blk-iocost: grab ioc->lock for debt handling
authorTejun Heo <tj@kernel.org>
Tue, 1 Sep 2020 18:52:42 +0000 (14:52 -0400)
committerJens Axboe <axboe@kernel.dk>
Wed, 2 Sep 2020 01:38:32 +0000 (19:38 -0600)
Currently, debt handling requires only iocg->waitq.lock. In the future, we
want to adjust and propagate inuse changes depending on debt status. Let's
grab ioc->lock in debt handling paths in preparation.

* Because ioc->lock nests outside iocg->waitq.lock, the decision to grab
  ioc->lock needs to be made before entering the critical sections.

* Add and use iocg_[un]lock() which handles the conditional double locking.

* Add @pay_debt to iocg_kick_waitq() so that debt payment happens only when
  the caller grabbed both locks.

This patch is prepatory and the comments contain references to future
changes.

Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-iocost.c

index f369886..23b173e 100644 (file)
@@ -680,6 +680,26 @@ static void iocg_commit_bio(struct ioc_gq *iocg, struct bio *bio, u64 cost)
        atomic64_add(cost, &iocg->vtime);
 }
 
+static void iocg_lock(struct ioc_gq *iocg, bool lock_ioc, unsigned long *flags)
+{
+       if (lock_ioc) {
+               spin_lock_irqsave(&iocg->ioc->lock, *flags);
+               spin_lock(&iocg->waitq.lock);
+       } else {
+               spin_lock_irqsave(&iocg->waitq.lock, *flags);
+       }
+}
+
+static void iocg_unlock(struct ioc_gq *iocg, bool unlock_ioc, unsigned long *flags)
+{
+       if (unlock_ioc) {
+               spin_unlock(&iocg->waitq.lock);
+               spin_unlock_irqrestore(&iocg->ioc->lock, *flags);
+       } else {
+               spin_unlock_irqrestore(&iocg->waitq.lock, *flags);
+       }
+}
+
 #define CREATE_TRACE_POINTS
 #include <trace/events/iocost.h>
 
@@ -1215,11 +1235,17 @@ static int iocg_wake_fn(struct wait_queue_entry *wq_entry, unsigned mode,
        return 0;
 }
 
-static void iocg_kick_waitq(struct ioc_gq *iocg, struct ioc_now *now)
+/*
+ * Calculate the accumulated budget, pay debt if @pay_debt and wake up waiters
+ * accordingly. When @pay_debt is %true, the caller must be holding ioc->lock in
+ * addition to iocg->waitq.lock.
+ */
+static void iocg_kick_waitq(struct ioc_gq *iocg, bool pay_debt,
+                           struct ioc_now *now)
 {
        struct ioc *ioc = iocg->ioc;
        struct iocg_wake_ctx ctx = { .iocg = iocg };
-       u64 vdebt, vshortage, expires, oexpires;
+       u64 vshortage, expires, oexpires;
        s64 vbudget;
        u32 hw_inuse;
 
@@ -1229,25 +1255,39 @@ static void iocg_kick_waitq(struct ioc_gq *iocg, struct ioc_now *now)
        vbudget = now->vnow - atomic64_read(&iocg->vtime);
 
        /* pay off debt */
-       vdebt = abs_cost_to_cost(iocg->abs_vdebt, hw_inuse);
-       if (vdebt && vbudget > 0) {
+       if (pay_debt && iocg->abs_vdebt && vbudget > 0) {
+               u64 vdebt = abs_cost_to_cost(iocg->abs_vdebt, hw_inuse);
                u64 delta = min_t(u64, vbudget, vdebt);
                u64 abs_delta = min(cost_to_abs_cost(delta, hw_inuse),
                                    iocg->abs_vdebt);
 
+               lockdep_assert_held(&ioc->lock);
+
                atomic64_add(delta, &iocg->vtime);
                atomic64_add(delta, &iocg->done_vtime);
                iocg->abs_vdebt -= abs_delta;
+               vbudget -= vdebt;
 
                iocg_kick_delay(iocg, now);
        }
 
        /*
+        * Debt can still be outstanding if we haven't paid all yet or the
+        * caller raced and called without @pay_debt. Shouldn't wake up waiters
+        * under debt. Make sure @vbudget reflects the outstanding amount and is
+        * not positive.
+        */
+       if (iocg->abs_vdebt) {
+               s64 vdebt = abs_cost_to_cost(iocg->abs_vdebt, hw_inuse);
+               vbudget = min_t(s64, 0, vbudget - vdebt);
+       }
+
+       /*
         * Wake up the ones which are due and see how much vtime we'll need
         * for the next one.
         */
        ctx.hw_inuse = hw_inuse;
-       ctx.vbudget = vbudget - vdebt;
+       ctx.vbudget = vbudget;
        __wake_up_locked_key(&iocg->waitq, TASK_NORMAL, &ctx);
        if (!waitqueue_active(&iocg->waitq))
                return;
@@ -1273,14 +1313,15 @@ static void iocg_kick_waitq(struct ioc_gq *iocg, struct ioc_now *now)
 static enum hrtimer_restart iocg_waitq_timer_fn(struct hrtimer *timer)
 {
        struct ioc_gq *iocg = container_of(timer, struct ioc_gq, waitq_timer);
+       bool pay_debt = READ_ONCE(iocg->abs_vdebt);
        struct ioc_now now;
        unsigned long flags;
 
        ioc_now(iocg->ioc, &now);
 
-       spin_lock_irqsave(&iocg->waitq.lock, flags);
-       iocg_kick_waitq(iocg, &now);
-       spin_unlock_irqrestore(&iocg->waitq.lock, flags);
+       iocg_lock(iocg, pay_debt, &flags);
+       iocg_kick_waitq(iocg, pay_debt, &now);
+       iocg_unlock(iocg, pay_debt, &flags);
 
        return HRTIMER_NORESTART;
 }
@@ -1396,7 +1437,7 @@ static void ioc_timer_fn(struct timer_list *timer)
 
                if (waitqueue_active(&iocg->waitq) || iocg->abs_vdebt) {
                        /* might be oversleeping vtime / hweight changes, kick */
-                       iocg_kick_waitq(iocg, &now);
+                       iocg_kick_waitq(iocg, true, &now);
                } else if (iocg_is_idle(iocg)) {
                        /* no waiter and idle, deactivate */
                        iocg->last_inuse = iocg->inuse;
@@ -1743,6 +1784,8 @@ static void ioc_rqos_throttle(struct rq_qos *rqos, struct bio *bio)
        struct iocg_wait wait;
        u32 hw_active, hw_inuse;
        u64 abs_cost, cost, vtime;
+       bool use_debt, ioc_locked;
+       unsigned long flags;
 
        /* bypass IOs if disabled or for root cgroup */
        if (!ioc->enabled || !iocg->level)
@@ -1786,15 +1829,26 @@ static void ioc_rqos_throttle(struct rq_qos *rqos, struct bio *bio)
        }
 
        /*
-        * We activated above but w/o any synchronization. Deactivation is
-        * synchronized with waitq.lock and we won't get deactivated as long
-        * as we're waiting or has debt, so we're good if we're activated
-        * here. In the unlikely case that we aren't, just issue the IO.
+        * We're over budget. This can be handled in two ways. IOs which may
+        * cause priority inversions are punted to @ioc->aux_iocg and charged as
+        * debt. Otherwise, the issuer is blocked on @iocg->waitq. Debt handling
+        * requires @ioc->lock, waitq handling @iocg->waitq.lock. Determine
+        * whether debt handling is needed and acquire locks accordingly.
         */
-       spin_lock_irq(&iocg->waitq.lock);
+       use_debt = bio_issue_as_root_blkg(bio) || fatal_signal_pending(current);
+       ioc_locked = use_debt || READ_ONCE(iocg->abs_vdebt);
 
+       iocg_lock(iocg, ioc_locked, &flags);
+
+       /*
+        * @iocg must stay activated for debt and waitq handling. Deactivation
+        * is synchronized against both ioc->lock and waitq.lock and we won't
+        * get deactivated as long as we're waiting or has debt, so we're good
+        * if we're activated here. In the unlikely cases that we aren't, just
+        * issue the IO.
+        */
        if (unlikely(list_empty(&iocg->active_list))) {
-               spin_unlock_irq(&iocg->waitq.lock);
+               iocg_unlock(iocg, ioc_locked, &flags);
                iocg_commit_bio(iocg, bio, cost);
                return;
        }
@@ -1816,12 +1870,12 @@ static void ioc_rqos_throttle(struct rq_qos *rqos, struct bio *bio)
         * clear them and leave @iocg inactive w/ dangling use_delay heavily
         * penalizing the cgroup and its descendants.
         */
-       if (bio_issue_as_root_blkg(bio) || fatal_signal_pending(current)) {
+       if (use_debt) {
                iocg->abs_vdebt += abs_cost;
                if (iocg_kick_delay(iocg, &now))
                        blkcg_schedule_throttle(rqos->q,
                                        (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
-               spin_unlock_irq(&iocg->waitq.lock);
+               iocg_unlock(iocg, ioc_locked, &flags);
                return;
        }
 
@@ -1845,9 +1899,9 @@ static void ioc_rqos_throttle(struct rq_qos *rqos, struct bio *bio)
        wait.committed = false; /* will be set true by waker */
 
        __add_wait_queue_entry_tail(&iocg->waitq, &wait.wait);
-       iocg_kick_waitq(iocg, &now);
+       iocg_kick_waitq(iocg, ioc_locked, &now);
 
-       spin_unlock_irq(&iocg->waitq.lock);
+       iocg_unlock(iocg, ioc_locked, &flags);
 
        while (true) {
                set_current_state(TASK_UNINTERRUPTIBLE);