elevator: Fix a race in elevator switching and md device initialization
[linux-3.10.git] / block / elevator.c
index 530fcfe..02d4390 100644 (file)
 #include <linux/slab.h>
 #include <linux/init.h>
 #include <linux/compiler.h>
-#include <linux/delay.h>
 #include <linux/blktrace_api.h>
-#include <trace/block.h>
 #include <linux/hash.h>
 #include <linux/uaccess.h>
+#include <linux/pm_runtime.h>
+
+#include <trace/events/block.h>
 
 #include "blk.h"
+#include "blk-cgroup.h"
 
 static DEFINE_SPINLOCK(elv_list_lock);
 static LIST_HEAD(elv_list);
@@ -45,13 +47,7 @@ static LIST_HEAD(elv_list);
 /*
  * Merge hash stuff.
  */
-static const int elv_hash_shift = 6;
-#define ELV_HASH_BLOCK(sec)    ((sec) >> 3)
-#define ELV_HASH_FN(sec)       \
-               (hash_long(ELV_HASH_BLOCK((sec)), elv_hash_shift))
-#define ELV_HASH_ENTRIES       (1 << elv_hash_shift)
-#define rq_hash_key(rq)                ((rq)->sector + (rq)->nr_sectors)
-#define ELV_ON_HASH(rq)                (!hlist_unhashed(&(rq)->hash))
+#define rq_hash_key(rq)                (blk_rq_pos(rq) + blk_rq_sectors(rq))
 
 /*
  * Query io scheduler to see if the current process issuing bio may be
@@ -60,10 +56,10 @@ static const int elv_hash_shift = 6;
 static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
 {
        struct request_queue *q = rq->q;
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
 
-       if (e->ops->elevator_allow_merge_fn)
-               return e->ops->elevator_allow_merge_fn(q, rq, bio);
+       if (e->type->ops.elevator_allow_merge_fn)
+               return e->type->ops.elevator_allow_merge_fn(q, rq, bio);
 
        return 1;
 }
@@ -71,33 +67,9 @@ static int elv_iosched_allow_merge(struct request *rq, struct bio *bio)
 /*
  * can we safely merge with this request?
  */
-int elv_rq_merge_ok(struct request *rq, struct bio *bio)
+bool elv_rq_merge_ok(struct request *rq, struct bio *bio)
 {
-       if (!rq_mergeable(rq))
-               return 0;
-
-       /*
-        * Don't merge file system requests and discard requests
-        */
-       if (bio_discard(bio) != bio_discard(rq->bio))
-               return 0;
-
-       /*
-        * different data direction or already started, don't merge
-        */
-       if (bio_data_dir(bio) != rq_data_dir(rq))
-               return 0;
-
-       /*
-        * must be same device and not a special request
-        */
-       if (rq->rq_disk != bio->bi_bdev->bd_disk || rq->special)
-               return 0;
-
-       /*
-        * only merge integrity protected bio into ditto rq
-        */
-       if (bio_integrity(bio) != blk_integrity_rq(rq))
+       if (!blk_rq_merge_ok(rq, bio))
                return 0;
 
        if (!elv_iosched_allow_merge(rq, bio))
@@ -107,23 +79,6 @@ int elv_rq_merge_ok(struct request *rq, struct bio *bio)
 }
 EXPORT_SYMBOL(elv_rq_merge_ok);
 
-static inline int elv_try_merge(struct request *__rq, struct bio *bio)
-{
-       int ret = ELEVATOR_NO_MERGE;
-
-       /*
-        * we can merge and sequence is ok, check if it's possible
-        */
-       if (elv_rq_merge_ok(__rq, bio)) {
-               if (__rq->sector + __rq->nr_sectors == bio->bi_sector)
-                       ret = ELEVATOR_BACK_MERGE;
-               else if (__rq->sector - bio_sectors(bio) == bio->bi_sector)
-                       ret = ELEVATOR_FRONT_MERGE;
-       }
-
-       return ret;
-}
-
 static struct elevator_type *elevator_find(const char *name)
 {
        struct elevator_type *e;
@@ -141,24 +96,16 @@ static void elevator_put(struct elevator_type *e)
        module_put(e->elevator_owner);
 }
 
-static struct elevator_type *elevator_get(const char *name)
+static struct elevator_type *elevator_get(const char *name, bool try_loading)
 {
        struct elevator_type *e;
 
        spin_lock(&elv_list_lock);
 
        e = elevator_find(name);
-       if (!e) {
-               char elv[ELV_NAME_MAX + strlen("-iosched")];
-
+       if (!e && try_loading) {
                spin_unlock(&elv_list_lock);
-
-               if (!strcmp(name, "anticipatory"))
-                       sprintf(elv, "as-iosched");
-               else
-                       sprintf(elv, "%s-iosched", name);
-
-               request_module("%s", elv);
+               request_module("%s-iosched", name);
                spin_lock(&elv_list_lock);
                e = elevator_find(name);
        }
@@ -171,20 +118,7 @@ static struct elevator_type *elevator_get(const char *name)
        return e;
 }
 
-static void *elevator_init_queue(struct request_queue *q,
-                                struct elevator_queue *eq)
-{
-       return eq->ops->elevator_init_fn(q);
-}
-
-static void elevator_attach(struct request_queue *q, struct elevator_queue *eq,
-                          void *data)
-{
-       q->elevator = eq;
-       eq->elevator_data = data;
-}
-
-static char chosen_elevator[16];
+static char chosen_elevator[ELV_NAME_MAX];
 
 static int __init elevator_setup(char *str)
 {
@@ -192,39 +126,43 @@ static int __init elevator_setup(char *str)
         * Be backwards-compatible with previous kernels, so users
         * won't get the wrong elevator.
         */
-       if (!strcmp(str, "as"))
-               strcpy(chosen_elevator, "anticipatory");
-       else
-               strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
+       strncpy(chosen_elevator, str, sizeof(chosen_elevator) - 1);
        return 1;
 }
 
 __setup("elevator=", elevator_setup);
 
+/* called during boot to load the elevator chosen by the elevator param */
+void __init load_default_elevator_module(void)
+{
+       struct elevator_type *e;
+
+       if (!chosen_elevator[0])
+               return;
+
+       spin_lock(&elv_list_lock);
+       e = elevator_find(chosen_elevator);
+       spin_unlock(&elv_list_lock);
+
+       if (!e)
+               request_module("%s-iosched", chosen_elevator);
+}
+
 static struct kobj_type elv_ktype;
 
-static elevator_t *elevator_alloc(struct request_queue *q,
+struct elevator_queue *elevator_alloc(struct request_queue *q,
                                  struct elevator_type *e)
 {
-       elevator_t *eq;
-       int i;
+       struct elevator_queue *eq;
 
-       eq = kmalloc_node(sizeof(elevator_t), GFP_KERNEL | __GFP_ZERO, q->node);
+       eq = kmalloc_node(sizeof(*eq), GFP_KERNEL | __GFP_ZERO, q->node);
        if (unlikely(!eq))
                goto err;
 
-       eq->ops = &e->ops;
-       eq->elevator_type = e;
+       eq->type = e;
        kobject_init(&eq->kobj, &elv_ktype);
        mutex_init(&eq->sysfs_lock);
-
-       eq->hash = kmalloc_node(sizeof(struct hlist_head) * ELV_HASH_ENTRIES,
-                                       GFP_KERNEL, q->node);
-       if (!eq->hash)
-               goto err;
-
-       for (i = 0; i < ELV_HASH_ENTRIES; i++)
-               INIT_HLIST_HEAD(&eq->hash[i]);
+       hash_init(eq->hash);
 
        return eq;
 err:
@@ -232,22 +170,30 @@ err:
        elevator_put(e);
        return NULL;
 }
+EXPORT_SYMBOL(elevator_alloc);
 
 static void elevator_release(struct kobject *kobj)
 {
-       elevator_t *e = container_of(kobj, elevator_t, kobj);
+       struct elevator_queue *e;
 
-       elevator_put(e->elevator_type);
-       kfree(e->hash);
+       e = container_of(kobj, struct elevator_queue, kobj);
+       elevator_put(e->type);
        kfree(e);
 }
 
 int elevator_init(struct request_queue *q, char *name)
 {
        struct elevator_type *e = NULL;
-       struct elevator_queue *eq;
-       int ret = 0;
-       void *data;
+       int err;
+
+       /*
+        * q->sysfs_lock must be held to provide mutual exclusion between
+        * elevator_switch() and here.
+        */
+       lockdep_assert_held(&q->sysfs_lock);
+
+       if (unlikely(q->elevator))
+               return 0;
 
        INIT_LIST_HEAD(&q->queue_head);
        q->last_merge = NULL;
@@ -255,74 +201,52 @@ int elevator_init(struct request_queue *q, char *name)
        q->boundary_rq = NULL;
 
        if (name) {
-               e = elevator_get(name);
+               e = elevator_get(name, true);
                if (!e)
                        return -EINVAL;
        }
 
+       /*
+        * Use the default elevator specified by config boot param or
+        * config option.  Don't try to load modules as we could be running
+        * off async and request_module() isn't allowed from async.
+        */
        if (!e && *chosen_elevator) {
-               e = elevator_get(chosen_elevator);
+               e = elevator_get(chosen_elevator, false);
                if (!e)
                        printk(KERN_ERR "I/O scheduler %s not found\n",
                                                        chosen_elevator);
        }
 
        if (!e) {
-               e = elevator_get(CONFIG_DEFAULT_IOSCHED);
+               e = elevator_get(CONFIG_DEFAULT_IOSCHED, false);
                if (!e) {
                        printk(KERN_ERR
                                "Default I/O scheduler not found. " \
                                "Using noop.\n");
-                       e = elevator_get("noop");
+                       e = elevator_get("noop", false);
                }
        }
 
-       eq = elevator_alloc(q, e);
-       if (!eq)
-               return -ENOMEM;
-
-       data = elevator_init_queue(q, eq);
-       if (!data) {
-               kobject_put(&eq->kobj);
-               return -ENOMEM;
-       }
-
-       elevator_attach(q, eq, data);
-       return ret;
+       err = e->ops.elevator_init_fn(q, e);
+       return 0;
 }
 EXPORT_SYMBOL(elevator_init);
 
-void elevator_exit(elevator_t *e)
+void elevator_exit(struct elevator_queue *e)
 {
        mutex_lock(&e->sysfs_lock);
-       if (e->ops->elevator_exit_fn)
-               e->ops->elevator_exit_fn(e);
-       e->ops = NULL;
+       if (e->type->ops.elevator_exit_fn)
+               e->type->ops.elevator_exit_fn(e);
        mutex_unlock(&e->sysfs_lock);
 
        kobject_put(&e->kobj);
 }
 EXPORT_SYMBOL(elevator_exit);
 
-static void elv_activate_rq(struct request_queue *q, struct request *rq)
-{
-       elevator_t *e = q->elevator;
-
-       if (e->ops->elevator_activate_req_fn)
-               e->ops->elevator_activate_req_fn(q, rq);
-}
-
-static void elv_deactivate_rq(struct request_queue *q, struct request *rq)
-{
-       elevator_t *e = q->elevator;
-
-       if (e->ops->elevator_deactivate_req_fn)
-               e->ops->elevator_deactivate_req_fn(q, rq);
-}
-
 static inline void __elv_rqhash_del(struct request *rq)
 {
-       hlist_del_init(&rq->hash);
+       hash_del(&rq->hash);
 }
 
 static void elv_rqhash_del(struct request_queue *q, struct request *rq)
@@ -333,10 +257,10 @@ static void elv_rqhash_del(struct request_queue *q, struct request *rq)
 
 static void elv_rqhash_add(struct request_queue *q, struct request *rq)
 {
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
 
        BUG_ON(ELV_ON_HASH(rq));
-       hlist_add_head(&rq->hash, &e->hash[ELV_HASH_FN(rq_hash_key(rq))]);
+       hash_add(e->hash, &rq->hash, rq_hash_key(rq));
 }
 
 static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
@@ -347,12 +271,11 @@ static void elv_rqhash_reposition(struct request_queue *q, struct request *rq)
 
 static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
 {
-       elevator_t *e = q->elevator;
-       struct hlist_head *hash_list = &e->hash[ELV_HASH_FN(offset)];
-       struct hlist_node *entry, *next;
+       struct elevator_queue *e = q->elevator;
+       struct hlist_node *next;
        struct request *rq;
 
-       hlist_for_each_entry_safe(rq, entry, next, hash_list, hash) {
+       hash_for_each_possible_safe(e->hash, rq, next, hash, offset) {
                BUG_ON(!ELV_ON_HASH(rq));
 
                if (unlikely(!rq_mergeable(rq))) {
@@ -371,7 +294,7 @@ static struct request *elv_rqhash_find(struct request_queue *q, sector_t offset)
  * RB-tree support functions for inserting/lookup/removal of requests
  * in a sorted RB tree.
  */
-struct request *elv_rb_add(struct rb_root *root, struct request *rq)
+void elv_rb_add(struct rb_root *root, struct request *rq)
 {
        struct rb_node **p = &root->rb_node;
        struct rb_node *parent = NULL;
@@ -381,17 +304,14 @@ struct request *elv_rb_add(struct rb_root *root, struct request *rq)
                parent = *p;
                __rq = rb_entry(parent, struct request, rb_node);
 
-               if (rq->sector < __rq->sector)
+               if (blk_rq_pos(rq) < blk_rq_pos(__rq))
                        p = &(*p)->rb_left;
-               else if (rq->sector > __rq->sector)
+               else if (blk_rq_pos(rq) >= blk_rq_pos(__rq))
                        p = &(*p)->rb_right;
-               else
-                       return __rq;
        }
 
        rb_link_node(&rq->rb_node, parent, p);
        rb_insert_color(&rq->rb_node, root);
-       return NULL;
 }
 EXPORT_SYMBOL(elv_rb_add);
 
@@ -411,9 +331,9 @@ struct request *elv_rb_find(struct rb_root *root, sector_t sector)
        while (n) {
                rq = rb_entry(n, struct request, rb_node);
 
-               if (sector < rq->sector)
+               if (sector < blk_rq_pos(rq))
                        n = n->rb_left;
-               else if (sector > rq->sector)
+               else if (sector > blk_rq_pos(rq))
                        n = n->rb_right;
                else
                        return rq;
@@ -442,24 +362,25 @@ void elv_dispatch_sort(struct request_queue *q, struct request *rq)
        q->nr_sorted--;
 
        boundary = q->end_sector;
-       stop_flags = REQ_SOFTBARRIER | REQ_HARDBARRIER | REQ_STARTED;
+       stop_flags = REQ_SOFTBARRIER | REQ_STARTED;
        list_for_each_prev(entry, &q->queue_head) {
                struct request *pos = list_entry_rq(entry);
 
-               if (blk_discard_rq(rq) != blk_discard_rq(pos))
+               if ((rq->cmd_flags & REQ_DISCARD) !=
+                   (pos->cmd_flags & REQ_DISCARD))
                        break;
                if (rq_data_dir(rq) != rq_data_dir(pos))
                        break;
                if (pos->cmd_flags & stop_flags)
                        break;
-               if (rq->sector >= boundary) {
-                       if (pos->sector < boundary)
+               if (blk_rq_pos(rq) >= boundary) {
+                       if (blk_rq_pos(pos) < boundary)
                                continue;
                } else {
-                       if (pos->sector >= boundary)
+                       if (blk_rq_pos(pos) >= boundary)
                                break;
                }
-               if (rq->sector >= pos->sector)
+               if (blk_rq_pos(rq) >= blk_rq_pos(pos))
                        break;
        }
 
@@ -489,22 +410,31 @@ EXPORT_SYMBOL(elv_dispatch_add_tail);
 
 int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
 {
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
        struct request *__rq;
        int ret;
 
        /*
+        * Levels of merges:
+        *      nomerges:  No merges at all attempted
+        *      noxmerges: Only simple one-hit cache try
+        *      merges:    All merge tries attempted
+        */
+       if (blk_queue_nomerges(q))
+               return ELEVATOR_NO_MERGE;
+
+       /*
         * First try one-hit cache.
         */
-       if (q->last_merge) {
-               ret = elv_try_merge(q->last_merge, bio);
+       if (q->last_merge && elv_rq_merge_ok(q->last_merge, bio)) {
+               ret = blk_try_merge(q->last_merge, bio);
                if (ret != ELEVATOR_NO_MERGE) {
                        *req = q->last_merge;
                        return ret;
                }
        }
 
-       if (blk_queue_nomerges(q))
+       if (blk_queue_noxmerges(q))
                return ELEVATOR_NO_MERGE;
 
        /*
@@ -516,18 +446,60 @@ int elv_merge(struct request_queue *q, struct request **req, struct bio *bio)
                return ELEVATOR_BACK_MERGE;
        }
 
-       if (e->ops->elevator_merge_fn)
-               return e->ops->elevator_merge_fn(q, req, bio);
+       if (e->type->ops.elevator_merge_fn)
+               return e->type->ops.elevator_merge_fn(q, req, bio);
 
        return ELEVATOR_NO_MERGE;
 }
 
+/*
+ * Attempt to do an insertion back merge. Only check for the case where
+ * we can append 'rq' to an existing request, so we can throw 'rq' away
+ * afterwards.
+ *
+ * Returns true if we merged, false otherwise
+ */
+static bool elv_attempt_insert_merge(struct request_queue *q,
+                                    struct request *rq)
+{
+       struct request *__rq;
+       bool ret;
+
+       if (blk_queue_nomerges(q))
+               return false;
+
+       /*
+        * First try one-hit cache.
+        */
+       if (q->last_merge && blk_attempt_req_merge(q, q->last_merge, rq))
+               return true;
+
+       if (blk_queue_noxmerges(q))
+               return false;
+
+       ret = false;
+       /*
+        * See if our hash lookup can find a potential backmerge.
+        */
+       while (1) {
+               __rq = elv_rqhash_find(q, blk_rq_pos(rq));
+               if (!__rq || !blk_attempt_req_merge(q, __rq, rq))
+                       break;
+
+               /* The merged request could be merged with others, try again */
+               ret = true;
+               rq = __rq;
+       }
+
+       return ret;
+}
+
 void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 {
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
 
-       if (e->ops->elevator_merged_fn)
-               e->ops->elevator_merged_fn(q, rq, type);
+       if (e->type->ops.elevator_merged_fn)
+               e->type->ops.elevator_merged_fn(q, rq, type);
 
        if (type == ELEVATOR_BACK_MERGE)
                elv_rqhash_reposition(q, rq);
@@ -538,18 +510,52 @@ void elv_merged_request(struct request_queue *q, struct request *rq, int type)
 void elv_merge_requests(struct request_queue *q, struct request *rq,
                             struct request *next)
 {
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
+       const int next_sorted = next->cmd_flags & REQ_SORTED;
 
-       if (e->ops->elevator_merge_req_fn)
-               e->ops->elevator_merge_req_fn(q, rq, next);
+       if (next_sorted && e->type->ops.elevator_merge_req_fn)
+               e->type->ops.elevator_merge_req_fn(q, rq, next);
 
        elv_rqhash_reposition(q, rq);
-       elv_rqhash_del(q, next);
 
-       q->nr_sorted--;
+       if (next_sorted) {
+               elv_rqhash_del(q, next);
+               q->nr_sorted--;
+       }
+
        q->last_merge = rq;
 }
 
+void elv_bio_merged(struct request_queue *q, struct request *rq,
+                       struct bio *bio)
+{
+       struct elevator_queue *e = q->elevator;
+
+       if (e->type->ops.elevator_bio_merged_fn)
+               e->type->ops.elevator_bio_merged_fn(q, rq, bio);
+}
+
+#ifdef CONFIG_PM_RUNTIME
+static void blk_pm_requeue_request(struct request *rq)
+{
+       if (rq->q->dev && !(rq->cmd_flags & REQ_PM))
+               rq->q->nr_pending--;
+}
+
+static void blk_pm_add_request(struct request_queue *q, struct request *rq)
+{
+       if (q->dev && !(rq->cmd_flags & REQ_PM) && q->nr_pending++ == 0 &&
+           (q->rpm_status == RPM_SUSPENDED || q->rpm_status == RPM_SUSPENDING))
+               pm_request_resume(q->dev);
+}
+#else
+static inline void blk_pm_requeue_request(struct request *rq) {}
+static inline void blk_pm_add_request(struct request_queue *q,
+                                     struct request *rq)
+{
+}
+#endif
+
 void elv_requeue_request(struct request_queue *q, struct request *rq)
 {
        /*
@@ -557,44 +563,56 @@ void elv_requeue_request(struct request_queue *q, struct request *rq)
         * in_flight count again
         */
        if (blk_account_rq(rq)) {
-               q->in_flight--;
-               if (blk_sorted_rq(rq))
+               q->in_flight[rq_is_sync(rq)]--;
+               if (rq->cmd_flags & REQ_SORTED)
                        elv_deactivate_rq(q, rq);
        }
 
        rq->cmd_flags &= ~REQ_STARTED;
 
-       elv_insert(q, rq, ELEVATOR_INSERT_REQUEUE);
+       blk_pm_requeue_request(rq);
+
+       __elv_add_request(q, rq, ELEVATOR_INSERT_REQUEUE);
 }
 
-static void elv_drain_elevator(struct request_queue *q)
+void elv_drain_elevator(struct request_queue *q)
 {
        static int printed;
-       while (q->elevator->ops->elevator_dispatch_fn(q, 1))
+
+       lockdep_assert_held(q->queue_lock);
+
+       while (q->elevator->type->ops.elevator_dispatch_fn(q, 1))
                ;
-       if (q->nr_sorted == 0)
-               return;
-       if (printed++ < 10) {
+       if (q->nr_sorted && printed++ < 10) {
                printk(KERN_ERR "%s: forced dispatching is broken "
                       "(nr_sorted=%u), please report this\n",
-                      q->elevator->elevator_type->elevator_name, q->nr_sorted);
+                      q->elevator->type->elevator_name, q->nr_sorted);
        }
 }
 
-void elv_insert(struct request_queue *q, struct request *rq, int where)
+void __elv_add_request(struct request_queue *q, struct request *rq, int where)
 {
-       struct list_head *pos;
-       unsigned ordseq;
-       int unplug_it = 1;
-
        trace_block_rq_insert(q, rq);
 
+       blk_pm_add_request(q, rq);
+
        rq->q = q;
 
+       if (rq->cmd_flags & REQ_SOFTBARRIER) {
+               /* barriers are scheduling boundary, update end_sector */
+               if (rq->cmd_type == REQ_TYPE_FS) {
+                       q->end_sector = rq_end_sector(rq);
+                       q->boundary_rq = rq;
+               }
+       } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
+                   (where == ELEVATOR_INSERT_SORT ||
+                    where == ELEVATOR_INSERT_SORT_MERGE))
+               where = ELEVATOR_INSERT_BACK;
+
        switch (where) {
+       case ELEVATOR_INSERT_REQUEUE:
        case ELEVATOR_INSERT_FRONT:
                rq->cmd_flags |= REQ_SOFTBARRIER;
-
                list_add(&rq->queuelist, &q->queue_head);
                break;
 
@@ -612,12 +630,19 @@ void elv_insert(struct request_queue *q, struct request *rq, int where)
                 *   with anything.  There's no point in delaying queue
                 *   processing.
                 */
-               blk_remove_plug(q);
-               blk_start_queueing(q);
+               __blk_run_queue(q);
                break;
 
+       case ELEVATOR_INSERT_SORT_MERGE:
+               /*
+                * If we succeed in merging this request with one in the
+                * queue already, we are done - rq has now been freed,
+                * so no need to do anything further.
+                */
+               if (elv_attempt_insert_merge(q, rq))
+                       break;
        case ELEVATOR_INSERT_SORT:
-               BUG_ON(!blk_fs_request(rq) && !blk_discard_rq(rq));
+               BUG_ON(rq->cmd_type != REQ_TYPE_FS);
                rq->cmd_flags |= REQ_SORTED;
                q->nr_sorted++;
                if (rq_mergeable(rq)) {
@@ -631,286 +656,73 @@ void elv_insert(struct request_queue *q, struct request *rq, int where)
                 * rq cannot be accessed after calling
                 * elevator_add_req_fn.
                 */
-               q->elevator->ops->elevator_add_req_fn(q, rq);
+               q->elevator->type->ops.elevator_add_req_fn(q, rq);
                break;
 
-       case ELEVATOR_INSERT_REQUEUE:
-               /*
-                * If ordered flush isn't in progress, we do front
-                * insertion; otherwise, requests should be requeued
-                * in ordseq order.
-                */
+       case ELEVATOR_INSERT_FLUSH:
                rq->cmd_flags |= REQ_SOFTBARRIER;
-
-               /*
-                * Most requeues happen because of a busy condition,
-                * don't force unplug of the queue for that case.
-                */
-               unplug_it = 0;
-
-               if (q->ordseq == 0) {
-                       list_add(&rq->queuelist, &q->queue_head);
-                       break;
-               }
-
-               ordseq = blk_ordered_req_seq(rq);
-
-               list_for_each(pos, &q->queue_head) {
-                       struct request *pos_rq = list_entry_rq(pos);
-                       if (ordseq <= blk_ordered_req_seq(pos_rq))
-                               break;
-               }
-
-               list_add_tail(&rq->queuelist, pos);
+               blk_insert_flush(rq);
                break;
-
        default:
                printk(KERN_ERR "%s: bad insertion point %d\n",
                       __func__, where);
                BUG();
        }
-
-       if (unplug_it && blk_queue_plugged(q)) {
-               int nrq = q->rq.count[READ] + q->rq.count[WRITE]
-                       - q->in_flight;
-
-               if (nrq >= q->unplug_thresh)
-                       __generic_unplug_device(q);
-       }
-}
-
-void __elv_add_request(struct request_queue *q, struct request *rq, int where,
-                      int plug)
-{
-       if (q->ordcolor)
-               rq->cmd_flags |= REQ_ORDERED_COLOR;
-
-       if (rq->cmd_flags & (REQ_SOFTBARRIER | REQ_HARDBARRIER)) {
-               /*
-                * toggle ordered color
-                */
-               if (blk_barrier_rq(rq))
-                       q->ordcolor ^= 1;
-
-               /*
-                * barriers implicitly indicate back insertion
-                */
-               if (where == ELEVATOR_INSERT_SORT)
-                       where = ELEVATOR_INSERT_BACK;
-
-               /*
-                * this request is scheduling boundary, update
-                * end_sector
-                */
-               if (blk_fs_request(rq) || blk_discard_rq(rq)) {
-                       q->end_sector = rq_end_sector(rq);
-                       q->boundary_rq = rq;
-               }
-       } else if (!(rq->cmd_flags & REQ_ELVPRIV) &&
-                   where == ELEVATOR_INSERT_SORT)
-               where = ELEVATOR_INSERT_BACK;
-
-       if (plug)
-               blk_plug_device(q);
-
-       elv_insert(q, rq, where);
 }
 EXPORT_SYMBOL(__elv_add_request);
 
-void elv_add_request(struct request_queue *q, struct request *rq, int where,
-                    int plug)
+void elv_add_request(struct request_queue *q, struct request *rq, int where)
 {
        unsigned long flags;
 
        spin_lock_irqsave(q->queue_lock, flags);
-       __elv_add_request(q, rq, where, plug);
+       __elv_add_request(q, rq, where);
        spin_unlock_irqrestore(q->queue_lock, flags);
 }
 EXPORT_SYMBOL(elv_add_request);
 
-static inline struct request *__elv_next_request(struct request_queue *q)
-{
-       struct request *rq;
-
-       while (1) {
-               while (!list_empty(&q->queue_head)) {
-                       rq = list_entry_rq(q->queue_head.next);
-                       if (blk_do_ordered(q, &rq))
-                               return rq;
-               }
-
-               if (!q->elevator->ops->elevator_dispatch_fn(q, 0))
-                       return NULL;
-       }
-}
-
-struct request *elv_next_request(struct request_queue *q)
-{
-       struct request *rq;
-       int ret;
-
-       while ((rq = __elv_next_request(q)) != NULL) {
-               /*
-                * Kill the empty barrier place holder, the driver must
-                * not ever see it.
-                */
-               if (blk_empty_barrier(rq)) {
-                       __blk_end_request(rq, 0, blk_rq_bytes(rq));
-                       continue;
-               }
-               if (!(rq->cmd_flags & REQ_STARTED)) {
-                       /*
-                        * This is the first time the device driver
-                        * sees this request (possibly after
-                        * requeueing).  Notify IO scheduler.
-                        */
-                       if (blk_sorted_rq(rq))
-                               elv_activate_rq(q, rq);
-
-                       /*
-                        * just mark as started even if we don't start
-                        * it, a request that has been delayed should
-                        * not be passed by new incoming requests
-                        */
-                       rq->cmd_flags |= REQ_STARTED;
-                       trace_block_rq_issue(q, rq);
-               }
-
-               if (!q->boundary_rq || q->boundary_rq == rq) {
-                       q->end_sector = rq_end_sector(rq);
-                       q->boundary_rq = NULL;
-               }
-
-               if (rq->cmd_flags & REQ_DONTPREP)
-                       break;
-
-               if (q->dma_drain_size && rq->data_len) {
-                       /*
-                        * make sure space for the drain appears we
-                        * know we can do this because max_hw_segments
-                        * has been adjusted to be one fewer than the
-                        * device can handle
-                        */
-                       rq->nr_phys_segments++;
-               }
-
-               if (!q->prep_rq_fn)
-                       break;
-
-               ret = q->prep_rq_fn(q, rq);
-               if (ret == BLKPREP_OK) {
-                       break;
-               } else if (ret == BLKPREP_DEFER) {
-                       /*
-                        * the request may have been (partially) prepped.
-                        * we need to keep this request in the front to
-                        * avoid resource deadlock.  REQ_STARTED will
-                        * prevent other fs requests from passing this one.
-                        */
-                       if (q->dma_drain_size && rq->data_len &&
-                           !(rq->cmd_flags & REQ_DONTPREP)) {
-                               /*
-                                * remove the space for the drain we added
-                                * so that we don't add it again
-                                */
-                               --rq->nr_phys_segments;
-                       }
-
-                       rq = NULL;
-                       break;
-               } else if (ret == BLKPREP_KILL) {
-                       rq->cmd_flags |= REQ_QUIET;
-                       __blk_end_request(rq, -EIO, blk_rq_bytes(rq));
-               } else {
-                       printk(KERN_ERR "%s: bad return=%d\n", __func__, ret);
-                       break;
-               }
-       }
-
-       return rq;
-}
-EXPORT_SYMBOL(elv_next_request);
-
-void elv_dequeue_request(struct request_queue *q, struct request *rq)
-{
-       BUG_ON(list_empty(&rq->queuelist));
-       BUG_ON(ELV_ON_HASH(rq));
-
-       list_del_init(&rq->queuelist);
-
-       /*
-        * the time frame between a request being removed from the lists
-        * and to it is freed is accounted as io that is in progress at
-        * the driver side.
-        */
-       if (blk_account_rq(rq))
-               q->in_flight++;
-
-       /*
-        * We are now handing the request to the hardware, add the
-        * timeout handler.
-        */
-       blk_add_timer(rq);
-}
-EXPORT_SYMBOL(elv_dequeue_request);
-
-int elv_queue_empty(struct request_queue *q)
-{
-       elevator_t *e = q->elevator;
-
-       if (!list_empty(&q->queue_head))
-               return 0;
-
-       if (e->ops->elevator_queue_empty_fn)
-               return e->ops->elevator_queue_empty_fn(q);
-
-       return 1;
-}
-EXPORT_SYMBOL(elv_queue_empty);
-
 struct request *elv_latter_request(struct request_queue *q, struct request *rq)
 {
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
 
-       if (e->ops->elevator_latter_req_fn)
-               return e->ops->elevator_latter_req_fn(q, rq);
+       if (e->type->ops.elevator_latter_req_fn)
+               return e->type->ops.elevator_latter_req_fn(q, rq);
        return NULL;
 }
 
 struct request *elv_former_request(struct request_queue *q, struct request *rq)
 {
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
 
-       if (e->ops->elevator_former_req_fn)
-               return e->ops->elevator_former_req_fn(q, rq);
+       if (e->type->ops.elevator_former_req_fn)
+               return e->type->ops.elevator_former_req_fn(q, rq);
        return NULL;
 }
 
-int elv_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
+int elv_set_request(struct request_queue *q, struct request *rq,
+                   struct bio *bio, gfp_t gfp_mask)
 {
-       elevator_t *e = q->elevator;
-
-       if (e->ops->elevator_set_req_fn)
-               return e->ops->elevator_set_req_fn(q, rq, gfp_mask);
+       struct elevator_queue *e = q->elevator;
 
-       rq->elevator_private = NULL;
+       if (e->type->ops.elevator_set_req_fn)
+               return e->type->ops.elevator_set_req_fn(q, rq, bio, gfp_mask);
        return 0;
 }
 
 void elv_put_request(struct request_queue *q, struct request *rq)
 {
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
 
-       if (e->ops->elevator_put_req_fn)
-               e->ops->elevator_put_req_fn(rq);
+       if (e->type->ops.elevator_put_req_fn)
+               e->type->ops.elevator_put_req_fn(rq);
 }
 
 int elv_may_queue(struct request_queue *q, int rw)
 {
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
 
-       if (e->ops->elevator_may_queue_fn)
-               return e->ops->elevator_may_queue_fn(q, rw);
+       if (e->type->ops.elevator_may_queue_fn)
+               return e->type->ops.elevator_may_queue_fn(q, rw);
 
        return ELV_MQUEUE_MAY;
 }
@@ -919,40 +731,34 @@ void elv_abort_queue(struct request_queue *q)
 {
        struct request *rq;
 
+       blk_abort_flushes(q);
+
        while (!list_empty(&q->queue_head)) {
                rq = list_entry_rq(q->queue_head.next);
                rq->cmd_flags |= REQ_QUIET;
                trace_block_rq_abort(q, rq);
-               __blk_end_request(rq, -EIO, blk_rq_bytes(rq));
+               /*
+                * Mark this request as started so we don't trigger
+                * any debug logic in the end I/O path.
+                */
+               blk_start_request(rq);
+               __blk_end_request_all(rq, -EIO);
        }
 }
 EXPORT_SYMBOL(elv_abort_queue);
 
 void elv_completed_request(struct request_queue *q, struct request *rq)
 {
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
 
        /*
         * request is released from the driver, io must be done
         */
        if (blk_account_rq(rq)) {
-               q->in_flight--;
-               if (blk_sorted_rq(rq) && e->ops->elevator_completed_req_fn)
-                       e->ops->elevator_completed_req_fn(q, rq);
-       }
-
-       /*
-        * Check if the queue is waiting for fs requests to be
-        * drained for flush sequence.
-        */
-       if (unlikely(q->ordseq)) {
-               struct request *first_rq = list_entry_rq(q->queue_head.next);
-               if (q->in_flight == 0 &&
-                   blk_ordered_cur_seq(q) == QUEUE_ORDSEQ_DRAIN &&
-                   blk_ordered_req_seq(first_rq) > QUEUE_ORDSEQ_DRAIN) {
-                       blk_ordered_complete_seq(q, QUEUE_ORDSEQ_DRAIN, 0);
-                       blk_start_queueing(q);
-               }
+               q->in_flight[rq_is_sync(rq)]--;
+               if ((rq->cmd_flags & REQ_SORTED) &&
+                   e->type->ops.elevator_completed_req_fn)
+                       e->type->ops.elevator_completed_req_fn(q, rq);
        }
 }
 
@@ -961,15 +767,16 @@ void elv_completed_request(struct request_queue *q, struct request *rq)
 static ssize_t
 elv_attr_show(struct kobject *kobj, struct attribute *attr, char *page)
 {
-       elevator_t *e = container_of(kobj, elevator_t, kobj);
        struct elv_fs_entry *entry = to_elv(attr);
+       struct elevator_queue *e;
        ssize_t error;
 
        if (!entry->show)
                return -EIO;
 
+       e = container_of(kobj, struct elevator_queue, kobj);
        mutex_lock(&e->sysfs_lock);
-       error = e->ops ? entry->show(e, page) : -ENOENT;
+       error = e->type ? entry->show(e, page) : -ENOENT;
        mutex_unlock(&e->sysfs_lock);
        return error;
 }
@@ -978,20 +785,21 @@ static ssize_t
 elv_attr_store(struct kobject *kobj, struct attribute *attr,
               const char *page, size_t length)
 {
-       elevator_t *e = container_of(kobj, elevator_t, kobj);
        struct elv_fs_entry *entry = to_elv(attr);
+       struct elevator_queue *e;
        ssize_t error;
 
        if (!entry->store)
                return -EIO;
 
+       e = container_of(kobj, struct elevator_queue, kobj);
        mutex_lock(&e->sysfs_lock);
-       error = e->ops ? entry->store(e, page, length) : -ENOENT;
+       error = e->type ? entry->store(e, page, length) : -ENOENT;
        mutex_unlock(&e->sysfs_lock);
        return error;
 }
 
-static struct sysfs_ops elv_sysfs_ops = {
+static const struct sysfs_ops elv_sysfs_ops = {
        .show   = elv_attr_show,
        .store  = elv_attr_store,
 };
@@ -1003,12 +811,12 @@ static struct kobj_type elv_ktype = {
 
 int elv_register_queue(struct request_queue *q)
 {
-       elevator_t *e = q->elevator;
+       struct elevator_queue *e = q->elevator;
        int error;
 
        error = kobject_add(&e->kobj, &q->kobj, "%s", "iosched");
        if (!error) {
-               struct elv_fs_entry *attr = e->elevator_type->elevator_attrs;
+               struct elv_fs_entry *attr = e->type->elevator_attrs;
                if (attr) {
                        while (attr->attr.name) {
                                if (sysfs_create_file(&e->kobj, &attr->attr))
@@ -1017,31 +825,54 @@ int elv_register_queue(struct request_queue *q)
                        }
                }
                kobject_uevent(&e->kobj, KOBJ_ADD);
+               e->registered = 1;
        }
        return error;
 }
-
-static void __elv_unregister_queue(elevator_t *e)
-{
-       kobject_uevent(&e->kobj, KOBJ_REMOVE);
-       kobject_del(&e->kobj);
-}
+EXPORT_SYMBOL(elv_register_queue);
 
 void elv_unregister_queue(struct request_queue *q)
 {
-       if (q)
-               __elv_unregister_queue(q->elevator);
+       if (q) {
+               struct elevator_queue *e = q->elevator;
+
+               kobject_uevent(&e->kobj, KOBJ_REMOVE);
+               kobject_del(&e->kobj);
+               e->registered = 0;
+       }
 }
+EXPORT_SYMBOL(elv_unregister_queue);
 
-void elv_register(struct elevator_type *e)
+int elv_register(struct elevator_type *e)
 {
        char *def = "";
 
+       /* create icq_cache if requested */
+       if (e->icq_size) {
+               if (WARN_ON(e->icq_size < sizeof(struct io_cq)) ||
+                   WARN_ON(e->icq_align < __alignof__(struct io_cq)))
+                       return -EINVAL;
+
+               snprintf(e->icq_cache_name, sizeof(e->icq_cache_name),
+                        "%s_io_cq", e->elevator_name);
+               e->icq_cache = kmem_cache_create(e->icq_cache_name, e->icq_size,
+                                                e->icq_align, 0, NULL);
+               if (!e->icq_cache)
+                       return -ENOMEM;
+       }
+
+       /* register, don't allow duplicate names */
        spin_lock(&elv_list_lock);
-       BUG_ON(elevator_find(e->elevator_name));
+       if (elevator_find(e->elevator_name)) {
+               spin_unlock(&elv_list_lock);
+               if (e->icq_cache)
+                       kmem_cache_destroy(e->icq_cache);
+               return -EBUSY;
+       }
        list_add_tail(&e->list, &elv_list);
        spin_unlock(&elv_list_lock);
 
+       /* print pretty message */
        if (!strcmp(e->elevator_name, chosen_elevator) ||
                        (!*chosen_elevator &&
                         !strcmp(e->elevator_name, CONFIG_DEFAULT_IOSCHED)))
@@ -1049,30 +880,26 @@ void elv_register(struct elevator_type *e)
 
        printk(KERN_INFO "io scheduler %s registered%s\n", e->elevator_name,
                                                                def);
+       return 0;
 }
 EXPORT_SYMBOL_GPL(elv_register);
 
 void elv_unregister(struct elevator_type *e)
 {
-       struct task_struct *g, *p;
+       /* unregister */
+       spin_lock(&elv_list_lock);
+       list_del_init(&e->list);
+       spin_unlock(&elv_list_lock);
 
        /*
-        * Iterate every thread in the process to remove the io contexts.
+        * Destroy icq_cache if it exists.  icq's are RCU managed.  Make
+        * sure all RCU operations are complete before proceeding.
         */
-       if (e->ops.trim) {
-               read_lock(&tasklist_lock);
-               do_each_thread(g, p) {
-                       task_lock(p);
-                       if (p->io_context)
-                               e->ops.trim(p->io_context);
-                       task_unlock(p);
-               } while_each_thread(g, p);
-               read_unlock(&tasklist_lock);
+       if (e->icq_cache) {
+               rcu_barrier();
+               kmem_cache_destroy(e->icq_cache);
+               e->icq_cache = NULL;
        }
-
-       spin_lock(&elv_list_lock);
-       list_del_init(&e->list);
-       spin_unlock(&elv_list_lock);
 }
 EXPORT_SYMBOL_GPL(elv_unregister);
 
@@ -1084,117 +911,112 @@ EXPORT_SYMBOL_GPL(elv_unregister);
  */
 static int elevator_switch(struct request_queue *q, struct elevator_type *new_e)
 {
-       elevator_t *old_elevator, *e;
-       void *data;
+       struct elevator_queue *old = q->elevator;
+       bool registered = old->registered;
+       int err;
 
        /*
-        * Allocate new elevator
+        * Turn on BYPASS and drain all requests w/ elevator private data.
+        * Block layer doesn't call into a quiesced elevator - all requests
+        * are directly put on the dispatch list without elevator data
+        * using INSERT_BACK.  All requests have SOFTBARRIER set and no
+        * merge happens either.
         */
-       e = elevator_alloc(q, new_e);
-       if (!e)
-               return 0;
+       blk_queue_bypass_start(q);
 
-       data = elevator_init_queue(q, e);
-       if (!data) {
-               kobject_put(&e->kobj);
-               return 0;
-       }
+       /* unregister and clear all auxiliary data of the old elevator */
+       if (registered)
+               elv_unregister_queue(q);
 
-       /*
-        * Turn on BYPASS and drain all requests w/ elevator private data
-        */
        spin_lock_irq(q->queue_lock);
-
-       queue_flag_set(QUEUE_FLAG_ELVSWITCH, q);
-
-       elv_drain_elevator(q);
-
-       while (q->rq.elvpriv) {
-               blk_start_queueing(q);
-               spin_unlock_irq(q->queue_lock);
-               msleep(10);
-               spin_lock_irq(q->queue_lock);
-               elv_drain_elevator(q);
-       }
-
-       /*
-        * Remember old elevator.
-        */
-       old_elevator = q->elevator;
-
-       /*
-        * attach and start new elevator
-        */
-       elevator_attach(q, e, data);
-
+       ioc_clear_queue(q);
        spin_unlock_irq(q->queue_lock);
 
-       __elv_unregister_queue(old_elevator);
+       /* allocate, init and register new elevator */
+       err = new_e->ops.elevator_init_fn(q, new_e);
+       if (err)
+               goto fail_init;
 
-       if (elv_register_queue(q))
-               goto fail_register;
+       if (registered) {
+               err = elv_register_queue(q);
+               if (err)
+                       goto fail_register;
+       }
 
-       /*
-        * finally exit old elevator and turn off BYPASS.
-        */
-       elevator_exit(old_elevator);
-       spin_lock_irq(q->queue_lock);
-       queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
-       spin_unlock_irq(q->queue_lock);
+       /* done, kill the old one and finish */
+       elevator_exit(old);
+       blk_queue_bypass_end(q);
 
-       blk_add_trace_msg(q, "elv switch: %s", e->elevator_type->elevator_name);
+       blk_add_trace_msg(q, "elv switch: %s", new_e->elevator_name);
 
-       return 1;
+       return 0;
 
 fail_register:
-       /*
-        * switch failed, exit the new io scheduler and reattach the old
-        * one again (along with re-adding the sysfs dir)
-        */
-       elevator_exit(e);
-       q->elevator = old_elevator;
+       elevator_exit(q->elevator);
+fail_init:
+       /* switch failed, restore and re-register old elevator */
+       q->elevator = old;
        elv_register_queue(q);
+       blk_queue_bypass_end(q);
 
-       spin_lock_irq(q->queue_lock);
-       queue_flag_clear(QUEUE_FLAG_ELVSWITCH, q);
-       spin_unlock_irq(q->queue_lock);
-
-       return 0;
+       return err;
 }
 
-ssize_t elv_iosched_store(struct request_queue *q, const char *name,
-                         size_t count)
+/*
+ * Switch this queue to the given IO scheduler.
+ */
+int elevator_change(struct request_queue *q, const char *name)
 {
        char elevator_name[ELV_NAME_MAX];
        struct elevator_type *e;
 
-       strlcpy(elevator_name, name, sizeof(elevator_name));
-       strstrip(elevator_name);
+       if (!q->elevator)
+               return -ENXIO;
 
-       e = elevator_get(elevator_name);
+       strlcpy(elevator_name, name, sizeof(elevator_name));
+       e = elevator_get(strstrip(elevator_name), true);
        if (!e) {
                printk(KERN_ERR "elevator: type %s not found\n", elevator_name);
                return -EINVAL;
        }
 
-       if (!strcmp(elevator_name, q->elevator->elevator_type->elevator_name)) {
+       if (!strcmp(elevator_name, q->elevator->type->elevator_name)) {
                elevator_put(e);
-               return count;
+               return 0;
        }
 
-       if (!elevator_switch(q, e))
-               printk(KERN_ERR "elevator: switch to %s failed\n",
-                                                       elevator_name);
-       return count;
+       return elevator_switch(q, e);
+}
+EXPORT_SYMBOL(elevator_change);
+
+ssize_t elv_iosched_store(struct request_queue *q, const char *name,
+                         size_t count)
+{
+       int ret;
+
+       if (!q->elevator)
+               return count;
+
+       ret = elevator_change(q, name);
+       if (!ret)
+               return count;
+
+       printk(KERN_ERR "elevator: switch to %s failed\n", name);
+       return ret;
 }
 
 ssize_t elv_iosched_show(struct request_queue *q, char *name)
 {
-       elevator_t *e = q->elevator;
-       struct elevator_type *elv = e->elevator_type;
+       struct elevator_queue *e = q->elevator;
+       struct elevator_type *elv;
        struct elevator_type *__e;
        int len = 0;
 
+       if (!q->elevator || !blk_queue_stackable(q))
+               return sprintf(name, "none\n");
+
+       elv = e->type;
+
        spin_lock(&elv_list_lock);
        list_for_each_entry(__e, &elv_list, list) {
                if (!strcmp(elv->elevator_name, __e->elevator_name))