Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / block / blk-tag.c
diff --git a/kernel/block/blk-tag.c b/kernel/block/blk-tag.c
new file mode 100644 (file)
index 0000000..f0344e6
--- /dev/null
@@ -0,0 +1,402 @@
+/*
+ * Functions related to tagged command queuing
+ */
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/bio.h>
+#include <linux/blkdev.h>
+#include <linux/slab.h>
+
+#include "blk.h"
+
+/**
+ * blk_queue_find_tag - find a request by its tag and queue
+ * @q:  The request queue for the device
+ * @tag: The tag of the request
+ *
+ * Notes:
+ *    Should be used when a device returns a tag and you want to match
+ *    it with a request.
+ *
+ *    no locks need be held.
+ **/
+struct request *blk_queue_find_tag(struct request_queue *q, int tag)
+{
+       return blk_map_queue_find_tag(q->queue_tags, tag);
+}
+EXPORT_SYMBOL(blk_queue_find_tag);
+
+/**
+ * blk_free_tags - release a given set of tag maintenance info
+ * @bqt:       the tag map to free
+ *
+ * Drop the reference count on @bqt and frees it when the last reference
+ * is dropped.
+ */
+void blk_free_tags(struct blk_queue_tag *bqt)
+{
+       if (atomic_dec_and_test(&bqt->refcnt)) {
+               BUG_ON(find_first_bit(bqt->tag_map, bqt->max_depth) <
+                                                       bqt->max_depth);
+
+               kfree(bqt->tag_index);
+               bqt->tag_index = NULL;
+
+               kfree(bqt->tag_map);
+               bqt->tag_map = NULL;
+
+               kfree(bqt);
+       }
+}
+EXPORT_SYMBOL(blk_free_tags);
+
+/**
+ * __blk_queue_free_tags - release tag maintenance info
+ * @q:  the request queue for the device
+ *
+ *  Notes:
+ *    blk_cleanup_queue() will take care of calling this function, if tagging
+ *    has been used. So there's no need to call this directly.
+ **/
+void __blk_queue_free_tags(struct request_queue *q)
+{
+       struct blk_queue_tag *bqt = q->queue_tags;
+
+       if (!bqt)
+               return;
+
+       blk_free_tags(bqt);
+
+       q->queue_tags = NULL;
+       queue_flag_clear_unlocked(QUEUE_FLAG_QUEUED, q);
+}
+
+/**
+ * blk_queue_free_tags - release tag maintenance info
+ * @q:  the request queue for the device
+ *
+ *  Notes:
+ *     This is used to disable tagged queuing to a device, yet leave
+ *     queue in function.
+ **/
+void blk_queue_free_tags(struct request_queue *q)
+{
+       queue_flag_clear_unlocked(QUEUE_FLAG_QUEUED, q);
+}
+EXPORT_SYMBOL(blk_queue_free_tags);
+
+static int
+init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth)
+{
+       struct request **tag_index;
+       unsigned long *tag_map;
+       int nr_ulongs;
+
+       if (q && depth > q->nr_requests * 2) {
+               depth = q->nr_requests * 2;
+               printk(KERN_ERR "%s: adjusted depth to %d\n",
+                      __func__, depth);
+       }
+
+       tag_index = kzalloc(depth * sizeof(struct request *), GFP_ATOMIC);
+       if (!tag_index)
+               goto fail;
+
+       nr_ulongs = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
+       tag_map = kzalloc(nr_ulongs * sizeof(unsigned long), GFP_ATOMIC);
+       if (!tag_map)
+               goto fail;
+
+       tags->real_max_depth = depth;
+       tags->max_depth = depth;
+       tags->tag_index = tag_index;
+       tags->tag_map = tag_map;
+
+       return 0;
+fail:
+       kfree(tag_index);
+       return -ENOMEM;
+}
+
+static struct blk_queue_tag *__blk_queue_init_tags(struct request_queue *q,
+                                               int depth, int alloc_policy)
+{
+       struct blk_queue_tag *tags;
+
+       tags = kmalloc(sizeof(struct blk_queue_tag), GFP_ATOMIC);
+       if (!tags)
+               goto fail;
+
+       if (init_tag_map(q, tags, depth))
+               goto fail;
+
+       atomic_set(&tags->refcnt, 1);
+       tags->alloc_policy = alloc_policy;
+       tags->next_tag = 0;
+       return tags;
+fail:
+       kfree(tags);
+       return NULL;
+}
+
+/**
+ * blk_init_tags - initialize the tag info for an external tag map
+ * @depth:     the maximum queue depth supported
+ * @alloc_policy: tag allocation policy
+ **/
+struct blk_queue_tag *blk_init_tags(int depth, int alloc_policy)
+{
+       return __blk_queue_init_tags(NULL, depth, alloc_policy);
+}
+EXPORT_SYMBOL(blk_init_tags);
+
+/**
+ * blk_queue_init_tags - initialize the queue tag info
+ * @q:  the request queue for the device
+ * @depth:  the maximum queue depth supported
+ * @tags: the tag to use
+ * @alloc_policy: tag allocation policy
+ *
+ * Queue lock must be held here if the function is called to resize an
+ * existing map.
+ **/
+int blk_queue_init_tags(struct request_queue *q, int depth,
+                       struct blk_queue_tag *tags, int alloc_policy)
+{
+       int rc;
+
+       BUG_ON(tags && q->queue_tags && tags != q->queue_tags);
+
+       if (!tags && !q->queue_tags) {
+               tags = __blk_queue_init_tags(q, depth, alloc_policy);
+
+               if (!tags)
+                       return -ENOMEM;
+
+       } else if (q->queue_tags) {
+               rc = blk_queue_resize_tags(q, depth);
+               if (rc)
+                       return rc;
+               queue_flag_set(QUEUE_FLAG_QUEUED, q);
+               return 0;
+       } else
+               atomic_inc(&tags->refcnt);
+
+       /*
+        * assign it, all done
+        */
+       q->queue_tags = tags;
+       queue_flag_set_unlocked(QUEUE_FLAG_QUEUED, q);
+       INIT_LIST_HEAD(&q->tag_busy_list);
+       return 0;
+}
+EXPORT_SYMBOL(blk_queue_init_tags);
+
+/**
+ * blk_queue_resize_tags - change the queueing depth
+ * @q:  the request queue for the device
+ * @new_depth: the new max command queueing depth
+ *
+ *  Notes:
+ *    Must be called with the queue lock held.
+ **/
+int blk_queue_resize_tags(struct request_queue *q, int new_depth)
+{
+       struct blk_queue_tag *bqt = q->queue_tags;
+       struct request **tag_index;
+       unsigned long *tag_map;
+       int max_depth, nr_ulongs;
+
+       if (!bqt)
+               return -ENXIO;
+
+       /*
+        * if we already have large enough real_max_depth.  just
+        * adjust max_depth.  *NOTE* as requests with tag value
+        * between new_depth and real_max_depth can be in-flight, tag
+        * map can not be shrunk blindly here.
+        */
+       if (new_depth <= bqt->real_max_depth) {
+               bqt->max_depth = new_depth;
+               return 0;
+       }
+
+       /*
+        * Currently cannot replace a shared tag map with a new
+        * one, so error out if this is the case
+        */
+       if (atomic_read(&bqt->refcnt) != 1)
+               return -EBUSY;
+
+       /*
+        * save the old state info, so we can copy it back
+        */
+       tag_index = bqt->tag_index;
+       tag_map = bqt->tag_map;
+       max_depth = bqt->real_max_depth;
+
+       if (init_tag_map(q, bqt, new_depth))
+               return -ENOMEM;
+
+       memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *));
+       nr_ulongs = ALIGN(max_depth, BITS_PER_LONG) / BITS_PER_LONG;
+       memcpy(bqt->tag_map, tag_map, nr_ulongs * sizeof(unsigned long));
+
+       kfree(tag_index);
+       kfree(tag_map);
+       return 0;
+}
+EXPORT_SYMBOL(blk_queue_resize_tags);
+
+/**
+ * blk_queue_end_tag - end tag operations for a request
+ * @q:  the request queue for the device
+ * @rq: the request that has completed
+ *
+ *  Description:
+ *    Typically called when end_that_request_first() returns %0, meaning
+ *    all transfers have been done for a request. It's important to call
+ *    this function before end_that_request_last(), as that will put the
+ *    request back on the free list thus corrupting the internal tag list.
+ *
+ *  Notes:
+ *   queue lock must be held.
+ **/
+void blk_queue_end_tag(struct request_queue *q, struct request *rq)
+{
+       struct blk_queue_tag *bqt = q->queue_tags;
+       unsigned tag = rq->tag; /* negative tags invalid */
+
+       BUG_ON(tag >= bqt->real_max_depth);
+
+       list_del_init(&rq->queuelist);
+       rq->cmd_flags &= ~REQ_QUEUED;
+       rq->tag = -1;
+
+       if (unlikely(bqt->tag_index[tag] == NULL))
+               printk(KERN_ERR "%s: tag %d is missing\n",
+                      __func__, tag);
+
+       bqt->tag_index[tag] = NULL;
+
+       if (unlikely(!test_bit(tag, bqt->tag_map))) {
+               printk(KERN_ERR "%s: attempt to clear non-busy tag (%d)\n",
+                      __func__, 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);
+}
+EXPORT_SYMBOL(blk_queue_end_tag);
+
+/**
+ * blk_queue_start_tag - find a free tag and assign it
+ * @q:  the request queue for the device
+ * @rq:  the block request that needs tagging
+ *
+ *  Description:
+ *    This can either be used as a stand-alone helper, or possibly be
+ *    assigned as the queue &prep_rq_fn (in which case &struct request
+ *    automagically gets a tag assigned). Note that this function
+ *    assumes that any type of request can be queued! if this is not
+ *    true for your device, you must check the request type before
+ *    calling this function.  The request will also be removed from
+ *    the request queue, so it's the drivers responsibility to readd
+ *    it if it should need to be restarted for some reason.
+ *
+ *  Notes:
+ *   queue lock must be held.
+ **/
+int blk_queue_start_tag(struct request_queue *q, struct request *rq)
+{
+       struct blk_queue_tag *bqt = q->queue_tags;
+       unsigned max_depth;
+       int tag;
+
+       if (unlikely((rq->cmd_flags & REQ_QUEUED))) {
+               printk(KERN_ERR
+                      "%s: request %p for device [%s] already tagged %d",
+                      __func__, rq,
+                      rq->rq_disk ? rq->rq_disk->disk_name : "?", rq->tag);
+               BUG();
+       }
+
+       /*
+        * 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) && max_depth > 1) {
+               switch (max_depth) {
+               case 2:
+                       max_depth = 1;
+                       break;
+               case 3:
+                       max_depth = 2;
+                       break;
+               default:
+                       max_depth -= 2;
+               }
+               if (q->in_flight[BLK_RW_ASYNC] > max_depth)
+                       return 1;
+       }
+
+       do {
+               if (bqt->alloc_policy == BLK_TAG_ALLOC_FIFO) {
+                       tag = find_first_zero_bit(bqt->tag_map, max_depth);
+                       if (tag >= max_depth)
+                               return 1;
+               } else {
+                       int start = bqt->next_tag;
+                       int size = min_t(int, bqt->max_depth, max_depth + start);
+                       tag = find_next_zero_bit(bqt->tag_map, size, start);
+                       if (tag >= size && start + size > bqt->max_depth) {
+                               size = start + size - bqt->max_depth;
+                               tag = find_first_zero_bit(bqt->tag_map, size);
+                       }
+                       if (tag >= size)
+                               return 1;
+               }
+
+       } while (test_and_set_bit_lock(tag, bqt->tag_map));
+       /*
+        * We need lock ordering semantics given by test_and_set_bit_lock.
+        * See blk_queue_end_tag for details.
+        */
+
+       bqt->next_tag = (tag + 1) % bqt->max_depth;
+       rq->cmd_flags |= REQ_QUEUED;
+       rq->tag = tag;
+       bqt->tag_index[tag] = rq;
+       blk_start_request(rq);
+       list_add(&rq->queuelist, &q->tag_busy_list);
+       return 0;
+}
+EXPORT_SYMBOL(blk_queue_start_tag);
+
+/**
+ * blk_queue_invalidate_tags - invalidate all pending tags
+ * @q:  the request queue for the device
+ *
+ *  Description:
+ *   Hardware conditions may dictate a need to stop all pending requests.
+ *   In this case, we will safely clear the block side of the tag queue and
+ *   readd all requests to the request queue in the right order.
+ *
+ *  Notes:
+ *   queue lock must be held.
+ **/
+void blk_queue_invalidate_tags(struct request_queue *q)
+{
+       struct list_head *tmp, *n;
+
+       list_for_each_safe(tmp, n, &q->tag_busy_list)
+               blk_requeue_request(q, list_entry_rq(tmp));
+}
+EXPORT_SYMBOL(blk_queue_invalidate_tags);