]> git.proxmox.com Git - mirror_qemu.git/blobdiff - block/throttle-groups.c
Merge remote-tracking branch 'remotes/berrange/tags/pull-qio-2017-04-04-1' into staging
[mirror_qemu.git] / block / throttle-groups.c
index efc462fbc556678118760e47b4ec29e92049a48f..b73e7a800b7140cac356e23457375147c2582a50 100644 (file)
  * along with this program; if not, see <http://www.gnu.org/licenses/>.
  */
 
+#include "qemu/osdep.h"
+#include "sysemu/block-backend.h"
 #include "block/throttle-groups.h"
 #include "qemu/queue.h"
 #include "qemu/thread.h"
 #include "sysemu/qtest.h"
 
 /* The ThrottleGroup structure (with its ThrottleState) is shared
- * among different BlockDriverState and it's independent from
+ * among different BlockBackends and it's independent from
  * AioContext, so in order to use it from different threads it needs
  * its own locking.
  *
  * This locking is however handled internally in this file, so it's
- * mostly transparent to outside users (but see the documentation in
- * throttle_groups_lock()).
+ * transparent to outside users.
  *
  * The whole ThrottleGroup structure is private and invisible to
  * outside users, that only use it through its ThrottleState.
  *
- * In addition to the ThrottleGroup structure, BlockDriverState has
+ * In addition to the ThrottleGroup structure, BlockBackendPublic has
  * fields that need to be accessed by other members of the group and
- * therefore also need to be protected by this lock. Once a BDS is
- * registered in a group those fields can be accessed by other threads
- * any time.
+ * therefore also need to be protected by this lock. Once a
+ * BlockBackend is registered in a group those fields can be accessed
+ * by other threads any time.
  *
  * Again, all this is handled internally and is mostly transparent to
  * the outside. The 'throttle_timers' field however has an additional
  * constraint because it may be temporarily invalid (see for example
  * bdrv_set_aio_context()). Therefore in this file a thread will
- * access some other BDS's timers only after verifying that that BDS
- * has throttled requests in the queue.
+ * access some other BlockBackend's timers only after verifying that
+ * that BlockBackend has throttled requests in the queue.
  */
 typedef struct ThrottleGroup {
     char *name; /* This is constant during the lifetime of the group */
 
     QemuMutex lock; /* This lock protects the following four fields */
     ThrottleState ts;
-    QLIST_HEAD(, BlockDriverState) head;
-    BlockDriverState *tokens[2];
+    QLIST_HEAD(, BlockBackendPublic) head;
+    BlockBackend *tokens[2];
     bool any_timer_armed[2];
 
     /* These two are protected by the global throttle_groups_lock */
@@ -76,9 +77,9 @@ static QTAILQ_HEAD(, ThrottleGroup) throttle_groups =
  * created.
  *
  * @name: the name of the ThrottleGroup
- * @ret:  the ThrottleGroup
+ * @ret:  the ThrottleState member of the ThrottleGroup
  */
-static ThrottleGroup *throttle_group_incref(const char *name)
+ThrottleState *throttle_group_incref(const char *name)
 {
     ThrottleGroup *tg = NULL;
     ThrottleGroup *iter;
@@ -108,7 +109,7 @@ static ThrottleGroup *throttle_group_incref(const char *name)
 
     qemu_mutex_unlock(&throttle_groups_lock);
 
-    return tg;
+    return &tg->ts;
 }
 
 /* Decrease the reference count of a ThrottleGroup.
@@ -116,10 +117,12 @@ static ThrottleGroup *throttle_group_incref(const char *name)
  * When the reference count reaches zero the ThrottleGroup is
  * destroyed.
  *
- * @tg:  The ThrottleGroup to unref
+ * @ts:  The ThrottleGroup to unref, given by its ThrottleState member
  */
-static void throttle_group_unref(ThrottleGroup *tg)
+void throttle_group_unref(ThrottleState *ts)
 {
+    ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
+
     qemu_mutex_lock(&throttle_groups_lock);
     if (--tg->refcount == 0) {
         QTAILQ_REMOVE(&throttle_groups, tg, list);
@@ -130,93 +133,117 @@ static void throttle_group_unref(ThrottleGroup *tg)
     qemu_mutex_unlock(&throttle_groups_lock);
 }
 
-/* Get the name from a BlockDriverState's ThrottleGroup. The name (and
- * the pointer) is guaranteed to remain constant during the lifetime
- * of the group.
+/* Get the name from a BlockBackend's ThrottleGroup. The name (and the pointer)
+ * is guaranteed to remain constant during the lifetime of the group.
  *
- * @bs:   a BlockDriverState that is member of a throttling group
+ * @blk:  a BlockBackend that is member of a throttling group
  * @ret:  the name of the group.
  */
-const char *throttle_group_get_name(BlockDriverState *bs)
+const char *throttle_group_get_name(BlockBackend *blk)
 {
-    ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
     return tg->name;
 }
 
-/* Return the next BlockDriverState in the round-robin sequence,
- * simulating a circular list.
+/* Return the next BlockBackend in the round-robin sequence, simulating a
+ * circular list.
  *
  * This assumes that tg->lock is held.
  *
- * @bs:  the current BlockDriverState
- * @ret: the next BlockDriverState in the sequence
+ * @blk: the current BlockBackend
+ * @ret: the next BlockBackend in the sequence
  */
-static BlockDriverState *throttle_group_next_bs(BlockDriverState *bs)
+static BlockBackend *throttle_group_next_blk(BlockBackend *blk)
 {
-    ThrottleState *ts = bs->throttle_state;
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleState *ts = blkp->throttle_state;
     ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
-    BlockDriverState *next = QLIST_NEXT(bs, round_robin);
+    BlockBackendPublic *next = QLIST_NEXT(blkp, round_robin);
 
     if (!next) {
-        return QLIST_FIRST(&tg->head);
+        next = QLIST_FIRST(&tg->head);
     }
 
-    return next;
+    return blk_by_public(next);
+}
+
+/*
+ * Return whether a BlockBackend has pending requests.
+ *
+ * This assumes that tg->lock is held.
+ *
+ * @blk: the BlockBackend
+ * @is_write:  the type of operation (read/write)
+ * @ret:       whether the BlockBackend has pending requests.
+ */
+static inline bool blk_has_pending_reqs(BlockBackend *blk,
+                                        bool is_write)
+{
+    const BlockBackendPublic *blkp = blk_get_public(blk);
+    return blkp->pending_reqs[is_write];
 }
 
-/* Return the next BlockDriverState in the round-robin sequence with
- * pending I/O requests.
+/* Return the next BlockBackend in the round-robin sequence with pending I/O
+ * requests.
  *
  * This assumes that tg->lock is held.
  *
- * @bs:        the current BlockDriverState
+ * @blk:       the current BlockBackend
  * @is_write:  the type of operation (read/write)
- * @ret:       the next BlockDriverState with pending requests, or bs
- *             if there is none.
+ * @ret:       the next BlockBackend with pending requests, or blk if there is
+ *             none.
  */
-static BlockDriverState *next_throttle_token(BlockDriverState *bs,
-                                             bool is_write)
+static BlockBackend *next_throttle_token(BlockBackend *blk, bool is_write)
 {
-    ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
-    BlockDriverState *token, *start;
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
+    BlockBackend *token, *start;
 
     start = token = tg->tokens[is_write];
 
     /* get next bs round in round robin style */
-    token = throttle_group_next_bs(token);
-    while (token != start && !token->pending_reqs[is_write]) {
-        token = throttle_group_next_bs(token);
+    token = throttle_group_next_blk(token);
+    while (token != start && !blk_has_pending_reqs(token, is_write)) {
+        token = throttle_group_next_blk(token);
     }
 
     /* If no IO are queued for scheduling on the next round robin token
      * then decide the token is the current bs because chances are
      * the current bs get the current request queued.
      */
-    if (token == start && !token->pending_reqs[is_write]) {
-        token = bs;
+    if (token == start && !blk_has_pending_reqs(token, is_write)) {
+        token = blk;
     }
 
+    /* Either we return the original BB, or one with pending requests */
+    assert(token == blk || blk_has_pending_reqs(token, is_write));
+
     return token;
 }
 
-/* Check if the next I/O request for a BlockDriverState needs to be
- * throttled or not. If there's no timer set in this group, set one
- * and update the token accordingly.
+/* Check if the next I/O request for a BlockBackend needs to be throttled or
+ * not. If there's no timer set in this group, set one and update the token
+ * accordingly.
  *
  * This assumes that tg->lock is held.
  *
- * @bs:         the current BlockDriverState
+ * @blk:        the current BlockBackend
  * @is_write:   the type of operation (read/write)
  * @ret:        whether the I/O request needs to be throttled or not
  */
-static bool throttle_group_schedule_timer(BlockDriverState *bs,
-                                          bool is_write)
+static bool throttle_group_schedule_timer(BlockBackend *blk, bool is_write)
 {
-    ThrottleState *ts = bs->throttle_state;
-    ThrottleTimers *tt = &bs->throttle_timers;
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleState *ts = blkp->throttle_state;
+    ThrottleTimers *tt = &blkp->throttle_timers;
     ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
     bool must_wait;
 
+    if (blkp->io_limits_disabled) {
+        return false;
+    }
+
     /* Check if any of the timers in this group is already armed */
     if (tg->any_timer_armed[is_write]) {
         return true;
@@ -224,9 +251,9 @@ static bool throttle_group_schedule_timer(BlockDriverState *bs,
 
     must_wait = throttle_schedule_timer(ts, tt, is_write);
 
-    /* If a timer just got armed, set bs as the current token */
+    /* If a timer just got armed, set blk as the current token */
     if (must_wait) {
-        tg->tokens[is_write] = bs;
+        tg->tokens[is_write] = blk;
         tg->any_timer_armed[is_write] = true;
     }
 
@@ -237,18 +264,19 @@ static bool throttle_group_schedule_timer(BlockDriverState *bs,
  *
  * This assumes that tg->lock is held.
  *
- * @bs:        the current BlockDriverState
+ * @blk:       the current BlockBackend
  * @is_write:  the type of operation (read/write)
  */
-static void schedule_next_request(BlockDriverState *bs, bool is_write)
+static void schedule_next_request(BlockBackend *blk, bool is_write)
 {
-    ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
     bool must_wait;
-    BlockDriverState *token;
+    BlockBackend *token;
 
     /* Check if there's any pending request to schedule next */
-    token = next_throttle_token(bs, is_write);
-    if (!token->pending_reqs[is_write]) {
+    token = next_throttle_token(blk, is_write);
+    if (!blk_has_pending_reqs(token, is_write)) {
         return;
     }
 
@@ -257,12 +285,12 @@ static void schedule_next_request(BlockDriverState *bs, bool is_write)
 
     /* If it doesn't have to wait, queue it for immediate execution */
     if (!must_wait) {
-        /* Give preference to requests from the current bs */
+        /* Give preference to requests from the current blk */
         if (qemu_in_coroutine() &&
-            qemu_co_queue_next(&bs->throttled_reqs[is_write])) {
-            token = bs;
+            qemu_co_queue_next(&blkp->throttled_reqs[is_write])) {
+            token = blk;
         } else {
-            ThrottleTimers *tt = &token->throttle_timers;
+            ThrottleTimers *tt = &blk_get_public(token)->throttle_timers;
             int64_t now = qemu_clock_get_ns(tt->clock_type);
             timer_mod(tt->timers[is_write], now + 1);
             tg->any_timer_armed[is_write] = true;
@@ -275,71 +303,94 @@ static void schedule_next_request(BlockDriverState *bs, bool is_write)
  * if necessary, and schedule the next request using a round robin
  * algorithm.
  *
- * @bs:        the current BlockDriverState
+ * @blk:       the current BlockBackend
  * @bytes:     the number of bytes for this I/O
  * @is_write:  the type of operation (read/write)
  */
-void coroutine_fn throttle_group_co_io_limits_intercept(BlockDriverState *bs,
+void coroutine_fn throttle_group_co_io_limits_intercept(BlockBackend *blk,
                                                         unsigned int bytes,
                                                         bool is_write)
 {
     bool must_wait;
-    BlockDriverState *token;
+    BlockBackend *token;
 
-    ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
     qemu_mutex_lock(&tg->lock);
 
     /* First we check if this I/O has to be throttled. */
-    token = next_throttle_token(bs, is_write);
+    token = next_throttle_token(blk, is_write);
     must_wait = throttle_group_schedule_timer(token, is_write);
 
     /* Wait if there's a timer set or queued requests of this type */
-    if (must_wait || bs->pending_reqs[is_write]) {
-        bs->pending_reqs[is_write]++;
+    if (must_wait || blkp->pending_reqs[is_write]) {
+        blkp->pending_reqs[is_write]++;
         qemu_mutex_unlock(&tg->lock);
-        qemu_co_queue_wait(&bs->throttled_reqs[is_write]);
+        qemu_co_queue_wait(&blkp->throttled_reqs[is_write], NULL);
         qemu_mutex_lock(&tg->lock);
-        bs->pending_reqs[is_write]--;
+        blkp->pending_reqs[is_write]--;
     }
 
     /* The I/O will be executed, so do the accounting */
-    throttle_account(bs->throttle_state, is_write, bytes);
+    throttle_account(blkp->throttle_state, is_write, bytes);
 
     /* Schedule the next request */
-    schedule_next_request(bs, is_write);
+    schedule_next_request(blk, is_write);
 
     qemu_mutex_unlock(&tg->lock);
 }
 
+void throttle_group_restart_blk(BlockBackend *blk)
+{
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    int i;
+
+    for (i = 0; i < 2; i++) {
+        while (qemu_co_enter_next(&blkp->throttled_reqs[i])) {
+            ;
+        }
+    }
+}
+
 /* Update the throttle configuration for a particular group. Similar
  * to throttle_config(), but guarantees atomicity within the
  * throttling group.
  *
- * @bs:  a BlockDriverState that is member of the group
+ * @blk: a BlockBackend that is a member of the group
  * @cfg: the configuration to set
  */
-void throttle_group_config(BlockDriverState *bs, ThrottleConfig *cfg)
+void throttle_group_config(BlockBackend *blk, ThrottleConfig *cfg)
 {
-    ThrottleTimers *tt = &bs->throttle_timers;
-    ThrottleState *ts = bs->throttle_state;
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleTimers *tt = &blkp->throttle_timers;
+    ThrottleState *ts = blkp->throttle_state;
     ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
     qemu_mutex_lock(&tg->lock);
-    throttle_config(ts, tt, cfg);
     /* throttle_config() cancels the timers */
-    tg->any_timer_armed[0] = tg->any_timer_armed[1] = false;
+    if (timer_pending(tt->timers[0])) {
+        tg->any_timer_armed[0] = false;
+    }
+    if (timer_pending(tt->timers[1])) {
+        tg->any_timer_armed[1] = false;
+    }
+    throttle_config(ts, tt, cfg);
     qemu_mutex_unlock(&tg->lock);
+
+    qemu_co_enter_next(&blkp->throttled_reqs[0]);
+    qemu_co_enter_next(&blkp->throttled_reqs[1]);
 }
 
 /* Get the throttle configuration from a particular group. Similar to
  * throttle_get_config(), but guarantees atomicity within the
  * throttling group.
  *
- * @bs:  a BlockDriverState that is member of the group
+ * @blk: a BlockBackend that is a member of the group
  * @cfg: the configuration will be written here
  */
-void throttle_group_get_config(BlockDriverState *bs, ThrottleConfig *cfg)
+void throttle_group_get_config(BlockBackend *blk, ThrottleConfig *cfg)
 {
-    ThrottleState *ts = bs->throttle_state;
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleState *ts = blkp->throttle_state;
     ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
     qemu_mutex_lock(&tg->lock);
     throttle_get_config(ts, cfg);
@@ -349,12 +400,13 @@ void throttle_group_get_config(BlockDriverState *bs, ThrottleConfig *cfg)
 /* ThrottleTimers callback. This wakes up a request that was waiting
  * because it had been throttled.
  *
- * @bs:        the BlockDriverState whose request had been throttled
+ * @blk:       the BlockBackend whose request had been throttled
  * @is_write:  the type of operation (read/write)
  */
-static void timer_cb(BlockDriverState *bs, bool is_write)
+static void timer_cb(BlockBackend *blk, bool is_write)
 {
-    ThrottleState *ts = bs->throttle_state;
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleState *ts = blkp->throttle_state;
     ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
     bool empty_queue;
 
@@ -364,13 +416,15 @@ static void timer_cb(BlockDriverState *bs, bool is_write)
     qemu_mutex_unlock(&tg->lock);
 
     /* Run the request that was waiting for this timer */
-    empty_queue = !qemu_co_enter_next(&bs->throttled_reqs[is_write]);
+    aio_context_acquire(blk_get_aio_context(blk));
+    empty_queue = !qemu_co_enter_next(&blkp->throttled_reqs[is_write]);
+    aio_context_release(blk_get_aio_context(blk));
 
     /* If the request queue was empty then we have to take care of
      * scheduling the next one */
     if (empty_queue) {
         qemu_mutex_lock(&tg->lock);
-        schedule_next_request(bs, is_write);
+        schedule_next_request(blk, is_write);
         qemu_mutex_unlock(&tg->lock);
     }
 }
@@ -385,18 +439,19 @@ static void write_timer_cb(void *opaque)
     timer_cb(opaque, true);
 }
 
-/* Register a BlockDriverState in the throttling group, also
- * initializing its timers and updating its throttle_state pointer to
- * point to it. If a throttling group with that name does not exist
- * yet, it will be created.
+/* Register a BlockBackend in the throttling group, also initializing its
+ * timers and updating its throttle_state pointer to point to it. If a
+ * throttling group with that name does not exist yet, it will be created.
  *
- * @bs:        the BlockDriverState to insert
+ * @blk:       the BlockBackend to insert
  * @groupname: the name of the group
  */
-void throttle_group_register_bs(BlockDriverState *bs, const char *groupname)
+void throttle_group_register_blk(BlockBackend *blk, const char *groupname)
 {
     int i;
-    ThrottleGroup *tg = throttle_group_incref(groupname);
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleState *ts = throttle_group_incref(groupname);
+    ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
     int clock_type = QEMU_CLOCK_REALTIME;
 
     if (qtest_enabled()) {
@@ -404,88 +459,67 @@ void throttle_group_register_bs(BlockDriverState *bs, const char *groupname)
         clock_type = QEMU_CLOCK_VIRTUAL;
     }
 
-    bs->throttle_state = &tg->ts;
+    blkp->throttle_state = ts;
 
     qemu_mutex_lock(&tg->lock);
-    /* If the ThrottleGroup is new set this BlockDriverState as the token */
+    /* If the ThrottleGroup is new set this BlockBackend as the token */
     for (i = 0; i < 2; i++) {
         if (!tg->tokens[i]) {
-            tg->tokens[i] = bs;
+            tg->tokens[i] = blk;
         }
     }
 
-    QLIST_INSERT_HEAD(&tg->head, bs, round_robin);
+    QLIST_INSERT_HEAD(&tg->head, blkp, round_robin);
 
-    throttle_timers_init(&bs->throttle_timers,
-                         bdrv_get_aio_context(bs),
+    throttle_timers_init(&blkp->throttle_timers,
+                         blk_get_aio_context(blk),
                          clock_type,
                          read_timer_cb,
                          write_timer_cb,
-                         bs);
+                         blk);
 
     qemu_mutex_unlock(&tg->lock);
 }
 
-/* Unregister a BlockDriverState from its group, removing it from the
- * list, destroying the timers and setting the throttle_state pointer
- * to NULL.
+/* Unregister a BlockBackend from its group, removing it from the list,
+ * destroying the timers and setting the throttle_state pointer to NULL.
+ *
+ * The BlockBackend must not have pending throttled requests, so the caller has
+ * to drain them first.
  *
  * The group will be destroyed if it's empty after this operation.
  *
- * @bs: the BlockDriverState to remove
+ * @blk: the BlockBackend to remove
  */
-void throttle_group_unregister_bs(BlockDriverState *bs)
+void throttle_group_unregister_blk(BlockBackend *blk)
 {
-    ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
+    BlockBackendPublic *blkp = blk_get_public(blk);
+    ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
     int i;
 
+    assert(blkp->pending_reqs[0] == 0 && blkp->pending_reqs[1] == 0);
+    assert(qemu_co_queue_empty(&blkp->throttled_reqs[0]));
+    assert(qemu_co_queue_empty(&blkp->throttled_reqs[1]));
+
     qemu_mutex_lock(&tg->lock);
     for (i = 0; i < 2; i++) {
-        if (tg->tokens[i] == bs) {
-            BlockDriverState *token = throttle_group_next_bs(bs);
-            /* Take care of the case where this is the last bs in the group */
-            if (token == bs) {
+        if (tg->tokens[i] == blk) {
+            BlockBackend *token = throttle_group_next_blk(blk);
+            /* Take care of the case where this is the last blk in the group */
+            if (token == blk) {
                 token = NULL;
             }
             tg->tokens[i] = token;
         }
     }
 
-    /* remove the current bs from the list */
-    QLIST_REMOVE(bs, round_robin);
-    throttle_timers_destroy(&bs->throttle_timers);
+    /* remove the current blk from the list */
+    QLIST_REMOVE(blkp, round_robin);
+    throttle_timers_destroy(&blkp->throttle_timers);
     qemu_mutex_unlock(&tg->lock);
 
-    throttle_group_unref(tg);
-    bs->throttle_state = NULL;
-}
-
-/* Acquire the lock of this throttling group.
- *
- * You won't normally need to use this. None of the functions from the
- * ThrottleGroup API require you to acquire the lock since all of them
- * deal with it internally.
- *
- * This should only be used in exceptional cases when you want to
- * access the protected fields of a BlockDriverState directly
- * (e.g. bdrv_swap()).
- *
- * @bs: a BlockDriverState that is member of the group
- */
-void throttle_group_lock(BlockDriverState *bs)
-{
-    ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
-    qemu_mutex_lock(&tg->lock);
-}
-
-/* Release the lock of this throttling group.
- *
- * See the comments in throttle_group_lock().
- */
-void throttle_group_unlock(BlockDriverState *bs)
-{
-    ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
-    qemu_mutex_unlock(&tg->lock);
+    throttle_group_unref(&tg->ts);
+    blkp->throttle_state = NULL;
 }
 
 static void throttle_groups_init(void)