request_queue to be
> + * synchronous, it should not be deferred.

This sounds mysterious. Which drivers? Why do these depend on this
function being synchronous? Anyway:

Reviewed-by: Bart Van Assche <bvanassche@acm.org>

Re: [PATCH v6 3/6] block: revert back to synchronous request_queue removal

[Luis Chamberlain]

On Fri, Jun 12, 2020 at 06:53:40PM -0700, Bart Van Assche wrote:
> On 2020-06-08 10:01, Luis Chamberlain wrote:
> > + * Drivers exist which depend on the release of the request_queue to be
> > + * synchronous, it should not be deferred.
> 
> This sounds mysterious. Which drivers? Why do these depend on this
> function being synchronous?

Sorry that should be "Userspace can exist". I've fixed that.

> Anyway:
> 
> Reviewed-by: Bart Van Assche <bvanassche@acm.org>

  Luis

[PATCH v6 4/6] blktrace: annotate required lock on do_blk_trace_setup()

[Luis Chamberlain]

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

Suggested-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
---
 kernel/trace/blktrace.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/kernel/trace/blktrace.c b/kernel/trace/blktrace.c
index 7f60029bdaff..7ff2ea5cd05e 100644
--- a/kernel/trace/blktrace.c
+++ b/kernel/trace/blktrace.c
@@ -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

Re: [PATCH v6 4/6] blktrace: annotate required lock on do_blk_trace_setup()

[Christoph Hellwig]

On Mon, Jun 08, 2020 at 05:01:24PM +0000, Luis Chamberlain wrote:
> Ensure it is clear which lock is required on do_blk_trace_setup().
> 
> Suggested-by: Bart Van Assche <bvanassche@acm.org>
> Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>

Looks good,

Reviewed-by: Christoph Hellwig <hch@lst.de>

Re: [PATCH v6 4/6] blktrace: annotate required lock on do_blk_trace_setup()

[Bart Van Assche]

On 2020-06-08 10:01, Luis Chamberlain wrote:
> Ensure it is clear which lock is required on do_blk_trace_setup().

Reviewed-by: Bart Van Assche <bvanassche@acm.org>

[PATCH v6 5/6] loop: be paranoid on exit and prevent new additions / removals

[Luis Chamberlain]

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

Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
---
 drivers/block/loop.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/drivers/block/loop.c b/drivers/block/loop.c
index c33bbbfd1bd9..d55e1b52f076 100644
--- a/drivers/block/loop.c
+++ b/drivers/block/loop.c
@@ -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

[PATCH v6 6/6] blktrace: fix debugfs use after free

[Luis Chamberlain]

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, but never for request-based block
drivers. Furthermore, that directory is only created on init for the
entire disk. This means that if you use blktrace on a partition, we'll
always be creating a new directory regardless of whether or not you
are doing blktrace on a make_request driver (multiqueue) or a
request-based block drivers.

These directory creations are only associated with a path, and so
when a debugfs_remove() is called it removes everything in its way.
A device removal will remove all blktrace files, and so if a blktrace
is still present a cleanup of blktrace files later will end up trying
to remove dentries pointing to NULL.

We can fix the UAF by using a debugfs directory which moving forward
will always be accessible if debugfs is enabled for both make_request
drivers (multiqueue) and request-based block drivers, *and* for all
partitions upon creation. This ensures that removal of the directories
only happens on device removal and removes the race of the files
underneath an active blktrace.

We special-case a solution for scsi-generic as well which got blktrace
support added by Christof via commit 6da127ad0918 ("blktrace: Add
blktrace ioctls to SCSI generic devices") so upstream since v2.6.25.
scsi-generic drives use a character device, however behind the scenes
we have a scsi device with a request_queue. How this is used varies
by class of driver (TYPE_DISK, TYPE_TAPE, etc).  We simply create its
a scsi-generic dedicated debugfs_dir and have blktrace use that when
this interface is used.

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.

Cc: Bart Van Assche <bvanassche@acm.org>
Cc: Omar Sandoval <osandov@fb.com>
Cc: Hannes Reinecke <hare@suse.com>
Cc: Nicolai Stange <nstange@suse.de>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: "Martin K. Petersen" <martin.petersen@oracle.com>
Cc: "James E.J. Bottomley" <jejb@linux.ibm.com>
Cc: yu kuai <yukuai3@huawei.com>
Reported-by: syzbot+603294af2d01acfdd6da@syzkaller.appspotmail.com
Fixes: 6ac93117ab00 ("blktrace: use existing disk debugfs directory")
Signed-off-by: Luis Chamberlain <mcgrof@kernel.org>
---
 block/blk-core.c             |  4 ---
 block/blk-mq-debugfs.c       |  5 ----
 block/blk-sysfs.c            | 40 +++++++++++++++++++++++++++
 block/blk.h                  |  2 --
 block/partitions/core.c      |  3 ++
 drivers/scsi/sg.c            |  3 ++
 include/linux/blkdev.h       |  4 ++-
 include/linux/blktrace_api.h |  1 -
 include/linux/genhd.h        |  1 +
 kernel/trace/blktrace.c      | 53 ++++++++++++++++++++++++++++--------
 10 files changed, 91 insertions(+), 25 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index a5126c0be777..fd850bebdf18 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -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);
@@ -1937,9 +1935,7 @@ 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;
 }
diff --git a/block/blk-mq-debugfs.c b/block/blk-mq-debugfs.c
index 15df3a36e9fa..a2800bc56fb4 100644
--- a/block/blk-mq-debugfs.c
+++ b/block/blk-mq-debugfs.c
@@ -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,
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 561624d4cc4e..70168435f079 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -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"
@@ -918,6 +919,9 @@ static void blk_release_queue(struct kobject *kobj)
 
 	blk_trace_shutdown(q);
 
+	debugfs_remove_recursive(q->debugfs_dir);
+	if (IS_ENABLED(CONFIG_CHR_DEV_SG))
+		debugfs_remove_recursive(q->sg_debugfs_dir);
 	if (queue_is_mq(q))
 		blk_mq_debugfs_unregister(q);
 
@@ -937,6 +941,20 @@ struct kobj_type blk_queue_ktype = {
 	.release	= blk_release_queue,
 };
 
+/**
+ * blk_sg_debugfs_init - initialize debugfs for scsi-generic
+ * @q: the associated queue
+ * @name: name of the scsi-generic device
+ *
+ * To be used by scsi-generic for allowing it to use blktrace.
+ */
+void blk_sg_debugfs_init(struct request_queue *q, const char *name)
+{
+	if (IS_ENABLED(CONFIG_CHR_DEV_SG))
+		q->sg_debugfs_dir = debugfs_create_dir(name, blk_debugfs_root);
+}
+EXPORT_SYMBOL_GPL(blk_sg_debugfs_init);
+
 /**
  * blk_register_queue - register a block layer queue with sysfs
  * @disk: Disk of which the request queue should be registered with sysfs.
@@ -989,6 +1007,28 @@ int blk_register_queue(struct gendisk *disk)
 		goto unlock;
 	}
 
+	/*
+	 * Blktrace needs a debugfs name even for queues that don't register
+	 * a gendisk, so it lazily registers the debugfs directory.  But that
+	 * can get us into a situation where a SCSI device is found, with no
+	 * driver for it (yet).  Then blktrace is used on the device, creating
+	 * the debugfs directory, and only after that a driver is loaded. In
+	 * that case we might already have a debugfs directory registered here.
+	 * Even worse we could be racing with blktrace to register it.
+	 */
+#ifdef CONFIG_BLK_DEV_IO_TRACE
+	mutex_lock(&q->blk_trace_mutex);
+	if (!q->debugfs_dir) {
+		q->debugfs_dir =
+			debugfs_create_dir(kobject_name(q->kobj.parent),
+				blk_debugfs_root);
+	}
+	mutex_unlock(&q->blk_trace_mutex);
+#else
+	q->debugfs_dir = debugfs_create_dir(kobject_name(q->kobj.parent),
+					    blk_debugfs_root);
+#endif
+
 	if (queue_is_mq(q)) {
 		__blk_mq_register_dev(dev, q);
 		blk_mq_debugfs_register(q);
diff --git a/block/blk.h b/block/blk.h
index b5d1f0fc6547..499308c6ab3b 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -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;
diff --git a/block/partitions/core.c b/block/partitions/core.c
index 78951e33b2d7..8387128364f1 100644
--- a/block/partitions/core.c
+++ b/block/partitions/core.c
@@ -10,6 +10,7 @@
 #include <linux/vmalloc.h>
 #include <linux/blktrace_api.h>
 #include <linux/raid/detect.h>
+#include <linux/debugfs.h>
 #include "check.h"
 
 static int (*check_part[])(struct parsed_partitions *) = {
@@ -322,6 +323,7 @@ void delete_partition(struct gendisk *disk, struct hd_struct *part)
 	get_device(disk_to_dev(part_to_disk(part)));
 	rcu_assign_pointer(ptbl->part[part->partno], NULL);
 	kobject_put(part->holder_dir);
+	debugfs_remove_recursive(part->debugfs_dir);
 	device_del(part_to_dev(part));
 
 	/*
@@ -444,6 +446,7 @@ static struct hd_struct *add_partition(struct gendisk *disk, int partno,
 	if (!p->holder_dir)
 		goto out_del;
 
+	p->debugfs_dir = debugfs_create_dir(dev_name(pdev), blk_debugfs_root);
 	dev_set_uevent_suppress(pdev, 0);
 	if (flags & ADDPART_FLAG_WHOLEDISK) {
 		err = device_create_file(pdev, &dev_attr_whole_disk);
diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
index 20472aaaf630..5f6ccf4ba4d9 100644
--- a/drivers/scsi/sg.c
+++ b/drivers/scsi/sg.c
@@ -1117,6 +1117,9 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,
 		return put_user(max_sectors_bytes(sdp->device->request_queue),
 				ip);
 	case BLKTRACESETUP:
+		if (!sdp->device->request_queue->sg_debugfs_dir)
+			blk_sg_debugfs_init(sdp->device->request_queue,
+					    sdp->disk->disk_name);
 		return blk_trace_setup(sdp->device->request_queue,
 				       sdp->disk->disk_name,
 				       MKDEV(SCSI_GENERIC_MAJOR, sdp->index),
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 2462b78c1013..afc43c8923c5 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -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;
+	struct dentry		*sg_debugfs_dir;
+#ifdef CONFIG_BLK_DEBUG_FS
 	struct dentry		*sched_debugfs_dir;
 	struct dentry		*rqos_debugfs_dir;
 #endif
@@ -858,6 +859,7 @@ static inline void rq_flush_dcache_pages(struct request *rq)
 
 extern int blk_register_queue(struct gendisk *disk);
 extern void blk_unregister_queue(struct gendisk *disk);
+extern void blk_sg_debugfs_init(struct request_queue *q, const char *name);
 extern blk_qc_t generic_make_request(struct bio *bio);
 extern blk_qc_t direct_make_request(struct bio *bio);
 extern void blk_rq_init(struct request_queue *q, struct request *rq);
diff --git a/include/linux/blktrace_api.h b/include/linux/blktrace_api.h
index 3b6ff5902edc..eb6db276e293 100644
--- a/include/linux/blktrace_api.h
+++ b/include/linux/blktrace_api.h
@@ -22,7 +22,6 @@ struct blk_trace {
 	u64 end_lba;
 	u32 pid;
 	u32 dev;
-	struct dentry *dir;
 	struct dentry *dropped_file;
 	struct dentry *msg_file;
 	struct list_head running_list;
diff --git a/include/linux/genhd.h b/include/linux/genhd.h
index 392aad5e29a2..902e50808bd9 100644
--- a/include/linux/genhd.h
+++ b/include/linux/genhd.h
@@ -76,6 +76,7 @@ struct hd_struct {
 	int make_it_fail;
 #endif
 	struct rcu_work rcu_work;
+	struct dentry *debugfs_dir;
 };
 
 /**
diff --git a/kernel/trace/blktrace.c b/kernel/trace/blktrace.c
index 7ff2ea5cd05e..4690d70e16a4 100644
--- a/kernel/trace/blktrace.c
+++ b/kernel/trace/blktrace.c
@@ -314,7 +314,6 @@ static void blk_trace_free(struct blk_trace *bt)
 	debugfs_remove(bt->msg_file);
 	debugfs_remove(bt->dropped_file);
 	relay_close(bt->rchan);
-	debugfs_remove(bt->dir);
 	free_percpu(bt->sequence);
 	free_percpu(bt->msg_data);
 	kfree(bt);
@@ -488,9 +487,6 @@ static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
 	if (!buts->buf_size || !buts->buf_nr)
 		return -EINVAL;
 
-	if (!blk_debugfs_root)
-		return -ENOENT;
-
 	strncpy(buts->name, name, BLKTRACE_BDEV_SIZE);
 	buts->name[BLKTRACE_BDEV_SIZE - 1] = '\0';
 
@@ -511,6 +507,47 @@ static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
 		return -EBUSY;
 	}
 
+	/*
+	 * We have to use a partition directory if a partition is being worked
+	 * on. The same request_queue is shared between all partitions.
+	 */
+	if (bdev && bdev != bdev->bd_contains) {
+		dir = bdev->bd_part->debugfs_dir;
+	} else if (IS_ENABLED(CONFIG_CHR_DEV_SG) &&
+		   MAJOR(dev) == SCSI_GENERIC_MAJOR) {
+		/*
+		 * scsi-generic exposes the request_queue through the /dev/sg*
+		 * interface but since that uses a different path than whatever
+		 * the respective scsi driver device name may expose and use
+		 * for the request_queue debugfs_dir. We have a dedicated
+		 * dentry for scsi-generic then.
+		 */
+		dir = q->sg_debugfs_dir;
+	} else {
+		/*
+		 * Drivers which do not use the *add_disk*() interfaces will
+		 * not have their debugfs_dir created for them automatically,
+		 * so we must create it for them.
+		 */
+		if (!q->debugfs_dir) {
+			q->debugfs_dir =
+				debugfs_create_dir(buts->name,
+						   blk_debugfs_root);
+		}
+		dir = q->debugfs_dir;
+	}
+
+	/*
+	 * 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);
+		return -ENOENT;
+	}
+
 	bt = kzalloc(sizeof(*bt), GFP_KERNEL);
 	if (!bt)
 		return -ENOMEM;
@@ -524,12 +561,6 @@ 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)
-		bt->dir = dir = debugfs_create_dir(buts->name, blk_debugfs_root);
-
 	bt->dev = dev;
 	atomic_set(&bt->dropped, 0);
 	INIT_LIST_HEAD(&bt->running_list);
@@ -565,8 +596,6 @@ 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

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Christoph Hellwig]

All the complicated code to deal with delayed creation of the debugfs
directory was only needed for sg.  With sg handled speparately that
can go away.  Also I think the sg handling can be cleaned up by:

 a) checking for a NULL block_devic instead of the major, especially as
    majors can be used independently for character and block devices
 b) just keeping the old method of creating the entry on demand for the
    sg devices, as no one shares the debugfs directory in that case.

Untested patch on top of your series below:

diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 70168435f07983..4e9909e1b25736 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -920,8 +920,6 @@ static void blk_release_queue(struct kobject *kobj)
 	blk_trace_shutdown(q);
 
 	debugfs_remove_recursive(q->debugfs_dir);
-	if (IS_ENABLED(CONFIG_CHR_DEV_SG))
-		debugfs_remove_recursive(q->sg_debugfs_dir);
 	if (queue_is_mq(q))
 		blk_mq_debugfs_unregister(q);
 
@@ -941,20 +939,6 @@ struct kobj_type blk_queue_ktype = {
 	.release	= blk_release_queue,
 };
 
-/**
- * blk_sg_debugfs_init - initialize debugfs for scsi-generic
- * @q: the associated queue
- * @name: name of the scsi-generic device
- *
- * To be used by scsi-generic for allowing it to use blktrace.
- */
-void blk_sg_debugfs_init(struct request_queue *q, const char *name)
-{
-	if (IS_ENABLED(CONFIG_CHR_DEV_SG))
-		q->sg_debugfs_dir = debugfs_create_dir(name, blk_debugfs_root);
-}
-EXPORT_SYMBOL_GPL(blk_sg_debugfs_init);
-
 /**
  * blk_register_queue - register a block layer queue with sysfs
  * @disk: Disk of which the request queue should be registered with sysfs.
@@ -1007,27 +991,8 @@ int blk_register_queue(struct gendisk *disk)
 		goto unlock;
 	}
 
-	/*
-	 * Blktrace needs a debugfs name even for queues that don't register
-	 * a gendisk, so it lazily registers the debugfs directory.  But that
-	 * can get us into a situation where a SCSI device is found, with no
-	 * driver for it (yet).  Then blktrace is used on the device, creating
-	 * the debugfs directory, and only after that a driver is loaded. In
-	 * that case we might already have a debugfs directory registered here.
-	 * Even worse we could be racing with blktrace to register it.
-	 */
-#ifdef CONFIG_BLK_DEV_IO_TRACE
-	mutex_lock(&q->blk_trace_mutex);
-	if (!q->debugfs_dir) {
-		q->debugfs_dir =
-			debugfs_create_dir(kobject_name(q->kobj.parent),
-				blk_debugfs_root);
-	}
-	mutex_unlock(&q->blk_trace_mutex);
-#else
 	q->debugfs_dir = debugfs_create_dir(kobject_name(q->kobj.parent),
 					    blk_debugfs_root);
-#endif
 
 	if (queue_is_mq(q)) {
 		__blk_mq_register_dev(dev, q);
diff --git a/drivers/scsi/sg.c b/drivers/scsi/sg.c
index 5f6ccf4ba4d907..20472aaaf630a4 100644
--- a/drivers/scsi/sg.c
+++ b/drivers/scsi/sg.c
@@ -1117,9 +1117,6 @@ sg_ioctl_common(struct file *filp, Sg_device *sdp, Sg_fd *sfp,
 		return put_user(max_sectors_bytes(sdp->device->request_queue),
 				ip);
 	case BLKTRACESETUP:
-		if (!sdp->device->request_queue->sg_debugfs_dir)
-			blk_sg_debugfs_init(sdp->device->request_queue,
-					    sdp->disk->disk_name);
 		return blk_trace_setup(sdp->device->request_queue,
 				       sdp->disk->disk_name,
 				       MKDEV(SCSI_GENERIC_MAJOR, sdp->index),
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 05a929329157fa..b49c7c741bc9f3 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -575,7 +575,6 @@ struct request_queue {
 	struct bio_set		bio_split;
 
 	struct dentry		*debugfs_dir;
-	struct dentry		*sg_debugfs_dir;
 #ifdef CONFIG_BLK_DEBUG_FS
 	struct dentry		*sched_debugfs_dir;
 	struct dentry		*rqos_debugfs_dir;
@@ -859,7 +858,6 @@ static inline void rq_flush_dcache_pages(struct request *rq)
 
 extern int blk_register_queue(struct gendisk *disk);
 extern void blk_unregister_queue(struct gendisk *disk);
-extern void blk_sg_debugfs_init(struct request_queue *q, const char *name);
 extern blk_qc_t generic_make_request(struct bio *bio);
 extern blk_qc_t direct_make_request(struct bio *bio);
 extern void blk_rq_init(struct request_queue *q, struct request *rq);
diff --git a/include/linux/blktrace_api.h b/include/linux/blktrace_api.h
index eb6db276e29310..3b6ff5902edce6 100644
--- a/include/linux/blktrace_api.h
+++ b/include/linux/blktrace_api.h
@@ -22,6 +22,7 @@ struct blk_trace {
 	u64 end_lba;
 	u32 pid;
 	u32 dev;
+	struct dentry *dir;
 	struct dentry *dropped_file;
 	struct dentry *msg_file;
 	struct list_head running_list;
diff --git a/kernel/trace/blktrace.c b/kernel/trace/blktrace.c
index 432fa60e7f8808..44239f603379d5 100644
--- a/kernel/trace/blktrace.c
+++ b/kernel/trace/blktrace.c
@@ -311,6 +311,7 @@ static void blk_trace_free(struct blk_trace *bt)
 	debugfs_remove(bt->msg_file);
 	debugfs_remove(bt->dropped_file);
 	relay_close(bt->rchan);
+	debugfs_remove(bt->dir);
 	free_percpu(bt->sequence);
 	free_percpu(bt->msg_data);
 	kfree(bt);
@@ -492,34 +493,23 @@ static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
 	 */
 	strreplace(buts->name, '/', '_');
 
-	/*
-	 * We have to use a partition directory if a partition is being worked
-	 * on. The same request_queue is shared between all partitions.
-	 */
-	if (bdev && bdev != bdev->bd_contains) {
-		dir = bdev->bd_part->debugfs_dir;
-	} else if (IS_ENABLED(CONFIG_CHR_DEV_SG) &&
-		   MAJOR(dev) == SCSI_GENERIC_MAJOR) {
+	bt = kzalloc(sizeof(*bt), GFP_KERNEL);
+	if (!bt)
+		return -ENOMEM;
+
+	if (unlikely(!bdev)) {
 		/*
-		 * scsi-generic exposes the request_queue through the /dev/sg*
-		 * interface but since that uses a different path than whatever
-		 * the respective scsi driver device name may expose and use
-		 * for the request_queue debugfs_dir. We have a dedicated
-		 * dentry for scsi-generic then.
+		 * When tracing something that is not a block device (e.g. the
+		 * /dev/sg nodes), create debugfs directory on demand.  This
+		 * directory will be remove when stopping the trace.
 		 */
-		dir = q->sg_debugfs_dir;
+		dir = debugfs_create_dir(buts->name, blk_debugfs_root);
+		bt->dir = dir;
 	} else {
-		/*
-		 * Drivers which do not use the *add_disk*() interfaces will
-		 * not have their debugfs_dir created for them automatically,
-		 * so we must create it for them.
-		 */
-		if (!q->debugfs_dir) {
-			q->debugfs_dir =
-				debugfs_create_dir(buts->name,
-						   blk_debugfs_root);
-		}
-		dir = q->debugfs_dir;
+		if (bdev != bdev->bd_contains)
+			dir = bdev->bd_part->debugfs_dir;
+		else
+			dir = q->debugfs_dir;
 	}
 
 	/*
@@ -530,12 +520,10 @@ static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
 	if (IS_ERR_OR_NULL(dir)) {
 		pr_warn("debugfs_dir not present for %s so skipping\n",
 			buts->name);
-		return -ENOENT;
+		ret = -ENOENT;
+		goto err;
 	}
 
-	bt = kzalloc(sizeof(*bt), GFP_KERNEL);
-	if (!bt)
-		return -ENOMEM;
 
 	ret = -ENOMEM;
 	bt->sequence = alloc_percpu(unsigned long);

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Luis Chamberlain]

I like this, more below.

On Tue, Jun 09, 2020 at 08:06:02AM -0700, Christoph Hellwig wrote:
> diff --git a/kernel/trace/blktrace.c b/kernel/trace/blktrace.c
> index 432fa60e7f8808..44239f603379d5 100644
> --- a/kernel/trace/blktrace.c
> +++ b/kernel/trace/blktrace.c
> @@ -492,34 +493,23 @@ static int do_blk_trace_setup(struct request_queue *q, char *name, dev_t dev,
>  	 */
>  	strreplace(buts->name, '/', '_');
>  
> -	/*
> -	 * We have to use a partition directory if a partition is being worked
> -	 * on. The same request_queue is shared between all partitions.
> -	 */
> -	if (bdev && bdev != bdev->bd_contains) {
> -		dir = bdev->bd_part->debugfs_dir;
> -	} else if (IS_ENABLED(CONFIG_CHR_DEV_SG) &&
> -		   MAJOR(dev) == SCSI_GENERIC_MAJOR) {
> +	bt = kzalloc(sizeof(*bt), GFP_KERNEL);
> +	if (!bt)
> +		return -ENOMEM;
> +
> +	if (unlikely(!bdev)) {
>  		/*
> -		 * scsi-generic exposes the request_queue through the /dev/sg*
> -		 * interface but since that uses a different path than whatever
> -		 * the respective scsi driver device name may expose and use
> -		 * for the request_queue debugfs_dir. We have a dedicated
> -		 * dentry for scsi-generic then.
> +		 * When tracing something that is not a block device (e.g. the
> +		 * /dev/sg nodes), create debugfs directory on demand.  This
> +		 * directory will be remove when stopping the trace.

Is scsi-generic is the only unwanted ugly child blktrace has to deal
with? For some reason I thought drivers/md/md.c was one but it seems
like it is not. Do we have an easy way to search for these? I think
this would just affect how we express the comment only.

>  		 */
> -		dir = q->sg_debugfs_dir;
> +		dir = debugfs_create_dir(buts->name, blk_debugfs_root);
> +		bt->dir = dir;

The other chicken and egg problem to consider at least in the comments
is that the debugfs directory for these types of devices *have* an
exposed path, but the data structure is rather opaque to the device and
even blktrace.  Fortunately given the recent set of changes around the
q->blk_trace and clarifications around its use we have made it clear now
that so long as hold the q->blk_trace_mutex *and* check q->blk_trace we
*should* not race against two separate creations of debugfs directories,
so I think this is safe, so long as these indpendent drivers don't end
up re-using the same path for some other things later in the future, and
since we have control over what goes under debugfsroot block / I think
we should be good.

But I think that the concern for race on names may still be worth
explaining a bit here.

  Luis

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Christoph Hellwig]

On Tue, Jun 09, 2020 at 05:29:22PM +0000, Luis Chamberlain wrote:
> Is scsi-generic is the only unwanted ugly child blktrace has to deal
> with? For some reason I thought drivers/md/md.c was one but it seems
> like it is not. Do we have an easy way to search for these? I think
> this would just affect how we express the comment only.

grep for blk_trace_setup.  For all blk devices that setup comes in
through the block device ioctl path, and that relies on having a
struct block_device and queue.  sg on the other hand calls
blk_trace_setup directly with a NULL bdev argument.

> >  		 */
> > -		dir = q->sg_debugfs_dir;
> > +		dir = debugfs_create_dir(buts->name, blk_debugfs_root);
> > +		bt->dir = dir;
> 
> The other chicken and egg problem to consider at least in the comments
> is that the debugfs directory for these types of devices *have* an
> exposed path, but the data structure is rather opaque to the device and
> even blktrace.  Fortunately given the recent set of changes around the
> q->blk_trace and clarifications around its use we have made it clear now
> that so long as hold the q->blk_trace_mutex *and* check q->blk_trace we
> *should* not race against two separate creations of debugfs directories,
> so I think this is safe, so long as these indpendent drivers don't end
> up re-using the same path for some other things later in the future, and
> since we have control over what goes under debugfsroot block / I think
> we should be good.
> 
> But I think that the concern for race on names may still be worth
> explaining a bit here.

Feel free to add more comments, but please try to keep them short
and crisp.  At the some point long comments really distract from what
is going on.

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Luis Chamberlain]

On Tue, Jun 09, 2020 at 10:32:18AM -0700, Christoph Hellwig wrote:
> On Tue, Jun 09, 2020 at 05:29:22PM +0000, Luis Chamberlain wrote:
> > Is scsi-generic is the only unwanted ugly child blktrace has to deal
> > with? For some reason I thought drivers/md/md.c was one but it seems
> > like it is not. Do we have an easy way to search for these? I think
> > this would just affect how we express the comment only.
> 
> grep for blk_trace_setup.  For all blk devices that setup comes in
> through the block device ioctl path, and that relies on having a
> struct block_device and queue.  sg on the other hand calls
> blk_trace_setup directly with a NULL bdev argument.

Alright, then we should be good.

> > >  		 */
> > > -		dir = q->sg_debugfs_dir;
> > > +		dir = debugfs_create_dir(buts->name, blk_debugfs_root);
> > > +		bt->dir = dir;
> > 
> > The other chicken and egg problem to consider at least in the comments
> > is that the debugfs directory for these types of devices *have* an
> > exposed path, but the data structure is rather opaque to the device and
> > even blktrace.  Fortunately given the recent set of changes around the
> > q->blk_trace and clarifications around its use we have made it clear now
> > that so long as hold the q->blk_trace_mutex *and* check q->blk_trace we
> > *should* not race against two separate creations of debugfs directories,
> > so I think this is safe, so long as these indpendent drivers don't end
> > up re-using the same path for some other things later in the future, and
> > since we have control over what goes under debugfsroot block / I think
> > we should be good.
> > 
> > But I think that the concern for race on names may still be worth
> > explaining a bit here.
> 
> Feel free to add more comments, but please try to keep them short
> and crisp.  At the some point long comments really distract from what
> is going on.

Sure.

Come to think of it, given the above, I think we can also do way with
the the partition stuff too, and rely on the buts->name too. I'll try
this out, and test it.

  Luis

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Christoph Hellwig]

On Tue, Jun 09, 2020 at 05:53:59PM +0000, Luis Chamberlain wrote:
> > Feel free to add more comments, but please try to keep them short
> > and crisp.  At the some point long comments really distract from what
> > is going on.
> 
> Sure.
> 
> Come to think of it, given the above, I think we can also do way with
> the the partition stuff too, and rely on the buts->name too. I'll try
> this out, and test it.

Yes, the sg path should work for partitions as well.  That should not
only simplify the code, but also the comments, we can do something like
the full patch below (incorporating your original one).  This doesn't
include the error check on the creation - I think that check probably
is a good idea for this case based on the comments in the old patch, but
also a separate issue that should go into another patch on top.

diff --git a/block/blk-mq-debugfs.c b/block/blk-mq-debugfs.c
index 15df3a36e9fa43..a2800bc56fb4d3 100644
--- a/block/blk-mq-debugfs.c
+++ b/block/blk-mq-debugfs.c
@@ -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,
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 561624d4cc4e7f..4e9909e1b25736 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -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"
@@ -918,6 +919,7 @@ static void blk_release_queue(struct kobject *kobj)
 
 	blk_trace_shutdown(q);
 
+	debugfs_remove_recursive(q->debugfs_dir);
 	if (queue_is_mq(q))
 		blk_mq_debugfs_unregister(q);
 
@@ -989,6 +991,9 @@ int blk_register_queue(struct gendisk *disk)
 		goto unlock;
 	}
 
+	q->debugfs_dir = debugfs_create_dir(kobject_name(q->kobj.parent),
+					    blk_debugfs_root);
+
 	if (queue_is_mq(q)) {
 		__blk_mq_register_dev(dev, q);
 		blk_mq_debugfs_register(q);
diff --git a/block/blk.h b/block/blk.h
index b5d1f0fc6547c7..499308c6ab3b0f 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -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;
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index fc88330a3d97ed..b49c7c741bc9f3 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -574,8 +574,8 @@ 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
diff --git a/kernel/trace/blktrace.c b/kernel/trace/blktrace.c
index fee5c8d8916690..6eb364b393714f 100644
--- a/kernel/trace/blktrace.c
+++ b/kernel/trace/blktrace.c
@@ -509,10 +509,15 @@ 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)
+	/*
+	 * When tracing a whole block device, reuse the existing debugfs
+	 * directory created by the block layer.  For partitions or character
+	 * devices (e.g. /dev/sg), create a new debugfs directory that will be
+	 * removed once the trace ends.
+	 */
+	if (bdev && bdev == bdev->bd_contains)
+		dir = q->debugfs_dir;
+	else
 		bt->dir = dir = debugfs_create_dir(buts->name, blk_debugfs_root);
 
 	bt->dev = dev;
@@ -553,8 +558,6 @@ 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;

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Luis Chamberlain]

On Tue, Jun 09, 2020 at 11:42:34PM -0700, Christoph Hellwig wrote:
> On Tue, Jun 09, 2020 at 05:53:59PM +0000, Luis Chamberlain wrote:
> > > Feel free to add more comments, but please try to keep them short
> > > and crisp.  At the some point long comments really distract from what
> > > is going on.
> > 
> > Sure.
> > 
> > Come to think of it, given the above, I think we can also do way with
> > the the partition stuff too, and rely on the buts->name too. I'll try
> > this out, and test it.
> 
> Yes, the sg path should work for partitions as well.  That should not
> only simplify the code, but also the comments, we can do something like
> the full patch below (incorporating your original one). 

Indeed I ended up with something similar, will use this variation.

> This doesn't
> include the error check on the creation - I think that check probably
> is a good idea for this case based on the comments in the old patch, but
> also a separate issue that should go into another patch on top.

Makes sense.

  Luis

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Luis Chamberlain]

So, upon updating the commit log, and moving the empty directory check
into another patch, I realized we might be able to simplify this now even
further still. Patch below. The key of the issue was that the use after free
happens when a recursive removal happens, and then later a specific
dentry removal happens. This happened for make_request block drivers
when using the whole disk, but since we *don't* have any other users of
the directory for the others cases, this in theory shuld not happen for
them either.

I'll try to shoot some bullets at this.

diff --git a/kernel/trace/blktrace.c b/kernel/trace/blktrace.c
index 7ff2ea5cd05e..5cea04c05e09 100644
--- a/kernel/trace/blktrace.c
+++ b/kernel/trace/blktrace.c
@@ -524,10 +524,16 @@ 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)
+	/*
+	 * 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))
+		dir = q->debugfs_dir;
+	else
 		bt->dir = dir = debugfs_create_dir(buts->name, blk_debugfs_root);
 
 	bt->dev = dev;
@@ -565,8 +571,6 @@ 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;

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Luis Chamberlain]

On Wed, Jun 10, 2020 at 09:52:13PM +0000, Luis Chamberlain wrote:
> diff --git a/kernel/trace/blktrace.c b/kernel/trace/blktrace.c
> index 7ff2ea5cd05e..5cea04c05e09 100644
> --- a/kernel/trace/blktrace.c
> +++ b/kernel/trace/blktrace.c
> @@ -524,10 +524,16 @@ 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)
> +	/*
> +	 * 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))

And this should be instead:

if (queue_is_mq(q) && bdev && bdev == bdev->bd_contains)

> +		dir = q->debugfs_dir;
> +	else
>  		bt->dir = dir = debugfs_create_dir(buts->name, blk_debugfs_root);
>  
>  	bt->dev = dev;
> @@ -565,8 +571,6 @@ 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;

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Christoph Hellwig]

On Wed, Jun 10, 2020 at 11:31:16PM +0000, Luis Chamberlain wrote:
> On Wed, Jun 10, 2020 at 09:52:13PM +0000, Luis Chamberlain wrote:
> > diff --git a/kernel/trace/blktrace.c b/kernel/trace/blktrace.c
> > index 7ff2ea5cd05e..5cea04c05e09 100644
> > --- a/kernel/trace/blktrace.c
> > +++ b/kernel/trace/blktrace.c
> > @@ -524,10 +524,16 @@ 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)
> > +	/*
> > +	 * 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))
> 
> And this should be instead:
> 
> if (queue_is_mq(q) && bdev && bdev == bdev->bd_contains)

Yes.  But I think keeping the queue_is_mq out and always creating the
debugfs dir is an improvement as we'll sooner or later grow more debugfs
files and than the whole set of problmes reappears.  But I'd be fine
with doing the above in the first patch that is tiny and backportable,
and then have another patch that always creates the debugfs directory.

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Bart Van Assche]

On 2020-06-08 10:01, Luis Chamberlain wrote:
> +	/*
> +	 * Blktrace needs a debugfs name even for queues that don't register
> +	 * a gendisk, so it lazily registers the debugfs directory.  But that
> +	 * can get us into a situation where a SCSI device is found, with no
> +	 * driver for it (yet).  Then blktrace is used on the device, creating
> +	 * the debugfs directory, and only after that a driver is loaded. In
> +	 * that case we might already have a debugfs directory registered here.
> +	 * Even worse we could be racing with blktrace to register it.
> +	 */

There are LLD and ULD drivers in the SCSI subsystem. Please mention the
driver type explicitly. I assume that you are referring to SCSI ULDs
since only SCSI ULD drivers call device_add_disk()?

Could the above comment be made shorter by only mentioning that blktrace
may have been set up before or concurrently with device_add_disk() and
that device_add_disk() calls blk_register_queue()?

>  	case BLKTRACESETUP:
> +		if (!sdp->device->request_queue->sg_debugfs_dir)
> +			blk_sg_debugfs_init(sdp->device->request_queue,
> +					    sdp->disk->disk_name);

How about moving the sg_debugfs_dir check into blk_sg_debugfs_init()?

Thanks,

Bart.

Re: [PATCH v6 6/6] blktrace: fix debugfs use after free

[Luis Chamberlain]

On Fri, Jun 12, 2020 at 07:42:12PM -0700, Bart Van Assche wrote:
> On 2020-06-08 10:01, Luis Chamberlain wrote:
> > +	/*
> > +	 * Blktrace needs a debugfs name even for queues that don't register
> > +	 * a gendisk, so it lazily registers the debugfs directory.  But that
> > +	 * can get us into a situation where a SCSI device is found, with no
> > +	 * driver for it (yet).  Then blktrace is used on the device, creating
> > +	 * the debugfs directory, and only after that a driver is loaded. In
> > +	 * that case we might already have a debugfs directory registered here.
> > +	 * Even worse we could be racing with blktrace to register it.
> > +	 */
> 
> There are LLD and ULD drivers in the SCSI subsystem. Please mention the
> driver type explicitly. I assume that you are referring to SCSI ULDs
> since only SCSI ULD drivers call device_add_disk()?

I've simplified this and so this is no longer a valid comment.

> >  	case BLKTRACESETUP:
> > +		if (!sdp->device->request_queue->sg_debugfs_dir)
> > +			blk_sg_debugfs_init(sdp->device->request_queue,
> > +					    sdp->disk->disk_name);
> 
> How about moving the sg_debugfs_dir check into blk_sg_debugfs_init()?

I found a way to not have to do any of this, the fix will be short and
sweet now.

  Luis