The WakeOvfIsDeferred code path in __call_rcu_nocb_wake() attempts to
wake rcuog when the callback count exceeds qhimark and callbacks aren't
done with their GP (newly queued or awaiting GP). However, a lot of
testing proves this wake is always redundant or useless.
In the flooding case, rcuog is always waiting for a GP to finish. So
waking up the rcuog thread is pointless. The timer wakeup adds overhead,
rcuog simply wakes up and goes back to sleep achieving nothing.
This path also adds a full memory barrier, and additional timer expiry
modifications unnecessarily.
The root cause is that WakeOvfIsDeferred fires when
!rcu_segcblist_ready_cbs() (GP not complete), but waking rcuog cannot
accelerate GP completion.
This commit therefore removes this path.
Tested with rcutorture scenarios: TREE01, TREE05, TREE08 (all NOCB
configurations) - all pass. Also stress tested using a kernel module
that floods call_rcu() to trigger the overload conditions and made the
observations confirming the findings.
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
The RCU grace period mechanism uses a two-phase FQS (Force Quiescent
State) design where the first FQS saves dyntick-idle snapshots and
the second FQS compares them. This results in long and unnecessary latency
for synchronize_rcu() on idle systems (two FQS waits of ~3ms each with
1000HZ) whenever one FQS wait sufficed.
Some investigations showed that the GP kthread's CPU is the holdout CPU
a lot of times after the first FQS as - it cannot be detected as "idle"
because it's actively running the FQS scan in the GP kthread.
Therefore, at the end of rcu_gp_init(), immediately report a quiescent
state for the GP kthread's CPU using rcu_qs() + rcu_report_qs_rdp(). The
GP kthread cannot be in an RCU read-side critical section while running
GP initialization, so this is safe and results in significant latency
improvements.
The following tests were performed:
(1) synchronize_rcu() benchmarking
100 synchronize_rcu() calls with 32 CPUs, 10 runs each (default fqs
jiffies settings):
Baseline (without fix):
| Run | Mean | Min | Max |
|-----|-----------|----------|-----------|
| 1 | 10.088 ms | 9.989 ms | 18.848 ms |
| 2 | 10.064 ms | 9.982 ms | 16.470 ms |
| 3 | 10.051 ms | 9.988 ms | 15.113 ms |
| 4 | 10.125 ms | 9.929 ms | 22.411 ms |
| 5 | 8.695 ms | 5.996 ms | 15.471 ms |
| 6 | 10.157 ms | 9.977 ms | 25.723 ms |
| 7 | 10.102 ms | 9.990 ms | 20.224 ms |
| 8 | 8.050 ms | 5.985 ms | 10.007 ms |
| 9 | 10.059 ms | 9.978 ms | 15.934 ms |
| 10 | 10.077 ms | 9.984 ms | 17.703 ms |
With fix:
| Run | Mean | Min | Max |
|-----|----------|----------|-----------|
| 1 | 6.027 ms | 5.915 ms | 8.589 ms |
| 2 | 6.032 ms | 5.984 ms | 9.241 ms |
| 3 | 6.010 ms | 5.986 ms | 7.004 ms |
| 4 | 6.076 ms | 5.993 ms | 10.001 ms |
| 5 | 6.084 ms | 5.893 ms | 10.250 ms |
| 6 | 6.034 ms | 5.908 ms | 9.456 ms |
| 7 | 6.051 ms | 5.993 ms | 10.000 ms |
| 8 | 6.057 ms | 5.941 ms | 10.001 ms |
| 9 | 6.016 ms | 5.927 ms | 7.540 ms |
| 10 | 6.036 ms | 5.993 ms | 9.579 ms |
Summary:
- Mean latency: 9.75 ms -> 6.04 ms (38% improvement)
- Max latency: 25.72 ms -> 10.25 ms (60% improvement)
(2) Bridge setup/teardown latency (Uladzislau Rezki)
x86_64 with 64 CPUs, 100 iterations of bridge add/configure/delete:
real time
1 - default: 24.221s
2 - this patch: 20.754s (14% faster)
3 - this patch + wake_from_gp: 15.895s (34% faster)
4 - wake_from_gp only: 18.947s (22% faster)
Per-synchronize_rcu() latency (in usec):
1 2 3 4
median: 37249.5 31540.5 15765 22480
min: 7881 7918 9803 7857
max: 63651 55639 31861 32040
This patch combined with rcu_normal_wake_from_gp reduces bridge
setup/teardown time from 24 seconds to 16 seconds.
(3) CPU overhead verification (Uladzislau Rezki)
System CPU time across 5 runs showed no measurable increase:
default: 1.698s - 1.937s
this patch: 1.667s - 1.930s
Conclusion: variations are within noise, no CPU overhead regression.
(4) rcutorture
Tested TREE and SRCU configurations - no regressions.
Reviewed-by: "Paul E. McKenney" <paulmck@kernel.org>
Tested-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Tested-by: Paul E. McKenney <paulmck@kernel.org>
Tested-by: Samir M <samir@linux.ibm.com>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Pull RCU updates from Frederic Weisbecker:
"SRCU:
- Properly handle SRCU readers within IRQ disabled sections in tiny
SRCU
- Preparation to reimplement RCU Tasks Trace on top of SRCU fast:
- Introduce API to expedite a grace period and test it through
rcutorture
- Split srcu-fast in two flavours: SRCU-fast and SRCU-fast-updown.
Both are still targeted toward faster readers (without full
barriers on LOCK and UNLOCK) at the expense of heavier write
side (using full RCU grace period ordering instead of simply
full ordering) as compared to "traditional" non-fast SRCU. But
those srcu-fast flavours are going to be optimized in two
different ways:
- SRCU-fast will become the reimplementation basis for
RCU-TASK-TRACE for consolidation. Since RCU-TASK-TRACE must
be NMI safe, SRCU-fast must be as well.
- SRCU-fast-updown will be needed for uretprobes code in order
to get rid of the read-side memory barriers while still
allowing entering the reader at task level while exiting it
in a timer handler. It is considered semaphore-like in that
it can have different owners between LOCK and UNLOCK.
However it is not NMI-safe.
The actual optimizations are work in progress for the next
cycle. Only the new interfaces are added for now, along with
related torture and scalability test code.
- Create/document/debug/torture new proper initializers for RCU fast:
DEFINE_SRCU_FAST() and init_srcu_struct_fast()
This allows for using right away the proper ordering on the write
side (either full ordering or full RCU grace period ordering)
without waiting for the read side to tell which to use.
This also optimizes the read side altogether with moving flavour
debug checks under debug config and with removing a costly RmW
operation on their first call.
- Make some diagnostic functions tracing safe
Refscale:
- Add performance testing for common context synchronizations
(Preemption, IRQ, Softirq) and per-cpu increments. Those are
relevant comparisons against SRCU-fast read side APIs, especially
as they are planned to synchronize further tracing fast-path code
Miscellanous:
- In order to prepare the layout for nohz_full work deferral to user
exit, the context tracking state must shrink the counter of
transitions to/from RCU not watching. The only possible hazard is
to trigger wrap-around more easily, delaying a bit grace periods
when that happens. This should be a rare event though. Yet add
debugging and torture code to test that assumption
- Fix memory leak on locktorture module
- Annotate accesses in rculist_nulls.h to prevent from KCSAN
warnings. On recent discussions, we also concluded that all those
WRITE_ONCE() and READ_ONCE() on list APIs deserve appropriate
comments. Something to be expected for the next cycle
- Provide a script to apply several configs to several commits with
torture
- Allow torture to reuse a build directory in order to save needless
rebuild time
- Various cleanups"
* tag 'rcu.release.v6.19' of git://git.kernel.org/pub/scm/linux/kernel/git/rcu/linux: (29 commits)
refscale: Add SRCU-fast-updown readers
refscale: Exercise DEFINE_STATIC_SRCU_FAST() and init_srcu_struct_fast()
rcutorture: Make srcu{,d}_torture_init() announce the SRCU type
srcu: Create an SRCU-fast-updown API
refscale: Do not disable interrupts for tests involving local_bh_enable()
refscale: Add non-atomic per-CPU increment readers
refscale: Add this_cpu_inc() readers
refscale: Add preempt_disable() readers
refscale: Add local_bh_disable() readers
refscale: Add local_irq_disable() and local_irq_save() readers
torture: Permit negative kvm.sh --kconfig numberic arguments
srcu: Add SRCU_READ_FLAVOR_FAST_UPDOWN CPP macro
rcu: Mark diagnostic functions as notrace
rcutorture: Make TREE04 use CONFIG_RCU_DYNTICKS_TORTURE
rcutorture: Remove redundant rcutorture_one_extend() from rcu_torture_one_read()
rcutorture: Permit kvm-again.sh to re-use the build directory
torture: Add kvm-series.sh to test commit/scenario combination
rcu: use WRITE_ONCE() for ->next and ->pprev of hlist_nulls
locktorture: Fix memory leak in param_set_cpumask()
doc: Update for SRCU-fast definitions and initialization
...
The rcu_lockdep_current_cpu_online(), rcu_read_lock_sched_held(),
rcu_read_lock_held(), rcu_read_lock_bh_held(), rcu_read_lock_any_held()
are used by tracing-related code paths, so putting traces on them is
unlikely to make anyone happy. This commit therefore marks them all
"notrace".
Reported-by: Leon Hwang <leon.hwang@linux.dev>
Reported-by: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Pull hyperv updates from Wei Liu:
- Unify guest entry code for KVM and MSHV (Sean Christopherson)
- Switch Hyper-V MSI domain to use msi_create_parent_irq_domain()
(Nam Cao)
- Add CONFIG_HYPERV_VMBUS and limit the semantics of CONFIG_HYPERV
(Mukesh Rathor)
- Add kexec/kdump support on Azure CVMs (Vitaly Kuznetsov)
- Deprecate hyperv_fb in favor of Hyper-V DRM driver (Prasanna
Kumar T S M)
- Miscellaneous enhancements, fixes and cleanups (Abhishek Tiwari,
Alok Tiwari, Nuno Das Neves, Wei Liu, Roman Kisel, Michael Kelley)
* tag 'hyperv-next-signed-20251006' of git://git.kernel.org/pub/scm/linux/kernel/git/hyperv/linux:
hyperv: Remove the spurious null directive line
MAINTAINERS: Mark hyperv_fb driver Obsolete
fbdev/hyperv_fb: deprecate this in favor of Hyper-V DRM driver
Drivers: hv: Make CONFIG_HYPERV bool
Drivers: hv: Add CONFIG_HYPERV_VMBUS option
Drivers: hv: vmbus: Fix typos in vmbus_drv.c
Drivers: hv: vmbus: Fix sysfs output format for ring buffer index
Drivers: hv: vmbus: Clean up sscanf format specifier in target_cpu_store()
x86/hyperv: Switch to msi_create_parent_irq_domain()
mshv: Use common "entry virt" APIs to do work in root before running guest
entry: Rename "kvm" entry code assets to "virt" to genericize APIs
entry/kvm: KVM: Move KVM details related to signal/-EINTR into KVM proper
mshv: Handle NEED_RESCHED_LAZY before transferring to guest
x86/hyperv: Add kexec/kdump support on Azure CVMs
Drivers: hv: Simplify data structures for VMBus channel close message
Drivers: hv: util: Cosmetic changes for hv_utils_transport.c
mshv: Add support for a new parent partition configuration
clocksource: hyper-v: Skip unnecessary checks for the root partition
hyperv: Add missing field to hv_output_map_device_interrupt
Rename the "kvm" entry code files and Kconfigs to use generic "virt"
nomenclature so that the code can be reused by other hypervisors (or
rather, their root/dom0 partition drivers), without incorrectly suggesting
the code somehow relies on and/or involves KVM.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Wei Liu <wei.liu@kernel.org>
Currently if a user enqueue a work item using schedule_delayed_work() the
used wq is "system_wq" (per-cpu wq) while queue_delayed_work() use
WORK_CPU_UNBOUND (used when a cpu is not specified). The same applies to
schedule_work() that is using system_wq and queue_work(), that makes use
again of WORK_CPU_UNBOUND.
This lack of consistentcy cannot be addressed without refactoring the API.
alloc_workqueue() treats all queues as per-CPU by default, while unbound
workqueues must opt-in via WQ_UNBOUND.
This default is suboptimal: most workloads benefit from unbound queues,
allowing the scheduler to place worker threads where they’re needed and
reducing noise when CPUs are isolated.
This change add the WQ_UNBOUND flag to sync_wq, to make explicit this
workqueue can be unbound and that it does not benefit from per-cpu work.
Once migration is complete, WQ_UNBOUND can be removed and unbound will
become the implicit default.
With the introduction of the WQ_PERCPU flag (equivalent to !WQ_UNBOUND),
any alloc_workqueue() caller that doesn’t explicitly specify WQ_UNBOUND
must now use WQ_PERCPU.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Marco Crivellari <marco.crivellari@suse.com>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Currently if a user enqueue a work item using schedule_delayed_work() the
used wq is "system_wq" (per-cpu wq) while queue_delayed_work() use
WORK_CPU_UNBOUND (used when a cpu is not specified). The same applies to
schedule_work() that is using system_wq and queue_work(), that makes use
again of WORK_CPU_UNBOUND.
This lack of consistentcy cannot be addressed without refactoring the API.
alloc_workqueue() treats all queues as per-CPU by default, while unbound
workqueues must opt-in via WQ_UNBOUND.
This default is suboptimal: most workloads benefit from unbound queues,
allowing the scheduler to place worker threads where they’re needed and
reducing noise when CPUs are isolated.
This patch adds a new WQ_PERCPU flag to explicitly request the use of
the per-CPU behavior. Both flags coexist for one release cycle to allow
callers to transition their calls.
Once migration is complete, WQ_UNBOUND can be removed and unbound will
become the implicit default.
With the introduction of the WQ_PERCPU flag (equivalent to !WQ_UNBOUND),
any alloc_workqueue() caller that doesn’t explicitly specify WQ_UNBOUND
must now use WQ_PERCPU.
All existing users have been updated accordingly.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Marco Crivellari <marco.crivellari@suse.com>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit adds to the rcu_barrier() kerneldoc header stating that this
function hurries lazy callbacks and that it does not normally result in
additional RCU grace periods.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
RCU re-initializes the deferred QS irq work everytime before attempting
to queue it. However there are situations where the irq work is
attempted to be queued even though it is already queued. In that case
re-initializing messes-up with the irq work queue that is about to be
handled.
The chances for that to happen are higher when the architecture doesn't
support self-IPIs and irq work are then all lazy, such as with the
following sequence:
1) rcu_read_unlock() is called when IRQs are disabled and there is a
grace period involving blocked tasks on the node. The irq work
is then initialized and queued.
2) The related tasks are unblocked and the CPU quiescent state
is reported. rdp->defer_qs_iw_pending is reset to DEFER_QS_IDLE,
allowing the irq work to be requeued in the future (note the previous
one hasn't fired yet).
3) A new grace period starts and the node has blocked tasks.
4) rcu_read_unlock() is called when IRQs are disabled again. The irq work
is re-initialized (but it's queued! and its node is cleared) and
requeued. Which means it's requeued to itself.
5) The irq work finally fires with the tick. But since it was requeued
to itself, it loops and hangs.
Fix this with initializing the irq work only once before the CPU boots.
Fixes: b41642c877 ("rcu: Fix rcu_read_unlock() deadloop due to IRQ work")
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202508071303.c1134cce-lkp@intel.com
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Reviewed-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Neeraj Upadhyay (AMD) <neeraj.upadhyay@kernel.org>
The synchronization of CPU offlining with GP initialization is confusing
to put it mildly (rightfully so as the issue it deals with is complex).
Recent discussions brought up a question -- what prevents the
rcu_implicit_dyntick_qs() from warning about QS reports for offline
CPUs (missing QS reports for offline CPUs causing indefinite hangs).
QS reporting for now-offline CPUs should only happen from:
- gp_init()
- rcutree_cpu_report_dead()
Add some documentation on this and refer to it from comments in the code
explaining how QS reporting is not missed when these functions are
concurrently running.
I referred heavily to this post [1] about the need for the ofl_lock.
[1] https://lore.kernel.org/all/20180924164443.GF4222@linux.ibm.com/
[ Applied paulmck feedback on moving documentation to Requirements.rst ]
Link: https://lore.kernel.org/all/01b4d228-9416-43f8-a62e-124b92e8741a@paulmck-laptop/
Co-developed-by: "Paul E. McKenney" <paulmck@kernel.org>
Signed-off-by: "Paul E. McKenney" <paulmck@kernel.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Neeraj Upadhyay (AMD) <neeraj.upadhyay@kernel.org>
The details of this are subtle and was discussed recently. Add a
quick-quiz about this and refer to it from the code, for more clarity.
Reviewed-by: "Paul E. McKenney" <paulmck@kernel.org>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Neeraj Upadhyay (AMD) <neeraj.upadhyay@kernel.org>
Add detailed comments explaining the critical ordering constraints
during RCU grace period initialization, based on discussions with
Frederic.
Reviewed-by: "Paul E. McKenney" <paulmck@kernel.org>
Co-developed-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Neeraj Upadhyay (AMD) <neeraj.upadhyay@kernel.org>
It is not possible to send an IPI to a dying CPU that has passed the
CPUHP_TEARDOWN_CPU stage. Remaining unhandled IPIs are handled later at
CPUHP_AP_SMPCFD_DYING stage by stop machine. This is the last
opportunity for RCU exp handler to request an expedited quiescent state.
And the upcoming final context switch between stop machine and idle must
have reported the requested context switch.
Therefore, it should not be possible to observe a pending requested
expedited quiescent state when RCU finally stops watching the outgoing
CPU. Once IPIs aren't possible anymore, the QS for the target CPU will
be reported on its behalf by the RCU exp kworker.
Provide an assertion to verify those expectations.
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Neeraj Upadhyay (AMD) <neeraj.upadhyay@kernel.org>
A CPU coming online checks for an ongoing grace period and reports
a quiescent state accordingly if needed. This special treatment that
shortcuts the expedited IPI finds its origin as an optimization purpose
on the following commit:
338b0f760e (rcu: Better hotplug handling for synchronize_sched_expedited()
The point is to avoid an IPI while waiting for a CPU to become online
or failing to become offline.
However this is pointless and even error prone for several reasons:
* If the CPU has been seen offline in the first round scanning offline
and idle CPUs, no IPI is even tried and the quiescent state is
reported on behalf of the CPU.
* This means that if the IPI fails, the CPU just became offline. So
it's unlikely to become online right away, unless the cpu hotplug
operation failed and rolled back, which is a rare event that can
wait a jiffy for a new IPI to be issued.
* But then the "optimization" applying on failing CPU hotplug down only
applies to !PREEMPT_RCU.
* This force reports a quiescent state even if ->cpu_no_qs.b.exp is not
set. As a result it can race with remote QS reports on the same rdp.
Fortunately it happens to be OK but an accident is waiting to happen.
For all those reasons, remove this optimization that doesn't look worthy
to keep around.
Reported-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Neeraj Upadhyay (AMD) <neeraj.upadhyay@kernel.org>
RCU relies on the context tracking nesting counter in order to determine
if it is running in extended quiescent state.
However the context tracking nesting counter is not completely
synchronized with the actual context tracking state:
* The nesting counter is set to 1 or incremented further _after_ the
actual state is set to RCU watching.
* The nesting counter is set to 0 or decremented further _before_ the
actual state is set to RCU not watching.
Therefore it is safe to assume that if ct_nesting() > 0, RCU is
watching. But if ct_nesting() <= 0, RCU is not watching except for tiny
windows.
This hasn't been a problem so far because rcu_is_cpu_rrupt_from_idle()
has only been called from interrupts. However the code is confusing
and abuses the role of the context tracking nesting counter while there
are more accurate indicators available.
Clarify and robustify accordingly.
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Neeraj Upadhyay (AMD) <neeraj.upadhyay@kernel.org>
Currently the call_rcu() API does not check whether a callback
pointer is NULL. If NULL is passed, rcu_core() will try to invoke
it, resulting in NULL pointer dereference and a kernel crash.
To prevent this and improve debuggability, this patch adds a check
for NULL and emits a kernel stack trace to help identify a faulty
caller.
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Currently, the ->gpwrap is not tested (at all per my testing) due to the
requirement of a large delta between a CPU's rdp->gp_seq and its node's
rnp->gpseq.
This results in no testing of ->gpwrap being set. This patch by default
adds 5 minutes of testing with ->gpwrap forced by lowering the delta
between rdp->gp_seq and rnp->gp_seq to just 8 GPs. All of this is
configurable, including the active time for the setting and a full
testing cycle.
By default, the first 25 minutes of a test will have the _default_
behavior there is right now (ULONG_MAX / 4) delta. Then for 5 minutes,
we switch to a smaller delta causing 1-2 wraps in 5 minutes. I believe
this is reasonable since we at least add a little bit of testing for
usecases where ->gpwrap is set.
[ Apply fix for Dan Carpenter's bug report on init path cleanup. ]
[ Apply kernel doc warning fix from Akira Yokosawa. ]
Tested-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
