block: introduce blk-iolatency io controller
authorJosef Bacik <jbacik@fb.com>
Tue, 3 Jul 2018 15:15:01 +0000 (11:15 -0400)
committerJens Axboe <axboe@kernel.dk>
Mon, 9 Jul 2018 15:07:54 +0000 (09:07 -0600)
Current IO controllers for the block layer are less than ideal for our
use case.  The io.max controller is great at hard limiting, but it is
not work conserving.  This patch introduces io.latency.  You provide a
latency target for your group and we monitor the io in short windows to
make sure we are not exceeding those latency targets.  This makes use of
the rq-qos infrastructure and works much like the wbt stuff.  There are
a few differences from wbt

 - It's bio based, so the latency covers the whole block layer in addition to
   the actual io.
 - We will throttle all IO types that comes in here if we need to.
 - We use the mean latency over the 100ms window.  This is because writes can
   be particularly fast, which could give us a false sense of the impact of
   other workloads on our protected workload.
 - By default there's no throttling, we set the queue_depth to INT_MAX so that
   we can have as many outstanding bio's as we're allowed to.  Only at
   throttle time do we pay attention to the actual queue depth.
 - We backcharge cgroups for root cg issued IO and induce artificial
   delays in order to deal with cases like metadata only or swap heavy
   workloads.

In testing this has worked out relatively well.  Protected workloads
will throttle noisy workloads down to 1 io at time if they are doing
normal IO on their own, or induce up to a 1 second delay per syscall if
they are doing a lot of root issued IO (metadata/swap IO).

Our testing has revolved mostly around our production web servers where
we have hhvm (the web server application) in a protected group and
everything else in another group.  We see slightly higher requests per
second (RPS) on the test tier vs the control tier, and much more stable
RPS across all machines in the test tier vs the control tier.

Another test we run is a slow memory allocator in the unprotected group.
Before this would eventually push us into swap and cause the whole box
to die and not recover at all.  With these patches we see slight RPS
drops (usually 10-15%) before the memory consumer is properly killed and
things recover within seconds.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/Kconfig
block/Makefile
block/blk-cgroup.c
block/blk-iolatency.c [new file with mode: 0644]
block/blk.h
include/linux/blk_types.h

index dfe7bc770fc9abb3d5e958a961cf8eb3eaf68984..1f2469a0123ceb1f36103c0db9eb71d140556193 100644 (file)
@@ -149,6 +149,18 @@ config BLK_WBT
        dynamically on an algorithm loosely based on CoDel, factoring in
        the realtime performance of the disk.
 
+config BLK_CGROUP_IOLATENCY
+       bool "Enable support for latency based cgroup IO protection"
+       depends on BLK_CGROUP=y
+       default n
+       ---help---
+       Enabling this option enables the .latency interface for IO throttling.
+       The IO controller will attempt to maintain average IO latencies below
+       the configured latency target, throttling anybody with a higher latency
+       target than the victimized group.
+
+       Note, this is an experimental interface and could be changed someday.
+
 config BLK_WBT_SQ
        bool "Single queue writeback throttling"
        default n
index 57d0f47ab05f944e969aec7df56a46681b339429..572b33f32c07cf7056fb1121abba753a9ba8a0ac 100644 (file)
@@ -17,6 +17,7 @@ obj-$(CONFIG_BLK_DEV_BSG)     += bsg.o
 obj-$(CONFIG_BLK_DEV_BSGLIB)   += bsg-lib.o
 obj-$(CONFIG_BLK_CGROUP)       += blk-cgroup.o
 obj-$(CONFIG_BLK_DEV_THROTTLING)       += blk-throttle.o
+obj-$(CONFIG_BLK_CGROUP_IOLATENCY)     += blk-iolatency.o
 obj-$(CONFIG_IOSCHED_NOOP)     += noop-iosched.o
 obj-$(CONFIG_IOSCHED_DEADLINE) += deadline-iosched.o
 obj-$(CONFIG_IOSCHED_CFQ)      += cfq-iosched.o
index d3310ec96c2abc81e848e58698f02349f02a72a4..7e2c19ce1a08eea0fd10a533afddda01537ad7fe 100644 (file)
@@ -1238,6 +1238,14 @@ int blkcg_init_queue(struct request_queue *q)
        if (preloaded)
                radix_tree_preload_end();
 
+       ret = blk_iolatency_init(q);
+       if (ret) {
+               spin_lock_irq(q->queue_lock);
+               blkg_destroy_all(q);
+               spin_unlock_irq(q->queue_lock);
+               return ret;
+       }
+
        ret = blk_throtl_init(q);
        if (ret) {
                spin_lock_irq(q->queue_lock);
diff --git a/block/blk-iolatency.c b/block/blk-iolatency.c
new file mode 100644 (file)
index 0000000..a35a1f5
--- /dev/null
@@ -0,0 +1,930 @@
+/*
+ * Block rq-qos base io controller
+ *
+ * This works similar to wbt with a few exceptions
+ *
+ * - It's bio based, so the latency covers the whole block layer in addition to
+ *   the actual io.
+ * - We will throttle all IO that comes in here if we need to.
+ * - We use the mean latency over the 100ms window.  This is because writes can
+ *   be particularly fast, which could give us a false sense of the impact of
+ *   other workloads on our protected workload.
+ * - By default there's no throttling, we set the queue_depth to INT_MAX so that
+ *   we can have as many outstanding bio's as we're allowed to.  Only at
+ *   throttle time do we pay attention to the actual queue depth.
+ *
+ * The hierarchy works like the cpu controller does, we track the latency at
+ * every configured node, and each configured node has it's own independent
+ * queue depth.  This means that we only care about our latency targets at the
+ * peer level.  Some group at the bottom of the hierarchy isn't going to affect
+ * a group at the end of some other path if we're only configred at leaf level.
+ *
+ * Consider the following
+ *
+ *                   root blkg
+ *             /                     \
+ *        fast (target=5ms)     slow (target=10ms)
+ *         /     \                  /        \
+ *       a        b          normal(15ms)   unloved
+ *
+ * "a" and "b" have no target, but their combined io under "fast" cannot exceed
+ * an average latency of 5ms.  If it does then we will throttle the "slow"
+ * group.  In the case of "normal", if it exceeds its 15ms target, we will
+ * throttle "unloved", but nobody else.
+ *
+ * In this example "fast", "slow", and "normal" will be the only groups actually
+ * accounting their io latencies.  We have to walk up the heirarchy to the root
+ * on every submit and complete so we can do the appropriate stat recording and
+ * adjust the queue depth of ourselves if needed.
+ *
+ * There are 2 ways we throttle IO.
+ *
+ * 1) Queue depth throttling.  As we throttle down we will adjust the maximum
+ * number of IO's we're allowed to have in flight.  This starts at (u64)-1 down
+ * to 1.  If the group is only ever submitting IO for itself then this is the
+ * only way we throttle.
+ *
+ * 2) Induced delay throttling.  This is for the case that a group is generating
+ * IO that has to be issued by the root cg to avoid priority inversion. So think
+ * REQ_META or REQ_SWAP.  If we are already at qd == 1 and we're getting a lot
+ * of work done for us on behalf of the root cg and are being asked to scale
+ * down more then we induce a latency at userspace return.  We accumulate the
+ * total amount of time we need to be punished by doing
+ *
+ * total_time += min_lat_nsec - actual_io_completion
+ *
+ * and then at throttle time will do
+ *
+ * throttle_time = min(total_time, NSEC_PER_SEC)
+ *
+ * This induced delay will throttle back the activity that is generating the
+ * root cg issued io's, wethere that's some metadata intensive operation or the
+ * group is using so much memory that it is pushing us into swap.
+ *
+ * Copyright (C) 2018 Josef Bacik
+ */
+#include <linux/kernel.h>
+#include <linux/blk_types.h>
+#include <linux/backing-dev.h>
+#include <linux/module.h>
+#include <linux/timer.h>
+#include <linux/memcontrol.h>
+#include <linux/sched/signal.h>
+#include <trace/events/block.h>
+#include "blk-rq-qos.h"
+#include "blk-stat.h"
+
+#define DEFAULT_SCALE_COOKIE 1000000U
+
+static struct blkcg_policy blkcg_policy_iolatency;
+struct iolatency_grp;
+
+struct blk_iolatency {
+       struct rq_qos rqos;
+       struct timer_list timer;
+       atomic_t enabled;
+};
+
+static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
+{
+       return container_of(rqos, struct blk_iolatency, rqos);
+}
+
+static inline bool blk_iolatency_enabled(struct blk_iolatency *blkiolat)
+{
+       return atomic_read(&blkiolat->enabled) > 0;
+}
+
+struct child_latency_info {
+       spinlock_t lock;
+
+       /* Last time we adjusted the scale of everybody. */
+       u64 last_scale_event;
+
+       /* The latency that we missed. */
+       u64 scale_lat;
+
+       /* Total io's from all of our children for the last summation. */
+       u64 nr_samples;
+
+       /* The guy who actually changed the latency numbers. */
+       struct iolatency_grp *scale_grp;
+
+       /* Cookie to tell if we need to scale up or down. */
+       atomic_t scale_cookie;
+};
+
+struct iolatency_grp {
+       struct blkg_policy_data pd;
+       struct blk_rq_stat __percpu *stats;
+       struct blk_iolatency *blkiolat;
+       struct rq_depth rq_depth;
+       struct rq_wait rq_wait;
+       atomic64_t window_start;
+       atomic_t scale_cookie;
+       u64 min_lat_nsec;
+       u64 cur_win_nsec;
+
+       /* total running average of our io latency. */
+       u64 total_lat_avg;
+       u64 total_lat_nr;
+
+       /* Our current number of IO's for the last summation. */
+       u64 nr_samples;
+
+       struct child_latency_info child_lat;
+};
+
+static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
+{
+       return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
+}
+
+static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
+{
+       return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
+}
+
+static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
+{
+       return pd_to_blkg(&iolat->pd);
+}
+
+static inline bool iolatency_may_queue(struct iolatency_grp *iolat,
+                                      wait_queue_entry_t *wait,
+                                      bool first_block)
+{
+       struct rq_wait *rqw = &iolat->rq_wait;
+
+       if (first_block && waitqueue_active(&rqw->wait) &&
+           rqw->wait.head.next != &wait->entry)
+               return false;
+       return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
+}
+
+static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
+                                      struct iolatency_grp *iolat,
+                                      spinlock_t *lock, bool issue_as_root,
+                                      bool use_memdelay)
+       __releases(lock)
+       __acquires(lock)
+{
+       struct rq_wait *rqw = &iolat->rq_wait;
+       unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
+       DEFINE_WAIT(wait);
+       bool first_block = true;
+
+       if (use_delay)
+               blkcg_schedule_throttle(rqos->q, use_memdelay);
+
+       /*
+        * To avoid priority inversions we want to just take a slot if we are
+        * issuing as root.  If we're being killed off there's no point in
+        * delaying things, we may have been killed by OOM so throttling may
+        * make recovery take even longer, so just let the IO's through so the
+        * task can go away.
+        */
+       if (issue_as_root || fatal_signal_pending(current)) {
+               atomic_inc(&rqw->inflight);
+               return;
+       }
+
+       if (iolatency_may_queue(iolat, &wait, first_block))
+               return;
+
+       do {
+               prepare_to_wait_exclusive(&rqw->wait, &wait,
+                                         TASK_UNINTERRUPTIBLE);
+
+               if (iolatency_may_queue(iolat, &wait, first_block))
+                       break;
+               first_block = false;
+
+               if (lock) {
+                       spin_unlock_irq(lock);
+                       io_schedule();
+                       spin_lock_irq(lock);
+               } else {
+                       io_schedule();
+               }
+       } while (1);
+
+       finish_wait(&rqw->wait, &wait);
+}
+
+#define SCALE_DOWN_FACTOR 2
+#define SCALE_UP_FACTOR 4
+
+static inline unsigned long scale_amount(unsigned long qd, bool up)
+{
+       return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
+}
+
+/*
+ * We scale the qd down faster than we scale up, so we need to use this helper
+ * to adjust the scale_cookie accordingly so we don't prematurely get
+ * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
+ *
+ * Each group has their own local copy of the last scale cookie they saw, so if
+ * the global scale cookie goes up or down they know which way they need to go
+ * based on their last knowledge of it.
+ */
+static void scale_cookie_change(struct blk_iolatency *blkiolat,
+                               struct child_latency_info *lat_info,
+                               bool up)
+{
+       unsigned long qd = blk_queue_depth(blkiolat->rqos.q);
+       unsigned long scale = scale_amount(qd, up);
+       unsigned long old = atomic_read(&lat_info->scale_cookie);
+       unsigned long max_scale = qd << 1;
+       unsigned long diff = 0;
+
+       if (old < DEFAULT_SCALE_COOKIE)
+               diff = DEFAULT_SCALE_COOKIE - old;
+
+       if (up) {
+               if (scale + old > DEFAULT_SCALE_COOKIE)
+                       atomic_set(&lat_info->scale_cookie,
+                                  DEFAULT_SCALE_COOKIE);
+               else if (diff > qd)
+                       atomic_inc(&lat_info->scale_cookie);
+               else
+                       atomic_add(scale, &lat_info->scale_cookie);
+       } else {
+               /*
+                * We don't want to dig a hole so deep that it takes us hours to
+                * dig out of it.  Just enough that we don't throttle/unthrottle
+                * with jagged workloads but can still unthrottle once pressure
+                * has sufficiently dissipated.
+                */
+               if (diff > qd) {
+                       if (diff < max_scale)
+                               atomic_dec(&lat_info->scale_cookie);
+               } else {
+                       atomic_sub(scale, &lat_info->scale_cookie);
+               }
+       }
+}
+
+/*
+ * Change the queue depth of the iolatency_grp.  We add/subtract 1/16th of the
+ * queue depth at a time so we don't get wild swings and hopefully dial in to
+ * fairer distribution of the overall queue depth.
+ */
+static void scale_change(struct iolatency_grp *iolat, bool up)
+{
+       unsigned long qd = blk_queue_depth(iolat->blkiolat->rqos.q);
+       unsigned long scale = scale_amount(qd, up);
+       unsigned long old = iolat->rq_depth.max_depth;
+       bool changed = false;
+
+       if (old > qd)
+               old = qd;
+
+       if (up) {
+               if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
+                       return;
+
+               if (old < qd) {
+                       changed = true;
+                       old += scale;
+                       old = min(old, qd);
+                       iolat->rq_depth.max_depth = old;
+                       wake_up_all(&iolat->rq_wait.wait);
+               }
+       } else if (old > 1) {
+               old >>= 1;
+               changed = true;
+               iolat->rq_depth.max_depth = max(old, 1UL);
+       }
+}
+
+/* Check our parent and see if the scale cookie has changed. */
+static void check_scale_change(struct iolatency_grp *iolat)
+{
+       struct iolatency_grp *parent;
+       struct child_latency_info *lat_info;
+       unsigned int cur_cookie;
+       unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
+       u64 scale_lat;
+       unsigned int old;
+       int direction = 0;
+
+       if (lat_to_blkg(iolat)->parent == NULL)
+               return;
+
+       parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
+       if (!parent)
+               return;
+
+       lat_info = &parent->child_lat;
+       cur_cookie = atomic_read(&lat_info->scale_cookie);
+       scale_lat = READ_ONCE(lat_info->scale_lat);
+
+       if (cur_cookie < our_cookie)
+               direction = -1;
+       else if (cur_cookie > our_cookie)
+               direction = 1;
+       else
+               return;
+
+       old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
+
+       /* Somebody beat us to the punch, just bail. */
+       if (old != our_cookie)
+               return;
+
+       if (direction < 0 && iolat->min_lat_nsec) {
+               u64 samples_thresh;
+
+               if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
+                       return;
+
+               /*
+                * Sometimes high priority groups are their own worst enemy, so
+                * instead of taking it out on some poor other group that did 5%
+                * or less of the IO's for the last summation just skip this
+                * scale down event.
+                */
+               samples_thresh = lat_info->nr_samples * 5;
+               samples_thresh = div64_u64(samples_thresh, 100);
+               if (iolat->nr_samples <= samples_thresh)
+                       return;
+       }
+
+       /* We're as low as we can go. */
+       if (iolat->rq_depth.max_depth == 1 && direction < 0) {
+               blkcg_use_delay(lat_to_blkg(iolat));
+               return;
+       }
+
+       /* We're back to the default cookie, unthrottle all the things. */
+       if (cur_cookie == DEFAULT_SCALE_COOKIE) {
+               blkcg_clear_delay(lat_to_blkg(iolat));
+               iolat->rq_depth.max_depth = INT_MAX;
+               wake_up_all(&iolat->rq_wait.wait);
+               return;
+       }
+
+       scale_change(iolat, direction > 0);
+}
+
+static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio,
+                                    spinlock_t *lock)
+{
+       struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
+       struct blkcg *blkcg;
+       struct blkcg_gq *blkg;
+       struct request_queue *q = rqos->q;
+       bool issue_as_root = bio_issue_as_root_blkg(bio);
+
+       if (!blk_iolatency_enabled(blkiolat))
+               return;
+
+       rcu_read_lock();
+       blkcg = bio_blkcg(bio);
+       bio_associate_blkcg(bio, &blkcg->css);
+       blkg = blkg_lookup(blkcg, q);
+       if (unlikely(!blkg)) {
+               if (!lock)
+                       spin_lock_irq(q->queue_lock);
+               blkg = blkg_lookup_create(blkcg, q);
+               if (IS_ERR(blkg))
+                       blkg = NULL;
+               if (!lock)
+                       spin_unlock_irq(q->queue_lock);
+       }
+       if (!blkg)
+               goto out;
+
+       bio_issue_init(&bio->bi_issue, bio_sectors(bio));
+       bio_associate_blkg(bio, blkg);
+out:
+       rcu_read_unlock();
+       while (blkg && blkg->parent) {
+               struct iolatency_grp *iolat = blkg_to_lat(blkg);
+               if (!iolat) {
+                       blkg = blkg->parent;
+                       continue;
+               }
+
+               check_scale_change(iolat);
+               __blkcg_iolatency_throttle(rqos, iolat, lock, issue_as_root,
+                                    (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
+               blkg = blkg->parent;
+       }
+       if (!timer_pending(&blkiolat->timer))
+               mod_timer(&blkiolat->timer, jiffies + HZ);
+}
+
+static void iolatency_record_time(struct iolatency_grp *iolat,
+                                 struct bio_issue *issue, u64 now,
+                                 bool issue_as_root)
+{
+       struct blk_rq_stat *rq_stat;
+       u64 start = bio_issue_time(issue);
+       u64 req_time;
+
+       if (now <= start)
+               return;
+
+       req_time = now - start;
+
+       /*
+        * We don't want to count issue_as_root bio's in the cgroups latency
+        * statistics as it could skew the numbers downwards.
+        */
+       if (unlikely(issue_as_root && iolat->rq_depth.max_depth != (u64)-1)) {
+               u64 sub = iolat->min_lat_nsec;
+               if (req_time < sub)
+                       blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
+               return;
+       }
+
+       rq_stat = get_cpu_ptr(iolat->stats);
+       blk_rq_stat_add(rq_stat, req_time);
+       put_cpu_ptr(rq_stat);
+}
+
+#define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
+#define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
+
+static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
+{
+       struct blkcg_gq *blkg = lat_to_blkg(iolat);
+       struct iolatency_grp *parent;
+       struct child_latency_info *lat_info;
+       struct blk_rq_stat stat;
+       unsigned long flags;
+       int cpu;
+
+       blk_rq_stat_init(&stat);
+       preempt_disable();
+       for_each_online_cpu(cpu) {
+               struct blk_rq_stat *s;
+               s = per_cpu_ptr(iolat->stats, cpu);
+               blk_rq_stat_sum(&stat, s);
+               blk_rq_stat_init(s);
+       }
+       preempt_enable();
+
+       /*
+        * Our average exceeded our window, scale up our window so we are more
+        * accurate, but not more than the global timer.
+        */
+       if (stat.mean > iolat->cur_win_nsec) {
+               iolat->cur_win_nsec <<= 1;
+               iolat->cur_win_nsec =
+                       max_t(u64, iolat->cur_win_nsec, NSEC_PER_SEC);
+       }
+
+       parent = blkg_to_lat(blkg->parent);
+       if (!parent)
+               return;
+
+       lat_info = &parent->child_lat;
+
+       iolat->total_lat_avg =
+               div64_u64((iolat->total_lat_avg * iolat->total_lat_nr) +
+                         stat.mean, iolat->total_lat_nr + 1);
+
+       iolat->total_lat_nr++;
+
+       /* Everything is ok and we don't need to adjust the scale. */
+       if (stat.mean <= iolat->min_lat_nsec &&
+           atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
+               return;
+
+       /* Somebody beat us to the punch, just bail. */
+       spin_lock_irqsave(&lat_info->lock, flags);
+       lat_info->nr_samples -= iolat->nr_samples;
+       lat_info->nr_samples += stat.nr_samples;
+       iolat->nr_samples = stat.nr_samples;
+
+       if ((lat_info->last_scale_event >= now ||
+           now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME) &&
+           lat_info->scale_lat <= iolat->min_lat_nsec)
+               goto out;
+
+       if (stat.mean <= iolat->min_lat_nsec &&
+           stat.nr_samples >= BLKIOLATENCY_MIN_GOOD_SAMPLES) {
+               if (lat_info->scale_grp == iolat) {
+                       lat_info->last_scale_event = now;
+                       scale_cookie_change(iolat->blkiolat, lat_info, true);
+               }
+       } else if (stat.mean > iolat->min_lat_nsec) {
+               lat_info->last_scale_event = now;
+               if (!lat_info->scale_grp ||
+                   lat_info->scale_lat > iolat->min_lat_nsec) {
+                       WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
+                       lat_info->scale_grp = iolat;
+               }
+               scale_cookie_change(iolat->blkiolat, lat_info, false);
+       }
+out:
+       spin_unlock_irqrestore(&lat_info->lock, flags);
+}
+
+static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
+{
+       struct blkcg_gq *blkg;
+       struct rq_wait *rqw;
+       struct iolatency_grp *iolat;
+       u64 window_start;
+       u64 now = ktime_to_ns(ktime_get());
+       bool issue_as_root = bio_issue_as_root_blkg(bio);
+       bool enabled = false;
+
+       blkg = bio->bi_blkg;
+       if (!blkg)
+               return;
+
+       iolat = blkg_to_lat(bio->bi_blkg);
+       if (!iolat)
+               return;
+
+       enabled = blk_iolatency_enabled(iolat->blkiolat);
+       while (blkg && blkg->parent) {
+               iolat = blkg_to_lat(blkg);
+               if (!iolat) {
+                       blkg = blkg->parent;
+                       continue;
+               }
+               rqw = &iolat->rq_wait;
+
+               atomic_dec(&rqw->inflight);
+               if (!enabled || iolat->min_lat_nsec == 0)
+                       goto next;
+               iolatency_record_time(iolat, &bio->bi_issue, now,
+                                     issue_as_root);
+               window_start = atomic64_read(&iolat->window_start);
+               if (now > window_start &&
+                   (now - window_start) >= iolat->cur_win_nsec) {
+                       if (atomic64_cmpxchg(&iolat->window_start,
+                                       window_start, now) == window_start)
+                               iolatency_check_latencies(iolat, now);
+               }
+next:
+               wake_up(&rqw->wait);
+               blkg = blkg->parent;
+       }
+}
+
+static void blkcg_iolatency_cleanup(struct rq_qos *rqos, struct bio *bio)
+{
+       struct blkcg_gq *blkg;
+
+       blkg = bio->bi_blkg;
+       while (blkg && blkg->parent) {
+               struct rq_wait *rqw;
+               struct iolatency_grp *iolat;
+
+               iolat = blkg_to_lat(blkg);
+               if (!iolat)
+                       goto next;
+
+               rqw = &iolat->rq_wait;
+               atomic_dec(&rqw->inflight);
+               wake_up(&rqw->wait);
+next:
+               blkg = blkg->parent;
+       }
+}
+
+static void blkcg_iolatency_exit(struct rq_qos *rqos)
+{
+       struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
+
+       del_timer_sync(&blkiolat->timer);
+       blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
+       kfree(blkiolat);
+}
+
+static struct rq_qos_ops blkcg_iolatency_ops = {
+       .throttle = blkcg_iolatency_throttle,
+       .cleanup = blkcg_iolatency_cleanup,
+       .done_bio = blkcg_iolatency_done_bio,
+       .exit = blkcg_iolatency_exit,
+};
+
+static void blkiolatency_timer_fn(struct timer_list *t)
+{
+       struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
+       struct blkcg_gq *blkg;
+       struct cgroup_subsys_state *pos_css;
+       u64 now = ktime_to_ns(ktime_get());
+
+       rcu_read_lock();
+       blkg_for_each_descendant_pre(blkg, pos_css,
+                                    blkiolat->rqos.q->root_blkg) {
+               struct iolatency_grp *iolat;
+               struct child_latency_info *lat_info;
+               unsigned long flags;
+               u64 cookie;
+
+               /*
+                * We could be exiting, don't access the pd unless we have a
+                * ref on the blkg.
+                */
+               if (!blkg_try_get(blkg))
+                       continue;
+
+               iolat = blkg_to_lat(blkg);
+               if (!iolat)
+                       continue;
+
+               lat_info = &iolat->child_lat;
+               cookie = atomic_read(&lat_info->scale_cookie);
+
+               if (cookie >= DEFAULT_SCALE_COOKIE)
+                       goto next;
+
+               spin_lock_irqsave(&lat_info->lock, flags);
+               if (lat_info->last_scale_event >= now)
+                       goto next_lock;
+
+               /*
+                * We scaled down but don't have a scale_grp, scale up and carry
+                * on.
+                */
+               if (lat_info->scale_grp == NULL) {
+                       scale_cookie_change(iolat->blkiolat, lat_info, true);
+                       goto next_lock;
+               }
+
+               /*
+                * It's been 5 seconds since our last scale event, clear the
+                * scale grp in case the group that needed the scale down isn't
+                * doing any IO currently.
+                */
+               if (now - lat_info->last_scale_event >=
+                   ((u64)NSEC_PER_SEC * 5))
+                       lat_info->scale_grp = NULL;
+next_lock:
+               spin_unlock_irqrestore(&lat_info->lock, flags);
+next:
+               blkg_put(blkg);
+       }
+       rcu_read_unlock();
+}
+
+int blk_iolatency_init(struct request_queue *q)
+{
+       struct blk_iolatency *blkiolat;
+       struct rq_qos *rqos;
+       int ret;
+
+       blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
+       if (!blkiolat)
+               return -ENOMEM;
+
+       rqos = &blkiolat->rqos;
+       rqos->id = RQ_QOS_CGROUP;
+       rqos->ops = &blkcg_iolatency_ops;
+       rqos->q = q;
+
+       rq_qos_add(q, rqos);
+
+       ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
+       if (ret) {
+               rq_qos_del(q, rqos);
+               kfree(blkiolat);
+               return ret;
+       }
+
+       timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
+
+       return 0;
+}
+
+static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
+{
+       struct iolatency_grp *iolat = blkg_to_lat(blkg);
+       struct blk_iolatency *blkiolat = iolat->blkiolat;
+       u64 oldval = iolat->min_lat_nsec;
+
+       iolat->min_lat_nsec = val;
+       iolat->cur_win_nsec = max_t(u64, val << 4, 100 * NSEC_PER_MSEC);
+       iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec, NSEC_PER_SEC);
+
+       if (!oldval && val)
+               atomic_inc(&blkiolat->enabled);
+       if (oldval && !val)
+               atomic_dec(&blkiolat->enabled);
+}
+
+static void iolatency_clear_scaling(struct blkcg_gq *blkg)
+{
+       if (blkg->parent) {
+               struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
+               struct child_latency_info *lat_info;
+               if (!iolat)
+                       return;
+
+               lat_info = &iolat->child_lat;
+               spin_lock(&lat_info->lock);
+               atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
+               lat_info->last_scale_event = 0;
+               lat_info->scale_grp = NULL;
+               lat_info->scale_lat = 0;
+               spin_unlock(&lat_info->lock);
+       }
+}
+
+static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
+                            size_t nbytes, loff_t off)
+{
+       struct blkcg *blkcg = css_to_blkcg(of_css(of));
+       struct blkcg_gq *blkg;
+       struct blk_iolatency *blkiolat;
+       struct blkg_conf_ctx ctx;
+       struct iolatency_grp *iolat;
+       char *p, *tok;
+       u64 lat_val = 0;
+       u64 oldval;
+       int ret;
+
+       ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
+       if (ret)
+               return ret;
+
+       iolat = blkg_to_lat(ctx.blkg);
+       blkiolat = iolat->blkiolat;
+       p = ctx.body;
+
+       ret = -EINVAL;
+       while ((tok = strsep(&p, " "))) {
+               char key[16];
+               char val[21];   /* 18446744073709551616 */
+
+               if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
+                       goto out;
+
+               if (!strcmp(key, "target")) {
+                       u64 v;
+
+                       if (!strcmp(val, "max"))
+                               lat_val = 0;
+                       else if (sscanf(val, "%llu", &v) == 1)
+                               lat_val = v * NSEC_PER_USEC;
+                       else
+                               goto out;
+               } else {
+                       goto out;
+               }
+       }
+
+       /* Walk up the tree to see if our new val is lower than it should be. */
+       blkg = ctx.blkg;
+       oldval = iolat->min_lat_nsec;
+
+       iolatency_set_min_lat_nsec(blkg, lat_val);
+       if (oldval != iolat->min_lat_nsec) {
+               iolatency_clear_scaling(blkg);
+       }
+
+       ret = 0;
+out:
+       blkg_conf_finish(&ctx);
+       return ret ?: nbytes;
+}
+
+static u64 iolatency_prfill_limit(struct seq_file *sf,
+                                 struct blkg_policy_data *pd, int off)
+{
+       struct iolatency_grp *iolat = pd_to_lat(pd);
+       const char *dname = blkg_dev_name(pd->blkg);
+
+       if (!dname || !iolat->min_lat_nsec)
+               return 0;
+       seq_printf(sf, "%s target=%llu\n",
+                  dname,
+                  (unsigned long long)iolat->min_lat_nsec / NSEC_PER_USEC);
+       return 0;
+}
+
+static int iolatency_print_limit(struct seq_file *sf, void *v)
+{
+       blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
+                         iolatency_prfill_limit,
+                         &blkcg_policy_iolatency, seq_cft(sf)->private, false);
+       return 0;
+}
+
+static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
+                               size_t size)
+{
+       struct iolatency_grp *iolat = pd_to_lat(pd);
+       unsigned long long avg_lat = div64_u64(iolat->total_lat_avg, NSEC_PER_USEC);
+
+       if (iolat->rq_depth.max_depth == (u64)-1)
+               return scnprintf(buf, size, " depth=max avg_lat=%llu",
+                                avg_lat);
+
+       return scnprintf(buf, size, " depth=%u avg_lat=%llu",
+                        iolat->rq_depth.max_depth, avg_lat);
+}
+
+
+static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp, int node)
+{
+       struct iolatency_grp *iolat;
+
+       iolat = kzalloc_node(sizeof(*iolat), gfp, node);
+       if (!iolat)
+               return NULL;
+       iolat->stats = __alloc_percpu_gfp(sizeof(struct blk_rq_stat),
+                                      __alignof__(struct blk_rq_stat), gfp);
+       if (!iolat->stats) {
+               kfree(iolat);
+               return NULL;
+       }
+       return &iolat->pd;
+}
+
+static void iolatency_pd_init(struct blkg_policy_data *pd)
+{
+       struct iolatency_grp *iolat = pd_to_lat(pd);
+       struct blkcg_gq *blkg = lat_to_blkg(iolat);
+       struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
+       struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
+       u64 now = ktime_to_ns(ktime_get());
+       int cpu;
+
+       for_each_possible_cpu(cpu) {
+               struct blk_rq_stat *stat;
+               stat = per_cpu_ptr(iolat->stats, cpu);
+               blk_rq_stat_init(stat);
+       }
+
+       rq_wait_init(&iolat->rq_wait);
+       spin_lock_init(&iolat->child_lat.lock);
+       iolat->rq_depth.queue_depth = blk_queue_depth(blkg->q);
+       iolat->rq_depth.max_depth = INT_MAX;
+       iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
+       iolat->blkiolat = blkiolat;
+       iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
+       atomic64_set(&iolat->window_start, now);
+
+       /*
+        * We init things in list order, so the pd for the parent may not be
+        * init'ed yet for whatever reason.
+        */
+       if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
+               struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
+               atomic_set(&iolat->scale_cookie,
+                          atomic_read(&parent->child_lat.scale_cookie));
+       } else {
+               atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
+       }
+
+       atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
+}
+
+static void iolatency_pd_offline(struct blkg_policy_data *pd)
+{
+       struct iolatency_grp *iolat = pd_to_lat(pd);
+       struct blkcg_gq *blkg = lat_to_blkg(iolat);
+
+       iolatency_set_min_lat_nsec(blkg, 0);
+       iolatency_clear_scaling(blkg);
+}
+
+static void iolatency_pd_free(struct blkg_policy_data *pd)
+{
+       struct iolatency_grp *iolat = pd_to_lat(pd);
+       free_percpu(iolat->stats);
+       kfree(iolat);
+}
+
+static struct cftype iolatency_files[] = {
+       {
+               .name = "latency",
+               .flags = CFTYPE_NOT_ON_ROOT,
+               .seq_show = iolatency_print_limit,
+               .write = iolatency_set_limit,
+       },
+       {}
+};
+
+static struct blkcg_policy blkcg_policy_iolatency = {
+       .dfl_cftypes    = iolatency_files,
+       .pd_alloc_fn    = iolatency_pd_alloc,
+       .pd_init_fn     = iolatency_pd_init,
+       .pd_offline_fn  = iolatency_pd_offline,
+       .pd_free_fn     = iolatency_pd_free,
+       .pd_stat_fn     = iolatency_pd_stat,
+};
+
+static int __init iolatency_init(void)
+{
+       return blkcg_policy_register(&blkcg_policy_iolatency);
+}
+
+static void __exit iolatency_exit(void)
+{
+       return blkcg_policy_unregister(&blkcg_policy_iolatency);
+}
+
+module_init(iolatency_init);
+module_exit(iolatency_exit);
index 8d23aea96ce9e73b100a8ee390344ba2b0c34bfd..69b14cd2bb225f87c8ed76be1409bbd4a969acd6 100644 (file)
@@ -412,4 +412,10 @@ static inline void blk_queue_bounce(struct request_queue *q, struct bio **bio)
 
 extern void blk_drain_queue(struct request_queue *q);
 
+#ifdef CONFIG_BLK_CGROUP_IOLATENCY
+extern int blk_iolatency_init(struct request_queue *q);
+#else
+static inline int blk_iolatency_init(struct request_queue *q) { return 0; }
+#endif
+
 #endif /* BLK_INTERNAL_H */
index 0ffc34c5cc83a900594c1706047f941a4b8ac979..e13449a379a172fccf536997a3ddf7ade389493c 100644 (file)
@@ -180,9 +180,7 @@ struct bio {
        struct io_context       *bi_ioc;
        struct cgroup_subsys_state *bi_css;
        struct blkcg_gq         *bi_blkg;
-#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
        struct bio_issue        bi_issue;
-#endif
 #endif
        union {
 #if defined(CONFIG_BLK_DEV_INTEGRITY)