scsi: sg: Fix occasional bogus elapsed time that exceeds timeout

A race condition was found in sg_proc_debug_helper(). It was observed on
a system using an IBM LTO-9 SAS Tape Drive (ULTRIUM-TD9) and monitoring
/proc/scsi/sg/debug every second. A very large elapsed time would
sometimes appear. This is caused by two race conditions.

We reproduced the issue with an IBM ULTRIUM-HH9 tape drive on an x86_64
architecture. A patched kernel was built, and the race condition could
not be observed anymore after the application of this patch. A
reproducer C program utilising the scsi_debug module was also built by
Changhui Zhong and can be viewed here:

https://github.com/MichaelRabek/linux-tests/blob/master/drivers/scsi/sg/sg_race_trigger.c

The first race happens between the reading of hp->duration in
sg_proc_debug_helper() and request completion in sg_rq_end_io().  The
hp->duration member variable may hold either of two types of
information:

 #1 - The start time of the request. This value is present while
      the request is not yet finished.

 #2 - The total execution time of the request (end_time - start_time).

If sg_proc_debug_helper() executes *after* the value of hp->duration was
changed from #1 to #2, but *before* srp->done is set to 1 in
sg_rq_end_io(), a fresh timestamp is taken in the else branch, and the
elapsed time (value type #2) is subtracted from a timestamp, which
cannot yield a valid elapsed time (which is a type #2 value as well).

To fix this issue, the value of hp->duration must change under the
protection of the sfp->rq_list_lock in sg_rq_end_io().  Since
sg_proc_debug_helper() takes this read lock, the change to srp->done and
srp->header.duration will happen atomically from the perspective of
sg_proc_debug_helper() and the race condition is thus eliminated.

The second race condition happens between sg_proc_debug_helper() and
sg_new_write(). Even though hp->duration is set to the current time
stamp in sg_add_request() under the write lock's protection, it gets
overwritten by a call to get_sg_io_hdr(), which calls copy_from_user()
to copy struct sg_io_hdr from userspace into kernel space. hp->duration
is set to the start time again in sg_common_write(). If
sg_proc_debug_helper() is called between these two calls, an arbitrary
value set by userspace (usually zero) is used to compute the elapsed
time.

To fix this issue, hp->duration must be set to the current timestamp
again after get_sg_io_hdr() returns successfully. A small race window
still exists between get_sg_io_hdr() and setting hp->duration, but this
window is only a few instructions wide and does not result in observable
issues in practice, as confirmed by testing.

Additionally, we fix the format specifier from %d to %u for printing
unsigned int values in sg_proc_debug_helper().

Signed-off-by: Michal Rábek <mrabek@redhat.com>
Suggested-by: Tomas Henzl <thenzl@redhat.com>
Tested-by: Changhui Zhong <czhong@redhat.com>
Reviewed-by: Ewan D. Milne <emilne@redhat.com>
Reviewed-by: John Meneghini <jmeneghi@redhat.com>
Reviewed-by: Tomas Henzl <thenzl@redhat.com>
Link: https://patch.msgid.link/20251212160900.64924-1-mrabek@redhat.com
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>

0 files changed, 0 insertions, 0 deletions