[PATCH 3.2 054/106] usb: gadget: f_mass_storage: Serialize wake and sleep execution

From: Ben Hutchings
Date: Sat Sep 09 2017 - 18:45:48 EST

3.2.93-rc1 review patch. If anyone has any objections, please let me know.


From: Thinh Nguyen <Thinh.Nguyen@xxxxxxxxxxxx>

commit dc9217b69dd6089dcfeb86ed4b3c671504326087 upstream.

f_mass_storage has a memorry barrier issue with the sleep and wake
functions that can cause a deadlock. This results in intermittent hangs
during MSC file transfer. The host will reset the device after receiving
no response to resume the transfer. This issue is seen when dwc3 is
processing 2 transfer-in-progress events at the same time, invoking
completion handlers for CSW and CBW. Also this issue occurs depending on
the system timing and latency.

To increase the chance to hit this issue, you can force dwc3 driver to
wait and process those 2 events at once by adding a small delay (~100us)
in dwc3_check_event_buf() whenever the request is for CSW and read the
event count again. Avoid debugging with printk and ftrace as extra
delays and memory barrier will mask this issue.

Scenario which can lead to failure:
1) The main thread sleeps and waits for the next command in
2) bulk_in_complete() wakes up main thread for CSW.
3) bulk_out_complete() tries to wake up the running main thread for CBW.
4) thread_wakeup_needed is not loaded with correct value in
5) Main thread goes to sleep again.

The pattern is shown below. Note the 2 critical variables.
* common->thread_wakeup_needed
* bh->state

CPU 0 (sleep_thread) CPU 1 (wakeup_thread)
============================== ===============================

bh->state = BH_STATE_FULL;
thread_wakeup_needed = 0; thread_wakeup_needed = 1;
if (bh->state != BH_STATE_FULL)
sleep again ...

As pointed out by Alan Stern, this is an R-pattern issue. The issue can
be seen when there are two wakeups in quick succession. The
thread_wakeup_needed can be overwritten in sleep_thread, and the read of
the bh->state maybe reordered before the write to thread_wakeup_needed.

This patch applies full memory barrier smp_mb() in both sleep_thread()
and wakeup_thread() to ensure the order which the thread_wakeup_needed
and bh->state are written and loaded.

However, a better solution in the future would be to use wait_queue
method that takes care of managing memory barrier between waker and

Acked-by: Alan Stern <stern@xxxxxxxxxxxxxxxxxxx>
Signed-off-by: Thinh Nguyen <thinhn@xxxxxxxxxxxx>
Signed-off-by: Felipe Balbi <felipe.balbi@xxxxxxxxxxxxxxx>
[bwh: Backported to 3.2: adjust filename]
Signed-off-by: Ben Hutchings <ben@xxxxxxxxxxxxxxx>
drivers/usb/gadget/f_mass_storage.c | 13 +++++++++++--
1 file changed, 11 insertions(+), 2 deletions(-)

--- a/drivers/usb/gadget/f_mass_storage.c
+++ b/drivers/usb/gadget/f_mass_storage.c
@@ -512,7 +512,11 @@ static int fsg_set_halt(struct fsg_dev *
/* Caller must hold fsg->lock */
static void wakeup_thread(struct fsg_common *common)
- smp_wmb(); /* ensure the write of bh->state is complete */
+ /*
+ * Ensure the reading of thread_wakeup_needed
+ * and the writing of bh->state are completed
+ */
+ smp_mb();
/* Tell the main thread that something has happened */
common->thread_wakeup_needed = 1;
if (common->thread_task)
@@ -732,7 +736,12 @@ static int sleep_thread(struct fsg_commo
common->thread_wakeup_needed = 0;
- smp_rmb(); /* ensure the latest bh->state is visible */
+ /*
+ * Ensure the writing of thread_wakeup_needed
+ * and the reading of bh->state are completed
+ */
+ smp_mb();
return rc;