Skip to contentSkip to conversation

zfcp: Fix spinlock imbalance in zfcp_qdio_sbal_get

PATCH Martin Peschke 1 participant 1 message May 22, 2013 1 message not in the archive s390

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): 20130515160356.GE4298@osiris

#1 lore raw
Quoted 26 lines

-------- Original Message --------
Subject: Re: [PATCH] zfcp: Fix spinlock imbalance in zfcp_qdio_sbal_get
Date: Wed, 15 May 2013 18:03:56 +0200
From: Heiko Carstens <[email protected]>
To: Mikulas Patocka <[email protected]>
CC: Steffen Maier <[email protected]>, [email protected], [email protected]

Quoted 16 lines · Mikulas Patocka

zfcp: Fix spinlock imbalance in zfcp_qdio_sbal_get

---
 drivers/s390/scsi/zfcp_qdio.c |    5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

Index: linux-2.6/drivers/s390/scsi/zfcp_qdio.c
===================================================================

drivers/s390/scsi/zfcp_qdio.c +0 -0
@@ -250,8 +250,11 @@ int zfcp_qdio_sbal_get(struct zfcp_qdio
 	ret = wait_event_interruptible_timeout(qdio->req_q_wq,
 			       zfcp_qdio_sbal_check(qdio), 5 * HZ);
 
Quoted 8 lines

drivers/s390/scsi/zfcp_qdio.c int zfcp_qdio_sbal_get(struct zfcp_qdio · line 257

Quoted 6 lines
drivers/s390/scsi/zfcp_qdio.c +4 -1
-	if (!(atomic_read(&qdio->adapter->status) & ZFCP_STATUS_ADAPTER_QDIOUP))
+	if (!(atomic_read(&qdio->adapter->status) & ZFCP_STATUS_ADAPTER_QDIOUP)) {
+		if (ret <= 0)
+			spin_lock_irq(&qdio->req_q_lock);
 		return -EIO;
+	}

Looks good to me. However it's Steffen's call.

Looking at the commit that introduced this race:

commit c2af7545aaff3495d9bf9a7608c52f0af86fb194

Author: Christof Schmitt <[email protected]>
Date:   Mon Jun 21 10:11:32 2010 +0200

[SCSI] zfcp: Do not wait for SBALs on stopped queue

Trying to read the FC host statistics on an offline adapter results in a 5 seconds wait. Reading the statistics tries to issue an exchange port data request which first waits up to 5 seconds for an entry in the request queue.

Change the strategy for getting a free SBAL to exit when the queue is stopped. Reading the statistics will then fail without the wait.

makes me think that it would be best to revert it.

A much simpler and less error-prone fix for the initial 5 sec delay usability issue would have been to check the ZFCP_STATUS_COMMON_UNBLOCKED flag when entering that sysfs code path, as done in other places (see zfcp_scsi_queuecommand(), for example).

I am reluctant to add more locking statements like the above. The entire use of req_q_lock sort of works - but the code has become crappy and should be cleaned up anyway. Just look at the lock-unlock-lock sequence done prior to entering the critical section.

Martin