block: reserve some tags just for sync IO

By only allowing async IO to consume 3/4 ths of the tag depth, we
always have slots free to serve sync IO. This is important to avoid
having writes fill the entire tag queue, thus starving reads.

Original patch and idea from Linus Torvalds <torvalds@linux-foundation.org>

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

diff --git a/block/blk-tag.c b/block/blk-tag.c
index 8a99688..c0d419e 100644 (file)
--- a/block/blk-tag.c
+++ b/block/blk-tag.c
@@ -337,6 +337,7 @@ EXPORT_SYMBOL(blk_queue_end_tag);
 int blk_queue_start_tag(struct request_queue *q, struct request *rq)
 {
        struct blk_queue_tag *bqt = q->queue_tags;
+       unsigned max_depth, offset;
        int tag;
 
        if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
@@ -350,10 +351,19 @@ int blk_queue_start_tag(struct request_queue *q, struct request *rq)
        /*
         * Protect against shared tag maps, as we may not have exclusive
         * access to the tag map.
+        *
+        * We reserve a few tags just for sync IO, since we don't want
+        * to starve sync IO on behalf of flooding async IO.
         */
+       max_depth = bqt->max_depth;
+       if (rq_is_sync(rq))
+               offset = 0;
+       else
+               offset = max_depth >> 2;
+
        do {
-               tag = find_first_zero_bit(bqt->tag_map, bqt->max_depth);
-               if (tag >= bqt->max_depth)
+               tag = find_next_zero_bit(bqt->tag_map, max_depth, offset);
+               if (tag >= max_depth)
                        return 1;
 
        } while (test_and_set_bit_lock(tag, bqt->tag_map));