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>
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 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>
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>
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>
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>
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>
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>
Currently if the user queues a new work item unintentionally
into a wq after the destroy_workqueue(wq), the work still can
be queued and scheduled without any noticeable kernel message
before the end of a RCU grace period.
As a debug-aid facility, this commit adds a new flag
__WQ_DESTROYING to spot that issue by triggering a kernel WARN
message.
Signed-off-by: Richard Clark <richard.xnu.clark@gmail.com>
Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
The syntax without dots is available since commit 43756e347f
("scripts/kernel-doc: Add support for named variable macro arguments").
The same HTML output is produced with and without this patch.
Signed-off-by: Jonathan Neuschäfer <j.neuschaefer@gmx.net>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Tejun Heo <tj@kernel.org>
Since flush operation synchronously waits for completion, flushing
system-wide WQs (e.g. system_wq) might introduce possibility of deadlock
due to unexpected locking dependency. Tejun Heo commented at [1] that it
makes no sense at all to call flush_workqueue() on the shared WQs as the
caller has no idea what it's gonna end up waiting for.
Although there is flush_scheduled_work() which flushes system_wq WQ with
"Think twice before calling this function! It's very easy to get into
trouble if you don't take great care." warning message, syzbot found a
circular locking dependency caused by flushing system_wq WQ [2].
Therefore, let's change the direction to that developers had better use
their local WQs if flush_scheduled_work()/flush_workqueue(system_*_wq) is
inevitable.
Steps for converting system-wide WQs into local WQs are explained at [3],
and a conversion to stop flushing system-wide WQs is in progress. Now we
want some mechanism for preventing developers who are not aware of this
conversion from again start flushing system-wide WQs.
Since I found that WARN_ON() is complete but awkward approach for teaching
developers about this problem, let's use __compiletime_warning() for
incomplete but handy approach. For completeness, we will also insert
WARN_ON() into __flush_workqueue() after all in-tree users stopped calling
flush_scheduled_work().
Link: https://lore.kernel.org/all/YgnQGZWT%2Fn3VAITX@slm.duckdns.org/ [1]
Link: https://syzkaller.appspot.com/bug?extid=bde0f89deacca7c765b8 [2]
Link: https://lkml.kernel.org/r/49925af7-78a8-a3dd-bce6-cfc02e1a9236@I-love.SAKURA.ne.jp [3]
Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Signed-off-by: Tejun Heo <tj@kernel.org>
Currently show_workqueue_state shows the state of all workqueues and of
all worker pools. In certain cases we may need to dump state of only a
specific workqueue or worker pool. For example in destroy_workqueue we
only need to show state of the workqueue which is getting destroyed.
So rename show_workqueue_state to show_all_workqueues(to signify it
dumps state of all busy workqueues) and divide it into more granular
functions (show_one_workqueue and show_one_worker_pool), that would show
states of individual workqueues and worker pools and can be used in
cases such as the one mentioned above.
Also, as mentioned earlier, make destroy_workqueue dump data pertaining
to only the workqueue that is being destroyed and make user(s) of
earlier interface(show_workqueue_state), use new interface
(show_all_workqueues).
Signed-off-by: Imran Khan <imran.f.khan@oracle.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
There are two kinds of "delayed" work items in workqueue subsystem.
One is for timer-delayed work items which are visible to workqueue users.
The other kind is for work items delayed by active management which can
not be directly visible to workqueue users. We mixed the word "delayed"
for both kinds and caused somewhat ambiguity.
This patch renames the later one (delayed by active management) to
"inactive", because it is used for workqueue active management and
most of its related symbols are named with "active" or "activate".
All "delayed" and "DELAYED" are carefully checked and renamed one by
one to avoid accidentally changing the name of the other kind for
timer-delayed.
No functional change intended.
Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
'wq_sysfs_register()' in annotation for 'WQ_SYSFS' is unavailable,
change it to 'workqueue_sysfs_register()'.
Signed-off-by: Menglong Dong <dong.menglong@zte.com.cn>
Reviewed-by: Lai Jiangshan <jiangshanlai@gmail.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
8a2e8e5dec7e("workqueue: fix cwq->nr_active underflow")
allocated one more bit from the work flags, and it updated
partial of the comments (128 bytes -> 256 bytes), but it
failed to update the info about the number of reserved bits.
Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
Pull workqueue updates from Tejun Heo:
"Nothing too interesting. Just two trivial patches"
* 'for-5.7' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq:
workqueue: Mark up unlocked access to wq->first_flusher
workqueue: Make workqueue_init*() return void
The return values of workqueue_init() and workqueue_early_int() are
always 0, and there is no usage of their return value. So just make
them return void.
Signed-off-by: Yu Chen <chen.yu@easystack.cn>
Signed-off-by: Tejun Heo <tj@kernel.org>
