Skip to content

Commit

Permalink
blk-iocost: implement vtime loss compensation
Browse files Browse the repository at this point in the history 
When an iocg accumulates too much vtime or gets deactivated, we throw away
some vtime, which lowers the overall device utilization. As the exact amount
which is being thrown away is known, we can compensate by accelerating the
vrate accordingly so that the extra vtime generated in the current period
matches what got lost.

This significantly improves work conservation when involving high weight
cgroups with intermittent and bursty IO patterns.

Signed-off-by: Tejun Heo <[email protected]>
Signed-off-by: Jens Axboe <[email protected]>
  • Loading branch information
@htejun @axboe
htejun authored and axboe committed Sep 2, 2020
1 parent dda1315 commit ac33e91
Showing 1 changed file with 90 additions and 42 deletions.
132 changes: 90 additions & 42 deletions block/blk-iocost.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -224,20 +224,12 @@ enum {
MARGIN_MIN_PCT = 10,
MARGIN_LOW_PCT = 20,
MARGIN_TARGET_PCT = 50,
MARGIN_MAX_PCT = 100,

INUSE_ADJ_STEP_PCT = 25,

/* Have some play in timer operations */
TIMER_SLACK_PCT = 1,

/*
* vtime can wrap well within a reasonable uptime when vrate is
* consistently raised. Don't trust recorded cgroup vtime if the
* period counter indicates that it's older than 5mins.
*/
VTIME_VALID_DUR = 300 * USEC_PER_SEC,

/* 1/64k is granular enough and can easily be handled w/ u32 */
WEIGHT_ONE = 1 << 16,

Expand Down Expand Up @@ -395,7 +387,6 @@ struct ioc_margins {
s64 min;
s64 low;
s64 target;
s64 max;
};

struct ioc_missed {
Expand Down Expand Up @@ -432,6 +423,8 @@ struct ioc {

enum ioc_running running;
atomic64_t vtime_rate;
u64 vtime_base_rate;
s64 vtime_err;

seqcount_spinlock_t period_seqcount;
u64 period_at; /* wallclock starttime */
Expand Down Expand Up @@ -760,12 +753,11 @@ static void ioc_refresh_margins(struct ioc *ioc)
{
struct ioc_margins *margins = &ioc->margins;
u32 period_us = ioc->period_us;
u64 vrate = atomic64_read(&ioc->vtime_rate);
u64 vrate = ioc->vtime_base_rate;

margins->min = (period_us * MARGIN_MIN_PCT / 100) * vrate;
margins->low = (period_us * MARGIN_LOW_PCT / 100) * vrate;
margins->target = (period_us * MARGIN_TARGET_PCT / 100) * vrate;
margins->max = (period_us * MARGIN_MAX_PCT / 100) * vrate;
}

/* latency Qos params changed, update period_us and all the dependent params */
Expand Down Expand Up @@ -831,8 +823,7 @@ static int ioc_autop_idx(struct ioc *ioc)
return idx;

/* step up/down based on the vrate */
vrate_pct = div64_u64(atomic64_read(&ioc->vtime_rate) * 100,
VTIME_PER_USEC);
vrate_pct = div64_u64(ioc->vtime_base_rate * 100, VTIME_PER_USEC);
now_ns = ktime_get_ns();

if (p->too_fast_vrate_pct && p->too_fast_vrate_pct <= vrate_pct) {
Expand Down Expand Up @@ -940,6 +931,43 @@ static bool ioc_refresh_params(struct ioc *ioc, bool force)
return true;
}

/*
* When an iocg accumulates too much vtime or gets deactivated, we throw away
* some vtime, which lowers the overall device utilization. As the exact amount
* which is being thrown away is known, we can compensate by accelerating the
* vrate accordingly so that the extra vtime generated in the current period
* matches what got lost.
*/
static void ioc_refresh_vrate(struct ioc *ioc, struct ioc_now *now)
{
s64 pleft = ioc->period_at + ioc->period_us - now->now;
s64 vperiod = ioc->period_us * ioc->vtime_base_rate;
s64 vcomp, vcomp_min, vcomp_max;

lockdep_assert_held(&ioc->lock);

/* we need some time left in this period */
if (pleft <= 0)
goto done;

/*
* Calculate how much vrate should be adjusted to offset the error.
* Limit the amount of adjustment and deduct the adjusted amount from
* the error.
*/
vcomp = -div64_s64(ioc->vtime_err, pleft);
vcomp_min = -(ioc->vtime_base_rate >> 1);
vcomp_max = ioc->vtime_base_rate;
vcomp = clamp(vcomp, vcomp_min, vcomp_max);

ioc->vtime_err += vcomp * pleft;

atomic64_set(&ioc->vtime_rate, ioc->vtime_base_rate + vcomp);
done:
/* bound how much error can accumulate */
ioc->vtime_err = clamp(ioc->vtime_err, -vperiod, vperiod);
}

/* take a snapshot of the current [v]time and vrate */
static void ioc_now(struct ioc *ioc, struct ioc_now *now)
{
Expand Down Expand Up @@ -1152,8 +1180,8 @@ static void weight_updated(struct ioc_gq *iocg, struct ioc_now *now)
static bool iocg_activate(struct ioc_gq *iocg, struct ioc_now *now)
{
struct ioc *ioc = iocg->ioc;
u64 last_period, cur_period, max_period_delta;
u64 vtime, vmin;
u64 last_period, cur_period;
u64 vtime, vtarget;
int i;

/*
Expand Down Expand Up @@ -1192,21 +1220,15 @@ static bool iocg_activate(struct ioc_gq *iocg, struct ioc_now *now)
goto fail_unlock;

/*
* vtime may wrap when vrate is raised substantially due to
* underestimated IO costs. Look at the period and ignore its
* vtime if the iocg has been idle for too long. Also, cap the
* budget it can start with to the margin.
* Always start with the target budget. On deactivation, we throw away
* anything above it.
*/
max_period_delta = DIV64_U64_ROUND_UP(VTIME_VALID_DUR, ioc->period_us);
vtarget = now->vnow - ioc->margins.target;
vtime = atomic64_read(&iocg->vtime);
vmin = now->vnow - ioc->margins.max;

if (last_period + max_period_delta < cur_period ||
time_before64(vtime, vmin)) {
atomic64_add(vmin - vtime, &iocg->vtime);
atomic64_add(vmin - vtime, &iocg->done_vtime);
vtime = vmin;
}
atomic64_add(vtarget - vtime, &iocg->vtime);
atomic64_add(vtarget - vtime, &iocg->done_vtime);
vtime = vtarget;

/*
* Activate, propagate weight and start period timer if not
Expand Down Expand Up @@ -1260,7 +1282,8 @@ static bool iocg_kick_delay(struct ioc_gq *iocg, struct ioc_now *now)
current_hweight(iocg, &hwa, NULL);
vover = atomic64_read(&iocg->vtime) +
abs_cost_to_cost(iocg->abs_vdebt, hwa) - now->vnow;
vover_pct = div64_s64(100 * vover, ioc->period_us * now->vrate);
vover_pct = div64_s64(100 * vover,
ioc->period_us * ioc->vtime_base_rate);

if (vover_pct <= MIN_DELAY_THR_PCT)
new_delay = 0;
Expand Down Expand Up @@ -1421,7 +1444,8 @@ static void iocg_kick_waitq(struct ioc_gq *iocg, bool pay_debt,
/* determine next wakeup, add a timer margin to guarantee chunking */
vshortage = -ctx.vbudget;
expires = now->now_ns +
DIV64_U64_ROUND_UP(vshortage, now->vrate) * NSEC_PER_USEC;
DIV64_U64_ROUND_UP(vshortage, ioc->vtime_base_rate) *
NSEC_PER_USEC;
expires += ioc->timer_slack_ns;

/* if already active and close enough, don't bother */
Expand Down Expand Up @@ -1536,6 +1560,7 @@ static void iocg_build_inner_walk(struct ioc_gq *iocg,
/* collect per-cpu counters and propagate the deltas to the parent */
static void iocg_flush_stat_one(struct ioc_gq *iocg, struct ioc_now *now)
{
struct ioc *ioc = iocg->ioc;
struct iocg_stat new_stat;
u64 abs_vusage = 0;
u64 vusage_delta;
Expand All @@ -1551,7 +1576,7 @@ static void iocg_flush_stat_one(struct ioc_gq *iocg, struct ioc_now *now)
vusage_delta = abs_vusage - iocg->last_stat_abs_vusage;
iocg->last_stat_abs_vusage = abs_vusage;

iocg->usage_delta_us = div64_u64(vusage_delta, now->vrate);
iocg->usage_delta_us = div64_u64(vusage_delta, ioc->vtime_base_rate);
iocg->local_stat.usage_us += iocg->usage_delta_us;

new_stat.usage_us =
Expand Down Expand Up @@ -1593,8 +1618,8 @@ static void iocg_flush_stat(struct list_head *target_iocgs, struct ioc_now *now)
* capacity. @hwm is the upper bound and used to signal no donation. This
* function also throws away @iocg's excess budget.
*/
static u32 hweight_after_donation(struct ioc_gq *iocg, u32 hwm, u32 usage,
struct ioc_now *now)
static u32 hweight_after_donation(struct ioc_gq *iocg, u32 old_hwi, u32 hwm,
u32 usage, struct ioc_now *now)
{
struct ioc *ioc = iocg->ioc;
u64 vtime = atomic64_read(&iocg->vtime);
Expand All @@ -1609,12 +1634,13 @@ static u32 hweight_after_donation(struct ioc_gq *iocg, u32 hwm, u32 usage,
time_after64(vtime, now->vnow - ioc->margins.min))
return hwm;

/* throw away excess above max */
excess = now->vnow - vtime - ioc->margins.max;
/* throw away excess above target */
excess = now->vnow - vtime - ioc->margins.target;
if (excess > 0) {
atomic64_add(excess, &iocg->vtime);
atomic64_add(excess, &iocg->done_vtime);
vtime += excess;
ioc->vtime_err -= div64_u64(excess * old_hwi, WEIGHT_ONE);
}

/*
Expand Down Expand Up @@ -1952,6 +1978,24 @@ static void ioc_timer_fn(struct timer_list *timer)
nr_debtors++;
} else if (iocg_is_idle(iocg)) {
/* no waiter and idle, deactivate */
u64 vtime = atomic64_read(&iocg->vtime);
s64 excess;

/*
* @iocg has been inactive for a full duration and will
* have a high budget. Account anything above target as
* error and throw away. On reactivation, it'll start
* with the target budget.
*/
excess = now.vnow - vtime - ioc->margins.target;
if (excess > 0) {
u32 old_hwi;

current_hweight(iocg, NULL, &old_hwi);
ioc->vtime_err -= div64_u64(excess * old_hwi,
WEIGHT_ONE);
}

__propagate_weights(iocg, 0, 0, false, &now);
list_del_init(&iocg->active_list);
}
Expand Down Expand Up @@ -1997,7 +2041,7 @@ static void ioc_timer_fn(struct timer_list *timer)
if (vdone != vtime) {
u64 inflight_us = DIV64_U64_ROUND_UP(
cost_to_abs_cost(vtime - vdone, hw_inuse),
now.vrate);
ioc->vtime_base_rate);
usage_us = max(usage_us, inflight_us);
}

Expand All @@ -2017,16 +2061,16 @@ static void ioc_timer_fn(struct timer_list *timer)
if (hw_inuse < hw_active ||
(!waitqueue_active(&iocg->waitq) &&
time_before64(vtime, now.vnow - ioc->margins.low))) {
u32 hwa, hwm, new_hwi;
u32 hwa, old_hwi, hwm, new_hwi;

/*
* Already donating or accumulated enough to start.
* Determine the donation amount.
*/
current_hweight(iocg, &hwa, NULL);
current_hweight(iocg, &hwa, &old_hwi);
hwm = current_hweight_max(iocg);
new_hwi = hweight_after_donation(iocg, hwm, usage,
&now);
new_hwi = hweight_after_donation(iocg, old_hwi, hwm,
usage, &now);
if (new_hwi < hwm) {
iocg->hweight_donating = hwa;
iocg->hweight_after_donation = new_hwi;
Expand Down Expand Up @@ -2130,7 +2174,7 @@ static void ioc_timer_fn(struct timer_list *timer)
ioc->busy_level = clamp(ioc->busy_level, -1000, 1000);

if (ioc->busy_level > 0 || (ioc->busy_level < 0 && !nr_lagging)) {
u64 vrate = atomic64_read(&ioc->vtime_rate);
u64 vrate = ioc->vtime_base_rate;
u64 vrate_min = ioc->vrate_min, vrate_max = ioc->vrate_max;

/* rq_wait signal is always reliable, ignore user vrate_min */
Expand Down Expand Up @@ -2167,7 +2211,7 @@ static void ioc_timer_fn(struct timer_list *timer)
trace_iocost_ioc_vrate_adj(ioc, vrate, missed_ppm, rq_wait_pct,
nr_lagging, nr_shortages);

atomic64_set(&ioc->vtime_rate, vrate);
ioc->vtime_base_rate = vrate;
ioc_refresh_margins(ioc);
} else if (ioc->busy_level != prev_busy_level || nr_lagging) {
trace_iocost_ioc_vrate_adj(ioc, atomic64_read(&ioc->vtime_rate),
Expand All @@ -2188,8 +2232,11 @@ static void ioc_timer_fn(struct timer_list *timer)
ioc_start_period(ioc, &now);
} else {
ioc->busy_level = 0;
ioc->vtime_err = 0;
ioc->running = IOC_IDLE;
}

ioc_refresh_vrate(ioc, &now);
}

spin_unlock_irq(&ioc->lock);
Expand Down Expand Up @@ -2628,6 +2675,7 @@ static int blk_iocost_init(struct request_queue *q)
INIT_LIST_HEAD(&ioc->active_iocgs);

ioc->running = IOC_IDLE;
ioc->vtime_base_rate = VTIME_PER_USEC;
atomic64_set(&ioc->vtime_rate, VTIME_PER_USEC);
seqcount_spinlock_init(&ioc->period_seqcount, &ioc->lock);
ioc->period_at = ktime_to_us(ktime_get());
Expand Down Expand Up @@ -2762,7 +2810,7 @@ static size_t ioc_pd_stat(struct blkg_policy_data *pd, char *buf, size_t size)

if (iocg->level == 0) {
unsigned vp10k = DIV64_U64_ROUND_CLOSEST(
atomic64_read(&ioc->vtime_rate) * 10000,
ioc->vtime_base_rate * 10000,
VTIME_PER_USEC);
pos += scnprintf(buf + pos, size - pos, " cost.vrate=%u.%02u",
vp10k / 100, vp10k % 100);
Expand Down

0 comments on commit ac33e91

Please sign in to comment.