[PATCH 2.5.17] Making SCSI not copy the request structure

[PATCH 2.5.17] Making SCSI not copy the request structure

[James Bottomley]

[-- Attachment #1: Type: text/plain, Size: 1055 bytes --]

The attached makes the request field in Scsi_Cmnd and Scsi_Request be pointers 
to the block layer request instead of copies of it.  The advantage of this is 
that it makes the scsi interface to the block layer conform much more closely 
to most other block device drivers.  It should also make it quite a lot easier 
to slot in the block layer tag command queueing functions.

The disadvantage is that you now _really_ need to use the scsi request 
interfaces instead of the command based ones (you will get a NULL deref if you 
allocate and use your own commands with no associated requests).  This also 
means that the following fibre drivers need some rework:  cpqfc, gdth, pluto, 
tmscsim.

It works on my system (at least it stays up during a full kernel compile).  
Any comments would be welcome.

After this, I plan to see if I can implement TCQ using the block level generic 
functions.  This would buy us the ability to support request barriers (modulo 
the already discussed edge cases) right down through the SCSI subsystem.

James Bottomley


[-- Attachment #2: scsi-req-2.5.17.diff --]
[-- Type: text/plain , Size: 38686 bytes --]

# This is a BitKeeper generated patch for the following project:
# Project Name: Linux kernel tree
# This patch format is intended for GNU patch command version 2.5 or higher.
# This patch includes the following deltas:
#	           ChangeSet	1.582   -> 1.583  
#	drivers/scsi/aha1542.c	1.13    -> 1.14   
#	drivers/scsi/scsi_error.c	1.10    -> 1.11   
#	drivers/scsi/ide-scsi.c	1.35    -> 1.36   
#	drivers/scsi/sun3_scsi.c	1.5     -> 1.6    
#	drivers/scsi/sun3_NCR5380.c	1.2     -> 1.3    
#	   drivers/scsi/sd.c	1.37    -> 1.38   
#	   drivers/scsi/sr.c	1.25    -> 1.26   
#	 drivers/scsi/scsi.h	1.12    -> 1.13   
#	drivers/scsi/u14-34f.c	1.10    -> 1.11   
#	 drivers/scsi/eata.c	1.12    -> 1.13   
#	   drivers/scsi/sg.c	1.15    -> 1.16   
#	 drivers/scsi/osst.c	1.13    -> 1.14   
#	drivers/scsi/scsi_lib.c	1.22    -> 1.23   
#	 drivers/scsi/scsi.c	1.28    -> 1.29   
#	drivers/scsi/scsi_debug.c	1.7     -> 1.8    
#	   drivers/scsi/st.c	1.15    -> 1.16   
#	drivers/scsi/sr_ioctl.c	1.12    -> 1.13   
#	drivers/scsi/constants.c	1.5     -> 1.6    
#	drivers/scsi/scsi_merge.c	1.19    -> 1.20   
#	drivers/scsi/qla1280.c	1.11    -> 1.12   
#
# The following is the BitKeeper ChangeSet Log
# --------------------------------------------
# 02/05/21	jejb@mulgrave.(none)	1.583
# make Scsi_Cmnd and Scsi_Request.request be a pointer to the
# block layer request instead of a copy.
# 
# Precursor to generic tag queueing work
# --------------------------------------------
#
diff -Nru a/drivers/scsi/aha1542.c b/drivers/scsi/aha1542.c
--- a/drivers/scsi/aha1542.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/aha1542.c	Tue May 21 18:22:18 2002
@@ -1635,14 +1635,14 @@
 		if (HOSTDATA(SCpnt->host)->SCint[i]) {
 			if (HOSTDATA(SCpnt->host)->SCint[i] == SCpnt) {
 				printk(KERN_ERR "Timed out command pending for %s\n",
-				       kdevname(SCpnt->request.rq_dev));
+				       kdevname(SCpnt->request->rq_dev));
 				if (HOSTDATA(SCpnt->host)->mb[i].status) {
 					printk(KERN_ERR "OGMB still full - restarting\n");
 					aha1542_out(SCpnt->host->io_port, &ahacmd, 1);
 				};
 			} else
 				printk(KERN_ERR "Other pending command %s\n",
-				       kdevname(SCpnt->request.rq_dev));
+				       kdevname(SCpnt->request->rq_dev));
 		}
 #endif
 
diff -Nru a/drivers/scsi/constants.c b/drivers/scsi/constants.c
--- a/drivers/scsi/constants.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/constants.c	Tue May 21 18:22:18 2002
@@ -807,13 +807,13 @@
 void print_sense(const char * devclass, Scsi_Cmnd * SCpnt)
 {
 	print_sense_internal(devclass, SCpnt->sense_buffer,
-			     SCpnt->request.rq_dev);
+			     SCpnt->request->rq_dev);
 }
 
 void print_req_sense(const char * devclass, Scsi_Request * SRpnt)
 {
 	print_sense_internal(devclass, SRpnt->sr_sense_buffer,
-			     SRpnt->sr_request.rq_dev);
+			     SRpnt->sr_request->rq_dev);
 }
 
 #if (CONSTANTS & CONST_MSG) 
diff -Nru a/drivers/scsi/eata.c b/drivers/scsi/eata.c
--- a/drivers/scsi/eata.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/eata.c	Tue May 21 18:22:18 2002
@@ -1563,7 +1563,7 @@
    if (linked_comm && SCpnt->device->queue_depth > 2
                                      && TLDEV(SCpnt->device->type)) {
       HD(j)->cp_stat[i] = READY;
-      flush_dev(SCpnt->device, SCpnt->request.sector, j, FALSE);
+      flush_dev(SCpnt->device, SCpnt->request->sector, j, FALSE);
       return 0;
       }
 
@@ -1875,11 +1875,11 @@
 
       if (!cpp->din) input_only = FALSE;
 
-      if (SCpnt->request.sector < minsec) minsec = SCpnt->request.sector;
-      if (SCpnt->request.sector > maxsec) maxsec = SCpnt->request.sector;
+      if (SCpnt->request->sector < minsec) minsec = SCpnt->request->sector;
+      if (SCpnt->request->sector > maxsec) maxsec = SCpnt->request->sector;
 
-      sl[n] = SCpnt->request.sector;
-      ioseek += SCpnt->request.nr_sectors;
+      sl[n] = SCpnt->request->sector;
+      ioseek += SCpnt->request->nr_sectors;
 
       if (!n) continue;
 
@@ -1907,7 +1907,7 @@
 
    if (!input_only) for (n = 0; n < n_ready; n++) {
       k = il[n]; cpp = &HD(j)->cp[k]; SCpnt = cpp->SCpnt;
-      ll[n] = SCpnt->request.nr_sectors; pl[n] = SCpnt->pid;
+      ll[n] = SCpnt->request->nr_sectors; pl[n] = SCpnt->pid;
 
       if (!n) continue;
 
@@ -1935,7 +1935,7 @@
                 " cur %ld s:%c r:%c rev:%c in:%c ov:%c xd %d.\n",
                 (ihdlr ? "ihdlr" : "qcomm"), SCpnt->channel, SCpnt->target,
                 SCpnt->lun, SCpnt->pid, k, flushcount, n_ready,
-                SCpnt->request.sector, SCpnt->request.nr_sectors, cursec,
+                SCpnt->request->sector, SCpnt->request->nr_sectors, cursec,
                 YESNO(s), YESNO(r), YESNO(rev), YESNO(input_only),
                 YESNO(overlap), cpp->din);
          }
@@ -2073,7 +2073,7 @@
 
    if (linked_comm && SCpnt->device->queue_depth > 2
                                      && TLDEV(SCpnt->device->type))
-      flush_dev(SCpnt->device, SCpnt->request.sector, j, TRUE);
+      flush_dev(SCpnt->device, SCpnt->request->sector, j, TRUE);
 
    tstatus = status_byte(spp->target_status);
 
diff -Nru a/drivers/scsi/ide-scsi.c b/drivers/scsi/ide-scsi.c
--- a/drivers/scsi/ide-scsi.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/ide-scsi.c	Tue May 21 18:22:18 2002
@@ -718,7 +718,7 @@
 {
 	idescsi_scsi_t *scsi = drive->driver_data;
 
-	if (major(cmd->request.rq_dev) == SCSI_GENERIC_MAJOR)
+	if (major(cmd->request->rq_dev) == SCSI_GENERIC_MAJOR)
 		return test_bit(IDESCSI_SG_TRANSFORM, &scsi->transform);
 	return test_bit(IDESCSI_TRANSFORM, &scsi->transform);
 }
diff -Nru a/drivers/scsi/osst.c b/drivers/scsi/osst.c
--- a/drivers/scsi/osst.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/osst.c	Tue May 21 18:22:18 2002
@@ -271,7 +271,7 @@
 /* Wakeup from interrupt */
 static void osst_sleep_done (Scsi_Cmnd * SCpnt)
 {
-	unsigned int dev = TAPE_NR(SCpnt->request.rq_dev);
+	unsigned int dev = TAPE_NR(SCpnt->request->rq_dev);
 	OS_Scsi_Tape * STp;
 
 	if (os_scsi_tapes && (STp = os_scsi_tapes[dev])) {
@@ -286,13 +286,13 @@
 		}
 		else
 			(STp->buffer)->midlevel_result = SCpnt->result;
-		SCpnt->request.rq_status = RQ_SCSI_DONE;
+		SCpnt->request->rq_status = RQ_SCSI_DONE;
 		(STp->buffer)->last_SRpnt = SCpnt->sc_request;
 
 #if DEBUG
 		STp->write_pending = 0;
 #endif
-		complete(SCpnt->request.waiting);
+		complete(SCpnt->request->waiting);
 	}
 #if DEBUG
 	else if (debugging)
@@ -314,7 +314,7 @@
 #endif
 	if (SRpnt == NULL) {
 		if ((SRpnt = scsi_allocate_request(STp->device)) == NULL) {
-			printk(KERN_ERR "osst%d:E: Can't get SCSI request.\n", TAPE_NR(STp->devt));
+			printk(KERN_ERR "osst%d:E: Can't get SCSI request->\n", TAPE_NR(STp->devt));
 			if (signal_pending(current))
 				(STp->buffer)->syscall_result = (-EINTR);
 			else
@@ -337,15 +337,15 @@
 		bp = (STp->buffer)->b_data;
 	SRpnt->sr_data_direction = direction;
 	SRpnt->sr_cmd_len = 0;
-	SRpnt->sr_request.waiting = &(STp->wait);
-	SRpnt->sr_request.rq_status = RQ_SCSI_BUSY;
-	SRpnt->sr_request.rq_dev = STp->devt;
+	SRpnt->sr_request->waiting = &(STp->wait);
+	SRpnt->sr_request->rq_status = RQ_SCSI_BUSY;
+	SRpnt->sr_request->rq_dev = STp->devt;
 
 	scsi_do_req(SRpnt, (void *)cmd, bp, bytes, osst_sleep_done, timeout, retries);
 
 	if (do_wait) {
-		wait_for_completion(SRpnt->sr_request.waiting);
-		SRpnt->sr_request.waiting = NULL;
+		wait_for_completion(SRpnt->sr_request->waiting);
+		SRpnt->sr_request->waiting = NULL;
 		STp->buffer->syscall_result = osst_chk_result(STp, SRpnt);
 #ifdef OSST_INJECT_ERRORS
 		if (STp->buffer->syscall_result == 0 &&
@@ -378,7 +378,7 @@
 		STp->nbr_finished++;
 #endif
 	wait_for_completion(&(STp->wait));
-	(STp->buffer)->last_SRpnt->sr_request.waiting = NULL;
+	(STp->buffer)->last_SRpnt->sr_request->waiting = NULL;
 
 	STp->buffer->syscall_result = osst_chk_result(STp, STp->buffer->last_SRpnt);
 
diff -Nru a/drivers/scsi/qla1280.c b/drivers/scsi/qla1280.c
--- a/drivers/scsi/qla1280.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/qla1280.c	Tue May 21 18:22:18 2002
@@ -4236,7 +4236,7 @@
 				/* Set transfer direction (READ and WRITE) */
 				/* Linux doesn't tell us                   */
 				/*
-				 * For block devices, cmd->request.cmd has the operation
+				 * For block devices, cmd->request->cmd has the operation
 				 * For character devices, this isn't always set properly, so
 				 * we need to check data_cmnd[0].  This catches the conditions
 				 * for st.c, but not sg. Generic commands are pass down to us.
@@ -6241,7 +6241,7 @@
 	       cmd->tag, cmd->flags, cmd->transfersize);
 	printk("  Pid=%li, SP=0x%p\n", cmd->pid, CMD_SP(cmd));
 	printk(" underflow size = 0x%x, direction=0x%x, req.cmd=0x%x \n",
-	       cmd->underflow, sp->dir, cmd->request.cmd);
+	       cmd->underflow, sp->dir, cmd->request->cmd);
 }
 
 /**************************************************************************
diff -Nru a/drivers/scsi/scsi.c b/drivers/scsi/scsi.c
--- a/drivers/scsi/scsi.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/scsi.c	Tue May 21 18:22:18 2002
@@ -256,7 +256,7 @@
 {
 	struct request *req;
 
-	req = &SCpnt->request;
+	req = SCpnt->request;
 	req->rq_status = RQ_SCSI_DONE;	/* Busy, but indicate request done */
 
 	if (req->waiting)
@@ -297,17 +297,19 @@
 Scsi_Request *scsi_allocate_request(Scsi_Device * device)
 {
   	Scsi_Request *SRpnt = NULL;
+        const int offset = ALIGN(sizeof(Scsi_Request), 4);
+        const int size = offset + sizeof(struct request);
   
   	if (!device)
   		panic("No device passed to scsi_allocate_request().\n");
   
-	SRpnt = (Scsi_Request *) kmalloc(sizeof(Scsi_Request), GFP_ATOMIC);
+        SRpnt = (Scsi_Request *) kmalloc(size, GFP_ATOMIC);
 	if( SRpnt == NULL )
 	{
 		return NULL;
 	}
-
-	memset(SRpnt, 0, sizeof(Scsi_Request));
+	memset(SRpnt, 0, size);
+        SRpnt->sr_request = (struct request *)(((char *)SRpnt) + offset);
 	SRpnt->sr_device = device;
 	SRpnt->sr_host = device->host;
 	SRpnt->sr_magic = SCSI_REQ_MAGIC;
@@ -435,7 +437,7 @@
 			 * Now we can check for a free command block for this device.
 			 */
 			for (SCpnt = device->device_queue; SCpnt; SCpnt = SCpnt->next) {
-				if (SCpnt->request.rq_status == RQ_INACTIVE)
+				if (SCpnt->request == NULL)
 					break;
 			}
 		}
@@ -504,9 +506,7 @@
 		}
 	}
 
-	SCpnt->request.rq_status = RQ_SCSI_BUSY;
-	SCpnt->request.waiting = NULL;	/* And no one is waiting for this
-					 * to complete */
+	SCpnt->request = NULL;
 	atomic_inc(&SCpnt->host->host_active);
 	atomic_inc(&SCpnt->device->device_active);
 
@@ -549,7 +549,7 @@
 
         SDpnt = SCpnt->device;
 
-	SCpnt->request.rq_status = RQ_INACTIVE;
+	SCpnt->request = NULL;
 	SCpnt->state = SCSI_STATE_UNUSED;
 	SCpnt->owner = SCSI_OWNER_NOBODY;
 	atomic_dec(&SCpnt->host->host_active);
@@ -772,13 +772,13 @@
 	DECLARE_COMPLETION(wait);
 	request_queue_t *q = &SRpnt->sr_device->request_queue;
 	
-	SRpnt->sr_request.waiting = &wait;
-	SRpnt->sr_request.rq_status = RQ_SCSI_BUSY;
+	SRpnt->sr_request->waiting = &wait;
+	SRpnt->sr_request->rq_status = RQ_SCSI_BUSY;
 	scsi_do_req (SRpnt, (void *) cmnd,
 		buffer, bufflen, scsi_wait_done, timeout, retries);
 	generic_unplug_device(q);
 	wait_for_completion(&wait);
-	SRpnt->sr_request.waiting = NULL;
+	SRpnt->sr_request->waiting = NULL;
 	if( SRpnt->sr_command != NULL )
 	{
 		scsi_release_command(SRpnt->sr_command);
@@ -929,8 +929,7 @@
 	SCpnt->cmd_len = SRpnt->sr_cmd_len;
 	SCpnt->use_sg = SRpnt->sr_use_sg;
 
-	memcpy((void *) &SCpnt->request, (const void *) &SRpnt->sr_request,
-	       sizeof(SRpnt->sr_request));
+        SCpnt->request = SRpnt->sr_request;
 	memcpy((void *) SCpnt->data_cmnd, (const void *) SRpnt->sr_cmnd, 
 	       sizeof(SCpnt->data_cmnd));
 	SCpnt->reset_chain = NULL;
@@ -1488,7 +1487,7 @@
 		SCpnt->target = SDpnt->id;
 		SCpnt->lun = SDpnt->lun;
 		SCpnt->channel = SDpnt->channel;
-		SCpnt->request.rq_status = RQ_INACTIVE;
+		SCpnt->request = NULL;
 		SCpnt->use_sg = 0;
 		SCpnt->old_use_sg = 0;
 		SCpnt->old_cmd_len = 0;
@@ -2060,16 +2059,16 @@
 			     SCpnt = SCpnt->next) {
 				online_status = SDpnt->online;
 				SDpnt->online = FALSE;
-				if (SCpnt->request.rq_status != RQ_INACTIVE) {
+				if (SCpnt->request && SCpnt->request->rq_status != RQ_INACTIVE) {
 					printk(KERN_ERR "SCSI device not inactive - rq_status=%d, target=%d, pid=%ld, state=%d, owner=%d.\n",
-					       SCpnt->request.rq_status, SCpnt->target, SCpnt->pid,
+					       SCpnt->request->rq_status, SCpnt->target, SCpnt->pid,
 					     SCpnt->state, SCpnt->owner);
 					for (SDpnt1 = shpnt->host_queue; SDpnt1;
 					     SDpnt1 = SDpnt1->next) {
 						for (SCpnt = SDpnt1->device_queue; SCpnt;
 						     SCpnt = SCpnt->next)
-							if (SCpnt->request.rq_status == RQ_SCSI_DISCONNECTING)
-								SCpnt->request.rq_status = RQ_INACTIVE;
+							if (SCpnt->request->rq_status == RQ_SCSI_DISCONNECTING)
+								SCpnt->request->rq_status = RQ_INACTIVE;
 					}
 					SDpnt->online = online_status;
 					printk(KERN_ERR "Device busy???\n");
@@ -2080,7 +2079,8 @@
 				 * continue on.
 				 */
 				SCpnt->state = SCSI_STATE_DISCONNECTING;
-				SCpnt->request.rq_status = RQ_SCSI_DISCONNECTING;	/* Mark as busy */
+                                if(SCpnt->request)
+                                        SCpnt->request->rq_status = RQ_SCSI_DISCONNECTING;	/* Mark as busy */
 			}
 		}
 	}
@@ -2390,11 +2390,11 @@
 				       SCpnt->target,
 				       SCpnt->lun,
 
-				       kdevname(SCpnt->request.rq_dev),
-				       SCpnt->request.sector,
-				       SCpnt->request.nr_sectors,
-				       (long)SCpnt->request.current_nr_sectors,
-				       SCpnt->request.rq_status,
+				       kdevname(SCpnt->request->rq_dev),
+				       SCpnt->request->sector,
+				       SCpnt->request->nr_sectors,
+				       (long)SCpnt->request->current_nr_sectors,
+				       SCpnt->request->rq_status,
 				       SCpnt->use_sg,
 
 				       SCpnt->retries,
@@ -2711,16 +2711,18 @@
 scsi_reset_provider(Scsi_Device *dev, int flag)
 {
 	Scsi_Cmnd SC, *SCpnt = &SC;
+        struct request req;
 	int rtn;
 
+        SCpnt->request = &req;
 	memset(&SCpnt->eh_timeout, 0, sizeof(SCpnt->eh_timeout));
 	SCpnt->host                    	= dev->host;
 	SCpnt->device                  	= dev;
 	SCpnt->target                  	= dev->id;
 	SCpnt->lun                     	= dev->lun;
 	SCpnt->channel                 	= dev->channel;
-	SCpnt->request.rq_status       	= RQ_SCSI_BUSY;
-	SCpnt->request.waiting        	= NULL;
+	SCpnt->request->rq_status      	= RQ_SCSI_BUSY;
+	SCpnt->request->waiting        	= NULL;
 	SCpnt->use_sg                  	= 0;
 	SCpnt->old_use_sg              	= 0;
 	SCpnt->old_cmd_len             	= 0;
diff -Nru a/drivers/scsi/scsi.h b/drivers/scsi/scsi.h
--- a/drivers/scsi/scsi.h	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/scsi.h	Tue May 21 18:22:18 2002
@@ -656,7 +656,7 @@
 	struct Scsi_Host *sr_host;
 	Scsi_Device *sr_device;
 	Scsi_Cmnd *sr_command;
-	struct request sr_request;	/* A copy of the command we are
+	struct request *sr_request;	/* A copy of the command we are
 				   working on */
 	unsigned sr_bufflen;	/* Size of data buffer */
 	void *sr_buffer;		/* Data buffer */
@@ -767,8 +767,8 @@
 				   transferred less actual number
 				   transferred (0 if not supported) */
 
-	struct request request;	/* A copy of the command we are
-				   working on */
+	struct request *request;	/* The command we are
+				   	   working on */
 
 	unsigned char sense_buffer[SCSI_SENSE_BUFFERSIZE];		/* obtained by REQUEST SENSE
 						 * when CHECK CONDITION is
diff -Nru a/drivers/scsi/scsi_debug.c b/drivers/scsi/scsi_debug.c
--- a/drivers/scsi/scsi_debug.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/scsi_debug.c	Tue May 21 18:22:18 2002
@@ -695,7 +695,7 @@
 	{
 		int delay = SCSI_SETUP_LATENCY;
 
-		delay += SCpnt->request.nr_sectors * SCSI_DATARATE;
+		delay += SCpnt->request->nr_sectors * SCSI_DATARATE;
 		if (delay)
 			usleep(delay);
 	}
diff -Nru a/drivers/scsi/scsi_error.c b/drivers/scsi/scsi_error.c
--- a/drivers/scsi/scsi_error.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/scsi_error.c	Tue May 21 18:22:18 2002
@@ -320,7 +320,7 @@
 		return;
 	}
 
-	SCpnt->request.rq_status = RQ_SCSI_DONE;
+	SCpnt->request->rq_status = RQ_SCSI_DONE;
 
 	SCpnt->owner = SCSI_OWNER_ERROR_HANDLER;
 	SCpnt->eh_state = SUCCESS;
@@ -347,7 +347,7 @@
 STATIC
 void scsi_eh_action_done(Scsi_Cmnd * SCpnt, int answer)
 {
-	SCpnt->request.rq_status = RQ_SCSI_DONE;
+	SCpnt->request->rq_status = RQ_SCSI_DONE;
 
 	SCpnt->owner = SCSI_OWNER_ERROR_HANDLER;
 	SCpnt->eh_state = (answer ? SUCCESS : FAILED);
@@ -602,7 +602,7 @@
 		 * Set up the semaphore so we wait for the command to complete.
 		 */
 		SCpnt->host->eh_action = &sem;
-		SCpnt->request.rq_status = RQ_SCSI_BUSY;
+		SCpnt->request->rq_status = RQ_SCSI_BUSY;
 
 		spin_lock_irqsave(SCpnt->host->host_lock, flags);
 		host->hostt->queuecommand(SCpnt, scsi_eh_done);
@@ -634,7 +634,7 @@
 				SCpnt->host->hostt->eh_abort_handler(SCpnt);
 			spin_unlock_irqrestore(SCpnt->host->host_lock, flags);
 			
-			SCpnt->request.rq_status = RQ_SCSI_DONE;
+			SCpnt->request->rq_status = RQ_SCSI_DONE;
 			SCpnt->owner = SCSI_OWNER_ERROR_HANDLER;
 			
 			SCpnt->eh_state = FAILED;
diff -Nru a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
--- a/drivers/scsi/scsi_lib.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/scsi_lib.c	Tue May 21 18:22:18 2002
@@ -79,9 +79,6 @@
 	rq->flags = REQ_SPECIAL | REQ_BARRIER;
 
 	rq->special = data;
-	rq->q = NULL;
-	rq->bio = rq->biotail = NULL;
-	rq->nr_phys_segments = 0;
 
 	/*
 	 * We have the option of inserting the head or the tail of the queue.
@@ -120,7 +117,7 @@
 {
 	request_queue_t *q = &SCpnt->device->request_queue;
 
-	__scsi_insert_special(q, &SCpnt->request, SCpnt, at_head);
+	__scsi_insert_special(q, SCpnt->request, SCpnt, at_head);
 	return 0;
 }
 
@@ -148,7 +145,7 @@
 {
 	request_queue_t *q = &SRpnt->sr_device->request_queue;
 
-	__scsi_insert_special(q, &SRpnt->sr_request, SRpnt, at_head);
+	__scsi_insert_special(q, SRpnt->sr_request, SRpnt, at_head);
 	return 0;
 }
 
@@ -259,8 +256,8 @@
 		 * in which case we need to request the blocks that come after
 		 * the bad sector.
 		 */
-		SCpnt->request.special = (void *) SCpnt;
-		_elv_add_request(q, &SCpnt->request, 0, 0);
+		SCpnt->request->special = (void *) SCpnt;
+		_elv_add_request(q, SCpnt->request, 0, 0);
 	}
 
 	/*
@@ -356,15 +353,18 @@
 				     int frequeue)
 {
 	request_queue_t *q = &SCpnt->device->request_queue;
-	struct request *req = &SCpnt->request;
+	struct request *req = SCpnt->request;
+	int flags;
 
 	ASSERT_LOCK(q->queue_lock, 0);
 
+	spin_lock_irqsave(q->queue_lock, flags);
 	/*
 	 * If there are blocks left over at the end, set up the command
 	 * to queue the remainder of them.
 	 */
 	if (end_that_request_first(req, uptodate, sectors)) {
+		spin_unlock_irqrestore(q->queue_lock, flags);
 		if (!requeue)
 			return SCpnt;
 
@@ -376,15 +376,12 @@
 		return SCpnt;
 	}
 
-	/*
-	 * This request is done.  If there is someone blocked waiting for this
-	 * request, wake them up.
-	 */
-	if (req->waiting)
-		complete(req->waiting);
-
 	add_blkdev_randomness(major(req->rq_dev));
 
+	end_that_request_last(req);
+
+	spin_unlock_irqrestore(q->queue_lock, flags);
+
 	/*
 	 * This will goose the queue request function at the end, so we don't
 	 * need to worry about launching another command.
@@ -441,7 +438,7 @@
  */
 static void scsi_release_buffers(Scsi_Cmnd * SCpnt)
 {
-	struct request *req = &SCpnt->request;
+	struct request *req = SCpnt->request;
 
 	ASSERT_LOCK(SCpnt->host->host_lock, 0);
 
@@ -491,7 +488,7 @@
 	int result = SCpnt->result;
 	int this_count = SCpnt->bufflen >> 9;
 	request_queue_t *q = &SCpnt->device->request_queue;
-	struct request *req = &SCpnt->request;
+	struct request *req = SCpnt->request;
 
 	/*
 	 * We must do one of several things here:
@@ -675,7 +672,7 @@
 	if (result) {
 		struct Scsi_Device_Template *STpnt;
 
-		STpnt = scsi_get_request_dev(&SCpnt->request);
+		STpnt = scsi_get_request_dev(SCpnt->request);
 		printk("SCSI %s error : host %d channel %d id %d lun %d return code = %x\n",
 		       (STpnt ? STpnt->name : "device"),
 		       SCpnt->device->host->host_no,
@@ -868,7 +865,7 @@
 		 * the remainder of a partially fulfilled request that can 
 		 * come up when there is a medium error.  We have to treat
 		 * these two cases differently.  We differentiate by looking
-		 * at request.cmd, as this tells us the real story.
+		 * at request->cmd, as this tells us the real story.
 		 */
 		if (req->flags & REQ_SPECIAL) {
 			STpnt = NULL;
@@ -904,6 +901,9 @@
 			 */
 			if (!SCpnt)
 				break;
+
+			/* pull a tag out of the request if we have one */
+			SCpnt->tag = req->tag;
 		} else {
 			blk_dump_rq_flags(req, "SCSI bad req");
 			break;
@@ -926,16 +926,8 @@
 		 */
 		blkdev_dequeue_request(req);
 
-		if (req != &SCpnt->request && req != &SRpnt->sr_request ) {
-			memcpy(&SCpnt->request, req, sizeof(struct request));
+		SCpnt->request = req;
 
-			/*
-			 * We have copied the data out of the request block -
-			 * it is now in a field in SCpnt.  Release the request
-			 * block.
-			 */
-			blkdev_release_request(req);
-		}
 		/*
 		 * Now it is finally safe to release the lock.  We are
 		 * not going to noodle the request list until this
@@ -945,7 +937,7 @@
 		req = NULL;
 		spin_unlock_irq(q->queue_lock);
 
-		if (SCpnt->request.flags & REQ_CMD) {
+		if (SCpnt->request->flags & REQ_CMD) {
 			/*
 			 * This will do a couple of things:
 			 *  1) Fill in the actual SCSI command.
@@ -959,7 +951,7 @@
 			 * request to be rejected immediately.
 			 */
 			if (STpnt == NULL)
-				STpnt = scsi_get_request_dev(&SCpnt->request);
+				STpnt = scsi_get_request_dev(SCpnt->request);
 
 			/* 
 			 * This sets up the scatter-gather table (allocating if
@@ -973,9 +965,9 @@
 					SDpnt->starved = 1;
 					SHpnt->some_device_starved = 1;
 				}
-				SCpnt->request.special = SCpnt;
-				SCpnt->request.flags |= REQ_SPECIAL;
-				_elv_add_request(q, &SCpnt->request, 0, 0);
+				SCpnt->request->special = SCpnt;
+				SCpnt->request->flags |= REQ_SPECIAL;
+				_elv_add_request(q, SCpnt->request, 0, 0);
 				break;
 			}
 
@@ -985,7 +977,7 @@
 			if (!STpnt->init_command(SCpnt)) {
 				scsi_release_buffers(SCpnt);
 				SCpnt = __scsi_end_request(SCpnt, 0, 
-							   SCpnt->request.nr_sectors, 0, 0);
+							   SCpnt->request->nr_sectors, 0, 0);
 				if( SCpnt != NULL )
 				{
 					panic("Should not have leftover blocks\n");
diff -Nru a/drivers/scsi/scsi_merge.c b/drivers/scsi/scsi_merge.c
--- a/drivers/scsi/scsi_merge.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/scsi_merge.c	Tue May 21 18:22:18 2002
@@ -58,7 +58,7 @@
  */
 int scsi_init_io(Scsi_Cmnd *SCpnt)
 {
-	struct request     *req = &SCpnt->request;
+	struct request     *req = SCpnt->request;
 	struct scatterlist *sgpnt;
 	int count, gfp_mask;
 
diff -Nru a/drivers/scsi/sd.c b/drivers/scsi/sd.c
--- a/drivers/scsi/sd.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/sd.c	Tue May 21 18:22:18 2002
@@ -324,13 +324,13 @@
 	/*
 	 * don't support specials for nwo
 	 */
-	if (!(SCpnt->request.flags & REQ_CMD))
+	if (!(SCpnt->request->flags & REQ_CMD))
 		return 0;
 
-	part_nr = SD_PARTITION(SCpnt->request.rq_dev);
-	dsk_nr = DEVICE_NR(SCpnt->request.rq_dev);
+	part_nr = SD_PARTITION(SCpnt->request->rq_dev);
+	dsk_nr = DEVICE_NR(SCpnt->request->rq_dev);
 
-	block = SCpnt->request.sector;
+	block = SCpnt->request->sector;
 	this_count = SCpnt->request_bufflen >> 9;
 
 	SCSI_LOG_HLQUEUE(1, printk("sd_command_init: dsk_nr=%d, block=%d, "
@@ -341,9 +341,9 @@
 	/* >>>>> this change is not in the lk 2.5 series */
 	if (part_nr >= (sd_template.dev_max << 4) || (part_nr & 0xf) ||
 	    !sdp || !sdp->online ||
- 	    block + SCpnt->request.nr_sectors > sd[part_nr].nr_sects) {
+ 	    block + SCpnt->request->nr_sectors > sd[part_nr].nr_sects) {
 		SCSI_LOG_HLQUEUE(2, printk("Finishing %ld sectors\n", 
-				 SCpnt->request.nr_sectors));
+				 SCpnt->request->nr_sectors));
 		SCSI_LOG_HLQUEUE(2, printk("Retry with 0x%p\n", SCpnt));
 		return 0;
 	}
@@ -372,7 +372,7 @@
 	 * for this.
 	 */
 	if (sdp->sector_size == 1024) {
-		if ((block & 1) || (SCpnt->request.nr_sectors & 1)) {
+		if ((block & 1) || (SCpnt->request->nr_sectors & 1)) {
 			printk(KERN_ERR "sd: Bad block number requested");
 			return 0;
 		} else {
@@ -381,7 +381,7 @@
 		}
 	}
 	if (sdp->sector_size == 2048) {
-		if ((block & 3) || (SCpnt->request.nr_sectors & 3)) {
+		if ((block & 3) || (SCpnt->request->nr_sectors & 3)) {
 			printk(KERN_ERR "sd: Bad block number requested");
 			return 0;
 		} else {
@@ -390,7 +390,7 @@
 		}
 	}
 	if (sdp->sector_size == 4096) {
-		if ((block & 7) || (SCpnt->request.nr_sectors & 7)) {
+		if ((block & 7) || (SCpnt->request->nr_sectors & 7)) {
 			printk(KERN_ERR "sd: Bad block number requested");
 			return 0;
 		} else {
@@ -398,25 +398,25 @@
 			this_count = this_count >> 3;
 		}
 	}
-	if (rq_data_dir(&SCpnt->request) == WRITE) {
+	if (rq_data_dir(SCpnt->request) == WRITE) {
 		if (!sdp->writeable) {
 			return 0;
 		}
 		SCpnt->cmnd[0] = WRITE_6;
 		SCpnt->sc_data_direction = SCSI_DATA_WRITE;
-	} else if (rq_data_dir(&SCpnt->request) == READ) {
+	} else if (rq_data_dir(SCpnt->request) == READ) {
 		SCpnt->cmnd[0] = READ_6;
 		SCpnt->sc_data_direction = SCSI_DATA_READ;
 	} else {
 		printk(KERN_ERR "sd: Unknown command %lx\n", 
-		       SCpnt->request.flags);
-/* overkill 	panic("Unknown sd command %lx\n", SCpnt->request.flags); */
+		       SCpnt->request->flags);
+/* overkill 	panic("Unknown sd command %lx\n", SCpnt->request->flags); */
 		return 0;
 	}
 
 	SCSI_LOG_HLQUEUE(2, printk("%s : %s %d/%ld 512 byte blocks.\n", 
-		nbuff, (rq_data_dir(&SCpnt->request) == WRITE) ? 
-		"writing" : "reading", this_count, SCpnt->request.nr_sectors));
+		nbuff, (rq_data_dir(SCpnt->request) == WRITE) ? 
+		"writing" : "reading", this_count, SCpnt->request->nr_sectors));
 
 	SCpnt->cmnd[1] = (SCpnt->device->scsi_level <= SCSI_2) ?
 			 ((SCpnt->lun << 5) & 0xe0) : 0;
@@ -662,7 +662,7 @@
 #if CONFIG_SCSI_LOGGING
 	char nbuff[6];
 
-	SCSI_LOG_HLCOMPLETE(1, sd_dskname(DEVICE_NR(SCpnt->request.rq_dev), 
+	SCSI_LOG_HLCOMPLETE(1, sd_dskname(DEVICE_NR(SCpnt->request->rq_dev), 
 			    nbuff));
 	SCSI_LOG_HLCOMPLETE(1, printk("sd_rw_intr: %s: res=0x%x\n", 
 				      nbuff, result));
@@ -688,8 +688,8 @@
 			(SCpnt->sense_buffer[4] << 16) |
 			(SCpnt->sense_buffer[5] << 8) |
 			SCpnt->sense_buffer[6];
-			if (SCpnt->request.bio != NULL)
-				block_sectors = bio_sectors(SCpnt->request.bio);
+			if (SCpnt->request->bio != NULL)
+				block_sectors = bio_sectors(SCpnt->request->bio);
 			switch (SCpnt->device->sector_size) {
 			case 1024:
 				error_sector <<= 1;
@@ -714,7 +714,7 @@
 			}
 
 			error_sector &= ~(block_sectors - 1);
-			good_sectors = error_sector - SCpnt->request.sector;
+			good_sectors = error_sector - SCpnt->request->sector;
 			if (good_sectors < 0 || good_sectors >= this_count)
 				good_sectors = 0;
 			break;
diff -Nru a/drivers/scsi/sg.c b/drivers/scsi/sg.c
--- a/drivers/scsi/sg.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/sg.c	Tue May 21 18:22:18 2002
@@ -695,7 +695,7 @@
 
     srp->my_cmdp = SRpnt;
     q = &SRpnt->sr_device->request_queue;
-    SRpnt->sr_request.rq_dev = sdp->i_rdev;
+    SRpnt->sr_request->rq_dev = sdp->i_rdev;
     SRpnt->sr_sense_buffer[0] = 0;
     SRpnt->sr_cmd_len = hp->cmd_len;
     if (! (hp->flags & SG_FLAG_LUN_INHIBIT)) {
@@ -1222,7 +1222,7 @@
     SRpnt->sr_bufflen = 0;
     SRpnt->sr_buffer = NULL;
     SRpnt->sr_underflow = 0;
-    SRpnt->sr_request.rq_dev = mk_kdev(0, 0);  /* "sg" _disowns_ request blk */
+    SRpnt->sr_request->rq_dev = mk_kdev(0, 0);  /* "sg" _disowns_ request blk */
 
     srp->my_cmdp = NULL;
     srp->done = 1;
diff -Nru a/drivers/scsi/sr.c b/drivers/scsi/sr.c
--- a/drivers/scsi/sr.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/sr.c	Tue May 21 18:22:18 2002
@@ -196,11 +196,11 @@
 	int this_count = SCpnt->bufflen >> 9;
 	int good_sectors = (result == 0 ? this_count : 0);
 	int block_sectors = 0;
-	int device_nr = DEVICE_NR(SCpnt->request.rq_dev);
+	int device_nr = DEVICE_NR(SCpnt->request->rq_dev);
 	Scsi_CD *SCp = &scsi_CDs[device_nr];
 
 #ifdef DEBUG
-	printk("sr.c done: %x %p\n", result, SCpnt->request.bh->b_data);
+	printk("sr.c done: %x %p\n", result, SCpnt->request->bh->b_data);
 #endif
 	/*
 	   Handle MEDIUM ERRORs or VOLUME OVERFLOWs that indicate partial success.
@@ -218,14 +218,14 @@
 		(SCpnt->sense_buffer[4] << 16) |
 		(SCpnt->sense_buffer[5] << 8) |
 		SCpnt->sense_buffer[6];
-		if (SCpnt->request.bio != NULL)
-			block_sectors = bio_sectors(SCpnt->request.bio);
+		if (SCpnt->request->bio != NULL)
+			block_sectors = bio_sectors(SCpnt->request->bio);
 		if (block_sectors < 4)
 			block_sectors = 4;
 		if (SCp->device->sector_size == 2048)
 			error_sector <<= 2;
 		error_sector &= ~(block_sectors - 1);
-		good_sectors = error_sector - SCpnt->request.sector;
+		good_sectors = error_sector - SCpnt->request->sector;
 		if (good_sectors < 0 || good_sectors >= this_count)
 			good_sectors = 0;
 		/*
@@ -265,14 +265,14 @@
 	int dev, devm, block=0, this_count, s_size;
 	Scsi_CD *SCp;
 
-	devm = minor(SCpnt->request.rq_dev);
-	dev = DEVICE_NR(SCpnt->request.rq_dev);
+	devm = minor(SCpnt->request->rq_dev);
+	dev = DEVICE_NR(SCpnt->request->rq_dev);
 	SCp = &scsi_CDs[dev];
 
 	SCSI_LOG_HLQUEUE(1, printk("Doing sr request, dev = %d, block = %d\n", devm, block));
 
 	if (dev >= sr_template.nr_dev || !SCp->device || !SCp->device->online) {
-		SCSI_LOG_HLQUEUE(2, printk("Finishing %ld sectors\n", SCpnt->request.nr_sectors));
+		SCSI_LOG_HLQUEUE(2, printk("Finishing %ld sectors\n", SCpnt->request->nr_sectors));
 		SCSI_LOG_HLQUEUE(2, printk("Retry with 0x%p\n", SCpnt));
 		return 0;
 	}
@@ -285,8 +285,8 @@
 		return 0;
 	}
 
-	if (!(SCpnt->request.flags & REQ_CMD)) {
-		blk_dump_rq_flags(&SCpnt->request, "sr unsup command");
+	if (!(SCpnt->request->flags & REQ_CMD)) {
+		blk_dump_rq_flags(SCpnt->request, "sr unsup command");
 		return 0;
 	}
 
@@ -307,23 +307,23 @@
 		return 0;
 	}
 
-	if (rq_data_dir(&SCpnt->request) == WRITE) {
+	if (rq_data_dir(SCpnt->request) == WRITE) {
 		if (!SCp->device->writeable)
 			return 0;
 		SCpnt->cmnd[0] = WRITE_10;
 		SCpnt->sc_data_direction = SCSI_DATA_WRITE;
-	} else if (rq_data_dir(&SCpnt->request) == READ) {
+	} else if (rq_data_dir(SCpnt->request) == READ) {
 		SCpnt->cmnd[0] = READ_10;
 		SCpnt->sc_data_direction = SCSI_DATA_READ;
 	} else {
-		blk_dump_rq_flags(&SCpnt->request, "Unknown sr command");
+		blk_dump_rq_flags(SCpnt->request, "Unknown sr command");
 		return 0;
 	}
 
 	/*
 	 * request doesn't start on hw block boundary, add scatter pads
 	 */
-	if ((SCpnt->request.sector % (s_size >> 9)) || (SCpnt->request_bufflen % s_size)) {
+	if ((SCpnt->request->sector % (s_size >> 9)) || (SCpnt->request_bufflen % s_size)) {
 		printk("sr: unaligned transfer\n");
 		return 0;
 	}
@@ -333,13 +333,13 @@
 
 	SCSI_LOG_HLQUEUE(2, printk("sr%d : %s %d/%ld 512 byte blocks.\n",
                                    devm,
-		   (rq_data_dir(&SCpnt->request) == WRITE) ? "writing" : "reading",
-				 this_count, SCpnt->request.nr_sectors));
+		   (rq_data_dir(SCpnt->request) == WRITE) ? "writing" : "reading",
+				 this_count, SCpnt->request->nr_sectors));
 
 	SCpnt->cmnd[1] = (SCpnt->device->scsi_level <= SCSI_2) ?
 			 ((SCpnt->lun << 5) & 0xe0) : 0;
 
-	block = SCpnt->request.sector / (s_size >> 9);
+	block = SCpnt->request->sector / (s_size >> 9);
 
 	if (this_count > 0xffff)
 		this_count = 0xffff;
@@ -495,7 +495,7 @@
 		cmd[1] = (SCp->device->scsi_level <= SCSI_2) ?
 			 ((SCp->device->lun << 5) & 0xe0) : 0;
 		memset((void *) &cmd[2], 0, 8);
-		SRpnt->sr_request.rq_status = RQ_SCSI_BUSY;	/* Mark as really busy */
+		SRpnt->sr_request->rq_status = RQ_SCSI_BUSY;	/* Mark as really busy */
 		SRpnt->sr_cmd_len = 0;
 
 		memset(buffer, 0, 8);
diff -Nru a/drivers/scsi/sr_ioctl.c b/drivers/scsi/sr_ioctl.c
--- a/drivers/scsi/sr_ioctl.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/sr_ioctl.c	Tue May 21 18:22:18 2002
@@ -92,7 +92,7 @@
 	SRpnt->sr_data_direction = readwrite;
 
 	/* use ISA DMA buffer if necessary */
-	SRpnt->sr_request.buffer = buffer;
+	SRpnt->sr_request->buffer = buffer;
 	if (buffer && SRpnt->sr_host->unchecked_isa_dma &&
 	    (virt_to_phys(buffer) + buflength - 1 > ISA_DMA_THRESHOLD)) {
 		bounce_buffer = (char *) kmalloc(buflength, GFP_DMA);
@@ -111,7 +111,7 @@
 	scsi_wait_req(SRpnt, (void *) sr_cmd, (void *) buffer, buflength,
 		      IOCTL_TIMEOUT, IOCTL_RETRIES);
 
-	req = &SRpnt->sr_request;
+	req = SRpnt->sr_request;
 	if (SRpnt->sr_buffer && req->buffer && SRpnt->sr_buffer != req->buffer) {
 		memcpy(req->buffer, SRpnt->sr_buffer, SRpnt->sr_bufflen);
 		kfree(SRpnt->sr_buffer);
diff -Nru a/drivers/scsi/st.c b/drivers/scsi/st.c
--- a/drivers/scsi/st.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/st.c	Tue May 21 18:22:18 2002
@@ -241,7 +241,7 @@
 		scode = 0;
 	}
 
-	dev = TAPE_NR(SRpnt->sr_request.rq_dev);
+	dev = TAPE_NR(SRpnt->sr_request->rq_dev);
         DEB(
         if (debugging) {
                 printk(ST_DEB_MSG "st%d: Error: %x, cmd: %x %x %x %x %x %x Len: %d\n",
@@ -318,7 +318,7 @@
 	int remainder;
 	Scsi_Tape *STp;
 
-	st_nbr = TAPE_NR(SCpnt->request.rq_dev);
+	st_nbr = TAPE_NR(SCpnt->request->rq_dev);
 	read_lock(&st_dev_arr_lock);
 	STp = scsi_tapes[st_nbr];
 	read_unlock(&st_dev_arr_lock);
@@ -340,11 +340,11 @@
 			(STp->buffer)->midlevel_result = INT_MAX;	/* OK */
 	} else
 		(STp->buffer)->midlevel_result = SCpnt->result;
-	SCpnt->request.rq_status = RQ_SCSI_DONE;
+	SCpnt->request->rq_status = RQ_SCSI_DONE;
 	(STp->buffer)->last_SRpnt = SCpnt->sc_request;
 	DEB( STp->write_pending = 0; )
 
-	complete(SCpnt->request.waiting);
+	complete(SCpnt->request->waiting);
 }
 
 
@@ -383,16 +383,16 @@
 		bp = (STp->buffer)->b_data;
 	SRpnt->sr_data_direction = direction;
 	SRpnt->sr_cmd_len = 0;
-	SRpnt->sr_request.waiting = &(STp->wait);
-	SRpnt->sr_request.rq_status = RQ_SCSI_BUSY;
-	SRpnt->sr_request.rq_dev = STp->devt;
+	SRpnt->sr_request->waiting = &(STp->wait);
+	SRpnt->sr_request->rq_status = RQ_SCSI_BUSY;
+	SRpnt->sr_request->rq_dev = STp->devt;
 
 	scsi_do_req(SRpnt, (void *) cmd, bp, bytes,
 		    st_sleep_done, timeout, retries);
 
 	if (do_wait) {
-		wait_for_completion(SRpnt->sr_request.waiting);
-		SRpnt->sr_request.waiting = NULL;
+		wait_for_completion(SRpnt->sr_request->waiting);
+		SRpnt->sr_request->waiting = NULL;
 		(STp->buffer)->syscall_result = st_chk_result(STp, SRpnt);
 	}
 	return SRpnt;
@@ -415,7 +415,7 @@
         ) /* end DEB */
 
 	wait_for_completion(&(STp->wait));
-	(STp->buffer)->last_SRpnt->sr_request.waiting = NULL;
+	(STp->buffer)->last_SRpnt->sr_request->waiting = NULL;
 
 	(STp->buffer)->syscall_result = st_chk_result(STp, (STp->buffer)->last_SRpnt);
 	scsi_release_request((STp->buffer)->last_SRpnt);
diff -Nru a/drivers/scsi/sun3_NCR5380.c b/drivers/scsi/sun3_NCR5380.c
--- a/drivers/scsi/sun3_NCR5380.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/sun3_NCR5380.c	Tue May 21 18:22:18 2002
@@ -1217,7 +1217,7 @@
 	       HOSTNO, NCR5380_read(BUS_AND_STATUS_REG),
 	       NCR5380_read(STATUS_REG));
 
-    if((sun3scsi_dma_finish(hostdata->connected->request.cmd))) {
+    if((sun3scsi_dma_finish(hostdata->connected->request->cmd))) {
 	    printk("scsi%d: overrun in UDC counter -- not prepared to deal with this!\n", HOSTNO);
 	    printk("please e-mail sammy@oh.verio.com with a description of how this\n");
 	    printk("error was produced.\n");
@@ -2016,9 +2016,9 @@
 		if((count > SUN3_DMA_MINSIZE) && (sun3_dma_setup_done
 						  != cmd))
 		{
-			if((cmd->request.cmd == 0) || (cmd->request.cmd == 1)) {
+			if((cmd->request->cmd == 0) || (cmd->request->cmd == 1)) {
 				sun3scsi_dma_setup(d, count,
-						   cmd->request.cmd);
+						   cmd->request->cmd);
 				sun3_dma_setup_done = cmd;
 			}
 		}
@@ -2628,7 +2628,7 @@
 						  != tmp))
 		{
 			sun3scsi_dma_setup(d, count,
-					   tmp->request.cmd);
+					   tmp->request->cmd);
 			sun3_dma_setup_done = tmp;
 		}
 #endif
diff -Nru a/drivers/scsi/sun3_scsi.c b/drivers/scsi/sun3_scsi.c
--- a/drivers/scsi/sun3_scsi.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/sun3_scsi.c	Tue May 21 18:22:18 2002
@@ -519,7 +519,7 @@
 static inline unsigned long sun3scsi_dma_xfer_len(unsigned long wanted, Scsi_Cmnd *cmd,
 				    int write_flag)
 {
-	if((cmd->request.cmd == 0) || (cmd->request.cmd == 1))
+	if((cmd->request->cmd == 0) || (cmd->request->cmd == 1))
  		return wanted;
 	else
 		return 0;
diff -Nru a/drivers/scsi/u14-34f.c b/drivers/scsi/u14-34f.c
--- a/drivers/scsi/u14-34f.c	Tue May 21 18:22:18 2002
+++ b/drivers/scsi/u14-34f.c	Tue May 21 18:22:18 2002
@@ -1206,7 +1206,7 @@
    if (linked_comm && SCpnt->device->queue_depth > 2
                                      && TLDEV(SCpnt->device->type)) {
       HD(j)->cp_stat[i] = READY;
-      flush_dev(SCpnt->device, SCpnt->request.sector, j, FALSE);
+      flush_dev(SCpnt->device, SCpnt->request->sector, j, FALSE);
       return 0;
       }
 
@@ -1529,11 +1529,11 @@
 
       if (!(cpp->xdir == DTD_IN)) input_only = FALSE;
 
-      if (SCpnt->request.sector < minsec) minsec = SCpnt->request.sector;
-      if (SCpnt->request.sector > maxsec) maxsec = SCpnt->request.sector;
+      if (SCpnt->request->sector < minsec) minsec = SCpnt->request->sector;
+      if (SCpnt->request->sector > maxsec) maxsec = SCpnt->request->sector;
 
-      sl[n] = SCpnt->request.sector;
-      ioseek += SCpnt->request.nr_sectors;
+      sl[n] = SCpnt->request->sector;
+      ioseek += SCpnt->request->nr_sectors;
 
       if (!n) continue;
 
@@ -1561,7 +1561,7 @@
 
    if (!input_only) for (n = 0; n < n_ready; n++) {
       k = il[n]; cpp = &HD(j)->cp[k]; SCpnt = cpp->SCpnt;
-      ll[n] = SCpnt->request.nr_sectors; pl[n] = SCpnt->pid;
+      ll[n] = SCpnt->request->nr_sectors; pl[n] = SCpnt->pid;
 
       if (!n) continue;
 
@@ -1589,7 +1589,7 @@
                 " cur %ld s:%c r:%c rev:%c in:%c ov:%c xd %d.\n",
                 (ihdlr ? "ihdlr" : "qcomm"), SCpnt->channel, SCpnt->target,
                 SCpnt->lun, SCpnt->pid, k, flushcount, n_ready,
-                SCpnt->request.sector, SCpnt->request.nr_sectors, cursec,
+                SCpnt->request->sector, SCpnt->request->nr_sectors, cursec,
                 YESNO(s), YESNO(r), YESNO(rev), YESNO(input_only),
                 YESNO(overlap), cpp->xdir);
          }
@@ -1718,7 +1718,7 @@
 
    if (linked_comm && SCpnt->device->queue_depth > 2
                                      && TLDEV(SCpnt->device->type))
-      flush_dev(SCpnt->device, SCpnt->request.sector, j, TRUE);
+      flush_dev(SCpnt->device, SCpnt->request->sector, j, TRUE);
 
    tstatus = status_byte(spp->target_status);
 

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Patrick Mansfield]

On Tue, May 21, 2002 at 07:11:34PM -0400, James Bottomley wrote:
> The attached makes the request field in Scsi_Cmnd and Scsi_Request be pointers 
> to the block layer request instead of copies of it.  The advantage of this is 
> that it makes the scsi interface to the block layer conform much more closely 
> to most other block device drivers.  It should also make it quite a lot easier 
> to slot in the block layer tag command queueing functions.
> 
> The disadvantage is that you now _really_ need to use the scsi request 
> interfaces instead of the command based ones (you will get a NULL deref if you 
> allocate and use your own commands with no associated requests).  This also 
> means that the following fibre drivers need some rework:  cpqfc, gdth, pluto, 
> tmscsim.
> 
> It works on my system (at least it stays up during a full kernel compile).  
> Any comments would be welcome.
> 
> After this, I plan to see if I can implement TCQ using the block level generic 
> functions.  This would buy us the ability to support request barriers (modulo 
> the already discussed edge cases) right down through the SCSI subsystem.

Hi - 

I applied your patch and successfuly ran with AIC + 2 disks (one the boot
disk), plus with qla (modified v6.0b20 to remove io request lock) drivers
attached to both Triton (disk array) and Seagate drives using block and raw io.

Do you think the queue depth on some of the adapters/devices should be
shrunk or the request queue increased with your patch? Some adapters set
device queue depths above 200 (for example, aic set mine to 253), this seems
like overkill, but today it means they can have 200 more IO's on the request
queue, where freeing the request after the IO completes means the request
queue (with your patch) means we sometimes would have 200 fewer entries.

I don't understand how/why the journaling file systems want to use a
barrier, and how it helps their IO.

Are the request barriers needed to prevent earlier IO from completing before
the barrier, or later IO from completing before the barrier, or both?

How do the request barriers work with volume managers or multi-path devices?

-- Patrick Mansfield

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Doug Ledford]

On Wed, May 22, 2002 at 03:44:06PM -0700, Patrick Mansfield wrote:
> Do you think the queue depth on some of the adapters/devices should be
> shrunk or the request queue increased with your patch? Some adapters set

Actually, more appropriate for future versions would be to make the queue
depth adjustable.  Justin's aic driver sets the depth to 253 because he
can't increase it later if he wants to, but he can queue things up
internally if he ever needs to decrease it.  If low level drivers had some
way of signalling up the stack "hey, this is how many queue slots I can
handle on this particular device now" at their leisure, then that would be
a far superior option to what we have now which is the mid layer saying to
the low level drivers "OK, tell me how many commands you can support on
this list of devices and you have to stick with that until the next
reboot".  So, instead of the mid layer calling into the low level driver 
to tell it to set queue depths, I think the low level driver should be 
allowed to call into the mid layer and have the mid layer adjust the queue 
depth on a device.  This should, of course, be properly locked so that a 
low level driver can change it's mind about a device's queue depth later 
if it wishes and call back into the mid layer with a new queue depth 
value without confusing the mid layer.

-- 
  Doug Ledford <dledford@redhat.com>     919-754-3700 x44233
         Red Hat, Inc. 
         1801 Varsity Dr.
         Raleigh, NC 27606
  

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Alan Cox]

> to tell it to set queue depths, I think the low level driver should be 
> allowed to call into the mid layer and have the mid layer adjust the queue 
> depth on a device.  This should, of course, be properly locked so that a 
> low level driver can change it's mind about a device's queue depth later 
> if it wishes and call back into the mid layer with a new queue depth 
> value without confusing the mid layer.

One thing I see consistently as a problem in the block layers in 2.4 is
that there is no way to say "thanks for the request you just threw at me
but Im busy right now" and to flag "busy/idle".

The network layer has a clean way to say "Im busy dont feed me",
"Im no longer busy, feed me" and also "thanks but I've just discovered I
am in fact busy have this request back anbd feed it me later"

It massively cleans up the interface handling compared to the wild hoops
things like qlogicisp jump through

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Doug Ledford]

On Thu, May 23, 2002 at 03:01:37AM +0100, Alan Cox wrote:
> One thing I see consistently as a problem in the block layers in 2.4 is
> that there is no way to say "thanks for the request you just threw at me
> but Im busy right now" and to flag "busy/idle".
> 
> The network layer has a clean way to say "Im busy dont feed me",
> "Im no longer busy, feed me" and also "thanks but I've just discovered I
> am in fact busy have this request back anbd feed it me later"

No, that does in fact exist.  At least in the new scsi eh code you are 
allowed to return a non-0 value from your low level driver's 
queuecommand() routine.  On a non-0 return, the mid layer throws the 
command onto the mlqueue and waits to send it until the mid layer 
processes a completion event from you (or there might also be a timeout, 
I'm not sure, but I would hope there is also a timeout since otherwise 
returning non-0 with no outstanding commands on a device because your 
driver is busy with something else could cause a queue to hang).  Anyway, 
it's there even if I never did personally use it since my driver used the 
old eh code ;-)

-- 
  Doug Ledford <dledford@redhat.com>     919-754-3700 x44233
         Red Hat, Inc. 
         1801 Varsity Dr.
         Raleigh, NC 27606
  

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Alan Cox]

> I'm not sure, but I would hope there is also a timeout since otherwise 
> returning non-0 with no outstanding commands on a device because your 
> driver is busy with something else could cause a queue to hang).  Anyway, 
> it's there even if I never did personally use it since my driver used the 
> old eh code ;-)

That does sound close to unusable on an SMP system. I don't see how you
guarantee that all your events do not complete between you executing the
non zero return to the scsi stack and it check for events

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[James Bottomley]

dledford@redhat.com said:
> On a non-0 return, the mid layer throws the  command onto the mlqueue
> and waits to send it until the mid layer  processes a completion event
> from you (or there might also be a timeout, 

Actually, there is no mlqueue any more, all commands are pushed back up to the 
block layer instead now.

dledford@redhat.com said:
> I'm not sure, but I would hope there is also a timeout since otherwise 
> returning non-0 with no outstanding commands on a device because your 
> driver is busy with something else could cause a queue to hang).  Anyway, 
> it's there even if I never did personally use it since my driver used the 
> old eh code ;-)

alan@lxorguk.ukuu.org.uk said:
> That does sound close to unusable on an SMP system. I don't see how
> you guarantee that all your events do not complete between you
> executing the non zero return to the scsi stack and it check for
> events 

Since we use the block queues and everything has to come back in through 
scsi_request_fn, we are SMP safe because of the per queue lock held over the 
request function as it takes work off the queue.

I think our actual problem is failure to signal plugging when we stop 
executing the request function under certain circumstances, so the block layer 
doesn't necessarily see when to call the request function again.

James

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[James Bottomley]

patmans@us.ibm.com said:
> I applied your patch and successfuly ran with AIC + 2 disks (one the
> boot disk), plus with qla (modified v6.0b20 to remove io request lock)
> drivers attached to both Triton (disk array) and Seagate drives using
> block and raw io. 

That's great, thanks for testing it.

> Do you think the queue depth on some of the adapters/devices should be
> shrunk or the request queue increased with your patch? Some adapters
> set device queue depths above 200 (for example, aic set mine to 253),
> this seems like overkill, but today it means they can have 200 more
> IO's on the request queue, where freeing the request after the IO
> completes means the request queue (with your patch) means we sometimes
> would have 200 fewer entries. 

That's a tough one.  There are differing schools of thought on queue depth.  I 
incline to the one that says that for modern scsi devices, 4-8 is probably a 
good figure, but there are definitely people who disagree.  The IDE code uses 
32 as the queue depth.  One of the things I hope to get from standardising the 
TCQ interface is the ability to adjust the queue depth from user land.

To move to a standard implementation in the generic layer, I think that 
practically the queue depths have to be lower (at least than 253). the current 
TCQ generic code uses an arbitrary length bitmap to track outstanding tags 
which means it would scale OK for high queue depths, but as you say, we are 
limited by the number of available requests.

> I don't understand how/why the journaling file systems want to use a
> barrier, and how it helps their IO.

> Are the request barriers needed to prevent earlier IO from completing
> before the barrier, or later IO from completing before the barrier, or
> both?

There were several discussion threads on the topic, but this is the only one I 
can find:

http://marc.theaimsgroup.com/?t=101360488200004&r=1&w=2

Essentially, journalled fs can function more efficiently if they can rely on 
transaction ordering (within ordering "barriers") making it all the way to the 
medium.  There was also a thought that this might speed up jfs operations, but 
no conclusive data was produced.

The elevator is allowed to re-order and merge requests within the barrier, but 
requests may not cross the barrier (REQ_BARRIER).

So, for instance, a jfs wants to write to a file, so it journals the write, 
performs the write and erases the journal.  The write cannot start until the 
journal entry is committed for the fs to maintain integrity on recovery, so 
currently you have to wait for the journal before beginning the fs write.  In 
the barrier abstraction, you simply send the journal entry and write down 
together with a barrier separating them.  The transaction integrity is 
maintained by the barrier ordering guarantee.

The idea for SCSI was that we translate the barrier to an ordered queue tag.  
There are, unfortunately, pathological error cases in SCSI where I/Os can 
cross the barrier, but I'm hoping that "works right almost all the time" is 
good enough.

James

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Aron Zeh]

Alas, there does not seem to be a timeout handler to restart IO when
queuecommand returned non 0.
The only way to retrigger IO seems to be a completion of an outstanding
command.
I have seen a lot of hangs in the FCP driver I am contributing to because
of this.
We finally decided never to let queuecommand fail and queue failed requests
internally while there might still be a chance to send them.
This is not too nice though, so it would probably be a good idea to use a
timeout as an alternative trigger to resend commands in the mlqueue.
Maybe a function to be called by the low-level driver would also be an
option.

Cheers,
Aron

On Thu, May 23, 2002 at 03:01:37AM +0100, Alan Cox wrote:
> One thing I see consistently as a problem in the block layers in 2.4 is
> that there is no way to say "thanks for the request you just threw at me
> but Im busy right now" and to flag "busy/idle".
>
> The network layer has a clean way to say "Im busy dont feed me",
> "Im no longer busy, feed me" and also "thanks but I've just discovered I
> am in fact busy have this request back anbd feed it me later"

No, that does in fact exist.  At least in the new scsi eh code you are
allowed to return a non-0 value from your low level driver's
queuecommand() routine.  On a non-0 return, the mid layer throws the
command onto the mlqueue and waits to send it until the mid layer
processes a completion event from you (or there might also be a timeout,
I'm not sure, but I would hope there is also a timeout since otherwise
returning non-0 with no outstanding commands on a device because your
driver is busy with something else could cause a queue to hang).  Anyway,
it's there even if I never did personally use it since my driver used the
old eh code ;-)

--
  Doug Ledford <dledford@redhat.com>     919-754-3700 x44233
         Red Hat, Inc.
         1801 Varsity Dr.
         Raleigh, NC 27606

-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[James Bottomley]

> Alas, there does not seem to be a timeout handler to restart IO when
> queuecommand returned non 0. The only way to retrigger IO seems to be
> a completion of an outstanding command. I have seen a lot of hangs in
> the FCP driver I am contributing to because of this. We finally
> decided never to let queuecommand fail and queue failed requests
> internally while there might still be a chance to send them. This is
> not too nice though, so it would probably be a good idea to use a
> timeout as an alternative trigger to resend commands in the mlqueue.
> Maybe a function to be called by the low-level driver would also be an
> option. 

I've seen it too occasionally, but I can't reproduce it consistently enough.  
I think the culprit may be our use of _elv_add_request with a no-plug 
argument.  Can you just change the last argument from 0 to 1 and see if this 
helps (in scsi_lib.c:__scsi_insert_special at about line 90)?

James

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Aron Zeh]

James,

I think we are talking about two different things.
If I am not mistaken, __scsi_insert_special is only called for character
devices or ioctl (at least that is what it states in the description).
My problem occurs during normal IO to a disk. I believe it is something of
this sort:
scsi_dispatch_command is called, host->can_queue is 4096 so the if
statement is true.
Then, the low-level driver's queuecommand returns with FAILED. The command
is put into the mlqueue to be resent when a command completes.
Due to the design of the zSeries FCP adapter, it is however possible that
temporarily, no command can be accepted at all.
Thus it is possible for queuecommand to fail on the one and only active
command, which will be put into the mlqueue.
There it will sit, waiting for the return of an outstanding command, but
there are non. Hence the system will hang.

I would think, that your problem only appears when using ioctls (e.g.
fdisk) or going out to character device (st, sg). Am I correct or am I
missing something?
Aron

> Alas, there does not seem to be a timeout handler to restart IO when
> queuecommand returned non 0. The only way to retrigger IO seems to be
> a completion of an outstanding command. I have seen a lot of hangs in
> the FCP driver I am contributing to because of this. We finally
> decided never to let queuecommand fail and queue failed requests
> internally while there might still be a chance to send them. This is
> not too nice though, so it would probably be a good idea to use a
> timeout as an alternative trigger to resend commands in the mlqueue.
> Maybe a function to be called by the low-level driver would also be an
> option.

I've seen it too occasionally, but I can't reproduce it consistently
enough.
I think the culprit may be our use of _elv_add_request with a no-plug
argument.  Can you just change the last argument from 0 to 1 and see if
this
helps (in scsi_lib.c:__scsi_insert_special at about line 90)?

James


-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[rakesh rakesh]

But the whole hypothesis is that at any point of time
mid level should be able to queue at least one command
to low level driver.Reference to doc by  Eric at
www.andante.org


--- Aron Zeh <ARZEH@de.ibm.com> wrote:
> 
> 
> James,
> 
> I think we are talking about two different things.
> If I am not mistaken, __scsi_insert_special is only
> called for character
> devices or ioctl (at least that is what it states in
> the description).
> My problem occurs during normal IO to a disk. I
> believe it is something of
> this sort:
> scsi_dispatch_command is called, host->can_queue is
> 4096 so the if
> statement is true.
> Then, the low-level driver's queuecommand returns
> with FAILED. The command
> is put into the mlqueue to be resent when a command
> completes.
> Due to the design of the zSeries FCP adapter, it is
> however possible that
> temporarily, no command can be accepted at all.
> Thus it is possible for queuecommand to fail on the
> one and only active
> command, which will be put into the mlqueue.
> There it will sit, waiting for the return of an
> outstanding command, but
> there are non. Hence the system will hang.
> 
> I would think, that your problem only appears when
> using ioctls (e.g.
> fdisk) or going out to character device (st, sg). Am
> I correct or am I
> missing something?
> Aron
> 
> > Alas, there does not seem to be a timeout handler
> to restart IO when
> > queuecommand returned non 0. The only way to
> retrigger IO seems to be
> > a completion of an outstanding command. I have
> seen a lot of hangs in
> > the FCP driver I am contributing to because of
> this. We finally
> > decided never to let queuecommand fail and queue
> failed requests
> > internally while there might still be a chance to
> send them. This is
> > not too nice though, so it would probably be a
> good idea to use a
> > timeout as an alternative trigger to resend
> commands in the mlqueue.
> > Maybe a function to be called by the low-level
> driver would also be an
> > option.
> 
> I've seen it too occasionally, but I can't reproduce
> it consistently
> enough.
> I think the culprit may be our use of
> _elv_add_request with a no-plug
> argument.  Can you just change the last argument
> from 0 to 1 and see if
> this
> helps (in scsi_lib.c:__scsi_insert_special at about
> line 90)?
> 
> James
> 
> 
> -
> To unsubscribe from this list: send the line
> "unsubscribe linux-scsi" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at 
> http://vger.kernel.org/majordomo-info.html
> 
> 
> 
> 
> -
> To unsubscribe from this list: send the line
> "unsubscribe linux-scsi" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at 
http://vger.kernel.org/majordomo-info.html


__________________________________________________
Do You Yahoo!?
LAUNCH - Your Yahoo! Music Experience
http://launch.yahoo.com

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[James Bottomley]

majordomo_linux@yahoo.com said:
> But the whole hypothesis is that at any point of time mid level should
> be able to queue at least one command to low level driver.Reference to
> doc by  Eric at www.andante.org 

That was the assumption years ago when the new error handler was created.  
Nowadays, I believe the block layer can stand to have all outstanding commands 
pushed back with no detrimental effect (otherwise blk_queue_invalidate_tags() 
wouldn't work) as long as we get the conditions for restart correct.

James

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[James Bottomley]

ARZEH@de.ibm.com said:
> I think we are talking about two different things. If I am not
> mistaken, __scsi_insert_special is only called for character devices
> or ioctl (at least that is what it states in the description).
[...]
> The command is put into the mlqueue to be resent when a command
> completes.

True, but look what happens to it after that:

scsi_mlqueue_insert() -> scsi_insert_special_cmd -> __scsi_insert_special().

__scsi_insert_special() is the ultimate point you always get to in the 
mid-layer for pushing anything back to the block layer.

My theory is that because we set host or device blocked in scsi_mlqueue_insert 
which causes the termination of our request processing loop.  But, the block 
layer only expects this termination on exhaustion or plugging, so it thinks 
we're still processing commands, however we're actually not.  Thus, the 
request function is never called again and we get stuck I/Os.  The way to 
correct this is to tell the block layer we have stopped processing by plugging 
the device queue.  Actually, I think host_blocked and device_blocked could be 
eliminated in favour of relying on the device queue plugging status.

The test of this theory is just the previously mentioned change of the 0 to 1, 
could you try it since you seem to be able to reproduce the problem reliably.

James

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Aron Zeh]

True, unfortunately the zSeries FCP adapter does not work that way. You
have no guarrantee that anything may be queued onto it at any time. I am
not certain if there are other adapters that work this way.

Aron


But the whole hypothesis is that at any point of time
mid level should be able to queue at least one command
to low level driver.Reference to doc by  Eric at
www.andante.org


--- Aron Zeh <ARZEH@de.ibm.com> wrote:
>
>
> James,
>
> I think we are talking about two different things.
> If I am not mistaken, __scsi_insert_special is only
> called for character
> devices or ioctl (at least that is what it states in
> the description).
> My problem occurs during normal IO to a disk. I
> believe it is something of
> this sort:
> scsi_dispatch_command is called, host->can_queue is
> 4096 so the if
> statement is true.
> Then, the low-level driver's queuecommand returns
> with FAILED. The command
> is put into the mlqueue to be resent when a command
> completes.
> Due to the design of the zSeries FCP adapter, it is
> however possible that
> temporarily, no command can be accepted at all.
> Thus it is possible for queuecommand to fail on the
> one and only active
> command, which will be put into the mlqueue.
> There it will sit, waiting for the return of an
> outstanding command, but
> there are non. Hence the system will hang.
>
> I would think, that your problem only appears when
> using ioctls (e.g.
> fdisk) or going out to character device (st, sg). Am
> I correct or am I
> missing something?
> Aron
>
> > Alas, there does not seem to be a timeout handler
> to restart IO when
> > queuecommand returned non 0. The only way to
> retrigger IO seems to be
> > a completion of an outstanding command. I have
> seen a lot of hangs in
> > the FCP driver I am contributing to because of
> this. We finally
> > decided never to let queuecommand fail and queue
> failed requests
> > internally while there might still be a chance to
> send them. This is
> > not too nice though, so it would probably be a
> good idea to use a
> > timeout as an alternative trigger to resend
> commands in the mlqueue.
> > Maybe a function to be called by the low-level
> driver would also be an
> > option.
>
> I've seen it too occasionally, but I can't reproduce
> it consistently
> enough.
> I think the culprit may be our use of
> _elv_add_request with a no-plug
> argument.  Can you just change the last argument
> from 0 to 1 and see if
> this
> helps (in scsi_lib.c:__scsi_insert_special at about
> line 90)?
>
> James
>
>
> -
> To unsubscribe from this list: send the line
> "unsubscribe linux-scsi" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at
> http://vger.kernel.org/majordomo-info.html
>
>
>
>
> -
> To unsubscribe from this list: send the line
> "unsubscribe linux-scsi" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at
http://vger.kernel.org/majordomo-info.html


__________________________________________________
Do You Yahoo!?
LAUNCH - Your Yahoo! Music Experience
http://launch.yahoo.com
-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Patrick Mansfield]

On Fri, May 24, 2002 at 11:35:40AM +0200, Aron Zeh wrote:
> 
> True, unfortunately the zSeries FCP adapter does not work that way. You
> have no guarrantee that anything may be queued onto it at any time. I am
> not certain if there are other adapters that work this way.
> 
> Aron

Here's a patch you might want to try out. It should at least prevent
the IO from hanging when the adapter queues up nothing, although it
will loop trying to send out IO. I've never been able to hit this,
it would be good to modify scsi_debug such that we can easily get
queue fulls and host busy's.

--- linux-2.4.18/drivers/scsi/scsi_queue.c-orig	Fri Feb  9 11:30:23 2001
+++ linux-2.4.18/drivers/scsi/scsi_queue.c	Thu May 23 09:50:49 2002
@@ -103,7 +103,7 @@
 		 * If a host is inactive and cannot queue any commands, I don't see
 		 * how things could possibly work anyways.
 		 */
-		if (host->host_busy == 0) {
+		if (host->host_busy == 1) {
 			if (scsi_retry_command(cmd) == 0) {
 				return 0;
 			}
@@ -118,7 +118,7 @@
 		 * If a host is inactive and cannot queue any commands, I don't see
 		 * how things could possibly work anyways.
 		 */
-		if (cmd->device->device_busy == 0) {
+		if (cmd->device->device_busy == 1) {
 			if (scsi_retry_command(cmd) == 0) {
 				return 0;
 			}

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Aron Zeh]

James,

I had hoped to be able to reproduce by now, but unfortunately my 2.5
kernels currently fail to boot on zSeries.
I'll try the repro and the bit change as soon as that little problem can be
alleviated.

Cheers,
Aron

ARZEH@de.ibm.com said:
> I think we are talking about two different things. If I am not
> mistaken, __scsi_insert_special is only called for character devices
> or ioctl (at least that is what it states in the description).
[...]
> The command is put into the mlqueue to be resent when a command
> completes.

True, but look what happens to it after that:

scsi_mlqueue_insert() -> scsi_insert_special_cmd -> __scsi_insert_special
().

__scsi_insert_special() is the ultimate point you always get to in the
mid-layer for pushing anything back to the block layer.

My theory is that because we set host or device blocked in
scsi_mlqueue_insert
which causes the termination of our request processing loop.  But, the
block
layer only expects this termination on exhaustion or plugging, so it thinks

we're still processing commands, however we're actually not.  Thus, the
request function is never called again and we get stuck I/Os.  The way to
correct this is to tell the block layer we have stopped processing by
plugging
the device queue.  Actually, I think host_blocked and device_blocked could
be
eliminated in favour of relying on the device queue plugging status.

The test of this theory is just the previously mentioned change of the 0 to
1,
could you try it since you seem to be able to reproduce the problem
reliably.

James


-
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[James Bottomley]

ARZEH@de.ibm.com said:
> I had hoped to be able to reproduce by now, but unfortunately my 2.5
> kernels currently fail to boot on zSeries. I'll try the repro and the
> bit change as soon as that little problem can be alleviated. 

Actually, I'm afraid there's more to it than my initial recommendation of just 
plugging the elevator queue.  Your problem lies in the device_blocked flag, 
which gets set when queuecommand returns non-zero and is not reset until an 
I/O comes back through scsi_finish_command().

I don't believe the host and device blocking serve any purpose anymore.  Also, 
with the advent of blocking in the blkdev layer, I think we can even get rid 
of the host_self_blocked flag (which is currently only used by the mesh 
driver).  So I'll look into getting rid of them all.

In the mean time, temporarily you may be able to fix it (and this should work 
for 2.4) by manually setting host_blocked to false and running the q->
request_fn(q) from your driver when you detect that the can't accept any 
commands condition is alleviated.

James

Re: [PATCH 2.5.17] Making SCSI not copy the request structure

[Aron Zeh]

James,

how about using this silly little patch of mine to do your testing. Should
be easy enough to put into other drivers and it simulates the zSeries
adapter behaviour on a heavily virtualised physical adapter relatively
well. (I admit that testing for one outstanding command is pushing it a
bit....)
I have tried out the bit-change with this. To no avail though, it still
hangs. By the way, blk_plug_device needs to be exported when setting the
bit to 1...so maybe it should always be on the list of exports?

Cheers,
Aron

--- drivers/scsi/aic7xxx/aic7xxx_linux.c        Wed Apr 17 17:31:28 2002
+++ drivers/scsi/aic7xxx/aic7xxx_linux_queuecommand_fails.c     Fri May 31
18:25:25 2002
@@ -1540,6 +1540,10 @@
        /*
         * Save the callback on completion function.
         */
+       if((cmd->host->host_busy==1) && ((jiffies & 0xf0)>0x80)) {
+               printk("**********+causing fail********\n");
+               return FAILED;
+       }
        cmd->scsi_done = scsi_done;

        dev = ahc_linux_get_device(ahc, cmd->channel, cmd->target,

ARZEH@de.ibm.com said:
> I had hoped to be able to reproduce by now, but unfortunately my 2.5
> kernels currently fail to boot on zSeries. I'll try the repro and the
> bit change as soon as that little problem can be alleviated.

Actually, I'm afraid there's more to it than my initial recommendation of
just
plugging the elevator queue.  Your problem lies in the device_blocked flag,

which gets set when queuecommand returns non-zero and is not reset until an

I/O comes back through scsi_finish_command().

I don't believe the host and device blocking serve any purpose anymore.
Also,
with the advent of blocking in the blkdev layer, I think we can even get
rid
of the host_self_blocked flag (which is currently only used by the mesh
driver).  So I'll look into getting rid of them all.

In the mean time, temporarily you may be able to fix it (and this should
work
for 2.4) by manually setting host_blocked to false and running the q->
request_fn(q) from your driver when you detect that the can't accept any
commands condition is alleviated.

James