[Qemu-devel] [PATCH 27/27] block/parallels: improve image writing performance further

["Denis V. Lunev" <den@openvz.org>]

Try to perform IO for the biggest continuous block possible.
All blocks abscent in the image are accounted in the same type
and preallocation is made for all of them at once.

The performance for sequential write is increased from 200 Gb/sec to
235 Gb/sec on my SSD HDD.

Signed-off-by: Denis V. Lunev <den@openvz.org>
Reviewed-by: Roman Kagan <rkagan@parallels.com>
CC: Kevin Wolf <kwolf@redhat.com>
CC: Stefan Hajnoczi <stefanha@redhat.com>
---
 block/parallels.c | 44 ++++++++++++++++++++++++--------------------
 1 file changed, 24 insertions(+), 20 deletions(-)

diff --git a/block/parallels.c b/block/parallels.c
index 2a1e31f..7531340 100644
--- a/block/parallels.c
+++ b/block/parallels.c
@@ -179,43 +179,47 @@ static int64_t block_status(BDRVParallelsState *s, int64_t sector_num,
     return start_off;
 }
 
-static int64_t allocate_cluster(BlockDriverState *bs, int64_t sector_num)
+static int64_t allocate_clusters(BlockDriverState *bs, int64_t sector_num,
+                                 int nb_sectors, int *pnum)
 {
     BDRVParallelsState *s = bs->opaque;
-    uint32_t idx, offset;
-    int64_t pos;
+    uint32_t idx, to_allocate, i;
+    int64_t pos, space;
 
-    idx = sector_num / s->tracks;
-    offset = sector_num % s->tracks;
+    pos = block_status(s, sector_num, nb_sectors, pnum);
+    if (pos > 0) {
+        return pos;
+    }
 
+    idx = sector_num / s->tracks;
     if (idx >= s->bat_size) {
         return -EINVAL;
     }
-    if (s->bat_bitmap[idx] != 0) {
-        return bat2sect(s, idx) + offset;
-    }
 
-    pos = bdrv_getlength(bs->file) >> BDRV_SECTOR_BITS;
-    if (s->data_end + s->tracks > pos) {
+    to_allocate = (sector_num + *pnum + s->tracks - 1) / s->tracks - idx;
+    space = to_allocate * s->tracks;
+    if (s->data_end + space > bdrv_getlength(bs->file) >> BDRV_SECTOR_BITS) {
         int ret;
+        space += s->prealloc_size;
         if (s->prealloc_mode == PRL_PREALLOC_MODE_FALLOCATE) {
-            ret = bdrv_write_zeroes(bs->file, s->data_end,
-                                    s->prealloc_size, 0);
+            ret = bdrv_write_zeroes(bs->file, s->data_end, space, 0);
         } else {
             ret = bdrv_truncate(bs->file,
-                    (s->data_end + s->prealloc_size) << BDRV_SECTOR_BITS);
+                                (s->data_end + space) << BDRV_SECTOR_BITS);
         }
         if (ret < 0) {
             return ret;
         }
     }
-    pos = s->data_end;
-    s->data_end += s->tracks;
 
-    s->bat_bitmap[idx] = cpu_to_le32(pos / s->off_multiplier);
+    for (i = 0; i < to_allocate; i++) {
+        s->bat_bitmap[idx + i] = cpu_to_le32(s->data_end / s->off_multiplier);
+        s->data_end += s->tracks;
+        bitmap_set(s->bat_dirty_bmap,
+                   bat_entry_off(idx) / s->bat_dirty_block, 1);
+    }
 
-    bitmap_set(s->bat_dirty_bmap, bat_entry_off(idx) / s->bat_dirty_block, 1);
-    return bat2sect(s, idx) + offset;
+    return bat2sect(s, idx) + sector_num % s->tracks;
 }
 
 
@@ -280,8 +284,8 @@ static coroutine_fn int parallels_co_writev(BlockDriverState *bs,
 
     qemu_co_mutex_lock(&s->lock);
     while (nb_sectors > 0) {
-        int64_t position = allocate_cluster(bs, sector_num);
-        int n = cluster_remainder(s, sector_num, nb_sectors);
+        int n;
+        int64_t position = allocate_clusters(bs, sector_num, nb_sectors, &n);
         int nbytes = n << BDRV_SECTOR_BITS;
 
         if (position < 0) {
-- 
1.9.1