raid5-merge-ios-rhel5.patch

lustre 1.6.5 source code

diff -pur linux-2.6.18-53.orig/drivers/md/raid5.c linux-2.6.18-53/drivers/md/raid5.c
--- linux-2.6.18-53.orig/drivers/md/raid5.c	2007-12-28 18:55:24.000000000 +0800
+++ linux-2.6.18-53/drivers/md/raid5.c	2007-12-28 19:08:15.000000000 +0800
@@ -1277,7 +1277,26 @@ static void compute_block_2(struct strip
 	}
 }
 
+/*
+ * The whole idea is to collect all bio's and then issue them
+ * disk by disk to assist merging a bit -bzzz
+ */
+static void raid5_flush_bios(raid5_conf_t *conf, struct bio *bios[], int raid_disks)
+{
+	struct bio *bio, *nbio;
+	int i;
 
+	for (i = 0; i < raid_disks; i++) {
+		bio = bios[i];
+		while (bio) {
+			nbio = bio->bi_next;
+			bio->bi_next = NULL;
+			generic_make_request(bio);
+			bio = nbio;
+		}
+		bios[i] = NULL;
+	}
+}
 
 /*
  * Each stripe/dev can have one or more bion attached.
@@ -1392,7 +1411,7 @@ static int stripe_to_pdidx(sector_t stri
  *
  */
  
-static void handle_stripe5(struct stripe_head *sh)
+static void handle_stripe5(struct stripe_head *sh, struct bio *bios[])
 {
 	raid5_conf_t *conf = sh->raid_conf;
 	int disks = sh->disks;
@@ -1939,7 +1958,11 @@ static void handle_stripe5(struct stripe
 			    test_bit(R5_ReWrite, &sh->dev[i].flags))
 				atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
 			atomic_inc(&conf->out_reqs_in_queue);
-			generic_make_request(bi);
+			if (bios) {
+				bi->bi_next = bios[i];
+				bios[i] = bi;
+			} else
+				generic_make_request(bi);
 		} else {
 			if (rw == 1)
 				set_bit(STRIPE_DEGRADED, &sh->state);
@@ -1951,7 +1974,7 @@ static void handle_stripe5(struct stripe
 	}
 }
 
-static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
+static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page, struct bio *bios[])
 {
 	raid6_conf_t *conf = sh->raid_conf;
 	int disks = conf->raid_disks;
@@ -2499,7 +2522,11 @@ static void handle_stripe6(struct stripe
 			if (rw == WRITE &&
 			    test_bit(R5_ReWrite, &sh->dev[i].flags))
 				atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
-			generic_make_request(bi);
+			if (bios) {
+				bi->bi_next = bios[i];
+				bios[i] = bi;
+			} else
+				generic_make_request(bi);
 			atomic_inc(&conf->out_reqs_in_queue);
 		} else {
 			if (rw == 1)
@@ -2512,12 +2539,12 @@ static void handle_stripe6(struct stripe
 	}
 }
 
-static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
+static void handle_stripe(struct stripe_head *sh, struct page *tmp_page, struct bio *bios[])
 {
 	if (sh->raid_conf->level == 6)
-		handle_stripe6(sh, tmp_page);
+		handle_stripe6(sh, tmp_page, bios);
 	else
-		handle_stripe5(sh);
+		handle_stripe5(sh, bios);
 }
 
 
@@ -2670,6 +2697,7 @@ static int make_request(request_queue_t 
 	int stripes_per_chunk, sectors_per_block;
 	int sectors_per_stripe;
 	int i, j;
+	struct bio *bios[MD_SB_DISKS];
 
 	DEFINE_WAIT(w);
 	int disks, data_disks;
@@ -2698,6 +2726,7 @@ static int make_request(request_queue_t 
 	sectors = bi->bi_size >> 9;
 	stripes_per_chunk = conf->chunk_size / STRIPE_SIZE;
 
+	memset(&bios, 0, sizeof(bios));
 redo_bio:
 	/* stripe by stripe handle needs a stable raid layout, so if this
 	 * reuqest covers the expanding region, wait it over. 
@@ -2756,8 +2785,10 @@ retry:
 					 * the raid layout has been changed, we have to redo the 
 					 * whole bio because we don't which sectors in it has been
 					 * done, and which is not done. -jay */
-					if (raid5_redo_bio(conf, bi, disks, logical_sector))
+					if (raid5_redo_bio(conf, bi, disks, logical_sector)) {
+						raid5_flush_bios(conf, bios, disks);
 						goto redo_bio;
+					}
 
 					if (test_bit(STRIPE_EXPANDING, &sh->state)) {
 						/* Stripe is busy expanding or
@@ -2766,6 +2797,7 @@ retry:
 						 */
 						release_stripe(sh);
 						sh = NULL;
+						raid5_flush_bios(conf, bios, disks);
 						raid5_unplug_device(mddev->queue);
 						schedule();
 						goto retry;
@@ -2784,17 +2816,19 @@ retry:
 			 */
 			if (r_sector >= mddev->suspend_lo &&
 			    r_sector < mddev->suspend_hi) {
-				handle_stripe(sh, NULL);
+				handle_stripe(sh, NULL, NULL);
 				release_stripe(sh);
 				sh = NULL;
+				raid5_flush_bios(conf, bios, disks);
 				schedule();
 				goto retry;
 			}
 
 			if (!add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
-				handle_stripe(sh, NULL);
+				handle_stripe(sh, NULL, NULL);
 				release_stripe(sh);
 				sh = NULL;
+				raid5_flush_bios(conf, bios, disks);
 				raid5_unplug_device(mddev->queue);
 				schedule();
 				goto retry;
@@ -2810,7 +2844,7 @@ retry:
 			r_sector += sectors_per_chunk;
 		}
 		if (sh) {
-			handle_stripe(sh, NULL);
+			handle_stripe(sh, NULL, NULL);
 			release_stripe(sh);
 			sh = NULL;
 		}
@@ -2820,6 +2854,9 @@ retry:
 	if (sectors > 0)
 		goto repeat;
 
+	/* flush all of the bios */
+	raid5_flush_bios(conf, bios, disks);
+
 	spin_lock_irq(&conf->device_lock);
 	remaining = --bi->bi_phys_segments;
 	spin_unlock_irq(&conf->device_lock);
@@ -3035,7 +3072,7 @@ static inline sector_t sync_request(mdde
 	clear_bit(STRIPE_INSYNC, &sh->state);
 	spin_unlock(&sh->lock);
 
-	handle_stripe(sh, NULL);
+	handle_stripe(sh, NULL, NULL);
 	release_stripe(sh);
 
 	return STRIPE_SECTORS;
@@ -3091,7 +3128,7 @@ static void raid5d (mddev_t *mddev)
 		
 		handled++;
 		atomic_inc(&conf->handled_in_raid5d);
-		handle_stripe(sh, conf->spare_page);
+		handle_stripe(sh, conf->spare_page, NULL);
 		release_stripe(sh);
 
 		spin_lock_irq(&conf->device_lock);