sg: fix EWOULDBLOCK errors with scsi-mq

With scsi-mq enabled, userspace programs can get unexpected EWOULDBLOCK (a.k.a. EAGAIN) errors when submitting commands to the SCSI generic driver. Fix by calling blk_get_request() with GFP_KERNEL instead of GFP_ATOMIC.

Note: to avoid introducing a potential deadlock, this patch should be applied after the patch titled "sg: fix unkillable I/O wait deadlock with scsi-mq".

Cc Douglas Gilbert <[email protected]>

Cc <[email protected]> # 3.17+

Signed-off-by Tony Battersby <[email protected]>

For inclusion in kernel 3.20.

The difference in behavior is due to bt_get() in block/blk-mq-tag.c checking for __GFP_WAIT.

The bsg driver already calls blk_get_request() with GFP_KERNEL, so there is no need for a change there.

drivers/scsi/sg.c renamed from drivers/scsi/sg.c.orig +16 -2

@@ -1695,7 +1695,22 @@ sg_start_req(Sg_request *srp, unsigned c
  			return -ENOMEM;
  	}
 
+	/*
+	 * NOTE
+	 *
+	 * With scsi-mq enabled, there are a fixed number of preallocated
+	 * requests equal in number to shost->can_queue.  If all of the
+	 * preallocated requests are already in use, then using GFP_ATOMIC with
+	 * blk_get_request() will return -EWOULDBLOCK, whereas using GFP_KERNEL
+	 * will cause blk_get_request() to sleep until an active command
+	 * completes, freeing up a request.  Neither option is ideal, but
+	 * GFP_KERNEL is the better choice to prevent userspace from getting an
+	 * unexpected EWOULDBLOCK.
+	 *
+	 * With scsi-mq disabled, blk_get_request() with GFP_KERNEL usually
+	 * does not sleep except under memory pressure.
+	 */
+	rq = blk_get_request(q, rw, GFP_KERNEL);
-	rq = blk_get_request(q, rw, GFP_ATOMIC);
  	if (IS_ERR(rq)) {
  		kfree(long_cmdp);
  		return PTR_ERR(rq);
 
--