zfcp_erp_adapter_reopen() schedules blocking of all of the adapter's
rports via zfcp_scsi_schedule_rports_block() and enqueues a reopen
adapter ERP action via zfcp_erp_action_enqueue(). Both are separately
processed asynchronously and concurrently.
Blocking of rports is done in a kworker by zfcp_scsi_rport_work(). It
calls zfcp_scsi_rport_block(), which then traces a DBF REC "scpdely" via
zfcp_dbf_rec_trig(). zfcp_dbf_rec_trig() acquires the DBF REC spin lock
and then iterates with list_for_each() over the adapter's ERP ready list
without holding the ERP lock. This opens a race window in which the
current list entry can be moved to another list, causing list_for_each()
to iterate forever on the wrong list, as the erp_ready_head is never
encountered as terminal condition.
Meanwhile the ERP action can be processed in the ERP thread by
zfcp_erp_thread(). It calls zfcp_erp_strategy(), which acquires the ERP
lock and then calls zfcp_erp_action_to_running() to move the ERP action
from the ready to the running list. zfcp_erp_action_to_running() can
move the ERP action using list_move() just during the aforementioned
race window. It then traces a REC RUN "erator1" via zfcp_dbf_rec_run().
zfcp_dbf_rec_run() tries to acquire the DBF REC spin lock. If this is
held by the infinitely looping kworker, it effectively spins forever.
Example Sequence Diagram:
Process ERP Thread rport_work
------------------- ------------------- -------------------
zfcp_erp_adapter_reopen()
zfcp_erp_adapter_block()
zfcp_scsi_schedule_rports_block()
lock ERP zfcp_scsi_rport_work()
zfcp_erp_action_enqueue(ZFCP_ERP_ACTION_REOPEN_ADAPTER)
list_add_tail() on ready !(rport_task==RPORT_ADD)
wake_up() ERP thread zfcp_scsi_rport_block()
zfcp_dbf_rec_trig() zfcp_erp_strategy() zfcp_dbf_rec_trig()
unlock ERP lock DBF REC
zfcp_erp_wait() lock ERP
| zfcp_erp_action_to_running()
| list_for_each() ready
| list_move() current entry
| ready to running
| zfcp_dbf_rec_run() endless loop over running
| zfcp_dbf_rec_run_lvl()
| lock DBF REC spins forever
Any adapter recovery can trigger this, such as setting the device offline
or reboot.
V4.9 commit 4eeaa4f3f1 ("zfcp: close window with unblocked rport
during rport gone") introduced additional tracing of (un)blocking of
rports. It missed that the adapter->erp_lock must be held when calling
zfcp_dbf_rec_trig().
This fix uses the approach formerly introduced by commit aa0fec6239
("[SCSI] zfcp: Fix sparse warning by providing new entry in dbf") that got
later removed by commit ae0904f60f ("[SCSI] zfcp: Redesign of the debug
tracing for recovery actions.").
Introduce zfcp_dbf_rec_trig_lock(), a wrapper for zfcp_dbf_rec_trig() that
acquires and releases the adapter->erp_lock for read.
Reported-by: Sebastian Ott <sebott@linux.ibm.com>
Signed-off-by: Jens Remus <jremus@linux.ibm.com>
Fixes: 4eeaa4f3f1 ("zfcp: close window with unblocked rport during rport gone")
Cc: <stable@vger.kernel.org> # 2.6.32+
Reviewed-by: Benjamin Block <bblock@linux.vnet.ibm.com>
Signed-off-by: Steffen Maier <maier@linux.ibm.com>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
fa89adba19