Skip to contentSkip to conversation

io_request_lock/queue_lock patch

PATCH Jonathan Lahr 5 participants 14 messages Aug 30, 2001 – Sep 6, 2001

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Included below is a snapshot of a patch I am developing to reduce io_request_lock contention in 2.4.

This patch changes __make_request to take the device-specific queue_lock instead of io_request_lock and changes generic_unplug_device to take queue_lock in addition to io_request_lock to serialize access to the request queue. Also, locking in various functions has been modified in accordance with this new scheme.

I have done some testing of this with aic7xxx, qla2x00, and lpfc adapters.

To apply it to 2.4.6 or 2.4.7, from linux/ "patch -p1 < iorlv0_247".

Comments and suggestions are welcome and invited.

Jonathan

----- patch -----

drivers/block/ll_rw_blk.c +13 -11
diff -Naur linux.base/drivers/block/ll_rw_blk.c linux.mod/drivers/block/ll_rw_blk.c
@@ -343,11 +343,13 @@
 void generic_unplug_device(void *data)
 {
 	request_queue_t *q = (request_queue_t *) data;
-	unsigned long flags;
+	unsigned long flags[2];
 
-	spin_lock_irqsave(&io_request_lock, flags);
+	spin_lock_irqsave(&io_request_lock, flags[0]);
+	spin_lock_irqsave(&q->queue_lock, flags[1]);
 	__generic_unplug_device(q);
-	spin_unlock_irqrestore(&io_request_lock, flags);
+	spin_unlock_irqrestore(&q->queue_lock, flags[1]);
+	spin_unlock_irqrestore(&io_request_lock, flags[0]);
 }
 
 static void blk_init_free_list(request_queue_t *q)
@@ -470,9 +472,9 @@
 	add_wait_queue_exclusive(&q->wait_for_request, &wait);
 	for (;;) {
 		__set_current_state(TASK_UNINTERRUPTIBLE);
-		spin_lock_irq(&io_request_lock);
+		spin_lock_irq(&q->queue_lock);
 		rq = get_request(q, rw);
-		spin_unlock_irq(&io_request_lock);
+		spin_unlock_irq(&q->queue_lock);
 		if (rq)
 			break;
 		generic_unplug_device(q);
@@ -487,9 +489,9 @@
 {
 	register struct request *rq;
 
-	spin_lock_irq(&io_request_lock);
+	spin_lock_irq(&q->queue_lock);
 	rq = get_request(q, rw);
-	spin_unlock_irq(&io_request_lock);
+	spin_unlock_irq(&q->queue_lock);
 	if (rq)
 		return rq;
 	return __get_request_wait(q, rw);
@@ -555,7 +557,7 @@
 	drive_stat_acct(req->rq_dev, req->cmd, req->nr_sectors, 1);
 
 	if (!q->plugged && q->head_active && insert_here == &q->queue_head) {
-		spin_unlock_irq(&io_request_lock);
+		spin_unlock_irq(&q->queue_lock);
 		BUG();
 	}
 
@@ -727,7 +729,7 @@
 	 * Now we acquire the request spinlock, we have to be mega careful
 	 * not to schedule or do something nonatomic
 	 */
-	spin_lock_irq(&io_request_lock);
+	spin_lock_irq(&q->queue_lock);
 
 	insert_here = head->prev;
 	if (list_empty(head)) {
@@ -794,7 +796,7 @@
 		req = freereq;
 		freereq = NULL;
 	} else if ((req = get_request(q, rw)) == NULL) {
-		spin_unlock_irq(&io_request_lock);
+		spin_unlock_irq(&q->queue_lock);
 		if (rw_ahead)
 			goto end_io;
 
@@ -821,7 +823,7 @@
 out:
 	if (freereq)
 		blkdev_release_request(freereq);
-	spin_unlock_irq(&io_request_lock);
+	spin_unlock_irq(&q->queue_lock);
 	return 0;
 end_io:
 	bh->b_end_io(bh, test_bit(BH_Uptodate, &bh->b_state));
drivers/scsi/scsi.c +0 -8
diff -Naur linux.base/drivers/scsi/scsi.c linux.mod/drivers/scsi/scsi.c
@@ -678,41 +678,33 @@
 		 * passes a meaningful return value.
 		 */
 		if (host->hostt->use_new_eh_code) {
-                        spin_lock_irqsave(&io_request_lock, flags);
 			rtn = host->hostt->queuecommand(SCpnt, scsi_done);
-                        spin_unlock_irqrestore(&io_request_lock, flags);
 			if (rtn != 0) {
 				scsi_delete_timer(SCpnt);
 				scsi_mlqueue_insert(SCpnt, SCSI_MLQUEUE_HOST_BUSY);
                                 SCSI_LOG_MLQUEUE(3, printk("queuecommand : request rejected\n"));                                
 			}
 		} else {
-                        spin_lock_irqsave(&io_request_lock, flags);
 			host->hostt->queuecommand(SCpnt, scsi_old_done);
-                        spin_unlock_irqrestore(&io_request_lock, flags);
 		}
 	} else {
 		int temp;
 
 		SCSI_LOG_MLQUEUE(3, printk("command() :  routine at %p\n", host->hostt->command));
-                spin_lock_irqsave(&io_request_lock, flags);
 		temp = host->hostt->command(SCpnt);
 		SCpnt->result = temp;
 #ifdef DEBUG_DELAY
-                spin_unlock_irqrestore(&io_request_lock, flags);
 		clock = jiffies + 4 * HZ;
 		while (time_before(jiffies, clock))
 			barrier();
 		printk("done(host = %d, result = %04x) : routine at %p\n",
 		       host->host_no, temp, host->hostt->command);
-                spin_lock_irqsave(&io_request_lock, flags);
 #endif
 		if (host->hostt->use_new_eh_code) {
 			scsi_done(SCpnt);
 		} else {
 			scsi_old_done(SCpnt);
 		}
-                spin_unlock_irqrestore(&io_request_lock, flags);
 	}
 	SCSI_LOG_MLQUEUE(3, printk("leaving scsi_dispatch_cmnd()\n"));
 	return rtn;
drivers/scsi/scsi_lib.c +15 -12
diff -Naur linux.base/drivers/scsi/scsi_lib.c linux.mod/drivers/scsi/scsi_lib.c
@@ -68,7 +68,7 @@
 static void __scsi_insert_special(request_queue_t *q, struct request *rq,
 				  void *data, int at_head)
 {
-	unsigned long flags;
+	unsigned long flags[2];
 
 	ASSERT_LOCK(&io_request_lock, 0);
 
@@ -84,7 +84,8 @@
 	 * head of the queue for things like a QUEUE_FULL message from a
 	 * device, or a host that is unable to accept a particular command.
 	 */
-	spin_lock_irqsave(&io_request_lock, flags);
+	spin_lock_irqsave(&io_request_lock, flags[0]);
+	spin_lock_irqsave(&q->queue_lock, flags[1]);
 
 	if (at_head)
 		list_add(&rq->queue, &q->queue_head);
@@ -92,7 +93,8 @@
 		list_add_tail(&rq->queue, &q->queue_head);
 
 	q->request_fn(q);
-	spin_unlock_irqrestore(&io_request_lock, flags);
+	spin_unlock_irqrestore(&q->queue_lock, flags[1]);
+	spin_unlock_irqrestore(&io_request_lock, flags[0]);
 }
 
 
@@ -246,13 +248,14 @@
 void scsi_queue_next_request(request_queue_t * q, Scsi_Cmnd * SCpnt)
 {
 	int all_clear;
-	unsigned long flags;
+	unsigned long flags[2];
 	Scsi_Device *SDpnt;
 	struct Scsi_Host *SHpnt;
 
 	ASSERT_LOCK(&io_request_lock, 0);
 
-	spin_lock_irqsave(&io_request_lock, flags);
+	spin_lock_irqsave(&io_request_lock, flags[0]);
+	spin_lock_irqsave(&q->queue_lock, flags[1]);
 	if (SCpnt != NULL) {
 
 		/*
@@ -328,7 +331,8 @@
 			SHpnt->some_device_starved = 0;
 		}
 	}
-	spin_unlock_irqrestore(&io_request_lock, flags);
+	spin_unlock_irqrestore(&q->queue_lock, flags[1]);
+	spin_unlock_irqrestore(&io_request_lock, flags[0]);
 }
 
 /*
@@ -821,6 +825,7 @@
 	struct Scsi_Device_Template *STpnt;
 
 	ASSERT_LOCK(&io_request_lock, 1);
+	ASSERT_LOCK(&q->queue_lock, 1);
 
 	SDpnt = (Scsi_Device *) q->queuedata;
 	if (!SDpnt) {
@@ -828,6 +833,7 @@
 	}
 	SHpnt = SDpnt->host;
 
+
 	/*
 	 * To start with, we keep looping until the queue is empty, or until
 	 * the host is no longer able to accept any more requests.
@@ -839,7 +845,7 @@
 		 * we need to check to see if the queue is plugged or not.
 		 */
 		if (SHpnt->in_recovery || q->plugged)
-			return;
+			break;
 
 		/*
 		 * If the device cannot accept another request, then quit.
@@ -887,9 +893,11 @@
 			 */
 			SDpnt->was_reset = 0;
 			if (SDpnt->removable && !in_interrupt()) {
+				spin_unlock_irq(&q->queue_lock);
 				spin_unlock_irq(&io_request_lock);
 				scsi_ioctl(SDpnt, SCSI_IOCTL_DOORLOCK, 0);
 				spin_lock_irq(&io_request_lock);
+				spin_lock_irq(&q->queue_lock);
 				continue;
 			}
 		}
@@ -998,8 +1006,6 @@
 		 * another.  
 		 */
 		req = NULL;
-		spin_unlock_irq(&io_request_lock);
-
 		if (SCpnt->request.cmd != SPECIAL) {
 			/*
 			 * This will do a couple of things:
@@ -1028,7 +1034,6 @@
 				{
 					panic("Should not have leftover blocks\n");
 				}
-				spin_lock_irq(&io_request_lock);
 				SHpnt->host_busy--;
 				SDpnt->device_busy--;
 				continue;
@@ -1044,7 +1049,6 @@
 				{
 					panic("Should not have leftover blocks\n");
 				}
-				spin_lock_irq(&io_request_lock);
 				SHpnt->host_busy--;
 				SDpnt->device_busy--;
 				continue;
@@ -1065,7 +1069,6 @@
 		 * Now we need to grab the lock again.  We are about to mess
 		 * with the request queue and try to find another command.
 		 */
-		spin_lock_irq(&io_request_lock);
 	}
 }
 
#1 lore raw

Here are my welcome comments. :)

In my opinion, it would well be a miracle if your patch does not introduce new race conditions, at least for drivers that still use the old scsi done method.

Gérard.

Quoted 276 lines · Jonathan Lahr

Included below is a snapshot of a patch I am developing to reduce io_request_lock contention in 2.4.

This patch changes __make_request to take the device-specific queue_lock instead of io_request_lock and changes generic_unplug_device to take queue_lock in addition to io_request_lock to serialize access to the request queue. Also, locking in various functions has been modified in accordance with this new scheme.

I have done some testing of this with aic7xxx, qla2x00, and lpfc adapters.

To apply it to 2.4.6 or 2.4.7, from linux/ "patch -p1 < iorlv0_247".

Comments and suggestions are welcome and invited.

Jonathan

----- patch -----
diff -Naur linux.base/drivers/block/ll_rw_blk.c linux.mod/drivers/block/ll_rw_blk.c

drivers/block/ll_rw_blk.c +13 -11
@@ -343,11 +343,13 @@
 void generic_unplug_device(void *data)
 {
 	request_queue_t *q = (request_queue_t *) data;
-	unsigned long flags;
+	unsigned long flags[2];
 
-	spin_lock_irqsave(&io_request_lock, flags);
+	spin_lock_irqsave(&io_request_lock, flags[0]);
+	spin_lock_irqsave(&q->queue_lock, flags[1]);
 	__generic_unplug_device(q);
-	spin_unlock_irqrestore(&io_request_lock, flags);
+	spin_unlock_irqrestore(&q->queue_lock, flags[1]);
+	spin_unlock_irqrestore(&io_request_lock, flags[0]);
 }
 
 static void blk_init_free_list(request_queue_t *q)
@@ -470,9 +472,9 @@
 	add_wait_queue_exclusive(&q->wait_for_request, &wait);
 	for (;;) {
 		__set_current_state(TASK_UNINTERRUPTIBLE);
-		spin_lock_irq(&io_request_lock);
+		spin_lock_irq(&q->queue_lock);
 		rq = get_request(q, rw);
-		spin_unlock_irq(&io_request_lock);
+		spin_unlock_irq(&q->queue_lock);
 		if (rq)
 			break;
 		generic_unplug_device(q);
@@ -487,9 +489,9 @@
 {
 	register struct request *rq;
 
-	spin_lock_irq(&io_request_lock);
+	spin_lock_irq(&q->queue_lock);
 	rq = get_request(q, rw);
-	spin_unlock_irq(&io_request_lock);
+	spin_unlock_irq(&q->queue_lock);
 	if (rq)
 		return rq;
 	return __get_request_wait(q, rw);
@@ -555,7 +557,7 @@
 	drive_stat_acct(req->rq_dev, req->cmd, req->nr_sectors, 1);
 
 	if (!q->plugged && q->head_active && insert_here == &q->queue_head) {
-		spin_unlock_irq(&io_request_lock);
+		spin_unlock_irq(&q->queue_lock);
 		BUG();
 	}
 
@@ -727,7 +729,7 @@
 	 * Now we acquire the request spinlock, we have to be mega careful
 	 * not to schedule or do something nonatomic
 	 */
-	spin_lock_irq(&io_request_lock);
+	spin_lock_irq(&q->queue_lock);
 
 	insert_here = head->prev;
 	if (list_empty(head)) {
@@ -794,7 +796,7 @@
 		req = freereq;
 		freereq = NULL;
 	} else if ((req = get_request(q, rw)) == NULL) {
-		spin_unlock_irq(&io_request_lock);
+		spin_unlock_irq(&q->queue_lock);
 		if (rw_ahead)
 			goto end_io;
 
@@ -821,7 +823,7 @@
 out:
 	if (freereq)
 		blkdev_release_request(freereq);
-	spin_unlock_irq(&io_request_lock);
+	spin_unlock_irq(&q->queue_lock);
 	return 0;
 end_io:
 	bh->b_end_io(bh, test_bit(BH_Uptodate, &bh->b_state));

diff -Naur linux.base/drivers/scsi/scsi.c linux.mod/drivers/scsi/scsi.c

drivers/scsi/scsi.c +0 -8
@@ -678,41 +678,33 @@
 		 * passes a meaningful return value.
 		 */
 		if (host->hostt->use_new_eh_code) {
-                        spin_lock_irqsave(&io_request_lock, flags);
 			rtn = host->hostt->queuecommand(SCpnt, scsi_done);
-                        spin_unlock_irqrestore(&io_request_lock, flags);
 			if (rtn != 0) {
 				scsi_delete_timer(SCpnt);
 				scsi_mlqueue_insert(SCpnt, SCSI_MLQUEUE_HOST_BUSY);
                                 SCSI_LOG_MLQUEUE(3, printk("queuecommand : request rejected\n"));
 			}
 		} else {
-                        spin_lock_irqsave(&io_request_lock, flags);
 			host->hostt->queuecommand(SCpnt, scsi_old_done);
-                        spin_unlock_irqrestore(&io_request_lock, flags);
 		}
 	} else {
 		int temp;
 
 		SCSI_LOG_MLQUEUE(3, printk("command() :  routine at %p\n", host->hostt->command));
-                spin_lock_irqsave(&io_request_lock, flags);
 		temp = host->hostt->command(SCpnt);
 		SCpnt->result = temp;
 #ifdef DEBUG_DELAY
-                spin_unlock_irqrestore(&io_request_lock, flags);
 		clock = jiffies + 4 * HZ;
 		while (time_before(jiffies, clock))
 			barrier();
 		printk("done(host = %d, result = %04x) : routine at %p\n",
 		       host->host_no, temp, host->hostt->command);
-                spin_lock_irqsave(&io_request_lock, flags);
 #endif
 		if (host->hostt->use_new_eh_code) {
 			scsi_done(SCpnt);
 		} else {
 			scsi_old_done(SCpnt);
 		}
-                spin_unlock_irqrestore(&io_request_lock, flags);
 	}
 	SCSI_LOG_MLQUEUE(3, printk("leaving scsi_dispatch_cmnd()\n"));
 	return rtn;

diff -Naur linux.base/drivers/scsi/scsi_lib.c linux.mod/drivers/scsi/scsi_lib.c

drivers/scsi/scsi_lib.c +15 -13
@@ -68,7 +68,7 @@
 static void __scsi_insert_special(request_queue_t *q, struct request *rq,
 				  void *data, int at_head)
 {
-	unsigned long flags;
+	unsigned long flags[2];
 
 	ASSERT_LOCK(&io_request_lock, 0);
 
@@ -84,7 +84,8 @@
 	 * head of the queue for things like a QUEUE_FULL message from a
 	 * device, or a host that is unable to accept a particular command.
 	 */
-	spin_lock_irqsave(&io_request_lock, flags);
+	spin_lock_irqsave(&io_request_lock, flags[0]);
+	spin_lock_irqsave(&q->queue_lock, flags[1]);
 
 	if (at_head)
 		list_add(&rq->queue, &q->queue_head);
@@ -92,7 +93,8 @@
 		list_add_tail(&rq->queue, &q->queue_head);
 
 	q->request_fn(q);
-	spin_unlock_irqrestore(&io_request_lock, flags);
+	spin_unlock_irqrestore(&q->queue_lock, flags[1]);
+	spin_unlock_irqrestore(&io_request_lock, flags[0]);
 }
 
 
@@ -246,13 +248,14 @@
 void scsi_queue_next_request(request_queue_t * q, Scsi_Cmnd * SCpnt)
 {
 	int all_clear;
-	unsigned long flags;
+	unsigned long flags[2];
 	Scsi_Device *SDpnt;
 	struct Scsi_Host *SHpnt;
 
 	ASSERT_LOCK(&io_request_lock, 0);
 
-	spin_lock_irqsave(&io_request_lock, flags);
+	spin_lock_irqsave(&io_request_lock, flags[0]);
+	spin_lock_irqsave(&q->queue_lock, flags[1]);
 	if (SCpnt != NULL) {
 
 		/*
@@ -328,7 +331,8 @@
 			SHpnt->some_device_starved = 0;
 		}
 	}
-	spin_unlock_irqrestore(&io_request_lock, flags);
+	spin_unlock_irqrestore(&q->queue_lock, flags[1]);
+	spin_unlock_irqrestore(&io_request_lock, flags[0]);
 }
 
 /*
@@ -821,6 +825,7 @@
 	struct Scsi_Device_Template *STpnt;
 
 	ASSERT_LOCK(&io_request_lock, 1);
+	ASSERT_LOCK(&q->queue_lock, 1);
 
 	SDpnt = (Scsi_Device *) q->queuedata;
 	if (!SDpnt) {
@@ -828,6 +833,7 @@
 	}
 	SHpnt = SDpnt->host;
 
+
 	/*
 	 * To start with, we keep looping until the queue is empty, or until
 	 * the host is no longer able to accept any more requests.
@@ -839,7 +845,7 @@
 		 * we need to check to see if the queue is plugged or not.
 		 */
 		if (SHpnt->in_recovery || q->plugged)
-			return;
+			break;
 
 		/*
 		 * If the device cannot accept another request, then quit.
@@ -887,9 +893,11 @@
 			 */
 			SDpnt->was_reset = 0;
 			if (SDpnt->removable && !in_interrupt()) {
+				spin_unlock_irq(&q->queue_lock);
 				spin_unlock_irq(&io_request_lock);
 				scsi_ioctl(SDpnt, SCSI_IOCTL_DOORLOCK, 0);
 				spin_lock_irq(&io_request_lock);
+				spin_lock_irq(&q->queue_lock);
 				continue;
 			}
 		}
@@ -998,8 +1006,6 @@
 		 * another.
 		 */
 		req = NULL;
-		spin_unlock_irq(&io_request_lock);
-
 		if (SCpnt->request.cmd != SPECIAL) {
 			/*
 			 * This will do a couple of things:
@@ -1028,7 +1034,6 @@
 				{
 					panic("Should not have leftover blocks\n");
 				}
-				spin_lock_irq(&io_request_lock);
 				SHpnt->host_busy--;
 				SDpnt->device_busy--;
 				continue;
@@ -1044,7 +1049,6 @@
 				{
 					panic("Should not have leftover blocks\n");
 				}
-				spin_lock_irq(&io_request_lock);
 				SHpnt->host_busy--;
 				SDpnt->device_busy--;
 				continue;
@@ -1065,7 +1069,6 @@
 		 * Now we need to grab the lock again.  We are about to mess
 		 * with the request queue and try to find another command.
 		 */
-		spin_lock_irq(&io_request_lock);
 	}
 }
 
-

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

#2 lore raw

I am afraid I would have to agree with Gérard. We were planning on
cleaning this mess up in the 2.5 kernel, and my inclination would be to leave this alone until then.

-Eric

----- Original Message -----
From: "Gérard Roudier" <[email protected]>
To: "Jonathan Lahr" <[email protected]>
Cc: <[email protected]>; <[email protected]>
Sent: Thursday, August 30, 2001 5:32 PM
Subject: Re: io_request_lock/queue_lock patch

Quoted 21 lines

Here are my welcome comments. :)

In my opinion, it would well be a miracle if your patch does not introduce new race conditions, at least for drivers that still use the old scsi done method.

Gérard.

Quoted 10 lines · Jonathan Lahr

Included below is a snapshot of a patch I am developing to reduce io_request_lock contention in 2.4.

This patch changes __make_request to take the device-specific queue_lock instead of io_request_lock and changes generic_unplug_device to take queue_lock in addition to io_request_lock to serialize access to the request queue. Also, locking in various functions has been modified in accordance with this new scheme.

I have done some testing of this with aic7xxx, qla2x00, and lpfc

adapters.

Quoted 9 lines
Quoted 9 lines

To apply it to 2.4.6 or 2.4.7, from linux/ "patch -p1 < iorlv0_247".

Comments and suggestions are welcome and invited.

Jonathan

----- patch -----
diff -Naur linux.base/drivers/block/ll_rw_blk.c

linux.mod/drivers/block/ll_rw_blk.c

Quoted 81 lines
Quoted 81 lines
drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 renamed from drivers/block/ll_rw_blk.c Thu Jul 19 20:51:23 2001 +13 -11
@@ -343,11 +343,13 @@
 void generic_unplug_device(void *data)
 {
 request_queue_t *q = (request_queue_t *) data;
- unsigned long flags;
+ unsigned long flags[2];
 
- spin_lock_irqsave(&io_request_lock, flags);
+ spin_lock_irqsave(&io_request_lock, flags[0]);
+ spin_lock_irqsave(&q->queue_lock, flags[1]);
 __generic_unplug_device(q);
- spin_unlock_irqrestore(&io_request_lock, flags);
+ spin_unlock_irqrestore(&q->queue_lock, flags[1]);
+ spin_unlock_irqrestore(&io_request_lock, flags[0]);
 }
 
 static void blk_init_free_list(request_queue_t *q)
@@ -470,9 +472,9 @@
 add_wait_queue_exclusive(&q->wait_for_request, &wait);
 for (;;) {
 __set_current_state(TASK_UNINTERRUPTIBLE);
- spin_lock_irq(&io_request_lock);
+ spin_lock_irq(&q->queue_lock);
 rq = get_request(q, rw);
- spin_unlock_irq(&io_request_lock);
+ spin_unlock_irq(&q->queue_lock);
 if (rq)
 break;
 generic_unplug_device(q);
@@ -487,9 +489,9 @@
 {
 register struct request *rq;
 
- spin_lock_irq(&io_request_lock);
+ spin_lock_irq(&q->queue_lock);
 rq = get_request(q, rw);
- spin_unlock_irq(&io_request_lock);
+ spin_unlock_irq(&q->queue_lock);
 if (rq)
 return rq;
 return __get_request_wait(q, rw);
@@ -555,7 +557,7 @@
 drive_stat_acct(req->rq_dev, req->cmd, req->nr_sectors, 1);
 
 if (!q->plugged && q->head_active && insert_here == &q->queue_head) {
- spin_unlock_irq(&io_request_lock);
+ spin_unlock_irq(&q->queue_lock);
 BUG();
 }
 
@@ -727,7 +729,7 @@
 * Now we acquire the request spinlock, we have to be mega careful
 * not to schedule or do something nonatomic
 */
- spin_lock_irq(&io_request_lock);
+ spin_lock_irq(&q->queue_lock);
 
 insert_here = head->prev;
 if (list_empty(head)) {
@@ -794,7 +796,7 @@
 req = freereq;
 freereq = NULL;
 } else if ((req = get_request(q, rw)) == NULL) {
- spin_unlock_irq(&io_request_lock);
+ spin_unlock_irq(&q->queue_lock);
 if (rw_ahead)
 goto end_io;
 
@@ -821,7 +823,7 @@
 out:
 if (freereq)
 blkdev_release_request(freereq);
- spin_unlock_irq(&io_request_lock);
+ spin_unlock_irq(&q->queue_lock);
 return 0;
 end_io:
 bh->b_end_io(bh, test_bit(BH_Uptodate, &bh->b_state));

diff -Naur linux.base/drivers/scsi/scsi.c linux.mod/drivers/scsi/scsi.c --- linux.base/drivers/scsi/scsi.c Thu Jul 19 21:07:04 2001

Quoted 8 lines

drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 line 683

Quoted 8 lines

+++ linux.mod/drivers/scsi/scsi.c Fri Aug 24 11:06:25 2001

drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 renamed from drivers/block/ll_rw_blk.c Thu Jul 19 20:51:23 2001 +0 -2
@@ -678,41 +678,33 @@
 * passes a meaningful return value.
 */
 if (host->hostt->use_new_eh_code) {
-                        spin_lock_irqsave(&io_request_lock, flags);
 rtn = host->hostt->queuecommand(SCpnt, scsi_done);
-                        spin_unlock_irqrestore(&io_request_lock,

flags);

Quoted 4 lines

drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 line 685

Quoted 4 lines
drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 renamed from drivers/block/ll_rw_blk.c Thu Jul 19 20:51:23 2001 +0 -0
 if (rtn != 0) {
 scsi_delete_timer(SCpnt);
 scsi_mlqueue_insert(SCpnt, SCSI_MLQUEUE_HOST_BUSY);
                                 SCSI_LOG_MLQUEUE(3,

printk("queuecommand : request rejected\n"));

Quoted 5 lines

drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 line 692

Quoted 5 lines
drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 renamed from drivers/block/ll_rw_blk.c Thu Jul 19 20:51:23 2001 +0 -2
 }
 } else {
-                        spin_lock_irqsave(&io_request_lock, flags);
 host->hostt->queuecommand(SCpnt, scsi_old_done);
-                        spin_unlock_irqrestore(&io_request_lock,

flags);

Quoted 5 lines

drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 line 693

Quoted 5 lines
drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 renamed from drivers/block/ll_rw_blk.c Thu Jul 19 20:51:23 2001 +0 -0
 }
 } else {
 int temp;
 
 SCSI_LOG_MLQUEUE(3, printk("command() :  routine at %p\n",

host->hostt->command));

Quoted 14 lines
Quoted 14 lines
drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 renamed from drivers/block/ll_rw_blk.c Thu Jul 19 20:51:23 2001 +0 -3
-                spin_lock_irqsave(&io_request_lock, flags);
 temp = host->hostt->command(SCpnt);
 SCpnt->result = temp;
 #ifdef DEBUG_DELAY
-                spin_unlock_irqrestore(&io_request_lock, flags);
 clock = jiffies + 4 * HZ;
 while (time_before(jiffies, clock))
 barrier();
 printk("done(host = %d, result = %04x) : routine at %p\n",
        host->host_no, temp, host->hostt->command);
-                spin_lock_irqsave(&io_request_lock, flags);
 #endif
 if (host->hostt->use_new_eh_code) {
 scsi_done(SCpnt);
Quoted 8 lines
Quoted 8 lines
drivers/block/ll_rw_blk.c Fri Aug 24 10:17:38 2001 renamed from drivers/block/ll_rw_blk.c Thu Jul 19 20:51:23 2001 +0 -1
 } else {
 scsi_old_done(SCpnt);
 }
-                spin_unlock_irqrestore(&io_request_lock, flags);
 }
 SCSI_LOG_MLQUEUE(3, printk("leaving scsi_dispatch_cmnd()\n"));
 return rtn;

diff -Naur linux.base/drivers/scsi/scsi_lib.c

linux.mod/drivers/scsi/scsi_lib.c

Quoted 4 lines

drivers/scsi/scsi_lib.c Fri Aug 24 10:19:50 2001 line 68

Quoted 4 lines
drivers/scsi/scsi_lib.c Fri Aug 24 10:19:50 2001 renamed from drivers/scsi/scsi_lib.c Thu Jul 19 20:48:04 2001 +0 -0
@@ -68,7 +68,7 @@
 static void __scsi_insert_special(request_queue_t *q, struct request

*rq,

Quoted 134 lines
Quoted 125 lines
drivers/scsi/scsi_lib.c Fri Aug 24 10:19:50 2001 renamed from drivers/scsi/scsi_lib.c Thu Jul 19 20:48:04 2001 +15 -13
   void *data, int at_head)
 {
- unsigned long flags;
+ unsigned long flags[2];
 
 ASSERT_LOCK(&io_request_lock, 0);
 
@@ -84,7 +84,8 @@
 * head of the queue for things like a QUEUE_FULL message from a
 * device, or a host that is unable to accept a particular command.
 */
- spin_lock_irqsave(&io_request_lock, flags);
+ spin_lock_irqsave(&io_request_lock, flags[0]);
+ spin_lock_irqsave(&q->queue_lock, flags[1]);
 
 if (at_head)
 list_add(&rq->queue, &q->queue_head);
@@ -92,7 +93,8 @@
 list_add_tail(&rq->queue, &q->queue_head);
 
 q->request_fn(q);
- spin_unlock_irqrestore(&io_request_lock, flags);
+ spin_unlock_irqrestore(&q->queue_lock, flags[1]);
+ spin_unlock_irqrestore(&io_request_lock, flags[0]);
 }
 
 
@@ -246,13 +248,14 @@
 void scsi_queue_next_request(request_queue_t * q, Scsi_Cmnd * SCpnt)
 {
 int all_clear;
- unsigned long flags;
+ unsigned long flags[2];
 Scsi_Device *SDpnt;
 struct Scsi_Host *SHpnt;
 
 ASSERT_LOCK(&io_request_lock, 0);
 
- spin_lock_irqsave(&io_request_lock, flags);
+ spin_lock_irqsave(&io_request_lock, flags[0]);
+ spin_lock_irqsave(&q->queue_lock, flags[1]);
 if (SCpnt != NULL) {
 
 /*
@@ -328,7 +331,8 @@
 SHpnt->some_device_starved = 0;
 }
 }
- spin_unlock_irqrestore(&io_request_lock, flags);
+ spin_unlock_irqrestore(&q->queue_lock, flags[1]);
+ spin_unlock_irqrestore(&io_request_lock, flags[0]);
 }
 
 /*
@@ -821,6 +825,7 @@
 struct Scsi_Device_Template *STpnt;
 
 ASSERT_LOCK(&io_request_lock, 1);
+ ASSERT_LOCK(&q->queue_lock, 1);
 
 SDpnt = (Scsi_Device *) q->queuedata;
 if (!SDpnt) {
@@ -828,6 +833,7 @@
 }
 SHpnt = SDpnt->host;
 
+
 /*
 * To start with, we keep looping until the queue is empty, or until
 * the host is no longer able to accept any more requests.
@@ -839,7 +845,7 @@
 * we need to check to see if the queue is plugged or not.
 */
 if (SHpnt->in_recovery || q->plugged)
- return;
+ break;
 
 /*
 * If the device cannot accept another request, then quit.
@@ -887,9 +893,11 @@
 */
 SDpnt->was_reset = 0;
 if (SDpnt->removable && !in_interrupt()) {
+ spin_unlock_irq(&q->queue_lock);
 spin_unlock_irq(&io_request_lock);
 scsi_ioctl(SDpnt, SCSI_IOCTL_DOORLOCK, 0);
 spin_lock_irq(&io_request_lock);
+ spin_lock_irq(&q->queue_lock);
 continue;
 }
 }
@@ -998,8 +1006,6 @@
 * another.
 */
 req = NULL;
- spin_unlock_irq(&io_request_lock);
-
 if (SCpnt->request.cmd != SPECIAL) {
 /*
 * This will do a couple of things:
@@ -1028,7 +1034,6 @@
 {
 panic("Should not have leftover blocks\n");
 }
- spin_lock_irq(&io_request_lock);
 SHpnt->host_busy--;
 SDpnt->device_busy--;
 continue;
@@ -1044,7 +1049,6 @@
 {
 panic("Should not have leftover blocks\n");
 }
- spin_lock_irq(&io_request_lock);
 SHpnt->host_busy--;
 SDpnt->device_busy--;
 continue;
@@ -1065,7 +1069,6 @@
 * Now we need to grab the lock again.  We are about to mess
 * with the request queue and try to find another command.
 */
- spin_lock_irq(&io_request_lock);
 }
 }
 
-
Quoted 3 lines

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

#3 lore raw
In reply to Gérard Roudier

Is there an estimate on when the 2.5 i/o subsystem will be released?

Jonathan

Quoted 22 lines · Eric Youngdale [[email protected]] wrote:

I am afraid I would have to agree with Gérard. We were planning on
cleaning this mess up in the 2.5 kernel, and my inclination would be to leave this alone until then.

-Eric

----- Original Message -----
From: "Gérard Roudier" <[email protected]>
To: "Jonathan Lahr" <[email protected]>
Cc: <[email protected]>; <[email protected]>
Sent: Thursday, August 30, 2001 5:32 PM
Subject: Re: io_request_lock/queue_lock patch

Quoted 8 lines

Here are my welcome comments. :)

In my opinion, it would well be a miracle if your patch does not introduce new race conditions, at least for drivers that still use the old scsi done method.

Gérard.

#5 lore raw
In reply to Jonathan Lahr
Quoted 3 lines · Jonathan Lahr

Included below is a snapshot of a patch I am developing to reduce io_request_lock contention in 2.4.

No no no, you are opening a serious can of worms. No offense, but did you really think this would fly?! This is already being taken care of for 2.5, lets leave 2.4 alone in this regard.

Signature · 1 line
Jens Axboe
#4 lore raw
In reply to Gérard Roudier
Quoted 2 lines · Jonathan Lahr

Is there an estimate on when the 2.5 i/o subsystem will be released?

I've at least put up incremental patches of what I'm doing here:

kernel.org/pub/linux/kernel/people/axboe/v2.5

bio-14 will be final next week.

Signature · 1 line
Jens Axboe
#6 lore raw
In reply to Jonathan Lahr

Jens,

Please elaborate on "no, no, no". Are you suggesting that no further
improvements can be made or should be attempted on the 2.4 i/o subsystem?

Jonathan

Quoted 11 lines · Jens Axboe [[email protected]] wrote:
Quoted 3 lines · Jonathan Lahr

Included below is a snapshot of a patch I am developing to reduce io_request_lock contention in 2.4.

No no no, you are opening a serious can of worms. No offense, but did you really think this would fly?! This is already being taken care of for 2.5, lets leave 2.4 alone in this regard.

Signature · 1 line
Jens Axboe
Signature · 5 lines
Jonathan Lahr
IBM Linux Technology Center
Beaverton, Oregon
[email protected]
503-578-3385
#7 lore raw
In reply to Jonathan Lahr
Quoted 5 lines · Jonathan Lahr

Jens,

Please elaborate on "no, no, no". Are you suggesting that no further
improvements can be made or should be attempted on the 2.4 i/o subsystem?

Of course not. The no no no just means that attempting to globally remove the io_request_lock at this point is a no-go, so don't even go there. The sledgehammer approach will not fly at this point, it's just way too risky.

Jens

#8 lore raw
In reply to Jonathan Lahr
Quoted 7 lines
Quoted 3 lines

Please elaborate on "no, no, no". Are you suggesting that no further
improvements can be made or should be attempted on the 2.4 i/o subsystem?

Of course not. The no no no just means that attempting to globally remove the io_request_lock at this point is a no-go, so don't even go there. The sledgehammer approach will not fly at this point, it's just way too risky.

I agree that reducing locking scope is often problematic. However, this patch does not globally remove the io_request_lock. The purpose of the patch is to protect request queue integrity with a per queue lock instead of the global io_request_lock. My intent was to leave other io_request_lock serialization intact. Any insight into whether the patch leaves data unprotected would be appreciated.

Jonathan

Signature · 5 lines
Jonathan Lahr
IBM Linux Technology Center
Beaverton, Oregon
[email protected]
503-578-3385
#9 lore raw
In reply to Jonathan Lahr
Quoted 8 lines · Jonathan Lahr
Quoted 7 lines
Quoted 3 lines

Please elaborate on "no, no, no". Are you suggesting that no further
improvements can be made or should be attempted on the 2.4 i/o subsystem?

Of course not. The no no no just means that attempting to globally remove the io_request_lock at this point is a no-go, so don't even go there. The sledgehammer approach will not fly at this point, it's just way too risky.

Quoted 6 lines

I agree that reducing locking scope is often problematic. However, this patch does not globally remove the io_request_lock. The purpose of the patch is to protect request queue integrity with a per queue lock instead of the global io_request_lock. My intent was to leave other io_request_lock serialization intact. Any insight into whether the patch leaves data unprotected would be appreciated.

You are now browsing the request list without agreeing on what lock is being held -- what happens to drivers assuming that io_request_lock protects the list? Boom. For 2.4 we simply cannot afford to muck around with this, it's jsut too dangerous. For 2.5 I already completely removed the io_request_lock (also helps to catch references to it from drivers).

I agree with your SCSI approach, it's the same we took. Low level drivers must be responsible for their own locking, the mid layer should not pre-grab anything for them.

Signature · 1 line
Jens Axboe
#10 lore raw
In reply to Jonathan Lahr
Quoted 6 lines

You are now browsing the request list without agreeing on what lock is being held -- what happens to drivers assuming that io_request_lock protects the list? Boom. For 2.4 we simply cannot afford to muck around with this, it's jsut too dangerous. For 2.5 I already completely removed the io_request_lock (also helps to catch references to it from drivers).

In this patch, io_request_lock and queue_lock are both acquired in generic_unplug_device, so request_fn invocations protect request queue integrity. __make_request acquires queue_lock instead of io_request_lock thus protecting queue integrity while allowing greater concurrency.

Nevertheless, I understand your unwillingness to change locking as pervasive as io_request_lock. Such changes would of course involve risk. I am simply trying to improve 2.4 i/o performance, since 2.4 could have a long time left to live.

Quoted 3 lines

I agree with your SCSI approach, it's the same we took. Low level drivers must be responsible for their own locking, the mid layer should not pre-grab anything for them.

Yes, calling out of subsystem scope with locks held can cause problems.

Thanks for your feedback.

Jonathan

Signature · 5 lines
Jonathan Lahr
IBM Linux Technology Center
Beaverton, Oregon
[email protected]
503-578-3385
#11 lore raw
In reply to Jonathan Lahr
Quoted 11 lines · Jonathan Lahr
Quoted 10 lines

You are now browsing the request list without agreeing on what lock is
being held -- what happens to drivers assuming that io_request_lock protects the list? Boom. For 2.4 we simply cannot afford to muck around
with this, it's jsut too dangerous. For 2.5 I already completely removed
the io_request_lock (also helps to catch references to it from drivers).

Quoted 5 lines

In this patch, io_request_lock and queue_lock are both acquired in generic_unplug_device, so request_fn invocations protect request queue integrity. __make_request acquires queue_lock instead of io_request_lock
thus protecting queue integrity while allowing greater concurrency.

You fixed SCSI for q->queue_head usage, that part looks ok. The low level call backs are a much bigger mess though. And you broke IDE, cciss, cpqarray, DAC960, etc etc in the process.

Quoted 4 lines

Nevertheless, I understand your unwillingness to change locking as pervasive as io_request_lock. Such changes would of course involve risk. I am simply trying to improve 2.4 i/o performance, since 2.4 could have a long time left to live.

I can certainly understand that, but I really hope you see what I mean that we cannot change this locking now.

Signature · 1 line
Jens Axboe
#12 lore raw
In reply to Jonathan Lahr
Quoted 5 lines · Jens Axboe

You are now browsing the request list without agreeing on what lock is being held -- what happens to drivers assuming that io_request_lock protects the list? Boom. For 2.4 we simply cannot afford to muck around with this, it's jsut too dangerous. For 2.5 I already completely removed the io_request_lock (also helps to catch references to it from drivers).

Is this part of the bio patches? (I confess, I haven't had the time to look yet.) If not, do you know when we'll be seeing sneak previews of this code? (Yes, it's me again! ;)

- Pete

#13 lore raw
In reply to Jonathan Lahr
Quoted 10 lines · Peter Rival
Quoted 5 lines · Jens Axboe

You are now browsing the request list without agreeing on what lock is being held -- what happens to drivers assuming that io_request_lock protects the list? Boom. For 2.4 we simply cannot afford to muck around with this, it's jsut too dangerous. For 2.5 I already completely removed the io_request_lock (also helps to catch references to it from drivers).

Is this part of the bio patches? (I confess, I haven't had the time to look yet.) If not, do you know when we'll be seeing sneak previews of this code? (Yes, it's me again! ;)

Yep it's part of the bio patch. bio-14 for 2.4.10-preX is pending today or tomorrow.

Signature · 1 line
Jens Axboe