Debt reduction was blocked if any iocg was short on budget in the past
period to avoid reducing debts while some iocgs are saturated. However, this
ends up unnecessarily blocking debt reduction due to temporary local
imbalances when the device is generally being underutilized, while also
failing to block when the underlying device is overwhelmed and the usage
becomes low from high latency.
Given that debt accumulation mostly happens with swapout bursts which can
significantly deteriorate the underlying device's latency response, the
current logic is not great.
Let's replace it with ioc->busy_level based condition so that we block debt
reduction when the underlying device is being saturated. ioc_forgive_debts()
call is moved after busy_level determination.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Debt reduction logic is going to be improved and expanded. Factor it out
into ioc_forgive_debts() and generalize the comment a bit. No functional
change.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
adjust_inuse_and_calc_cost() is responsible for reducing the amount of
donated weights dynamically in period as the budget runs low. Because we
don't want to do full donation calculation in period, we keep latching up
inuse by INUSE_ADJ_STEP_PCT of the active weight of the cgroup until the
resulting hweight_inuse is satisfactory.
Unfortunately, the adj_step calculation was reading the active weight before
acquiring ioc->lock. Because the current thread could have lost race to
activate the iocg to another thread before entering this function, it may
read the active weight as zero before acquiring ioc->lock. When this
happens, the adj_step is calculated as zero and the incremental adjustment
loop becomes an infinite one.
Fix it by fetching the active weight after acquiring ioc->lock.
Fixes: b0853ab4a2 ("blk-iocost: revamp in-period donation snapbacks")
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Conceptually, root_iocg->hweight_donating must be less than WEIGHT_ONE but
all hweight calculations round up and thus it may end up >= WEIGHT_ONE
triggering divide-by-zero and other issues. Bound the value to avoid
surprises.
Fixes: e08d02aa5f ("blk-iocost: implement Andy's method for donation weight updates")
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
These are really cheap to collect and can be useful in debugging iocost
behavior. Add them as debug stats for now.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
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 <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A low weight iocg can amass a large amount of debt, for example, when
anonymous memory gets reclaimed aggressively. If the system has a lot of
memory paired with a slow IO device, the debt can span multiple seconds or
more. If there are no other subsequent IO issuers, the in-debt iocg may end
up blocked paying its debt while the IO device is idle.
This patch implements a mechanism to protect against such pathological
cases. If the device has been sufficiently idle for a substantial amount of
time, the debts are halved. The criteria are on the conservative side as we
want to resolve the rare extreme cases without impacting regular operation
by forgiving debts too readily.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Curently, iocost syncs the delay duration to the outstanding debt amount,
which seemed enough to protect the system from anon memory hogs. However,
that was mostly because the delay calcuation was using hweight_inuse which
quickly converges towards zero under debt for delay duration calculation,
often pusnishing debtors overly harshly for longer than deserved.
The previous patch fixed the delay calcuation and now the protection against
anonymous memory hogs isn't enough because the effect of delay is indirect
and non-linear and a huge amount of future debt can accumulate abruptly
while unthrottled.
This patch implements delay hysteresis so that delay is decayed
exponentially over time instead of getting cleared immediately as debt is
paid off. While the overall behavior is similar to the blk-cgroup
implementation used by blk-iolatency, a lot of the details are different and
due to the empirical nature of the mechanism, it's challenging to adapt the
mechanism for one controller without negatively impacting the other.
As the delay is gradually decayed now, there's no point in running it from
its own hrtimer. Periodic updates are now performed from ioc_timer_fn() and
the dedicated hrtimer is removed.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Debt handling had several issues.
* How much inuse a debtor carries wasn't clearly defined. inuse would be
driven down over time from not issuing IOs but it'd be better to clamp it
to minimum immediately once in debt.
* How much can be paid off was determined by hweight_inuse. As inuse was
driven down, the payment amount would fall together regardless of the
debtor's active weight. This means that the debtors were punished harshly.
* ioc_rqos_merge() wasn't calling blkcg_schedule_throttle() after
iocg_kick_delay().
This patch revamps debt handling so that
* Debt handling owns inuse for iocgs in debt and keeps them at zero.
* Payment amount is determined by hweight_active. This is more deterministic
and safer than hweight_inuse but still far from ideal in that it doesn't
factor in possible donations from other iocgs for debt payments. This
likely needs further improvements in the future.
* iocg_rqos_merge() now calls blkcg_schedule_throttle() as necessary.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Andy Newell <newella@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When the margin drops below the minimum on a donating iocg, donation is
immediately canceled in full. There are a couple shortcomings with the
current behavior.
* It's abrupt. A small temporary budget deficit can lead to a wide swing in
weight allocation and a large surplus.
* It's open coded in the issue path but not implemented for the merge path.
A series of merges at a low inuse can make the iocg incur debts and stall
incorrectly.
This patch reimplements in-period donation snapbacks so that
* inuse adjustment and cost calculations are factored into
adjust_inuse_and_calc_cost() which is called from both the issue and merge
paths.
* Snapbacks are more gradual. It occurs in quarter steps.
* A snapback triggers if the margin goes below the low threshold and is
lower than the budget at the time of the last adjustment.
* For the above, __propagate_weights() stores the margin in
iocg->saved_margin. Move iocg->last_inuse storing together into
__propagate_weights() for consistency.
* Full snapback is guaranteed when there are waiters.
* With precise donation and gradual snapbacks, inuse adjustments are now a
lot more effective and the value of scaling inuse on weight changes isn't
clear. Removed inuse scaling from weight_update().
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
iocost has various safety nets to combat inuse adjustment calculation
inaccuracies. With Andy's method implemented in transfer_surpluses(), inuse
adjustment calculations are now accurate and we can make donation amount
determinations accurate too.
* Stop keeping track of past usage history and using the maximum. Act on the
immediate usage information.
* Remove donation constraints defined by SURPLUS_* constants. Donate
whatever isn't used.
* Determine the donation amount so that the iocg will end up with
MARGIN_TARGET_PCT budget at the end of the coming period assuming the same
usage as the previous period. TARGET is set at 50% of period, which is the
previous maximum. This provides smooth convergence for most repetitive IO
patterns.
* Apply donation logic early at 20% budget. There's no risk in doing so as
the calculation is based on the delta between the current budget and the
target budget at the end of the coming period.
* Remove preemptive iocg activation for zero cost IOs. As donation can reach
near zero now, the mere activation doesn't provide any protection anymore.
In the unlikely case that this becomes a problem, the right solution is
assigning appropriate costs for such IOs.
This significantly improves the donation determination logic while also
simplifying it. Now all donations are immediate, exact and smooth.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Andy Newell <newella@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
iocost implements work conservation by reducing iocg->inuse and propagating
the adjustment upwards proportionally. However, while I knew the target
absolute hierarchical proportion - adjusted hweight_inuse, I couldn't figure
out how to determine the iocg->inuse adjustment to achieve that and
approximated the adjustment by scaling iocg->inuse using the proportion of
the needed hweight_inuse changes.
When nested, these scalings aren't accurate even when adjusting a single
node as the donating node also receives the benefit of the donated portion.
When multiple nodes are donating as they often do, they can be wildly wrong.
iocost employed various safety nets to combat the inaccuracies. There are
ample buffers in determining how much to donate, the adjustments are
conservative and gradual. While it can achieve a reasonable level of work
conservation in simple scenarios, the inaccuracies can easily add up leading
to significant loss of total work. This in turn makes it difficult to
closely cap vrate as vrate adjustment is needed to compensate for the loss
of work. The combination of inaccurate donation calculations and vrate
adjustments can lead to wide fluctuations and clunky overall behaviors.
Andy Newell devised a method to calculate the needed ->inuse updates to
achieve the target hweight_inuse's. The method is compatible with the
proportional inuse adjustment propagation which allows all hot path
operations to be local to each iocg.
To roughly summarize, Andy's method divides the tree into donating and
non-donating parts, calculates global donation rate which is used to
determine the target hweight_inuse for each node, and then derives per-level
proportions. There's non-trivial amount of math involved. Please refer to
the following pdfs for detailed descriptions.
https://drive.google.com/file/d/1PsJwxPFtjUnwOY1QJ5AeICCcsL7BM3bohttps://drive.google.com/file/d/1vONz1-fzVO7oY5DXXsLjSxEtYYQbOvsEhttps://drive.google.com/file/d/1WcrltBOSPN0qXVdBgnKm4mdp9FhuEFQN
This patch implements Andy's method in transfer_surpluses(). This makes the
donation calculations accurate per cycle and enables further improvements in
other parts of the donation logic.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Andy Newell <newella@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The way the surplus donation logic is structured isn't great. There are two
separate paths for starting/increasing donations and decreasing them making
the logic harder to follow and is prone to unnecessary behavior differences.
In preparation for improved donation handling, this patch restructures the
code so that
* All donors - new, increasing and decreasing - are funneled through the
same code path.
* The target donation calculation is factored into hweight_after_donation()
which is called once from the same spot for all possible donors.
* Actual inuse adjustment is factored into trasnfer_surpluses().
This change introduces a few behavior differences - e.g. donation amount
reduction now uses the max usage of the recent three periods just like new
and increasing donations, and inuse now gets adjusted upwards the same way
it gets downwards. These differences are unlikely to have severely negative
implications and the whole logic will be revamped soon.
This patch also removes two tracepoints. The existing TPs don't quite fit
the new implementation. A later patch will update and reinstate them.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Budget donations are inaccurate and could take multiple periods to converge.
To prevent triggering vrate adjustments while surplus transfers were
catching up, vrate adjustment was suppressed if donations were increasing,
which was indicated by non-zero nr_surpluses.
This entangling won't be necessary with the scheduled rewrite of donation
mechanism which will make it precise and immediate. Let's decouple the two
in preparation.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Instead of marking iocgs with surplus with a flag and filtering for them
while walking all active iocgs, build a surpluses list. This doesn't make
much difference now but will help implementing improved donation logic which
will iterate iocgs with surplus multiple times.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, iocg->usages[] which are used to guide inuse adjustments are
calculated from vtime deltas. This, however, assumes that the hierarchical
inuse weight at the time of calculation held for the entire period, which
often isn't true and can lead to significant errors.
Now that we have absolute usage information collected, we can derive
iocg->usages[] from iocg->local_stat.usage_us so that inuse adjustment
decisions are made based on actual absolute usage. The calculated usage is
clamped between 1 and WEIGHT_ONE and WEIGHT_ONE is also used to signal
saturation regardless of the current hierarchical inuse weight.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, iocost doesn't collect or expose any statistics punting off all
monitoring duties to drgn based iocost_monitor.py. While it works for some
scenarios, there are some usability and data availability challenges. For
example, accurate per-cgroup usage information can't be tracked by vtime
progression at all and the number available in iocg->usages[] are really
short-term snapshots used for control heuristics with possibly significant
errors.
This patch implements per-cgroup absolute usage stat counter and exposes it
through io.stat along with the current vrate. Usage stat collection and
flushing employ the same method as cgroup rstat on the active iocg's and the
only hot path overhead is preemption toggling and adding to a percpu
counter.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Currently, debt handling requires only iocg->waitq.lock. In the future, we
want to adjust and propagate inuse changes depending on debt status. Let's
grab ioc->lock in debt handling paths in preparation.
* Because ioc->lock nests outside iocg->waitq.lock, the decision to grab
ioc->lock needs to be made before entering the critical sections.
* Add and use iocg_[un]lock() which handles the conditional double locking.
* Add @pay_debt to iocg_kick_waitq() so that debt payment happens only when
the caller grabbed both locks.
This patch is prepatory and the comments contain references to future
changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The margin handling was pretty inconsistent.
* ioc->margin_us and ioc->inuse_margin_vtime were used as vtime margin
thresholds. However, the two are in different units with the former
requiring conversion to vtime on use.
* iocg_kick_waitq() was using a quarter of WAITQ_TIMER_MARGIN_PCT of
period_us as the timer slack - ~1.2%. While iocg_kick_delay() was using a
quarter of ioc->margin_us - ~12.5%. There aren't strong reasons to use
different values for the two.
This patch cleans up margin and timer slack handling:
* vtime margins are now recorded in ioc->margins.{min, max} on period
duration changes and used consistently.
* Timer slack is now 1% of period_us and recorded in ioc->timer_slack_ns and
used consistently for iocg_kick_waitq() and iocg_kick_delay().
The only functional change is shortening of timer slack. No meaningful
visible change is expected.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
They are in microseconds and wrap in around 1.2 hours with u32. While
unlikely, confusions from wraparounds are still possible. We aren't saving
anything meaningful by keeping these u32. Let's make them u64.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
To improve weight donations, we want to able to scale inuse with a greater
accuracy and down below 1. Let's make non-hierarchical weights to use
WEIGHT_ONE based fixed point numbers too like hierarchical ones.
This doesn't cause any behavior changes yet.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We're gonna use HWEIGHT_WHOLE for regular weights too. Let's rename it to
WEIGHT_ONE.
Pure rename.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
iocg_kick_waitq() is the function which pays debt and iocg_kick_delay()
updates the actual delay status accordingly. If iocg_kick_delay() is not
called after iocg_kick_delay() updated debt, unnecessarily large delays can
be applied temporarily.
Let's make sure such conditions don't occur by making iocg_kick_waitq()
always call iocg_kick_delay() after paying debt.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We'll make iocg_kick_waitq() call iocg_kick_delay(). Reorder them in
preparation. This is pure code reorganization.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
__propagate_weights() currently expects the callers to clamp inuse within
[1, active], which is needlessly fragile. The inuse adjustment logic is
going to be revamped, in preparation, let's make __propagate_weights() clamp
inuse on entry.
Also, make it avoid weight updates altogether if neither active or inuse is
changed.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
It already propagates two weights - active and inuse - and there will be
another soon. Let's drop the confusing misnomers. Rename
[__]propagate_active_weights() to [__]propagate_weights() and
commit_active_weights() to commit_weights().
This is pure rename.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
blk-iocost has been reading percpu stat counters from remote cpus which on
some archs can lead to torn reads in really rare occassions. Use local[64]_t
for those counters.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
ioc_pd_free() grabs irq-safe ioc->lock without ensuring that irq is disabled
when it can be called with irq disabled or enabled. This has a small chance
of causing A-A deadlocks and triggers lockdep splats. Use irqsave operations
instead.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: 7caa47151a ("blkcg: implement blk-iocost")
Cc: stable@vger.kernel.org # v5.4+
Signed-off-by: Jens Axboe <axboe@kernel.dk>
- Untangle the header spaghetti which causes build failures in various
situations caused by the lockdep additions to seqcount to validate that
the write side critical sections are non-preemptible.
- The seqcount associated lock debug addons which were blocked by the
above fallout.
seqcount writers contrary to seqlock writers must be externally
serialized, which usually happens via locking - except for strict per
CPU seqcounts. As the lock is not part of the seqcount, lockdep cannot
validate that the lock is held.
This new debug mechanism adds the concept of associated locks.
sequence count has now lock type variants and corresponding
initializers which take a pointer to the associated lock used for
writer serialization. If lockdep is enabled the pointer is stored and
write_seqcount_begin() has a lockdep assertion to validate that the
lock is held.
Aside of the type and the initializer no other code changes are
required at the seqcount usage sites. The rest of the seqcount API is
unchanged and determines the type at compile time with the help of
_Generic which is possible now that the minimal GCC version has been
moved up.
Adding this lockdep coverage unearthed a handful of seqcount bugs which
have been addressed already independent of this.
While generaly useful this comes with a Trojan Horse twist: On RT
kernels the write side critical section can become preemtible if the
writers are serialized by an associated lock, which leads to the well
known reader preempts writer livelock. RT prevents this by storing the
associated lock pointer independent of lockdep in the seqcount and
changing the reader side to block on the lock when a reader detects
that a writer is in the write side critical section.
- Conversion of seqcount usage sites to associated types and initializers.
-----BEGIN PGP SIGNATURE-----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=7Gj+
-----END PGP SIGNATURE-----
Merge tag 'locking-urgent-2020-08-10' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
Pull locking updates from Thomas Gleixner:
"A set of locking fixes and updates:
- Untangle the header spaghetti which causes build failures in
various situations caused by the lockdep additions to seqcount to
validate that the write side critical sections are non-preemptible.
- The seqcount associated lock debug addons which were blocked by the
above fallout.
seqcount writers contrary to seqlock writers must be externally
serialized, which usually happens via locking - except for strict
per CPU seqcounts. As the lock is not part of the seqcount, lockdep
cannot validate that the lock is held.
This new debug mechanism adds the concept of associated locks.
sequence count has now lock type variants and corresponding
initializers which take a pointer to the associated lock used for
writer serialization. If lockdep is enabled the pointer is stored
and write_seqcount_begin() has a lockdep assertion to validate that
the lock is held.
Aside of the type and the initializer no other code changes are
required at the seqcount usage sites. The rest of the seqcount API
is unchanged and determines the type at compile time with the help
of _Generic which is possible now that the minimal GCC version has
been moved up.
Adding this lockdep coverage unearthed a handful of seqcount bugs
which have been addressed already independent of this.
While generally useful this comes with a Trojan Horse twist: On RT
kernels the write side critical section can become preemtible if
the writers are serialized by an associated lock, which leads to
the well known reader preempts writer livelock. RT prevents this by
storing the associated lock pointer independent of lockdep in the
seqcount and changing the reader side to block on the lock when a
reader detects that a writer is in the write side critical section.
- Conversion of seqcount usage sites to associated types and
initializers"
* tag 'locking-urgent-2020-08-10' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (25 commits)
locking/seqlock, headers: Untangle the spaghetti monster
locking, arch/ia64: Reduce <asm/smp.h> header dependencies by moving XTP bits into the new <asm/xtp.h> header
x86/headers: Remove APIC headers from <asm/smp.h>
seqcount: More consistent seqprop names
seqcount: Compress SEQCNT_LOCKNAME_ZERO()
seqlock: Fold seqcount_LOCKNAME_init() definition
seqlock: Fold seqcount_LOCKNAME_t definition
seqlock: s/__SEQ_LOCKDEP/__SEQ_LOCK/g
hrtimer: Use sequence counter with associated raw spinlock
kvm/eventfd: Use sequence counter with associated spinlock
userfaultfd: Use sequence counter with associated spinlock
NFSv4: Use sequence counter with associated spinlock
iocost: Use sequence counter with associated spinlock
raid5: Use sequence counter with associated spinlock
vfs: Use sequence counter with associated spinlock
timekeeping: Use sequence counter with associated raw spinlock
xfrm: policy: Use sequence counters with associated lock
netfilter: nft_set_rbtree: Use sequence counter with associated rwlock
netfilter: conntrack: Use sequence counter with associated spinlock
sched: tasks: Use sequence counter with associated spinlock
...
We shouldn't skip iocg when its abs_vdebt is not zero.
Fixes: 0b80f9866e ("iocost: protect iocg->abs_vdebt with iocg->waitq.lock")
Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
A sequence counter write side critical section must be protected by some
form of locking to serialize writers. A plain seqcount_t does not
contain the information of which lock must be held when entering a write
side critical section.
Use the new seqcount_spinlock_t data type, which allows to associate a
spinlock with the sequence counter. This enables lockdep to verify that
the spinlock used for writer serialization is held when the write side
critical section is entered.
If lockdep is disabled this lock association is compiled out and has
neither storage size nor runtime overhead.
Signed-off-by: Ahmed S. Darwish <a.darwish@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Daniel Wagner <dwagner@suse.de>
Link: https://lkml.kernel.org/r/20200720155530.1173732-21-a.darwish@linutronix.de
Make use of the struct_size() helper instead of an open-coded version
in order to avoid any potential type mistakes.
This code was detected with the help of Coccinelle and, audited and
fixed manually.
Signed-off-by: Gustavo A. R. Silva <gustavoars@kernel.org>
Addresses-KSPP-ID: https://github.com/KSPP/linux/issues/83
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When the QoS targets are met and nothing is being throttled, there's
no way to tell how saturated the underlying device is - it could be
almost entirely idle, at the cusp of saturation or anywhere inbetween.
Given that there's no information, it's best to keep vrate as-is in
this state. Before 7cd806a9a9 ("iocost: improve nr_lagging
handling"), this was the case - if the device isn't missing QoS
targets and nothing is being throttled, busy_level was reset to zero.
While fixing nr_lagging handling, 7cd806a9a9 ("iocost: improve
nr_lagging handling") broke this. Now, while the device is hitting
QoS targets and nothing is being throttled, vrate keeps getting
adjusted according to the existing busy_level.
This led to vrate keeping climing till it hits max when there's an IO
issuer with limited request concurrency if the vrate started low.
vrate starts getting adjusted upwards until the issuer can issue IOs
w/o being throttled. From then on, QoS targets keeps getting met and
nothing on the system needs throttling and vrate keeps getting
increased due to the existing busy_level.
This patch makes the following changes to the busy_level logic.
* Reset busy_level if nr_shortages is zero to avoid the above
scenario.
* Make non-zero nr_lagging block lowering nr_level but still clear
positive busy_level if there's clear non-saturation signal - QoS
targets are met and nr_shortages is non-zero. nr_lagging's role is
preventing adjusting vrate upwards while there are long-running
commands and it shouldn't keep busy_level positive while there's
clear non-saturation signal.
* Restructure code for clarity and add comments.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Andy Newell <newella@fb.com>
Fixes: 7cd806a9a9 ("iocost: improve nr_lagging handling")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull in block-5.7 fixes for 5.8. Mostly to resolve a conflict with
the blk-iocost changes, but we also need the base of the bdi
use-after-free as well as we build on top of it.
* block-5.7:
nvme: fix possible hang when ns scanning fails during error recovery
nvme-pci: fix "slimmer CQ head update"
bdi: add a ->dev_name field to struct backing_dev_info
bdi: use bdi_dev_name() to get device name
bdi: move bdi_dev_name out of line
vboxsf: don't use the source name in the bdi name
iocost: protect iocg->abs_vdebt with iocg->waitq.lock
block: remove the bd_openers checks in blk_drop_partitions
nvme: prevent double free in nvme_alloc_ns() error handling
null_blk: Cleanup zoned device initialization
null_blk: Fix zoned command handling
block: remove unused header
blk-iocost: Fix error on iocost_ioc_vrate_adj
bdev: Reduce time holding bd_mutex in sync in blkdev_close()
buffer: remove useless comment and WB_REASON_FREE_MORE_MEM, reason.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
abs_vdebt is an atomic_64 which tracks how much over budget a given cgroup
is and controls the activation of use_delay mechanism. Once a cgroup goes
over budget from forced IOs, it has to pay it back with its future budget.
The progress guarantee on debt paying comes from the iocg being active -
active iocgs are processed by the periodic timer, which ensures that as time
passes the debts dissipate and the iocg returns to normal operation.
However, both iocg activation and vdebt handling are asynchronous and a
sequence like the following may happen.
1. The iocg is in the process of being deactivated by the periodic timer.
2. A bio enters ioc_rqos_throttle(), calls iocg_activate() which returns
without anything because it still sees that the iocg is already active.
3. The iocg is deactivated.
4. The bio from #2 is over budget but needs to be forced. It increases
abs_vdebt and goes over the threshold and enables use_delay.
5. IO control is enabled for the iocg's subtree and now IOs are attributed
to the descendant cgroups and the iocg itself no longer issues IOs.
This leaves the iocg with stuck abs_vdebt - it has debt but inactive and no
further IOs which can activate it. This can end up unduly punishing all the
descendants cgroups.
The usual throttling path has the same issue - the iocg must be active while
throttled to ensure that future event will wake it up - and solves the
problem by synchronizing the throttling path with a spinlock. abs_vdebt
handling is another form of overage handling and shares a lot of
characteristics including the fact that it isn't in the hottest path.
This patch fixes the above and other possible races by strictly
synchronizing abs_vdebt and use_delay handling with iocg->waitq.lock.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Vlad Dmitriev <vvd@fb.com>
Cc: stable@vger.kernel.org # v5.4+
Fixes: e1518f63f2 ("blk-iocost: Don't let merges push vtime into the future")
Signed-off-by: Jens Axboe <axboe@kernel.dk>
On each IO completion, iocost decides whether the IO met or missed its latency
target. Currently, the targets are fixed numbers per IO type. While this can be
good enough for loose latency targets way higher than typical completion
latencies, the effect of IO size makes it difficult to tighten the latency
target - a target adequate for 4k IOs might be too tight for 512k IOs and
vice-versa.
iocost already has all the necessary information to account for different IO
sizes when testing whether the latency target is met as iocost can calculate the
size vtime cost of a given IO. This patch updates the completion path to
calculate the size vtime cost of the IO, deduct the nsec equivalent from the
observed latency and use the adjusted value to decide whether the target is met.
This makes latency targets independent from IO size and enables determining
adequate latency targets with fixed size fio runs.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Andy Newell <newella@fb.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The use_delay mechanism was introduced by blk-iolatency to hold memory
allocators accountable for the reclaim and other shared IOs they cause. The
duration of the delay is dynamically balanced between iolatency increasing the
value on each target miss and it auto-decaying as time passes and threads get
delayed on it.
While this works well for iolatency, iocost's control model isn't compatible
with it. There is no repeated "violation" events which can be balanced against
auto-decaying. iocost instead knows how much a given cgroup is over budget and
wants to prevent that cgroup from issuing IOs while over budget. Until now,
iocost has been adding the cost of force-issued IOs. However, this doesn't
reflect the amount which is already over budget and is simply not enough to
counter the auto-decaying allowing anon-memory leaking low priority cgroup to
go over its alloted share of IOs.
As auto-decaying doesn't make much sense for iocost, this patch introduces a
different mode of operation for use_delay - when blkcg_set_delay() are used
insted of blkcg_add/use_delay(), the delay duration is not auto-decayed until it
is explicitly cleared with blkcg_clear_delay(). iocost is updated to keep the
delay duration synchronized to the budget overage amount.
With this change, iocost can effectively police cgroups which generate
significant amount of force-issued IOs.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Systemtap 4.2 is unable to correctly interpret the "u32 (*missed_ppm)[2]"
argument of the iocost_ioc_vrate_adj trace entry defined in
include/trace/events/iocost.h leading to the following error:
/tmp/stapAcz0G0/stap_c89c58b83cea1724e26395efa9ed4939_6321_aux_6.c:78:8:
error: expected ‘;’, ‘,’ or ‘)’ before ‘*’ token
, u32[]* __tracepoint_arg_missed_ppm
That argument type is indeed rather complex and hard to read. Looking
at block/blk-iocost.c. It is just a 2-entry u32 array. By simplifying
the argument to a simple "u32 *missed_ppm" and adjusting the trace
entry accordingly, the compilation error was gone.
Fixes: 7caa47151a ("blkcg: implement blk-iocost")
Acked-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
-----BEGIN PGP SIGNATURE-----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=6Glq
-----END PGP SIGNATURE-----
Merge tag 'for-5.7/block-2020-03-29' of git://git.kernel.dk/linux-block
Pull block updates from Jens Axboe:
- Online capacity resizing (Balbir)
- Number of hardware queue change fixes (Bart)
- null_blk fault injection addition (Bart)
- Cleanup of queue allocation, unifying the node/no-node API
(Christoph)
- Cleanup of genhd, moving code to where it makes sense (Christoph)
- Cleanup of the partition handling code (Christoph)
- disk stat fixes/improvements (Konstantin)
- BFQ improvements (Paolo)
- Various fixes and improvements
* tag 'for-5.7/block-2020-03-29' of git://git.kernel.dk/linux-block: (72 commits)
block: return NULL in blk_alloc_queue() on error
block: move bio_map_* to blk-map.c
Revert "blkdev: check for valid request queue before issuing flush"
block: simplify queue allocation
bcache: pass the make_request methods to blk_queue_make_request
null_blk: use blk_mq_init_queue_data
block: add a blk_mq_init_queue_data helper
block: move the ->devnode callback to struct block_device_operations
block: move the part_stat* helpers from genhd.h to a new header
block: move block layer internals out of include/linux/genhd.h
block: move guard_bio_eod to bio.c
block: unexport get_gendisk
block: unexport disk_map_sector_rcu
block: unexport disk_get_part
block: mark part_in_flight and part_in_flight_rw static
block: mark block_depr static
block: factor out requeue handling from dispatch code
block/diskstats: replace time_in_queue with sum of request times
block/diskstats: accumulate all per-cpu counters in one pass
block/diskstats: more accurate approximation of io_ticks for slow disks
...
vtimes may wrap and time_before/after64() should be used to determine
whether a given vtime is before or after another. iocg_is_idle() was
incorrectly using plain "<" comparison do determine whether done_vtime
is before vtime. Here, the only thing we're interested in is whether
done_vtime matches vtime which indicates that there's nothing in
flight. Let's test for inequality instead.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: 7caa47151a ("blkcg: implement blk-iocost")
Cc: stable@vger.kernel.org # v5.4+
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When over-budget IOs are force-issued through root cgroup,
iocg_kick_delay() adjusts the async delay accordingly but doesn't
actually schedule async throttle for the issuing task. This bug is
pretty well masked because sooner or later the offending threads are
gonna get directly throttled on regular IOs or have async delay
scheduled by mem_cgroup_throttle_swaprate().
However, it can affect control quality on filesystem metadata heavy
operations. Let's fix it by invoking blkcg_schedule_throttle() when
iocg_kick_delay() says async delay is needed.
Signed-off-by: Tejun Heo <tj@kernel.org>
Fixes: 7caa47151a ("blkcg: implement blk-iocost")
Cc: stable@vger.kernel.org
Reported-by: Josef Bacik <josef@toxicpanda.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
There is a bug that checking the same active_list over and over again
in iocg_activate(). The intention of the code was checking whether all
the ancestors and self have already been activated. So fix it.
Fixes: 7caa47151a ("blkcg: implement blk-iocost")
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jiufei Xue <jiufei.xue@linux.alibaba.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This code causes a static analysis warning:
block/blk-iocost.c:2113 ioc_weight_write() error: double lock 'irq'
We disable IRQs in blkg_conf_prep() and re-enable them in
blkg_conf_finish(). IRQ disable/enable should not be nested because
that means the IRQs will be enabled at the first unlock instead of the
second one.
Fixes: 7caa47151a ("blkcg: implement blk-iocost")
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The default hard disk param sets latency targets at 50ms. As the
default target percentiles are zero, these don't directly regulate
vrate; however, they're still used to calculate the period length -
100ms in this case.
This is excessively low. A SATA drive with QD32 saturated with random
IOs can easily reach avg completion latency of several hundred msecs.
A period duration which is substantially lower than avg completion
latency can lead to wildly fluctuating vrate.
Let's bump up the default latency targets to 250ms so that the period
duration is sufficiently long.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Some IOs may span multiple periods. As latencies are collected on
completion, the inbetween periods won't register them and may
incorrectly decide to increase vrate. nr_lagging tracks these IOs to
avoid those situations. Currently, whenever there are IOs which are
spanning from the previous period, busy_level is reset to 0 if
negative thus suppressing vrate increase.
This has the following two problems.
* When latency target percentiles aren't set, vrate adjustment should
only be governed by queue depth depletion; however, the current code
keeps nr_lagging active which pulls in latency results and can keep
down vrate unexpectedly.
* When lagging condition is detected, it resets the entire negative
busy_level. This turned out to be way too aggressive on some
devices which sometimes experience extended latencies on a small
subset of commands. In addition, a lagging IO will be accounted as
latency target miss on completion anyway and resetting busy_level
amplifies its impact unnecessarily.
This patch fixes the above two problems by disabling nr_lagging
counting when latency target percentiles aren't set and blocking vrate
increases when there are lagging IOs while leaving busy_level as-is.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
vrate_adj tracepoint traces vrate changes; however, it does so only
when busy_level is non-zero. busy_level turning to zero can sometimes
be as interesting an event. This patch also enables vrate_adj
tracepoint on other vrate related events - busy_level changes and
non-zero nr_lagging.
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Merges have the same problem that forced-bios had which is fixed by
the previous patch. The cost of a merge is calculated at the time of
issue and force-advances vtime into the future. Until global vtime
catches up, how the cgroup's hweight changes in the meantime doesn't
matter and it often leads to situations where the cost is calculated
at one hweight and paid at a very different one. See the previous
patch for more details.
Fix it by never advancing vtime into the future for merges. If budget
is available, vtime is advanced. Otherwise, the cost is charged as
debt.
This brings merge cost handling in line with issue cost handling in
ioc_rqos_throttle().
Signed-off-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Tejun Heo