raid6-merge-ios.patch

来自「lustre 1.6.5 source code」· PATCH 代码 · 共 127 行

diff -pur linux-2.6.9.orig/drivers/md/raid6main.c linux-2.6.9/drivers/md/raid6main.c
--- linux-2.6.9.orig/drivers/md/raid6main.c	2008-01-10 13:51:32.000000000 +0800
+++ linux-2.6.9/drivers/md/raid6main.c	2008-01-10 13:52:20.000000000 +0800
@@ -956,6 +956,26 @@ static void add_stripe_bio (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 raid6_flush_bios(raid6_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;
+	}
+}
 
 /*
  * handle_stripe - do things to a stripe.
@@ -975,7 +995,7 @@ static void add_stripe_bio (struct strip
  *
  */
 
-static void handle_stripe(struct stripe_head *sh)
+static void handle_stripe(struct stripe_head *sh, struct bio *bios[])
 {
 	raid6_conf_t *conf = sh->raid_conf;
 	int disks = conf->raid_disks;
@@ -1452,7 +1472,11 @@ static void handle_stripe(struct stripe_
 			bi->bi_size = STRIPE_SIZE;
 			bi->bi_next = NULL;
 			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 {
 			PRINTK("skip op %ld on disc %d for sector %llu\n",
 				bi->bi_rw, i, (unsigned long long)sh->sector);
@@ -1575,6 +1599,7 @@ static int make_request (request_queue_t
 	int sectors_per_chunk;
 	int stripes_per_chunk, sectors_per_block;
 	int sectors_per_stripe;
+	struct bio *bios[MD_SB_DISKS];
 	int i, j;
 
 	atomic_inc(&conf->in_reqs_in_queue);
@@ -1611,6 +1636,7 @@ static int make_request (request_queue_t
 	sector_div(block, sectors_per_block);
 	sectors = bi->bi_size >> 9;
  
+	memset(&bios, 0, sizeof(bios));
  repeat:
 	stripe = block * (sectors_per_block / data_disks);
 	b_sector = stripe * data_disks;
@@ -1630,9 +1656,17 @@ static int make_request (request_queue_t
 			new_sector = raid6_compute_sector(r_sector, raid_disks,
 							data_disks, &dd_idx, 
 							&pd_idx, conf);
-			if (sh == NULL)
-				sh = get_active_stripe(conf, new_sector, pd_idx,
-							(bi->bi_rw&RWA_MASK));
+			if (sh == NULL) {
+				/* first, try to get stripe w/o blocking
+				 * if we can't, then it's time to submit
+				 * all collected bio's in order to free
+				 * some space in the cache -bzzz */
+				sh = get_active_stripe(conf, new_sector, pd_idx, 1);
+				if (!sh && !(bi->bi_rw&RWA_MASK)) {
+					raid6_flush_bios(conf, bios, raid_disks);
+					sh = get_active_stripe(conf, new_sector, pd_idx, 0);
+				}
+			}
 			if (sh) {
 				add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK));
 			} else {
@@ -1653,7 +1687,7 @@ static int make_request (request_queue_t
 
 		if (sh) {
 			raid6_plug_device(conf);
-			handle_stripe(sh);
+			handle_stripe(sh, bios);
 			release_stripe(sh);
 			sh = NULL;
 		}
@@ -1664,6 +1698,9 @@ static int make_request (request_queue_t
 	if(sectors > 0)
 		goto repeat;
 
+	/* now flush all bio's */
+	raid6_flush_bios(conf, bios, raid_disks);
+
 	spin_lock_irq(&conf->device_lock);
 	if (--bi->bi_phys_segments == 0) {
 		int bytes = bi->bi_size;
@@ -1719,7 +1756,7 @@ static int sync_request (mddev_t *mddev,
 	clear_bit(STRIPE_INSYNC, &sh->state);
 	spin_unlock(&sh->lock);
 
-	handle_stripe(sh);
+	handle_stripe(sh, NULL);
 	release_stripe(sh);
 
 	return STRIPE_SECTORS;
@@ -1769,7 +1806,7 @@ static void raid6d (mddev_t *mddev)
 		handled++;
 
 		atomic_inc(&conf->handled_in_raid5d);
-		handle_stripe(sh);
+		handle_stripe(sh, NULL);
 		release_stripe(sh);
 
 		spin_lock_irq(&conf->device_lock);