[BLOCK] Fix bad sharing of tag busy list on queues with shared tag maps
[safe/jmp/linux-2.6] / block / ll_rw_blk.c
index 3935469..56f2646 100644 (file)
@@ -39,7 +39,7 @@
 
 static void blk_unplug_work(struct work_struct *work);
 static void blk_unplug_timeout(unsigned long data);
-static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io);
+static void drive_stat_acct(struct request *rq, int new_io);
 static void init_request_from_bio(struct request *req, struct bio *bio);
 static int __make_request(struct request_queue *q, struct bio *bio);
 static struct io_context *current_io_context(gfp_t gfp_flags, int node);
@@ -791,7 +791,6 @@ static int __blk_free_tags(struct blk_queue_tag *bqt)
        retval = atomic_dec_and_test(&bqt->refcnt);
        if (retval) {
                BUG_ON(bqt->busy);
-               BUG_ON(!list_empty(&bqt->busy_list));
 
                kfree(bqt->tag_index);
                bqt->tag_index = NULL;
@@ -903,7 +902,6 @@ static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
        if (init_tag_map(q, tags, depth))
                goto fail;
 
-       INIT_LIST_HEAD(&tags->busy_list);
        tags->busy = 0;
        atomic_set(&tags->refcnt, 1);
        return tags;
@@ -954,6 +952,7 @@ int blk_queue_init_tags(struct request_queue *q, int depth,
         */
        q->queue_tags = tags;
        q->queue_flags |= (1 << QUEUE_FLAG_QUEUED);
+       INIT_LIST_HEAD(&q->tag_busy_list);
        return 0;
 fail:
        kfree(tags);
@@ -1057,18 +1056,16 @@ void blk_queue_end_tag(struct request_queue *q, struct request *rq)
 
        bqt->tag_index[tag] = NULL;
 
-       /*
-        * We use test_and_clear_bit's memory ordering properties here.
-        * The tag_map bit acts as a lock for tag_index[bit], so we need
-        * a barrer before clearing the bit (precisely: release semantics).
-        * Could use clear_bit_unlock when it is merged.
-        */
-       if (unlikely(!test_and_clear_bit(tag, bqt->tag_map))) {
+       if (unlikely(!test_bit(tag, bqt->tag_map))) {
                printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
                       __FUNCTION__, tag);
                return;
        }
-
+       /*
+        * The tag_map bit acts as a lock for tag_index[bit], so we need
+        * unlock memory barrier semantics.
+        */
+       clear_bit_unlock(tag, bqt->tag_map);
        bqt->busy--;
 }
 
@@ -1114,17 +1111,17 @@ int blk_queue_start_tag(struct request_queue *q, struct request *rq)
                if (tag >= bqt->max_depth)
                        return 1;
 
-       } while (test_and_set_bit(tag, bqt->tag_map));
+       } while (test_and_set_bit_lock(tag, bqt->tag_map));
        /*
-        * We rely on test_and_set_bit providing lock memory ordering semantics
-        * (could use test_and_set_bit_lock when it is merged).
+        * We need lock ordering semantics given by test_and_set_bit_lock.
+        * See blk_queue_end_tag for details.
         */
 
        rq->cmd_flags |= REQ_QUEUED;
        rq->tag = tag;
        bqt->tag_index[tag] = rq;
        blkdev_dequeue_request(rq);
-       list_add(&rq->queuelist, &bqt->busy_list);
+       list_add(&rq->queuelist, &q->tag_busy_list);
        bqt->busy++;
        return 0;
 }
@@ -1145,11 +1142,10 @@ EXPORT_SYMBOL(blk_queue_start_tag);
  **/
 void blk_queue_invalidate_tags(struct request_queue *q)
 {
-       struct blk_queue_tag *bqt = q->queue_tags;
        struct list_head *tmp, *n;
        struct request *rq;
 
-       list_for_each_safe(tmp, n, &bqt->busy_list) {
+       list_for_each_safe(tmp, n, &q->tag_busy_list) {
                rq = list_entry_rq(tmp);
 
                if (rq->tag == -1) {
@@ -1351,18 +1347,30 @@ int blk_rq_map_sg(struct request_queue *q, struct request *rq,
 new_segment:
                        if (!sg)
                                sg = sglist;
-                       else
+                       else {
+                               /*
+                                * If the driver previously mapped a shorter
+                                * list, we could see a termination bit
+                                * prematurely unless it fully inits the sg
+                                * table on each mapping. We KNOW that there
+                                * must be more entries here or the driver
+                                * would be buggy, so force clear the
+                                * termination bit to avoid doing a full
+                                * sg_init_table() in drivers for each command.
+                                */
+                               sg->page_link &= ~0x02;
                                sg = sg_next(sg);
+                       }
 
-                       memset(sg, 0, sizeof(*sg));
-                       sg->page = bvec->bv_page;
-                       sg->length = nbytes;
-                       sg->offset = bvec->bv_offset;
+                       sg_set_page(sg, bvec->bv_page, nbytes, bvec->bv_offset);
                        nsegs++;
                }
                bvprv = bvec;
        } /* segments in rq */
 
+       if (sg)
+               __sg_mark_end(sg);
+
        return nsegs;
 }
 
@@ -1726,6 +1734,7 @@ EXPORT_SYMBOL(blk_stop_queue);
 void blk_sync_queue(struct request_queue *q)
 {
        del_timer_sync(&q->unplug_timer);
+       kblockd_flush_work(&q->unplug_work);
 }
 EXPORT_SYMBOL(blk_sync_queue);
 
@@ -2329,7 +2338,7 @@ void blk_insert_request(struct request_queue *q, struct request *rq,
        if (blk_rq_tagged(rq))
                blk_queue_end_tag(q, rq);
 
-       drive_stat_acct(rq, rq->nr_sectors, 1);
+       drive_stat_acct(rq, 1);
        __elv_add_request(q, rq, where, 0);
        blk_start_queueing(q);
        spin_unlock_irqrestore(q->queue_lock, flags);
@@ -2724,7 +2733,7 @@ int blkdev_issue_flush(struct block_device *bdev, sector_t *error_sector)
 
 EXPORT_SYMBOL(blkdev_issue_flush);
 
-static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
+static void drive_stat_acct(struct request *rq, int new_io)
 {
        int rw = rq_data_dir(rq);
 
@@ -2746,7 +2755,7 @@ static void drive_stat_acct(struct request *rq, int nr_sectors, int new_io)
  */
 static inline void add_request(struct request_queue * q, struct request * req)
 {
-       drive_stat_acct(req, req->nr_sectors, 1);
+       drive_stat_acct(req, 1);
 
        /*
         * elevator indicated where it wants this request to be
@@ -3003,7 +3012,7 @@ static int __make_request(struct request_queue *q, struct bio *bio)
                        req->biotail = bio;
                        req->nr_sectors = req->hard_nr_sectors += nr_sectors;
                        req->ioprio = ioprio_best(req->ioprio, prio);
-                       drive_stat_acct(req, nr_sectors, 0);
+                       drive_stat_acct(req, 0);
                        if (!attempt_back_merge(q, req))
                                elv_merged_request(q, req, el_ret);
                        goto out;
@@ -3030,7 +3039,7 @@ static int __make_request(struct request_queue *q, struct bio *bio)
                        req->sector = req->hard_sector = bio->bi_sector;
                        req->nr_sectors = req->hard_nr_sectors += nr_sectors;
                        req->ioprio = ioprio_best(req->ioprio, prio);
-                       drive_stat_acct(req, nr_sectors, 0);
+                       drive_stat_acct(req, 0);
                        if (!attempt_front_merge(q, req))
                                elv_merged_request(q, req, el_ret);
                        goto out;
@@ -3367,7 +3376,7 @@ void submit_bio(int rw, struct bio *bio)
                if (unlikely(block_dump)) {
                        char b[BDEVNAME_SIZE];
                        printk(KERN_DEBUG "%s(%d): %s block %Lu on %s\n",
-                               current->comm, current->pid,
+                       current->comm, task_pid_nr(current),
                                (rw & WRITE) ? "WRITE" : "READ",
                                (unsigned long long)bio->bi_sector,
                                bdevname(bio->bi_bdev,b));
@@ -3739,7 +3748,7 @@ EXPORT_SYMBOL(end_dequeued_request);
 
 /**
  * end_request - end I/O on the current segment of the request
- * @rq:                the request being processed
+ * @req:       the request being processed
  * @uptodate:  error value or 0/1 uptodate flag
  *
  * Description: