md: raid0 :Enables chunk size other than powers of 2.
authorraz ben yehuda <raziebe@gmail.com>
Tue, 16 Jun 2009 07:02:05 +0000 (17:02 +1000)
committerNeilBrown <neilb@suse.de>
Tue, 16 Jun 2009 07:02:05 +0000 (17:02 +1000)
Maintain two flows, one for pow2 chunk sizes (which uses masks and
shift), and a flow for the general case (which uses sector_div).
This is for the sake of performance.

 - introduce map_sector and is_io_in_chunk_boundary to encapsulate
   those two flows better for raid0_make_request
 - fix blk_mergeable to support the two flows.

Signed-off-by: raziebe@gmail.com
Signed-off-by: NeilBrown <neilb@suse.de>
drivers/md/raid0.c

index 39936a2..7cd2671 100644 (file)
@@ -273,7 +273,12 @@ static int raid0_mergeable_bvec(struct request_queue *q,
        unsigned int chunk_sectors = mddev->chunk_size >> 9;
        unsigned int bio_sectors = bvm->bi_size >> 9;
 
-       max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
+       if (is_power_of_2(mddev->chunk_size))
+               max =  (chunk_sectors - ((sector & (chunk_sectors-1))
+                                               + bio_sectors)) << 9;
+       else
+               max =  (chunk_sectors - (sector_div(sector, chunk_sectors)
+                                               + bio_sectors)) << 9;
        if (max < 0) max = 0; /* bio_add cannot handle a negative return */
        if (max <= biovec->bv_len && bio_sectors == 0)
                return biovec->bv_len;
@@ -299,9 +304,8 @@ static int raid0_run(mddev_t *mddev)
 {
        int ret;
 
-       if (mddev->chunk_size == 0 ||
-           !is_power_of_2(mddev->chunk_size)) {
-               printk(KERN_ERR "md/raid0: chunk size must be a power of 2.\n");
+       if (mddev->chunk_size == 0) {
+               printk(KERN_ERR "md/raid0: chunk size must be set.\n");
                return -EINVAL;
        }
        blk_queue_max_sectors(mddev->queue, mddev->chunk_size >> 9);
@@ -367,15 +371,65 @@ static struct strip_zone *find_zone(struct raid0_private_data *conf,
        BUG();
 }
 
-static int raid0_make_request (struct request_queue *q, struct bio *bio)
+/*
+ * remaps the bio to the target device. we separate two flows.
+ * power 2 flow and a general flow for the sake of perfromance
+*/
+static mdk_rdev_t *map_sector(mddev_t *mddev, struct strip_zone *zone,
+                               sector_t sector, sector_t *sector_offset)
 {
-       mddev_t *mddev = q->queuedata;
-       unsigned int sect_in_chunk, chunksect_bits, chunk_sects;
+       unsigned int sect_in_chunk;
+       sector_t chunk;
        raid0_conf_t *conf = mddev->private;
+       unsigned int chunk_sects = mddev->chunk_size >> 9;
+
+       if (is_power_of_2(mddev->chunk_size)) {
+               int chunksect_bits = ffz(~chunk_sects);
+               /* find the sector offset inside the chunk */
+               sect_in_chunk  = sector & (chunk_sects - 1);
+               sector >>= chunksect_bits;
+               /* chunk in zone */
+               chunk = *sector_offset;
+               /* quotient is the chunk in real device*/
+               sector_div(chunk, zone->nb_dev << chunksect_bits);
+       } else{
+               sect_in_chunk = sector_div(sector, chunk_sects);
+               chunk = *sector_offset;
+               sector_div(chunk, chunk_sects * zone->nb_dev);
+       }
+       /*
+       *  position the bio over the real device
+       *  real sector = chunk in device + starting of zone
+       *       + the position in the chunk
+       */
+       *sector_offset = (chunk * chunk_sects) + sect_in_chunk;
+       return conf->devlist[(zone - conf->strip_zone)*mddev->raid_disks
+                            + sector_div(sector, zone->nb_dev)];
+}
+
+/*
+ * Is io distribute over 1 or more chunks ?
+*/
+static inline int is_io_in_chunk_boundary(mddev_t *mddev,
+                       unsigned int chunk_sects, struct bio *bio)
+{
+       if (likely(is_power_of_2(mddev->chunk_size))) {
+               return chunk_sects >= ((bio->bi_sector & (chunk_sects-1))
+                                       + (bio->bi_size >> 9));
+       } else{
+               sector_t sector = bio->bi_sector;
+               return chunk_sects >= (sector_div(sector, chunk_sects)
+                                               + (bio->bi_size >> 9));
+       }
+}
+
+static int raid0_make_request(struct request_queue *q, struct bio *bio)
+{
+       mddev_t *mddev = q->queuedata;
+       unsigned int chunk_sects;
+       sector_t sector_offset;
        struct strip_zone *zone;
        mdk_rdev_t *tmp_dev;
-       sector_t chunk;
-       sector_t sector, rsect, sector_offset;
        const int rw = bio_data_dir(bio);
        int cpu;
 
@@ -391,10 +445,8 @@ static int raid0_make_request (struct request_queue *q, struct bio *bio)
        part_stat_unlock();
 
        chunk_sects = mddev->chunk_size >> 9;
-       chunksect_bits = ffz(~chunk_sects);
-       sector = bio->bi_sector;
-
-       if (unlikely(chunk_sects < (bio->bi_sector & (chunk_sects - 1)) + (bio->bi_size >> 9))) {
+       if (unlikely(!is_io_in_chunk_boundary(mddev, chunk_sects, bio))) {
+               sector_t sector = bio->bi_sector;
                struct bio_pair *bp;
                /* Sanity check -- queue functions should prevent this happening */
                if (bio->bi_vcnt != 1 ||
@@ -403,7 +455,12 @@ static int raid0_make_request (struct request_queue *q, struct bio *bio)
                /* This is a one page bio that upper layers
                 * refuse to split for us, so we need to split it.
                 */
-               bp = bio_split(bio, chunk_sects - (bio->bi_sector & (chunk_sects - 1)));
+               if (likely(is_power_of_2(mddev->chunk_size)))
+                       bp = bio_split(bio, chunk_sects - (sector &
+                                                          (chunk_sects-1)));
+               else
+                       bp = bio_split(bio, chunk_sects -
+                                      sector_div(sector, chunk_sects));
                if (raid0_make_request(q, &bp->bio1))
                        generic_make_request(&bp->bio1);
                if (raid0_make_request(q, &bp->bio2))
@@ -412,24 +469,14 @@ static int raid0_make_request (struct request_queue *q, struct bio *bio)
                bio_pair_release(bp);
                return 0;
        }
-       sector_offset = sector;
-       zone = find_zone(conf, &sector_offset);
-       sect_in_chunk = bio->bi_sector & (chunk_sects - 1);
-       {
-               sector_t x = sector_offset >> chunksect_bits;
-
-               sector_div(x, zone->nb_dev);
-               chunk = x;
 
-               x = sector >> chunksect_bits;
-               tmp_dev = conf->devlist[(zone - conf->strip_zone)*mddev->raid_disks
-                                       + sector_div(x, zone->nb_dev)];
-       }
-       rsect = (chunk << chunksect_bits) + zone->dev_start + sect_in_chunk;
+       sector_offset = bio->bi_sector;
+       zone =  find_zone(mddev->private, &sector_offset);
+       tmp_dev = map_sector(mddev, zone, bio->bi_sector,
+                            &sector_offset);
        bio->bi_bdev = tmp_dev->bdev;
-       bio->bi_sector = rsect + tmp_dev->data_offset;
-
+       bio->bi_sector = sector_offset + zone->dev_start +
+               tmp_dev->data_offset;
        /*
         * Let the main block layer submit the IO and resolve recursion:
         */