raid5-merge-ios.patch

非常经典的一个分布式系统

Merge IO requests to try and get larger requests on underlying drives.

Index: linux-2.6.9/drivers/md/raid5.c
===================================================================
--- linux-2.6.9.orig/drivers/md/raid5.c	2006-05-22 00:10:04.000000000 +0400
+++ linux-2.6.9/drivers/md/raid5.c	2006-05-22 00:10:06.000000000 +0400
@@ -934,6 +934,26 @@ 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 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;
+	}
+}
 
 /*
  * handle_stripe - do things to a stripe.
@@ -953,7 +973,7 @@ 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[])
 {
 	raid5_conf_t *conf = sh->raid_conf;
 	int disks = conf->raid_disks;
@@ -1376,7 +1396,11 @@ 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);
@@ -1501,6 +1525,7 @@ static int make_request (request_queue_t
 	int sectors_per_chunk = conf->chunk_size >> 9;
 	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);
@@ -1530,6 +1555,7 @@ 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;
@@ -1549,9 +1575,17 @@ repeat:
 			new_sector = raid5_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)) {
+					raid5_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 {
@@ -1571,7 +1605,7 @@ repeat:
 		}
 		if (sh) {
 			raid5_plug_device(conf);
-			handle_stripe(sh);
+			handle_stripe(sh, bios);
 			release_stripe(sh);
 			sh = NULL;
 		}
@@ -1581,6 +1615,9 @@ repeat:
 	if (sectors > 0)
 		goto repeat;
 
+	/* now flush all bio's */
+	raid5_flush_bios(conf, bios, raid_disks);
+
 	spin_lock_irq(&conf->device_lock);
 	if (--bi->bi_phys_segments == 0) {
 		int bytes = bi->bi_size;
@@ -1636,7 +1673,7 @@ 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;
@@ -1685,7 +1722,7 @@ static void raid5d (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);