AuxXxilium/linux_dsm_epyc7002
1
0
Fork 0
mirror of https://github.com/AuxXxilium/linux_dsm_epyc7002.git synced 2024-12-16 10:06:42 +07:00
Code Issues Actions Packages Projects Releases Wiki Activity
5200ab6a32
linux_dsm_epyc7002/block/blk-iolatency.c

1070 lines
28 KiB
C
Raw Normal View History

block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
/*
* 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.
blk-iolatency: fix max_depth comparisons max_depth used to be a u64, but I changed it to a unsigned int but didn't convert my comparisons over everywhere. Fix by using UINT_MAX everywhere instead of (u64)-1. Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-11 21:34:42 +07:00
* - By default there's no throttling, we set the queue_depth to UINT_MAX so
* that we can have as many outstanding bio's as we're allowed to. Only at
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
* 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>
block: make iolatency avg_lat exponentially decay Currently, avg_lat is calculated by accumulating the mean of every window in a long running cumulative average. As time goes on, the metric becomes less and less useful due to the accumulated history. This patch reuses the same calculation done in load averages to make the avg_lat metric more lively. Unlike load averages, the avg only advances when a window elapses (due to an io). Idle periods extend the most recent window. Bucketing is used to limit the history of avg_lat by binding it to the window size. So, the window range for 1/exp (decay rate) is [1 min, 2.5 min) when windows elapse immediately. The current sample window size is exposed in the debug info to enable calculation of the window range. Signed-off-by: Dennis Zhou <dennisszhou@gmail.com> Acked-by: Tejun Heo <tj@kernel.org> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-08-02 13:15:41 +07:00
#include <linux/sched/loadavg.h>
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
#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;
};
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
struct percentile_stats {
u64 total;
u64 missed;
};
struct latency_stat {
union {
struct percentile_stats ps;
struct blk_rq_stat rqs;
};
};
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
struct iolatency_grp {
struct blkg_policy_data pd;
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
struct latency_stat __percpu *stats;
blk-iolatency: keep track of previous windows stats We apply a smoothing to the scale changes in order to keep sawtoothy behavior from occurring. However our window for checking if we've missed our target can sometimes be lower than the smoothing interval (500ms), especially on faster drives like ssd's. In order to deal with this keep track of the running tally of the previous intervals that we threw away because we had already done a scale event recently. This is needed for the ssd case as these low latency drives will have bursts of latency, and if it happens to be ok for the window that directly follows the opening of the scale window we could unthrottle when previous windows we were missing our target. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:43 +07:00
struct latency_stat cur_stat;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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. */
block: make iolatency avg_lat exponentially decay Currently, avg_lat is calculated by accumulating the mean of every window in a long running cumulative average. As time goes on, the metric becomes less and less useful due to the accumulated history. This patch reuses the same calculation done in load averages to make the avg_lat metric more lively. Unlike load averages, the avg only advances when a window elapses (due to an io). Idle periods extend the most recent window. Bucketing is used to limit the history of avg_lat by binding it to the window size. So, the window range for 1/exp (decay rate) is [1 min, 2.5 min) when windows elapse immediately. The current sample window size is exposed in the debug info to enable calculation of the window range. Signed-off-by: Dennis Zhou <dennisszhou@gmail.com> Acked-by: Tejun Heo <tj@kernel.org> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-08-02 13:15:41 +07:00
u64 lat_avg;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
/* Our current number of IO's for the last summation. */
u64 nr_samples;
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
bool ssd;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
struct child_latency_info child_lat;
};
block: make iolatency avg_lat exponentially decay Currently, avg_lat is calculated by accumulating the mean of every window in a long running cumulative average. As time goes on, the metric becomes less and less useful due to the accumulated history. This patch reuses the same calculation done in load averages to make the avg_lat metric more lively. Unlike load averages, the avg only advances when a window elapses (due to an io). Idle periods extend the most recent window. Bucketing is used to limit the history of avg_lat by binding it to the window size. So, the window range for 1/exp (decay rate) is [1 min, 2.5 min) when windows elapse immediately. The current sample window size is exposed in the debug info to enable calculation of the window range. Signed-off-by: Dennis Zhou <dennisszhou@gmail.com> Acked-by: Tejun Heo <tj@kernel.org> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-08-02 13:15:41 +07:00
#define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
#define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
/*
* These are the constants used to fake the fixed-point moving average
sched: loadavg: consolidate LOAD_INT, LOAD_FRAC, CALC_LOAD There are several definitions of those functions/macros in places that mess with fixed-point load averages. Provide an official version. [akpm@linux-foundation.org: fix missed conversion in block/blk-iolatency.c] Link: http://lkml.kernel.org/r/20180828172258.3185-5-hannes@cmpxchg.org Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Suren Baghdasaryan <surenb@google.com> Tested-by: Daniel Drake <drake@endlessm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Johannes Weiner <jweiner@fb.com> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Enderborg <peter.enderborg@sony.com> Cc: Randy Dunlap <rdunlap@infradead.org> Cc: Shakeel Butt <shakeelb@google.com> Cc: Tejun Heo <tj@kernel.org> Cc: Vinayak Menon <vinmenon@codeaurora.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 05:06:11 +07:00
* calculation just like load average. The call to calc_load() folds
block: make iolatency avg_lat exponentially decay Currently, avg_lat is calculated by accumulating the mean of every window in a long running cumulative average. As time goes on, the metric becomes less and less useful due to the accumulated history. This patch reuses the same calculation done in load averages to make the avg_lat metric more lively. Unlike load averages, the avg only advances when a window elapses (due to an io). Idle periods extend the most recent window. Bucketing is used to limit the history of avg_lat by binding it to the window size. So, the window range for 1/exp (decay rate) is [1 min, 2.5 min) when windows elapse immediately. The current sample window size is exposed in the debug info to enable calculation of the window range. Signed-off-by: Dennis Zhou <dennisszhou@gmail.com> Acked-by: Tejun Heo <tj@kernel.org> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-08-02 13:15:41 +07:00
* (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg. The sampling
* window size is bucketed to try to approximately calculate average
* latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
* elapse immediately. Note, windows only elapse with IO activity. Idle
* periods extend the most recent window.
*/
#define BLKIOLATENCY_NR_EXP_FACTORS 5
#define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
(BLKIOLATENCY_NR_EXP_FACTORS - 1))
static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
2045, // exp(1/600) - 600 samples
2039, // exp(1/240) - 240 samples
2031, // exp(1/120) - 120 samples
2023, // exp(1/80) - 80 samples
2014, // exp(1/60) - 60 samples
};
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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);
}
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
static inline void latency_stat_init(struct iolatency_grp *iolat,
struct latency_stat *stat)
{
if (iolat->ssd) {
stat->ps.total = 0;
stat->ps.missed = 0;
} else
blk_rq_stat_init(&stat->rqs);
}
static inline void latency_stat_sum(struct iolatency_grp *iolat,
struct latency_stat *sum,
struct latency_stat *stat)
{
if (iolat->ssd) {
sum->ps.total += stat->ps.total;
sum->ps.missed += stat->ps.missed;
} else
blk_rq_stat_sum(&sum->rqs, &stat->rqs);
}
static inline void latency_stat_record_time(struct iolatency_grp *iolat,
u64 req_time)
{
struct latency_stat *stat = get_cpu_ptr(iolat->stats);
if (iolat->ssd) {
if (req_time >= iolat->min_lat_nsec)
stat->ps.missed++;
stat->ps.total++;
} else
blk_rq_stat_add(&stat->rqs, req_time);
put_cpu_ptr(stat);
}
static inline bool latency_sum_ok(struct iolatency_grp *iolat,
struct latency_stat *stat)
{
if (iolat->ssd) {
u64 thresh = div64_u64(stat->ps.total, 10);
thresh = max(thresh, 1ULL);
return stat->ps.missed < thresh;
}
return stat->rqs.mean <= iolat->min_lat_nsec;
}
static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
struct latency_stat *stat)
{
if (iolat->ssd)
return stat->ps.total;
return stat->rqs.nr_samples;
}
static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
struct latency_stat *stat)
{
int exp_idx;
if (iolat->ssd)
return;
/*
sched: loadavg: consolidate LOAD_INT, LOAD_FRAC, CALC_LOAD There are several definitions of those functions/macros in places that mess with fixed-point load averages. Provide an official version. [akpm@linux-foundation.org: fix missed conversion in block/blk-iolatency.c] Link: http://lkml.kernel.org/r/20180828172258.3185-5-hannes@cmpxchg.org Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Suren Baghdasaryan <surenb@google.com> Tested-by: Daniel Drake <drake@endlessm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Johannes Weiner <jweiner@fb.com> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Enderborg <peter.enderborg@sony.com> Cc: Randy Dunlap <rdunlap@infradead.org> Cc: Shakeel Butt <shakeelb@google.com> Cc: Tejun Heo <tj@kernel.org> Cc: Vinayak Menon <vinmenon@codeaurora.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 05:06:11 +07:00
* calc_load() takes in a number stored in fixed point representation.
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
* Because we are using this for IO time in ns, the values stored
* are significantly larger than the FIXED_1 denominator (2048).
* Therefore, rounding errors in the calculation are negligible and
* can be ignored.
*/
exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
div64_u64(iolat->cur_win_nsec,
BLKIOLATENCY_EXP_BUCKET_SIZE));
sched: loadavg: consolidate LOAD_INT, LOAD_FRAC, CALC_LOAD There are several definitions of those functions/macros in places that mess with fixed-point load averages. Provide an official version. [akpm@linux-foundation.org: fix missed conversion in block/blk-iolatency.c] Link: http://lkml.kernel.org/r/20180828172258.3185-5-hannes@cmpxchg.org Signed-off-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Tested-by: Suren Baghdasaryan <surenb@google.com> Tested-by: Daniel Drake <drake@endlessm.com> Cc: Christopher Lameter <cl@linux.com> Cc: Ingo Molnar <mingo@redhat.com> Cc: Johannes Weiner <jweiner@fb.com> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Enderborg <peter.enderborg@sony.com> Cc: Randy Dunlap <rdunlap@infradead.org> Cc: Shakeel Butt <shakeelb@google.com> Cc: Tejun Heo <tj@kernel.org> Cc: Vinayak Menon <vinmenon@codeaurora.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-10-27 05:06:11 +07:00
iolat->lat_avg = calc_load(iolat->lat_avg,
iolatency_exp_factors[exp_idx],
stat->rqs.mean);
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
}
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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)
{
blk-iolatency: use q->nr_requests directly We were using blk_queue_depth() assuming that it would return nr_requests, but we hit a case in production on drives that had to have NCQ turned off in order for them to not shit the bed which resulted in a qd of 1, even though the nr_requests was much larger. iolatency really only cares about requests we are allowed to queue up, as any io that get's onto the request list is going to be serviced soonish, so we want to be throttling before the bio gets onto the request list. To make iolatency work as expected, simply use q->nr_requests instead of blk_queue_depth() as that is what we actually care about. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:39 +07:00
unsigned long qd = blkiolat->rqos.q->nr_requests;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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)
{
blk-iolatency: use q->nr_requests directly We were using blk_queue_depth() assuming that it would return nr_requests, but we hit a case in production on drives that had to have NCQ turned off in order for them to not shit the bed which resulted in a qd of 1, even though the nr_requests was much larger. iolatency really only cares about requests we are allowed to queue up, as any io that get's onto the request list is going to be serviced soonish, so we want to be throttling before the bio gets onto the request list. To make iolatency work as expected, simply use q->nr_requests instead of blk_queue_depth() as that is what we actually care about. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:39 +07:00
unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
unsigned long scale = scale_amount(qd, up);
unsigned long old = iolat->rq_depth.max_depth;
if (old > qd)
old = qd;
if (up) {
if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
return;
if (old < qd) {
old += scale;
old = min(old, qd);
iolat->rq_depth.max_depth = old;
wake_up_all(&iolat->rq_wait.wait);
}
blk-iolatency: deal with nr_requests == 1 Hitting the case where blk_queue_depth() returned 1 uncovered the fact that iolatency doesn't actually handle this case properly, it simply doesn't scale down anybody. For this case we should go straight into applying the time delay, which we weren't doing. Since we already limit the floor at 1 request this if statement is not needed, and this allows us to set our depth to 1 which allows us to apply the delay if needed. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:40 +07:00
} else {
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
old >>= 1;
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;
blk-iolatency: deal with small samples There is logic to keep cgroups that haven't done a lot of IO in the most recent scale window from being punished for over-active higher priority groups. However for things like ssd's where the windows are pretty short we'll end up with small numbers of samples, so 5% of samples will come out to 0 if there aren't enough. Make the floor 1 sample to keep us from improperly bailing out of scaling down. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:41 +07:00
samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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));
blk-iolatency: fix max_depth comparisons max_depth used to be a u64, but I changed it to a unsigned int but didn't convert my comparisons over everywhere. Fix by using UINT_MAX everywhere instead of (u64)-1. Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-11 21:34:42 +07:00
iolat->rq_depth.max_depth = UINT_MAX;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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);
blkcg: revert blkcg cleanups series This reverts a series committed earlier due to null pointer exception bug report in [1]. It seems there are edge case interactions that I did not consider and will need some time to understand what causes the adverse interactions. The original series can be found in [2] with a follow up series in [3]. [1] https://www.spinics.net/lists/cgroups/msg20719.html [2] https://lore.kernel.org/lkml/20180911184137.35897-1-dennisszhou@gmail.com/ [3] https://lore.kernel.org/lkml/20181020185612.51587-1-dennis@kernel.org/ This reverts the following commits: d459d853c2ed, b2c3fa546705, 101246ec02b5, b3b9f24f5fcc, e2b0989954ae, f0fcb3ec89f3, c839e7a03f92, bdc2491708c4, 74b7c02a9bc1, 5bf9a1f3b4ef, a7b39b4e961c, 07b05bcc3213, 49f4c2dc2b50, 27e6fa996c53 Signed-off-by: Dennis Zhou <dennis@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-11-02 04:24:10 +07:00
struct blkcg *blkcg;
struct blkcg_gq *blkg;
struct request_queue *q = rqos->q;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
bool issue_as_root = bio_issue_as_root_blkg(bio);
if (!blk_iolatency_enabled(blkiolat))
return;
blkcg: revert blkcg cleanups series This reverts a series committed earlier due to null pointer exception bug report in [1]. It seems there are edge case interactions that I did not consider and will need some time to understand what causes the adverse interactions. The original series can be found in [2] with a follow up series in [3]. [1] https://www.spinics.net/lists/cgroups/msg20719.html [2] https://lore.kernel.org/lkml/20180911184137.35897-1-dennisszhou@gmail.com/ [3] https://lore.kernel.org/lkml/20181020185612.51587-1-dennis@kernel.org/ This reverts the following commits: d459d853c2ed, b2c3fa546705, 101246ec02b5, b3b9f24f5fcc, e2b0989954ae, f0fcb3ec89f3, c839e7a03f92, bdc2491708c4, 74b7c02a9bc1, 5bf9a1f3b4ef, a7b39b4e961c, 07b05bcc3213, 49f4c2dc2b50, 27e6fa996c53 Signed-off-by: Dennis Zhou <dennis@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-11-02 04:24:10 +07:00
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();
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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)
{
u64 start = bio_issue_time(issue);
u64 req_time;
blk-iolatency: truncate our current time In our longer tests we noticed that some boxes would degrade to the point of uselessness. This is because we truncate the current time when saving it in our bio, but I was using the raw current time to subtract from. So once the box had been up a certain amount of time it would appear as if our IO's were taking several years to complete. Fix this by truncating the current time so it matches the issue time. Verified this worked by running with this patch for a week on our test tier. Signed-off-by: Josef Bacik <jbacik@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-16 23:12:23 +07:00
/*
* Have to do this so we are truncated to the correct time that our
* issue is truncated to.
*/
now = __bio_issue_time(now);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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.
*/
blk-iolatency: fix max_depth comparisons max_depth used to be a u64, but I changed it to a unsigned int but didn't convert my comparisons over everywhere. Fix by using UINT_MAX everywhere instead of (u64)-1. Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-11 21:34:42 +07:00
if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
u64 sub = iolat->min_lat_nsec;
if (req_time < sub)
blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
return;
}
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
latency_stat_record_time(iolat, req_time);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
}
#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;
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
struct latency_stat stat;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
unsigned long flags;
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
int cpu;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
latency_stat_init(iolat, &stat);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
preempt_disable();
for_each_online_cpu(cpu) {
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
struct latency_stat *s;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
s = per_cpu_ptr(iolat->stats, cpu);
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
latency_stat_sum(iolat, &stat, s);
latency_stat_init(iolat, s);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
}
preempt_enable();
parent = blkg_to_lat(blkg->parent);
if (!parent)
return;
lat_info = &parent->child_lat;
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
iolat_update_total_lat_avg(iolat, &stat);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
/* Everything is ok and we don't need to adjust the scale. */
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
if (latency_sum_ok(iolat, &stat) &&
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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);
blk-iolatency: keep track of previous windows stats We apply a smoothing to the scale changes in order to keep sawtoothy behavior from occurring. However our window for checking if we've missed our target can sometimes be lower than the smoothing interval (500ms), especially on faster drives like ssd's. In order to deal with this keep track of the running tally of the previous intervals that we threw away because we had already done a scale event recently. This is needed for the ssd case as these low latency drives will have bursts of latency, and if it happens to be ok for the window that directly follows the opening of the scale window we could unthrottle when previous windows we were missing our target. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:43 +07:00
latency_stat_sum(iolat, &iolat->cur_stat, &stat);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
lat_info->nr_samples -= iolat->nr_samples;
blk-iolatency: keep track of previous windows stats We apply a smoothing to the scale changes in order to keep sawtoothy behavior from occurring. However our window for checking if we've missed our target can sometimes be lower than the smoothing interval (500ms), especially on faster drives like ssd's. In order to deal with this keep track of the running tally of the previous intervals that we threw away because we had already done a scale event recently. This is needed for the ssd case as these low latency drives will have bursts of latency, and if it happens to be ok for the window that directly follows the opening of the scale window we could unthrottle when previous windows we were missing our target. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:43 +07:00
lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
if ((lat_info->last_scale_event >= now ||
blk-iolatency: keep track of previous windows stats We apply a smoothing to the scale changes in order to keep sawtoothy behavior from occurring. However our window for checking if we've missed our target can sometimes be lower than the smoothing interval (500ms), especially on faster drives like ssd's. In order to deal with this keep track of the running tally of the previous intervals that we threw away because we had already done a scale event recently. This is needed for the ssd case as these low latency drives will have bursts of latency, and if it happens to be ok for the window that directly follows the opening of the scale window we could unthrottle when previous windows we were missing our target. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:43 +07:00
now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
goto out;
blk-iolatency: keep track of previous windows stats We apply a smoothing to the scale changes in order to keep sawtoothy behavior from occurring. However our window for checking if we've missed our target can sometimes be lower than the smoothing interval (500ms), especially on faster drives like ssd's. In order to deal with this keep track of the running tally of the previous intervals that we threw away because we had already done a scale event recently. This is needed for the ssd case as these low latency drives will have bursts of latency, and if it happens to be ok for the window that directly follows the opening of the scale window we could unthrottle when previous windows we were missing our target. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:43 +07:00
if (latency_sum_ok(iolat, &iolat->cur_stat) &&
latency_sum_ok(iolat, &stat)) {
if (latency_stat_samples(iolat, &iolat->cur_stat) <
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
BLKIOLATENCY_MIN_GOOD_SAMPLES)
goto out;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
if (lat_info->scale_grp == iolat) {
lat_info->last_scale_event = now;
scale_cookie_change(iolat->blkiolat, lat_info, true);
}
blk-iolatency: keep track of previous windows stats We apply a smoothing to the scale changes in order to keep sawtoothy behavior from occurring. However our window for checking if we've missed our target can sometimes be lower than the smoothing interval (500ms), especially on faster drives like ssd's. In order to deal with this keep track of the running tally of the previous intervals that we threw away because we had already done a scale event recently. This is needed for the ssd case as these low latency drives will have bursts of latency, and if it happens to be ok for the window that directly follows the opening of the scale window we could unthrottle when previous windows we were missing our target. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:43 +07:00
} else if (lat_info->scale_lat == 0 ||
lat_info->scale_lat >= iolat->min_lat_nsec) {
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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);
}
blk-iolatency: keep track of previous windows stats We apply a smoothing to the scale changes in order to keep sawtoothy behavior from occurring. However our window for checking if we've missed our target can sometimes be lower than the smoothing interval (500ms), especially on faster drives like ssd's. In order to deal with this keep track of the running tally of the previous intervals that we threw away because we had already done a scale event recently. This is needed for the ssd case as these low latency drives will have bursts of latency, and if it happens to be ok for the window that directly follows the opening of the scale window we could unthrottle when previous windows we were missing our target. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:43 +07:00
latency_stat_init(iolat, &iolat->cur_stat);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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.
*/
blkcg: revert blkcg cleanups series This reverts a series committed earlier due to null pointer exception bug report in [1]. It seems there are edge case interactions that I did not consider and will need some time to understand what causes the adverse interactions. The original series can be found in [2] with a follow up series in [3]. [1] https://www.spinics.net/lists/cgroups/msg20719.html [2] https://lore.kernel.org/lkml/20180911184137.35897-1-dennisszhou@gmail.com/ [3] https://lore.kernel.org/lkml/20181020185612.51587-1-dennis@kernel.org/ This reverts the following commits: d459d853c2ed, b2c3fa546705, 101246ec02b5, b3b9f24f5fcc, e2b0989954ae, f0fcb3ec89f3, c839e7a03f92, bdc2491708c4, 74b7c02a9bc1, 5bf9a1f3b4ef, a7b39b4e961c, 07b05bcc3213, 49f4c2dc2b50, 27e6fa996c53 Signed-off-by: Dennis Zhou <dennis@kernel.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-11-02 04:24:10 +07:00
if (!blkg_try_get(blkg))
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
continue;
iolat = blkg_to_lat(blkg);
if (!iolat)
blk-iolatency: fix blkg leak in timer_fn At this point we have a ref on the blkg, we need to drop it if we don't have a iolat. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-31 23:39:02 +07:00
goto next;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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;
block: make iolatency avg_lat exponentially decay Currently, avg_lat is calculated by accumulating the mean of every window in a long running cumulative average. As time goes on, the metric becomes less and less useful due to the accumulated history. This patch reuses the same calculation done in load averages to make the avg_lat metric more lively. Unlike load averages, the avg only advances when a window elapses (due to an io). Idle periods extend the most recent window. Bucketing is used to limit the history of avg_lat by binding it to the window size. So, the window range for 1/exp (decay rate) is [1 min, 2.5 min) when windows elapse immediately. The current sample window size is exposed in the debug info to enable calculation of the window range. Signed-off-by: Dennis Zhou <dennisszhou@gmail.com> Acked-by: Tejun Heo <tj@kernel.org> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-08-02 13:15:41 +07:00
iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
BLKIOLATENCY_MAX_WIN_SIZE);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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 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);
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",
block: iolatency: avoid 64-bit division On 32-bit architectures, dividing a 64-bit number needs to use the do_div() function or something like it to avoid a link failure: block/blk-iolatency.o: In function `iolatency_prfill_limit': blk-iolatency.c:(.text+0x8cc): undefined reference to `__aeabi_uldivmod' Using div_u64() gives us the best output and avoids the need for an explicit cast. Fixes: d70675121546 ("block: introduce blk-iolatency io controller") Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-10 22:21:34 +07:00
dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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;
}
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf,
size_t size)
{
struct latency_stat stat;
int cpu;
latency_stat_init(iolat, &stat);
preempt_disable();
for_each_online_cpu(cpu) {
struct latency_stat *s;
s = per_cpu_ptr(iolat->stats, cpu);
latency_stat_sum(iolat, &stat, s);
}
preempt_enable();
if (iolat->rq_depth.max_depth == UINT_MAX)
return scnprintf(buf, size, " missed=%llu total=%llu depth=max",
(unsigned long long)stat.ps.missed,
(unsigned long long)stat.ps.total);
return scnprintf(buf, size, " missed=%llu total=%llu depth=%u",
(unsigned long long)stat.ps.missed,
(unsigned long long)stat.ps.total,
iolat->rq_depth.max_depth);
}
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
size_t size)
{
struct iolatency_grp *iolat = pd_to_lat(pd);
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
unsigned long long avg_lat;
unsigned long long cur_win;
if (iolat->ssd)
return iolatency_ssd_stat(iolat, buf, size);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
blk-iolatency: fix max_depth comparisons max_depth used to be a u64, but I changed it to a unsigned int but didn't convert my comparisons over everywhere. Fix by using UINT_MAX everywhere instead of (u64)-1. Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-11 21:34:42 +07:00
if (iolat->rq_depth.max_depth == UINT_MAX)
block: make iolatency avg_lat exponentially decay Currently, avg_lat is calculated by accumulating the mean of every window in a long running cumulative average. As time goes on, the metric becomes less and less useful due to the accumulated history. This patch reuses the same calculation done in load averages to make the avg_lat metric more lively. Unlike load averages, the avg only advances when a window elapses (due to an io). Idle periods extend the most recent window. Bucketing is used to limit the history of avg_lat by binding it to the window size. So, the window range for 1/exp (decay rate) is [1 min, 2.5 min) when windows elapse immediately. The current sample window size is exposed in the debug info to enable calculation of the window range. Signed-off-by: Dennis Zhou <dennisszhou@gmail.com> Acked-by: Tejun Heo <tj@kernel.org> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-08-02 13:15:41 +07:00
return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu",
avg_lat, cur_win);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
block: make iolatency avg_lat exponentially decay Currently, avg_lat is calculated by accumulating the mean of every window in a long running cumulative average. As time goes on, the metric becomes less and less useful due to the accumulated history. This patch reuses the same calculation done in load averages to make the avg_lat metric more lively. Unlike load averages, the avg only advances when a window elapses (due to an io). Idle periods extend the most recent window. Bucketing is used to limit the history of avg_lat by binding it to the window size. So, the window range for 1/exp (decay rate) is [1 min, 2.5 min) when windows elapse immediately. The current sample window size is exposed in the debug info to enable calculation of the window range. Signed-off-by: Dennis Zhou <dennisszhou@gmail.com> Acked-by: Tejun Heo <tj@kernel.org> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Acked-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-08-02 13:15:41 +07:00
return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
iolat->rq_depth.max_depth, avg_lat, cur_win);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
}
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;
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
__alignof__(struct latency_stat), gfp);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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;
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
if (blk_queue_nonrot(blkg->q))
iolat->ssd = true;
else
iolat->ssd = false;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
for_each_possible_cpu(cpu) {
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
struct latency_stat *stat;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
stat = per_cpu_ptr(iolat->stats, cpu);
blk-iolatency: use a percentile approache for ssd's We use an average latency approach for determining if we're missing our latency target. This works well for rotational storage where we have generally consistent latencies, but for ssd's and other low latency devices you have more of a spikey behavior, which means we often won't throttle misbehaving groups because a lot of IO completes at drastically faster times than our latency target. Instead keep track of how many IO's miss our target and how many IO's are done in our time window. If the p(90) latency is above our target then we know we need to throttle. With this change in place we are seeing the same throttling behavior with our testcase on ssd's as we see with rotational drives. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:42 +07:00
latency_stat_init(iolat, stat);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
}
blk-iolatency: keep track of previous windows stats We apply a smoothing to the scale changes in order to keep sawtoothy behavior from occurring. However our window for checking if we've missed our target can sometimes be lower than the smoothing interval (500ms), especially on faster drives like ssd's. In order to deal with this keep track of the running tally of the previous intervals that we threw away because we had already done a scale event recently. This is needed for the ssd case as these low latency drives will have bursts of latency, and if it happens to be ok for the window that directly follows the opening of the scale window we could unthrottle when previous windows we were missing our target. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:43 +07:00
latency_stat_init(iolat, &iolat->cur_stat);
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
rq_wait_init(&iolat->rq_wait);
spin_lock_init(&iolat->child_lat.lock);
blk-iolatency: use q->nr_requests directly We were using blk_queue_depth() assuming that it would return nr_requests, but we hit a case in production on drives that had to have NCQ turned off in order for them to not shit the bed which resulted in a qd of 1, even though the nr_requests was much larger. iolatency really only cares about requests we are allowed to queue up, as any io that get's onto the request list is going to be serviced soonish, so we want to be throttling before the bio gets onto the request list. To make iolatency work as expected, simply use q->nr_requests instead of blk_queue_depth() as that is what we actually care about. Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-09-29 00:45:39 +07:00
iolat->rq_depth.queue_depth = blkg->q->nr_requests;
blk-iolatency: fix max_depth comparisons max_depth used to be a u64, but I changed it to a unsigned int but didn't convert my comparisons over everywhere. Fix by using UINT_MAX everywhere instead of (u64)-1. Reported-by: Dan Carpenter <dan.carpenter@oracle.com> Signed-off-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
2018-07-11 21:34:42 +07:00
iolat->rq_depth.max_depth = UINT_MAX;
block: introduce blk-iolatency io controller 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>
2018-07-03 22:15:01 +07:00
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);
Reference in New Issue Copy Permalink