Each CPU-specific and unbound kworker kthread conforms to a particular
naming scheme. However, this does not extend to the rescuer kworker.
At present, a rescuer kworker is simply named according to its
workqueue's name. This can be cryptic.
This patch modifies a rescuer to follow the kworker naming scheme.
The "R" is indicative of a rescuer and after "-" is its workqueue's
name e.g. "kworker/R-ext4-rsv-conver".
tj: Use "R" instead of "r" as the prefix to make it more distinctive and
consistent with how highpri pools are marked.
Signed-off-by: Aaron Tomlin <atomlin@atomlin.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
While workqueue.default_affinity_scope is writable, it only affects
workqueues which are created afterwards and isn't very useful. Instead,
let's introduce explicit "default" scope and update the effective scope
dynamically when workqueue.default_affinity_scope is changed.
Signed-off-by: Tejun Heo <tj@kernel.org>
An unbound workqueue can be served by multiple worker_pools to improve
locality. The segmentation is achieved by grouping CPUs into pods. By
default, the cache boundaries according to cpus_share_cache() define the
CPUs are grouped. Let's a workqueue is allowed to run on all CPUs and the
system has two L3 caches. The workqueue would be mapped to two worker_pools
each serving one L3 cache domains.
While this improves locality, because the pod boundaries are strict, it
limits the total bandwidth a given issuer can consume. For example, let's
say there is a thread pinned to a CPU issuing enough work items to saturate
the whole machine. With the machine segmented into two pods, no matter how
many work items it issues, it can only use half of the CPUs on the system.
While this limitation has existed for a very long time, it wasn't very
pronounced because the affinity grouping used to be always by NUMA nodes.
With cache boundaries as the default and support for even finer grained
scopes (smt and cpu), it is now an a lot more pressing problem.
This patch implements non-strict affinity scope where the pod boundaries
aren't enforced strictly. Going back to the previous example, the workqueue
would still be mapped to two worker_pools; however, the affinity enforcement
would be soft. The workers in both pools would have their cpus_allowed set
to the whole machine thus allowing the scheduler to migrate them anywhere on
the machine. However, whenever an idle worker is woken up, the workqueue
code asks the scheduler to bring back the task within the pod if the worker
is outside. ie. work items start executing within its affinity scope but can
be migrated outside as the scheduler sees fit. This removes the hard cap on
utilization while maintaining the benefits of affinity scopes.
After the earlier ->__pod_cpumask changes, the implementation is pretty
simple. When non-strict which is the new default:
* pool_allowed_cpus() returns @pool->attrs->cpumask instead of
->__pod_cpumask so that the workers are allowed to run on any CPU that
the associated workqueues allow.
* If the idle worker task's ->wake_cpu is outside the pod, kick_pool() sets
the field to a CPU within the pod.
This would be the first use of task_struct->wake_cpu outside scheduler
proper, so it isn't clear whether this would be acceptable. However, other
methods of migrating tasks are significantly more expensive and are likely
prohibitively so if we want to do this on every work item. This needs
discussion with scheduler folks.
There is also a race window where setting ->wake_cpu wouldn't be effective
as the target task is still on CPU. However, the window is pretty small and
this being a best-effort optimization, it doesn't seem to warrant more
complexity at the moment.
While the non-strict cache affinity scopes seem to be the best option, the
performance picture interacts with the affinity scope and is a bit
complicated to fully discuss in this patch, so the behavior is made easily
selectable through wqattrs and sysfs and the next patch will add
documentation to discuss performance implications.
v2: pool->attrs->affn_strict is set to true for per-cpu worker_pools.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
workqueue_attrs has two uses:
* to specify the required unouned workqueue properties by users
* to match worker_pool's properties to workqueues by core code
For example, if the user wants to restrict a workqueue to run only CPUs 0
and 2, and the two CPUs are on different affinity scopes, the workqueue's
attrs->cpumask would contains CPUs 0 and 2, and the workqueue would be
associated with two worker_pools, one with attrs->cpumask containing just
CPU 0 and the other CPU 2.
Workqueue wants to support non-strict affinity scopes where work items are
started in their matching affinity scopes but the scheduler is free to
migrate them outside the starting scopes, which can enable utilizing the
whole machine while maintaining most of the locality benefits from affinity
scopes.
To enable that, worker_pools need to distinguish the strict affinity that it
has to follow (because that's the restriction coming from the user) and the
soft affinity that it wants to apply when dispatching work items. Note that
two worker_pools with different soft dispatching requirements have to be
separate; otherwise, for example, we'd be ping-ponging worker threads across
NUMA boundaries constantly.
This patch adds workqueue_attrs->__pod_cpumask. The new field is double
underscored as it's only used internally to distinguish worker_pools. A
worker_pool's ->cpumask is now always the same as the online subset of
allowed CPUs of the associated workqueues, and ->__pod_cpumask is the pod's
subset of that ->cpumask. Going back to the example above, both worker_pools
would have ->cpumask containing both CPUs 0 and 2 but one's ->__pod_cpumask
would contain 0 while the other's 2.
* pool_allowed_cpus() is added. It returns the worker_pool's strict cpumask
that the pool's workers must stay within. This is currently always
->__pod_cpumask as all boundaries are still strict.
* As a workqueue_attrs can now track both the associated workqueues' cpumask
and its per-pod subset, wq_calc_pod_cpumask() no longer needs an external
out-argument. Drop @cpumask and instead store the result in
->__pod_cpumask.
* The above also simplifies apply_wqattrs_prepare() as the same
workqueue_attrs can be used to create all pods associated with a
workqueue. tmp_attrs is dropped.
* wq_update_pod() is updated to use wqattrs_equal() to test whether a pwq
update is needed instead of only comparing ->cpumask so that
->__pod_cpumask is compared too. It can directly compare ->__pod_cpumaks
but the code is easier to understand and more robust this way.
The only user-visible behavior change is that two workqueues with different
cpumasks no longer can share worker_pools even when their pod subsets
coincide. Going back to the example, let's say there's another workqueue
with cpumask 0, 2, 3, where 2 and 3 are in the same pod. It would be mapped
to two worker_pools - one with CPU 0, the other with 2 and 3. The former has
the same cpumask as the first pod of the earlier example and would have
shared the same worker_pool but that's no longer the case after this patch.
The worker_pools would have the same ->__pod_cpumask but their ->cpumask's
wouldn't match.
While this is necessary to support non-strict affinity scopes, there can be
further optimizations to maintain sharing among strict affinity scopes.
However, non-strict affinity scopes are going to be preferable for most use
cases and we don't see very diverse mixture of unbound workqueue cpumasks
anyway, so the additional overhead doesn't seem to justify the extra
complexity.
v2: - wq_update_pod() was incorrectly comparing target_attrs->__pod_cpumask
to pool->attrs->cpumask instead of its ->__pod_cpumask. Fix it by
using wqattrs_equal() for comparison instead.
- Per-cpu worker pools weren't initializing ->__pod_cpumask which caused
a subtle problem later on. Set it to cpumask_of(cpu) like ->cpumask.
Signed-off-by: Tejun Heo <tj@kernel.org>
Checking need_more_worker() and calling wake_up_worker() is a repeated
pattern. Let's add kick_pool(), which checks need_more_worker() and
open-code wake_up_worker(), and replace wake_up_worker() uses. The following
conversions aren't one-to-one:
* __queue_work() was using __need_more_work() because it knows that
pool->worklist isn't empty. Switching to kick_pool() adds an extra
list_empty() test.
* create_worker() always needs to wake up the newly minted worker whether
there's more work to do or not to avoid triggering hung task check on the
new task. Keep the current wake_up_process() and still add kick_pool().
This may lead to an extra wakeup which isn't harmful.
* pwq_adjust_max_active() was explicitly checking whether it needs to wake
up a worker or not to avoid spurious wakeups. As kick_pool() only wakes up
a worker when necessary, this explicit check is no longer necessary and
dropped.
* unbind_workers() now calls kick_pool() instead of wake_up_worker() adding
a need_more_worker() test. This avoids spurious wakeups and shouldn't
break anything.
wake_up_worker() is dropped as kick_pool() replaces all its users. After
this patch, all paths that wakes up a non-rescuer worker to initiate work
item execution use kick_pool(). This will enable future changes to improve
locality.
Signed-off-by: Tejun Heo <tj@kernel.org>
The two work execution paths in worker_thread() and rescuer_thread() use
move_linked_works() to claim work items from @pool->worklist. Once claimed,
process_schedule_works() is called which invokes process_one_work() on each
work item. process_one_work() then uses find_worker_executing_work() to
detect and handle collisions - situations where the work item to be executed
is still running on another worker.
This works fine, but, to improve work execution locality, we want to
establish work to worker association earlier and know for sure that the
worker is going to excute the work once asssigned, which requires performing
collision handling earlier while trying to assign the work item to the
worker.
This patch introduces assign_work() which assigns a work item to a worker
using move_linked_works() and then performs collision handling. As collision
handling is handled earlier, process_one_work() no longer needs to worry
about them.
After the this patch, collision checks for linked work items are skipped,
which should be fine as they can't be queued multiple times concurrently.
For work items running from rescuers, the timing of collision handling may
change but the invariant that the work items go through collision handling
before starting execution does not.
This patch shouldn't cause noticeable behavior changes, especially given
that worker_thread() behavior remains the same.
Signed-off-by: Tejun Heo <tj@kernel.org>
Add three more affinity scopes - WQ_AFFN_CPU, SMT and CACHE - and make CACHE
the default. The code changes to actually add the additional scopes are
trivial.
Also add module parameter "workqueue.default_affinity_scope" to override the
default scope and "affinity_scope" sysfs file to configure it per workqueue.
wq_dump.py and documentations are updated accordingly.
This enables significant flexibility in configuring how unbound workqueues
behave. If affinity scope is set to "cpu", it'll behave close to a per-cpu
workqueue. On the other hand, "system" removes all locality boundaries.
Many modern machines have multiple L3 caches often while being mostly
uniform in terms of memory access. Thus, workqueue's previous behavior of
spreading work items in each NUMA node had negative performance implications
from unncessarily crossing L3 boundaries between issue and execution.
However, picking a finer grained affinity scope also has a downside in that
an issuer in one group can't utilize CPUs in other groups.
While dependent on the specifics of workload, there's usually a noticeable
penalty in crossing L3 boundaries, so let's default to CACHE. This issue
will be further addressed and documented with examples in future patches.
Signed-off-by: Tejun Heo <tj@kernel.org>
While wq_pod_type[] can now group CPUs in any aribitrary way, WQ_AFFN_NUM
init is hard coded into workqueue_init_topology(). This patch modularizes
the init path by introducing init_pod_type() which takes a callback to
determine whether two CPUs should share a pod as an argument.
init_pod_type() first scans the CPU combinations testing for sharing to
assign consecutive pod IDs and initialize pod_type->cpu_pod[]. Once
->cpu_pod[] is determined, ->pod_cpus[] and ->pod_node[] are initialized
accordingly. WQ_AFFN_NUMA is now initialized by calling init_pod_type() with
cpus_share_numa() which tests whether the CPU belongs to the same NUMA node.
This patch may change the pod ID assigned to each NUMA node but that
shouldn't cause any behavior changes as the NUMA node to use for allocations
are tracked separately in pod_type->pod_node[]. This makes adding new
affinty types pretty easy.
Signed-off-by: Tejun Heo <tj@kernel.org>
While renamed to pod, the code still assumes that the pods are defined by
NUMA boundaries. Let's generalize it:
* workqueue_attrs->affn_scope is added. Each enum represents the type of
boundaries that define the pods. There are currently two scopes -
WQ_AFFN_NUMA and WQ_AFFN_SYSTEM. The former is the same behavior as before
- one pod per NUMA node. The latter defines one global pod across the
whole system.
* struct wq_pod_type is added which describes how pods are configured for
each affnity scope. For each pod, it lists the member CPUs and the
preferred NUMA node for memory allocations. The reverse mapping from CPU
to pod is also available.
* wq_pod_enabled is dropped. Pod is now always enabled. The previously
disabled behavior is now implemented through WQ_AFFN_SYSTEM.
* get_unbound_pool() wants to determine the NUMA node to allocate memory
from for the new pool. The variables are renamed from node to pod but the
logic still assumes they're one and the same. Clearly distinguish them -
walk the WQ_AFFN_NUMA pods to find the matching pod and then use the pod's
NUMA node.
* wq_calc_pod_cpumask() was taking @pod but assumed that it was the NUMA
node. Take @cpu instead and determine the cpumask to use from the pod_type
matching @attrs.
* apply_wqattrs_prepare() is update to return ERR_PTR() on error instead of
NULL so that it can indicate -EINVAL on invalid affinity scopes.
This patch allows CPUs to be grouped into pods however desired per type.
While this patch causes some internal behavior changes, nothing material
should change for workqueue users.
v2: Trigger WARN_ON_ONCE() in wqattrs_pod_type() if affn_scope is
WQ_AFFN_NR_TYPES which indicates that the function is called with a
worker_pool's attrs instead of a workqueue's.
Signed-off-by: Tejun Heo <tj@kernel.org>
workqueue_attrs can be used for both workqueues and worker_pools. However,
some fields, currently only ->ordered, only apply to workqueues and should
be cleared to the default / invalid values.
Currently, an unbound workqueue explicitly clears attrs->ordered in
get_unbound_pool() after copying the source workqueue attrs, while per-cpu
workqueues rely on the fact that zeroing on allocation gives us the desired
default value for pool->attrs->ordered.
This is fragile. Let's add wqattrs_clear_for_pool() which clears
attrs->ordered and is called from both init_worker_pool() and
get_unbound_pool(). This will ease adding more workqueue-only attrs fields.
In get_unbound_pool(), pool->node initialization is moved upwards for
readability. This shouldn't cause any behavior changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
For an unbound pool, multiple cpumasks are involved.
U: The user-specified cpumask (may be filtered with cpu_possible_mask).
A: The actual cpumask filtered by wq_unbound_cpumask. If the filtering
leaves no CPU, wq_unbound_cpumask is used.
P: Per-pod subsets of #A.
wq->attrs stores #U, wq->dfl_pwq->pool->attrs->cpumask #A, and
wq->cpu_pwq[CPU]->pool->attrs->cpumask #P.
wq_update_pod() is called to update per-pod pwq's during CPU hotplug. To
calculate the new #P for each workqueue, it needs to call
wq_calc_pod_cpumask() with @attrs that contains #A. Currently,
wq_update_pod() achieves this by calling wq_calc_pod_cpumask() with
wq->dfl_pwq->pool->attrs.
This is rather fragile because we're calling wq_calc_pod_cpumask() with
@attrs of a worker_pool rather than the workqueue's actual attrs when what
we want to calculate is the workqueue's cpumask on the pod. While this works
fine currently, future changes will add fields which are used differently
between workqueues and worker_pools and this subtlety will bite us.
This patch factors out #U -> #A calculation from apply_wqattrs_prepare()
into wqattrs_actualize_cpumask and updates wq_update_pod() to copy
wq->unbound_attrs and use the new helper to obtain #A freshly instead of
abusing wq->dfl_pwq->pool_attrs.
This shouldn't cause any behavior changes in the current code.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: K Prateek Nayak <kprateek.nayak@amd.com>
Reference: http://lkml.kernel.org/r/30625cdd-4d61-594b-8db9-6816b017dde3@amd.com
During boot, to initialize unbound CPU pods, wq_pod_init() was called from
workqueue_init(). This is early enough for NUMA nodes to be set up but
before SMP is brought up and CPU topology information is populated.
Workqueue is in the process of improving CPU locality for unbound workqueues
and will need access to topology information during pod init. This adds a
new init function workqueue_init_topology() which is called after CPU
topology information is available and replaces wq_pod_init().
As unbound CPU pods are now initialized after workqueues are activated, we
need to revisit the workqueues to apply the pod configuration. Workqueues
which are created before workqueue_init_topology() are set up so that they
always use the default worker pool. After pods are set up in
workqueue_init_topology(), wq_update_pod() is called on all existing
workqueues to update the pool associations accordingly.
Note that wq_update_pod_attrs_buf allocation is moved to
workqueue_init_early(). This isn't necessary right now but enables further
generalization of pod handling in the future.
This patch changes the initialization sequence but the end result should be
the same.
Signed-off-by: Tejun Heo <tj@kernel.org>
wq_pod_init() is called from workqueue_init() and responsible for
initializing unbound CPU pods according to NUMA node. Workqueue is in the
process of improving affinity awareness and wants to use other topology
information to initialize unbound CPU pods; however, unlike NUMA nodes,
other topology information isn't yet available in workqueue_init().
The next patch will introduce a later stage init function for workqueue
which will be responsible for initializing unbound CPU pods. Relocate
wq_pod_init() below workqueue_init() where the new init function is going to
be located so that the diff can show the content differences.
Just a relocation. No functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Workqueue is in the process of improving CPU affinity awareness. It will
become more flexible and won't be tied to NUMA node boundaries. This patch
renames all NUMA related names in workqueue.c to use "pod" instead.
While "pod" isn't a very common term, it short and captures the grouping of
CPUs well enough. These names are only going to be used within workqueue
implementation proper, so the specific naming doesn't matter that much.
* wq_numa_possible_cpumask -> wq_pod_cpus
* wq_numa_enabled -> wq_pod_enabled
* wq_update_unbound_numa_attrs_buf -> wq_update_pod_attrs_buf
* workqueue_select_cpu_near -> select_numa_node_cpu
This rename is different from others. The function is only used by
queue_work_node() and specifically tries to find a CPU in the specified
NUMA node. As workqueue affinity will become more flexible and untied from
NUMA, this function's name should specifically describe that it's for
NUMA.
* wq_calc_node_cpumask -> wq_calc_pod_cpumask
* wq_update_unbound_numa -> wq_update_pod
* wq_numa_init -> wq_pod_init
* node -> pod in local variables
Only renames. No functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
With the recent removal of NUMA related module param and sysfs knob,
workqueue_attrs->no_numa is now only used to implement ordered workqueues.
Let's rename the field so that it's less confusing especially with the
planned CPU affinity awareness improvements.
Just a rename. No functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
A pwq (pool_workqueue) represents an association between a workqueue and a
worker_pool. When a work item is queued, the workqueue selects the pwq to
use, which in turn determines the pool, and queues the work item to the pool
through the pwq. pwq is also what implements the maximum concurrency limit -
@max_active.
As a per-cpu workqueue should be assocaited with a different worker_pool on
each CPU, it always had per-cpu pwq's that are accessed through wq->cpu_pwq.
However, unbound workqueues were sharing a pwq within each NUMA node by
default. The sharing has several downsides:
* Because @max_active is per-pwq, the meaning of @max_active changes
depending on the machine configuration and whether workqueue NUMA locality
support is enabled.
* Makes per-cpu and unbound code deviate.
* Gets in the way of making workqueue CPU locality awareness more flexible.
This patch makes unbound workqueues use per-cpu pwq's the same way per-cpu
workqueues do by making the following changes:
* wq->numa_pwq_tbl[] is removed and unbound workqueues now use wq->cpu_pwq
just like per-cpu workqueues. wq->cpu_pwq is now RCU protected for unbound
workqueues.
* numa_pwq_tbl_install() is renamed to install_unbound_pwq() and installs
the specified pwq to the target CPU's wq->cpu_pwq.
* apply_wqattrs_prepare() now always allocates a separate pwq for each CPU
unless the workqueue is ordered. If ordered, all CPUs use wq->dfl_pwq.
This makes the return value of wq_calc_node_cpumask() unnecessary. It now
returns void.
* @max_active now means the same thing for both per-cpu and unbound
workqueues. WQ_UNBOUND_MAX_ACTIVE now equals WQ_MAX_ACTIVE and
documentation is updated accordingly. WQ_UNBOUND_MAX_ACTIVE is no longer
used in workqueue implementation and will be removed later.
* All unbound pwq operations which used to be per-numa-node are now per-cpu.
For most unbound workqueue users, this shouldn't cause noticeable changes.
Work item issue and completion will be a small bit faster, flush_workqueue()
would become a bit more expensive, and the total concurrency limit would
likely become higher. All @max_active==1 use cases are currently being
audited for conversion into alloc_ordered_workqueue() and they shouldn't be
affected once the audit and conversion is complete.
One area where the behavior change may be more noticeable is
workqueue_congested() as the reported congestion state is now per CPU
instead of NUMA node. There are only two users of this interface -
drivers/infiniband/hw/hfi1 and net/smc. Maintainers of both subsystems are
cc'd. Inputs on the behavior change would be very much appreciated.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Leon Romanovsky <leon@kernel.org>
Cc: Karsten Graul <kgraul@linux.ibm.com>
Cc: Wenjia Zhang <wenjia@linux.ibm.com>
Cc: Jan Karcher <jaka@linux.ibm.com>
When a CPU went online or offline, wq_update_unbound_numa() was called only
on the CPU which was going up or down. This works fine because all CPUs on
the same NUMA node share the same pool_workqueue slot - one CPU updating it
updates it for everyone in the node.
However, future changes will make each CPU use a separate pool_workqueue
even when they're sharing the same worker_pool, which requires updating
pool_workqueue's for all CPUs which may be sharing the same pool_workqueue
on hotplug.
To accommodate the planned changes, this patch updates
workqueue_on/offline_cpu() so that they call wq_update_unbound_numa() for
all CPUs sharing the same NUMA node as the CPU going up or down. In the
current code, the second+ calls would be noops and there shouldn't be any
behavior changes.
* As wq_update_unbound_numa() is now called on multiple CPUs per each
hotplug event, @cpu is renamed to @hotplug_cpu and another @cpu argument
is added. The former indicates the CPU being hot[un]plugged and the latter
the CPU whose pool_workqueue is being updated.
* In wq_update_unbound_numa(), cpu_off is renamed to off_cpu for consistency
with the new @hotplug_cpu.
Signed-off-by: Tejun Heo <tj@kernel.org>
Currently, all per-cpu pwq's (pool_workqueue's) are allocated directly
through a per-cpu allocation and thus, unlike unbound workqueues, not
reference counted. This difference in lifetime management between the two
types is a bit confusing.
Unbound workqueues are currently accessed through wq->numa_pwq_tbl[] which
isn't suitiable for the planned CPU locality related improvements. The plan
is to unify pwq handling across per-cpu and unbound workqueues so that
they're always accessed through wq->cpu_pwq.
In preparation, this patch makes per-cpu pwq's to be allocated, reference
counted and released the same way as unbound pwq's. wq->cpu_pwq now holds
pointers to pwq's instead of containing them directly.
pwq_unbound_release_workfn() is renamed to pwq_release_workfn() as it's now
also used for per-cpu work items.
Signed-off-by: Tejun Heo <tj@kernel.org>
pool_workqueue release path is currently bounced to system_wq; however, this
is a bit tricky because this bouncing occurs while holding a pool lock and
thus has risk of causing a A-A deadlock. This is currently addressed by the
fact that only unbound workqueues use this bouncing path and system_wq is a
per-cpu workqueue.
While this works, it's brittle and requires a work-around like setting the
lockdep subclass for the lock of unbound pools. Besides, future changes will
use the bouncing path for per-cpu workqueues too making the current approach
unusable.
Let's just use a dedicated kthread_worker to untangle the dependency. This
is just one more kthread for all workqueues and makes the pwq release logic
simpler and more robust.
Signed-off-by: Tejun Heo <tj@kernel.org>
Unbound workqueue CPU affinity is going to receive an overhaul and the NUMA
specific knobs won't make sense anymore. Remove them. Also, the pool_ids
knob was used for debugging and not really meaningful given that there is no
visibility into the pools associated with those IDs. Remove it too. A future
patch will improve overall visibility.
Signed-off-by: Tejun Heo <tj@kernel.org>
Collect first_idle_worker(), worker_enter/leave_idle(),
find_worker_executing_work(), move_linked_works() and wake_up_worker() into
one place. These functions will later be used to implement higher level
worker management logic.
No functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
wq->cpu_pwqs is a percpu variable carraying one pointer to a pool_workqueue.
The field name being plural is unusual and confusing. Rename it to singular.
This patch doesn't cause any functional changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
insert_work() always tried to wake up a worker; however, the only time it
needs to try to wake up a worker is when a new active work item is queued.
When a work item goes on the inactive list or queueing a flush work item,
there's no reason to try to wake up a worker.
This patch moves the worker wakeup logic out of insert_work() and places it
in the active new work item queueing path in __queue_work().
While at it:
* __queue_work() is dereferencing pwq->pool repeatedly. Add local variable
pool.
* Every caller of insert_work() calls debug_work_activate(). Consolidate the
invocations into insert_work().
* In __queue_work() pool->watchdog_ts update is relocated slightly. This is
to better accommodate future changes.
This makes wakeups more precise and will help the planned change to assign
work items to workers before waking them up. No behavior changes intended.
v2: WARN_ON_ONCE(pool != last_pool) added in __queue_work() to clarify as
suggested by Lai.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Lai Jiangshan <jiangshanlai@gmail.com>
* Drop the trivial optimization in worker_thread() where it bypasses calling
process_scheduled_works() if the first work item isn't linked. This is a
mostly pointless micro optimization and gets in the way of improving the
work processing path.
* Consolidate pool->watchdog_ts updates in the two callers into
process_scheduled_works().
Signed-off-by: Tejun Heo <tj@kernel.org>
worker->flags used to be accessed from scheduler hooks without grabbing
pool->lock for concurrency management. This is no longer true since
6d25be5782 ("sched/core, workqueues: Distangle worker accounting from rq
lock"). Also, it's unclear why worker_pool->flags was using the "X" rule.
All relevant users are accessing it under the pool lock.
Let's drop the special "X" rule and use the "L" rule for these flag fields
instead. While at it, replace the CONTEXT comment with
lockdep_assert_held().
This allows worker_set/clr_flags() to be used from context which isn't the
worker itself. This will be used later to implement assinging work items to
workers before waking them up so that workqueue can have better control over
which worker executes which work item on which CPU.
The only actual changes are sanity checks. There shouldn't be any visible
behavior changes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Unbound workqueue execution locality improvement patchset is about to
applied which will cause merge conflicts with changes in for-6.5-fixes.
Let's avoid future merge conflict by pulling in for-6.5-fixes.
Signed-off-by: Tejun Heo <tj@kernel.org>
Use LIST_HEAD() to initialize cull_list instead of open-coding it.
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
wq_cpu_intensive_thresh_us is used to detect CPU-hogging per-cpu work items.
Once detected, they're excluded from concurrency management to prevent them
from blocking other per-cpu work items. If CONFIG_WQ_CPU_INTENSIVE_REPORT is
enabled, repeat offenders are also reported so that the code can be updated.
The default threshold is 10ms which is long enough to do fair bit of work on
modern CPUs while short enough to be usually not noticeable. This
unfortunately leads to a lot of, arguable spurious, detections on very slow
CPUs. Using the same threshold across CPUs whose performance levels may be
apart by multiple levels of magnitude doesn't make whole lot of sense.
This patch scales up wq_cpu_intensive_thresh_us upto 1 second when BogoMIPS
is below 4000. This is obviously very inaccurate but it doesn't have to be
accurate to be useful. The mechanism is still useful when the threshold is
fully scaled up and the benefits of reports are usually shared with everyone
regardless of who's reporting, so as long as there are sufficient number of
fast machines reporting, we don't lose much.
Some (or is it all?) ARM CPUs systemtically report significantly lower
BogoMIPS. While this doesn't break anything, given how widespread ARM CPUs
are, it's at least a missed opportunity and it probably would be a good idea
to teach workqueue about it.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-and-Tested-by: Geert Uytterhoeven <geert@linux-m68k.org>
Motivation of doing this is to better improve boot times for devices when
we want to prevent our workqueue works from running on some specific CPUs,
e,g, some CPUs are busy with interrupts.
Signed-off-by: tiozhang <tiozhang@didiglobal.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Based on commit c4f135d643 ("workqueue: Wrap flush_workqueue() using
a macro"), all in-tree users stopped flushing system-wide workqueues.
Therefore, start emitting runtime message so that all out-of-tree users
will understand that they need to update their code.
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Tejun Heo <tj@kernel.org>
* Concurrency-managed per-cpu work items that hog CPUs and delay the
execution of other work items are now automatically detected and excluded
from concurrency management. Reporting on such work items can also be
enabled through a config option.
* Added tools/workqueue/wq_monitor.py which improves visibility into
workqueue usages and behaviors.
* Includes Arnd's minimal fix for gcc-13 enum warning on 32bit compiles.
This conflicts with afa4bb778e ("workqueue: clean up WORK_* constant
types, clarify masking") in master. Can be resolved by picking the master
version.
-----BEGIN PGP SIGNATURE-----
iIQEABYIACwWIQTfIjM1kS57o3GsC/uxYfJx3gVYGQUCZJoGvw4cdGpAa2VybmVs
Lm9yZwAKCRCxYfJx3gVYGZu0AP9IGK2opAzO9i3i1/Ys81b3sHi9PwrYWH3g252T
Oe3O6QD/Wh0wYBVl0o7IdW6BGdd5iNwIEs420G53UmmPrATqsgQ=
=TffY
-----END PGP SIGNATURE-----
Merge tag 'wq-for-6.5' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq
Pull workqueue updates from Tejun Heo:
- Concurrency-managed per-cpu work items that hog CPUs and delay the
execution of other work items are now automatically detected and
excluded from concurrency management. Reporting on such work items
can also be enabled through a config option.
- Added tools/workqueue/wq_monitor.py which improves visibility into
workqueue usages and behaviors.
- Arnd's minimal fix for gcc-13 enum warning on 32bit compiles,
superseded by commit afa4bb778e in mainline.
* tag 'wq-for-6.5' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq:
workqueue: Disable per-cpu CPU hog detection when wq_cpu_intensive_thresh_us is 0
workqueue: Fix WARN_ON_ONCE() triggers in worker_enter_idle()
workqueue: fix enum type for gcc-13
workqueue: Track and monitor per-workqueue CPU time usage
workqueue: Report work funcs that trigger automatic CPU_INTENSIVE mechanism
workqueue: Automatically mark CPU-hogging work items CPU_INTENSIVE
workqueue: Improve locking rule description for worker fields
workqueue: Move worker_set/clr_flags() upwards
workqueue: Re-order struct worker fields
workqueue: Add pwq->stats[] and a monitoring script
Further upgrade queue_work_on() comment
Dave Airlie reports that gcc-13.1.1 has started complaining about some
of the workqueue code in 32-bit arm builds:
kernel/workqueue.c: In function ‘get_work_pwq’:
kernel/workqueue.c:713:24: error: cast to pointer from integer of different size [-Werror=int-to-pointer-cast]
713 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
| ^
[ ... a couple of other cases ... ]
and while it's not immediately clear exactly why gcc started complaining
about it now, I suspect it's some C23-induced enum type handlign fixup in
gcc-13 is the cause.
Whatever the reason for starting to complain, the code and data types
are indeed disgusting enough that the complaint is warranted.
The wq code ends up creating various "helper constants" (like that
WORK_STRUCT_WQ_DATA_MASK) using an enum type, which is all kinds of
confused. The mask needs to be 'unsigned long', not some unspecified
enum type.
To make matters worse, the actual "mask and cast to a pointer" is
repeated a couple of times, and the cast isn't even always done to the
right pointer, but - as the error case above - to a 'void *' with then
the compiler finishing the job.
That's now how we roll in the kernel.
So create the masks using the proper types rather than some ambiguous
enumeration, and use a nice helper that actually does the type
conversion in one well-defined place.
Incidentally, this magically makes clang generate better code. That,
admittedly, is really just a sign of clang having been seriously
confused before, and cleaning up the typing unconfuses the compiler too.
Reported-by: Dave Airlie <airlied@gmail.com>
Link: https://lore.kernel.org/lkml/CAPM=9twNnV4zMCvrPkw3H-ajZOH-01JVh_kDrxdPYQErz8ZTdA@mail.gmail.com/
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Tejun Heo <tj@kernel.org>
Cc: Nick Desaulniers <ndesaulniers@google.com>
Cc: Nathan Chancellor <nathan@kernel.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
If workqueue.cpu_intensive_thresh_us is set to 0, the detection mechanism
for CPU-hogging per-cpu work item will keep triggering spuriously:
workqueue: process_srcu hogged CPU for >0us 4 times, consider switching to WQ_UNBOUND
workqueue: gc_worker hogged CPU for >0us 4 times, consider switching to WQ_UNBOUND
workqueue: gc_worker hogged CPU for >0us 8 times, consider switching to WQ_UNBOUND
workqueue: wait_rcu_exp_gp hogged CPU for >0us 4 times, consider switching to WQ_UNBOUND
workqueue: kfree_rcu_monitor hogged CPU for >0us 4 times, consider switching to WQ_UNBOUND
workqueue: kfree_rcu_monitor hogged CPU for >0us 8 times, consider switching to WQ_UNBOUND
workqueue: reg_todo hogged CPU for >0us 4 times, consider switching to WQ_UNBOUND
This commit therefore disables the CPU-hog detection mechanism when
workqueue.cpu_intensive_thresh_us is set to 0.
tj: Patch description updated and the condition check on
cpu_intensive_thresh_us separated into a separate if statement for
readability.
Signed-off-by: Zqiang <qiang.zhang1211@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Now that wq_worker_tick() is there, we can easily track the rough CPU time
consumption of each workqueue by charging the whole tick whenever a tick
hits an active workqueue. While not super accurate, it provides reasonable
visibility into the workqueues that consume a lot of CPU cycles.
wq_monitor.py is updated to report the per-workqueue CPU times.
v2: wq_monitor.py was using "cputime" as the key when outputting in json
format. Use "cpu_time" instead for consistency with other fields.
Signed-off-by: Tejun Heo <tj@kernel.org>
Workqueue now automatically marks per-cpu work items that hog CPU for too
long as CPU_INTENSIVE, which excludes them from concurrency management and
prevents stalling other concurrency-managed work items. If a work function
keeps running over the thershold, it likely needs to be switched to use an
unbound workqueue.
This patch adds a debug mechanism which tracks the work functions which
trigger the automatic CPU_INTENSIVE mechanism and report them using
pr_warn() with exponential backoff.
v3: Documentation update.
v2: Drop bouncing to kthread_worker for printing messages. It was to avoid
introducing circular locking dependency through printk but not effective
as it still had pool lock -> wci_lock -> printk -> pool lock loop. Let's
just print directly using printk_deferred().
Signed-off-by: Tejun Heo <tj@kernel.org>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
If a per-cpu work item hogs the CPU, it can prevent other work items from
starting through concurrency management. A per-cpu workqueue which intends
to host such CPU-hogging work items can choose to not participate in
concurrency management by setting %WQ_CPU_INTENSIVE; however, this can be
error-prone and difficult to debug when missed.
This patch adds an automatic CPU usage based detection. If a
concurrency-managed work item consumes more CPU time than the threshold
(10ms by default) continuously without intervening sleeps, wq_worker_tick()
which is called from scheduler_tick() will detect the condition and
automatically mark it CPU_INTENSIVE.
The mechanism isn't foolproof:
* Detection depends on tick hitting the work item. Getting preempted at the
right timings may allow a violating work item to evade detection at least
temporarily.
* nohz_full CPUs may not be running ticks and thus can fail detection.
* Even when detection is working, the 10ms detection delays can add up if
many CPU-hogging work items are queued at the same time.
However, in vast majority of cases, this should be able to detect violations
reliably and provide reasonable protection with a small increase in code
complexity.
If some work items trigger this condition repeatedly, the bigger problem
likely is the CPU being saturated with such per-cpu work items and the
solution would be making them UNBOUND. The next patch will add a debug
mechanism to help spot such cases.
v4: Documentation for workqueue.cpu_intensive_thresh_us added to
kernel-parameters.txt.
v3: Switch to use wq_worker_tick() instead of hooking into preemptions as
suggested by Peter.
v2: Lai pointed out that wq_worker_stopping() also needs to be called from
preemption and rtlock paths and an earlier patch was updated
accordingly. This patch adds a comment describing the risk of infinte
recursions and how they're avoided.
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Lai Jiangshan <jiangshanlai@gmail.com>
* Some worker fields are modified only by the worker itself while holding
pool->lock thus making them safe to read from self, IRQ context if the CPU
is running the worker or while holding pool->lock. Add 'K' locking rule
for them.
* worker->sleeping is currently marked "None" which isn't very descriptive.
It's used only by the worker itself. Add 'S' locking rule for it.
A future patch will depend on the 'K' rule to access worker->current_* from
the scheduler ticks.
Signed-off-by: Tejun Heo <tj@kernel.org>
They are going to be used in wq_worker_stopping(). Move them upwards.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Lai Jiangshan <jiangshanlai@gmail.com>
Currently, the only way to peer into workqueue operations is through
tracing. While possible, it isn't easy or convenient to monitor
per-workqueue behaviors over time this way. Let's add pwq->stats[] that
track relevant events and a drgn monitoring script -
tools/workqueue/wq_monitor.py.
It's arguable whether this needs to be configurable. However, it currently
only has several counters and the runtime overhead shouldn't be noticeable
given that they're on pwq's which are per-cpu on per-cpu workqueues and
per-numa-node on unbound ones. Let's keep it simple for the time being.
v2: Patch reordered to earlier with fewer fields. Field will be added back
gradually. Help message improved.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Lai Jiangshan <jiangshanlai@gmail.com>
The current queue_work_on() docbook comment says that the caller must
ensure that the specified CPU can't go away, and further says that the
penalty for failing to nail down the specified CPU is that the workqueue
handler might find itself executing on some other CPU. This is true
as far as it goes, but fails to note what happens if the specified CPU
never was online. Therefore, further expand this comment to say that
specifying a CPU that was never online will result in a splat.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Cc: Lai Jiangshan <jiangshanlai@gmail.com>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
Mostly changes from Petr to improve warning and error reporting. Workqueue
now reports more of the relevant failures with better context which should
help debugging.
-----BEGIN PGP SIGNATURE-----
iIQEABYIACwWIQTfIjM1kS57o3GsC/uxYfJx3gVYGQUCZErbOg4cdGpAa2VybmVs
Lm9yZwAKCRCxYfJx3gVYGfipAP9BXrUBI99uZL7XjSEe6rLWa0STODyWh67ikR+0
nUO22QEA+ttt5ecLZyopYB18Le06VBBW7J5cs0Cg83bCBTGb4wE=
=gHbc
-----END PGP SIGNATURE-----
Merge tag 'wq-for-6.4' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq
Pull workqueue updates from Tejun Heo:
"Mostly changes from Petr to improve warning and error reporting.
Workqueue now reports more of the relevant failures with better
context which should help debugging"
* tag 'wq-for-6.4' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq:
workqueue: Introduce show_freezable_workqueues
workqueue: Print backtraces from CPUs with hung CPU bound workqueues
workqueue: Warn when a rescuer could not be created
workqueue: Interrupted create_worker() is not a repeated event
workqueue: Warn when a new worker could not be created
workqueue: Fix hung time report of worker pools
workqueue: Simplify a pr_warn() call in wq_select_unbound_cpu()
MAINTAINERS: Add workqueue_internal.h to the WORKQUEUE entry
Currently show_all_workqueue is called if freeze fails at the time of
freeze the workqueues, which shows the status of all workqueues and of
all worker pools. In this cases we may only need to dump state of only
workqueues that are freezable and busy.
This patch defines show_freezable_workqueues, which uses
show_one_workqueue, a granular function that shows the state of individual
workqueues, so that dump only the state of freezable workqueues
at that time.
tj: Minor message adjustment.
Signed-off-by: Jungseung Lee <js07.lee@samsung.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The workqueue watchdog reports a lockup when there was not any progress
in the worker pool for a long time. The progress means that a pending
work item starts being proceed.
Worker pools for unbound workqueues always wake up an idle worker and
try to process the work immediately. The last idle worker has to create
new worker first. The stall might happen only when a new worker could
not be created in which case an error should get printed. Another problem
might be too high load. In this case, workers are victims of a global
system problem.
Worker pools for CPU bound workqueues are designed for lightweight
work items that do not need much CPU time. They are proceed one by
one on a single worker. New worker is used only when a work is sleeping.
It creates one additional scenario. The stall might happen when
the CPU-bound workqueue is used for CPU-intensive work.
More precisely, the stall is detected when a CPU-bound worker is in
the TASK_RUNNING state for too long. In this case, it might be useful
to see the backtrace from the problematic worker.
The information how long a worker is in the running state is not available.
But the CPU-bound worker pools do not have many workers in the running
state by definition. And only few pools are typically blocked.
It should be acceptable to print backtraces from all workers in
TASK_RUNNING state in the stalled worker pools. The number of false
positives should be very low.
Signed-off-by: Petr Mladek <pmladek@suse.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Rescuers are created when a workqueue with WQ_MEM_RECLAIM is allocated.
It typically happens during the system boot.
systemd switches the root filesystem from initrd to the booted system
during boot. It kills processes that block the switch for too long.
One of the process might be modprobe that tries to create a workqueue.
These problems are hard to reproduce. Also alloc_workqueue() does not
pass the error code. Make the debugging easier by printing an error,
similar to create_worker().
Signed-off-by: Petr Mladek <pmladek@suse.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
kthread_create_on_node() might get interrupted(). It is rare but realistic.
For example, when an unbound workqueue is allocated in module_init()
callback. It is done in the context of the "modprobe" process. And,
for example, systemd might kill pending processes when switching root
from initrd to the booted system.
The interrupt is a one-off event and the race might be hard to reproduce.
It is always worth printing.
Signed-off-by: Petr Mladek <pmladek@suse.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The workqueue watchdog reports a lockup when there was not any progress
in the worker pool for a long time. The progress means that a pending
work item starts being proceed.
The progress is guaranteed by using idle workers or creating new workers
for pending work items.
There are several reasons why a new worker could not be created:
+ there is not enough memory
+ there is no free pool ID (IDR API)
+ the system reached PID limit
+ the process creating the new worker was interrupted
+ the last idle worker (manager) has not been scheduled for a long
time. It was not able to even start creating the kthread.
None of these failures is reported at the moment. The only clue is that
show_one_worker_pool() prints that there is a manager. It is the last
idle worker that is responsible for creating a new one. But it is not
clear if create_worker() is failing and why.
Make the debugging easier by printing errors in create_worker().
The error code is important, especially from kthread_create_on_node().
It helps to distinguish the various reasons. For example, reaching
memory limit (-ENOMEM), other system limits (-EAGAIN), or process
interrupted (-EINTR).
Use pr_once() to avoid repeating the same error every CREATE_COOLDOWN
for each stuck worker pool.
Ratelimited printk() might be better. It would help to know if the problem
remains. It would be more clear if the create_worker() errors and workqueue
stalls are related. Also old messages might get lost when the internal log
buffer is full. The problem is that printk() might touch the watchdog.
For example, see touch_nmi_watchdog() in serial8250_console_write().
It would require synchronization of the begin and length of the ratelimit
interval with the workqueue watchdog. Otherwise, the error messages
might break the watchdog. This does not look worth the complexity.
Signed-off-by: Petr Mladek <pmladek@suse.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The workqueue watchdog prints a warning when there is no progress in
a worker pool. Where the progress means that the pool started processing
a pending work item.
Note that it is perfectly fine to process work items much longer.
The progress should be guaranteed by waking up or creating idle
workers.
show_one_worker_pool() prints state of non-idle worker pool. It shows
a delay since the last pool->watchdog_ts.
The timestamp is updated when a first pending work is queued in
__queue_work(). Also it is updated when a work is dequeued for
processing in worker_thread() and rescuer_thread().
The delay is misleading when there is no pending work item. In this
case it shows how long the last work item is being proceed. Show
zero instead. There is no stall if there is no pending work.
Fixes: 82607adcf9 ("workqueue: implement lockup detector")
Signed-off-by: Petr Mladek <pmladek@suse.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Use pr_warn_once() to achieve the same thing. It's simpler.
Signed-off-by: Ammar Faizi <ammarfaizi2@gnuweeb.org>
Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
