raid5-optimize-memcpy.patch

非常经典的一个分布式系统

In case of full-stripe writes don't copy data into internal cache.
This optimization reduces CPU load by 30% rougly.

Index: linux-2.6.9/include/linux/raid/raid5.h
===================================================================
--- linux-2.6.9.orig/include/linux/raid/raid5.h	2006-05-21 17:57:25.000000000 +0400
+++ linux-2.6.9/include/linux/raid/raid5.h	2006-05-22 00:10:04.000000000 +0400
@@ -152,6 +152,7 @@ struct stripe_head {
 #define	R5_Wantread	4	/* want to schedule a read */
 #define	R5_Wantwrite	5
 #define	R5_Syncio	6	/* this io need to be accounted as resync io */
+#define	R5_Direct	7	/* use page fom passed bio to avoid memcpy */
 
 /*
  * Write method
Index: linux-2.6.9/drivers/md/raid5.c
===================================================================
--- linux-2.6.9.orig/drivers/md/raid5.c	2006-05-22 00:10:01.000000000 +0400
+++ linux-2.6.9/drivers/md/raid5.c	2006-05-22 00:10:04.000000000 +0400
@@ -411,6 +411,8 @@ static int raid5_end_read_request (struc
 		clear_buffer_uptodate(bh);
 	}
 #endif
+	if (test_bit(R5_Direct, &sh->dev[i].flags))
+		printk("R5_Direct for READ ?!\n");
 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
 	set_bit(STRIPE_HANDLE, &sh->state);
 	release_stripe(sh);
@@ -449,6 +451,10 @@ static int raid5_end_write_request (stru
 
 	rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
 	
+	if (test_bit(R5_Direct, &sh->dev[i].flags)) {
+		BUG_ON(sh->dev[i].req.bi_io_vec[0].bv_page == sh->dev[i].page);
+		sh->dev[i].req.bi_io_vec[0].bv_page = sh->dev[i].page;
+	}
 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
 	set_bit(STRIPE_HANDLE, &sh->state);
 	__release_stripe(conf, sh);
@@ -673,6 +679,49 @@ static void copy_data(int frombio, struc
 	}
 }
 
+static struct page *zero_copy_data(struct bio *bio, sector_t sector)
+{
+	struct bio_vec *bvl;
+	int i;
+
+	for (;bio && bio->bi_sector < sector+STRIPE_SECTORS;
+	      bio = r5_next_bio(bio, sector) ) {
+		int page_offset;
+		if (bio->bi_sector >= sector)
+			page_offset = (signed)(bio->bi_sector - sector) * 512;
+		else 
+			page_offset = (signed)(sector - bio->bi_sector) * -512;
+		bio_for_each_segment(bvl, bio, i) {
+			int len = bio_iovec_idx(bio,i)->bv_len;
+			int clen;
+			int b_offset = 0;			
+
+			if (page_offset < 0) {
+				b_offset = -page_offset;
+				page_offset += b_offset;
+				len -= b_offset;
+			}
+
+			if (len > 0 && page_offset + len > STRIPE_SIZE)
+				clen = STRIPE_SIZE - page_offset;	
+			else clen = len;
+			
+			if (clen > 0) {
+				BUG_ON(clen < STRIPE_SIZE);
+				/*printk("  sector %lu: page %p from index %u\n",
+					(unsigned long) sector,
+					bio_iovec_idx(bio, i)->bv_page, i);*/
+				return bio_iovec_idx(bio, i)->bv_page;
+			}	
+			if (clen < len) /* hit end of page */
+				break;
+			page_offset +=  len;
+		}
+	}
+	BUG();
+	return NULL;
+}
+
 #define check_xor() 	do { 						\
 			   if (count == MAX_XOR_BLOCKS) {		\
 				xor_block(count, STRIPE_SIZE, ptr);	\
@@ -717,6 +766,8 @@ static void compute_parity(struct stripe
 	int i, pd_idx = sh->pd_idx, disks = conf->raid_disks, count;
 	void *ptr[MAX_XOR_BLOCKS];
 	struct bio *chosen;
+	struct page *page;
+	int zerocopy = 0;
 
 	PRINTK("compute_parity, stripe %llu, method %d\n",
 		(unsigned long long)sh->sector, method);
@@ -743,13 +794,17 @@ static void compute_parity(struct stripe
 		break;
 	case RECONSTRUCT_WRITE:
 		memset(ptr[0], 0, STRIPE_SIZE);
-		for (i= disks; i-- ;)
+		zerocopy = 1;
+		for (i= disks; i-- ;) {
+			if (i != pd_idx && !sh->dev[i].towrite)
+				zerocopy = 0;
 			if (i!=pd_idx && sh->dev[i].towrite) {
 				chosen = sh->dev[i].towrite;
 				sh->dev[i].towrite = NULL;
 				if (sh->dev[i].written) BUG();
 				sh->dev[i].written = chosen;
 			}
+		}
 		break;
 	case CHECK_PARITY:
 		break;
@@ -759,34 +814,62 @@ static void compute_parity(struct stripe
 		count = 1;
 	}
 	
-	for (i = disks; i--;)
-		if (sh->dev[i].written) {
-			sector_t sector = sh->dev[i].sector;
-			struct bio *wbi = sh->dev[i].written;
-			while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
-				copy_data(1, wbi, sh->dev[i].page, sector);
-				wbi = r5_next_bio(wbi, sector);
-			}
+	for (i = disks; i--;) {
+		struct bio *wbi = sh->dev[i].written;
+		sector_t sector;
+
+		if (!wbi)
+			continue;
+
+		sector = sh->dev[i].sector;
+		set_bit(R5_LOCKED, &sh->dev[i].flags);
+		BUG_ON(test_bit(R5_Direct, &sh->dev[i].flags));
+
+		/* check if it's covered by a single page
+		   and whole stripe is written at once.
+		 * in this case we can avoid memcpy() */
+		if (zerocopy && wbi && wbi->bi_next == NULL && 
+				test_bit(R5_OVERWRITE, &sh->dev[i].flags)) {
+			page = zero_copy_data(wbi, sector);
+			BUG_ON(PageHighMem(page));
+			sh->dev[i].req.bi_io_vec[0].bv_page = page;
+			set_bit(R5_Direct, &sh->dev[i].flags);
+			clear_bit(R5_UPTODATE, &sh->dev[i].flags);
+			continue;
+		}
 
-			set_bit(R5_LOCKED, &sh->dev[i].flags);
-			set_bit(R5_UPTODATE, &sh->dev[i].flags);
+		set_bit(R5_UPTODATE, &sh->dev[i].flags);
+		while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
+			copy_data(1, wbi, sh->dev[i].page, sector);
+			wbi = r5_next_bio(wbi, sector);
 		}
+	}
 
 	switch(method) {
 	case RECONSTRUCT_WRITE:
 	case CHECK_PARITY:
-		for (i=disks; i--;)
-			if (i != pd_idx) {
-				ptr[count++] = page_address(sh->dev[i].page);
-				check_xor();
-			}
+		for (i=disks; i--;) {
+			if (i == pd_idx)
+				continue;
+			if (test_bit(R5_Direct, &sh->dev[i].flags))
+				page = sh->dev[i].req.bi_io_vec[0].bv_page;
+			else
+				page = sh->dev[i].page;
+			ptr[count++] = page_address(page);
+			check_xor();
+		}
 		break;
 	case READ_MODIFY_WRITE:
-		for (i = disks; i--;)
-			if (sh->dev[i].written) {
-				ptr[count++] = page_address(sh->dev[i].page);
-				check_xor();
-			}
+		for (i = disks; i--;) {
+			if (!sh->dev[i].written)
+				continue;
+			if (test_bit(R5_Direct, &sh->dev[i].flags))
+				page = sh->dev[i].req.bi_io_vec[0].bv_page;
+			else
+				page = sh->dev[i].page;
+			ptr[count++] = page_address(page);
+			check_xor();
+		}
 	}
 	if (count != 1)
 		xor_block(count, STRIPE_SIZE, ptr);
@@ -1012,7 +1094,7 @@ static void handle_stripe(struct stripe_
 	dev = &sh->dev[sh->pd_idx];
 	if ( written &&
 	     ( (test_bit(R5_Insync, &dev->flags) && !test_bit(R5_LOCKED, &dev->flags) &&
-		test_bit(R5_UPTODATE, &dev->flags))
+		(test_bit(R5_UPTODATE, &dev->flags) || test_bit(R5_Direct, &dev->flags)))
 	       || (failed == 1 && failed_num == sh->pd_idx))
 	    ) {
 	    /* any written block on an uptodate or failed drive can be returned.
@@ -1023,13 +1105,16 @@ static void handle_stripe(struct stripe_
 		if (sh->dev[i].written) {
 		    dev = &sh->dev[i];
 		    if (!test_bit(R5_LOCKED, &dev->flags) &&
-			 test_bit(R5_UPTODATE, &dev->flags) ) {
+			 (test_bit(R5_UPTODATE, &dev->flags) ||
+			  	test_bit(R5_Direct, &dev->flags)) ) {
 			/* We can return any write requests */
 			    struct bio *wbi, *wbi2;
 			    PRINTK("Return write for disc %d\n", i);
 			    spin_lock_irq(&conf->device_lock);
 			    wbi = dev->written;
 			    dev->written = NULL;
+			    if (test_bit(R5_Direct, &dev->flags))
+				    clear_bit(R5_Direct, &dev->flags);
 			    while (wbi && wbi->bi_sector < dev->sector + STRIPE_SECTORS) {
 				    wbi2 = r5_next_bio(wbi, dev->sector);
 				    if (--wbi->bi_phys_segments == 0) {