Currently, __call_rcu_nocb_wake() advances callbacks each time that it
detects excessive numbers of callbacks, though only if it succeeds in
conditionally acquiring its leaf rcu_node structure's ->lock. Despite
the conditional acquisition of ->lock, this does increase contention.
This commit therefore avoids advancing callbacks unless there are
callbacks in ->cblist whose grace period has completed and advancing
has not yet been done during this jiffy.
Note that this decision does not take the presence of new callbacks
into account. That is because on this code path, there will always be
at least one new callback, namely the one we just enqueued.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Currently, nocb_cb_wait() advances callbacks on each pass through its
loop, though only if it succeeds in conditionally acquiring its leaf
rcu_node structure's ->lock. Despite the conditional acquisition of
->lock, this does increase contention. This commit therefore avoids
advancing callbacks unless there are callbacks in ->cblist whose grace
period has completed.
Note that nocb_cb_wait() doesn't worry about callbacks that have not
yet been assigned a grace period. The idea is that the only reason for
nocb_cb_wait() to advance callbacks is to allow it to continue invoking
callbacks. Time will tell whether this is the correct choice.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
When callbacks are in full flow, the common case is waiting for a
grace period, and this grace period will normally take a few jiffies to
complete. It therefore isn't all that helpful for __call_rcu_nocb_wake()
to do a synchronous wakeup in this case. This commit therefore turns this
into a timer-based deferred wakeup of the no-CBs grace-period kthread.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit causes locking, sleeping, and callback state to be printed
for no-CBs CPUs when the rcutorture writer is delayed sufficiently for
rcutorture to complain.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Use of the rcu_data structure's segmented ->cblist for no-CBs CPUs
takes advantage of unrelated grace periods, thus reducing the memory
footprint in the face of floods of call_rcu() invocations. However,
the ->cblist field is a more-complex rcu_segcblist structure which must
be protected via locking. Even though there are only three entities
which can acquire this lock (the CPU invoking call_rcu(), the no-CBs
grace-period kthread, and the no-CBs callbacks kthread), the contention
on this lock is excessive under heavy stress.
This commit therefore greatly reduces contention by provisioning
an rcu_cblist structure field named ->nocb_bypass within the
rcu_data structure. Each no-CBs CPU is permitted only a limited
number of enqueues onto the ->cblist per jiffy, controlled by a new
nocb_nobypass_lim_per_jiffy kernel boot parameter that defaults to
about 16 enqueues per millisecond (16 * 1000 / HZ). When that limit is
exceeded, the CPU instead enqueues onto the new ->nocb_bypass.
The ->nocb_bypass is flushed into the ->cblist every jiffy or when
the number of callbacks on ->nocb_bypass exceeds qhimark, whichever
happens first. During call_rcu() floods, this flushing is carried out
by the CPU during the course of its call_rcu() invocations. However,
a CPU could simply stop invoking call_rcu() at any time. The no-CBs
grace-period kthread therefore carries out less-aggressive flushing
(every few jiffies or when the number of callbacks on ->nocb_bypass
exceeds (2 * qhimark), whichever comes first). This means that the
no-CBs grace-period kthread cannot be permitted to do unbounded waits
while there are callbacks on ->nocb_bypass. A ->nocb_bypass_timer is
used to provide the needed wakeups.
[ paulmck: Apply Coverity feedback reported by Colin Ian King. ]
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
When there are excessive numbers of callbacks, and when either the
corresponding no-CBs callback kthread is asleep or there is no more
ready-to-invoke callbacks, and when least one callback is pending,
__call_rcu_nocb_wake() will advance the callbacks, but refrain from
awakening the corresponding no-CBs grace-period kthread. However,
because rcu_advance_cbs_nowake() is used, it is possible (if a bit
unlikely) that the needed advancement could not happen due to a grace
period not being in progress. Plus there will always be at least one
pending callback due to one having just now been enqueued.
This commit therefore attempts to advance callbacks and awakens the
no-CBs grace-period kthread when there are excessive numbers of callbacks
posted and when the no-CBs callback kthread is not in a position to do
anything helpful.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
The sleep/wakeup of the no-CBs grace-period kthreads is synchronized
using the ->nocb_lock of the first CPU corresponding to that kthread.
This commit provides a separate ->nocb_gp_lock for this purpose, thus
reducing contention on ->nocb_lock.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Currently, nocb_cb_wait() unconditionally acquires the leaf rcu_node
->lock to advance callbacks when done invoking the previous batch.
It does this while holding ->nocb_lock, which means that contention on
the leaf rcu_node ->lock visits itself on the ->nocb_lock. This commit
therefore makes this lock acquisition conditional, forgoing callback
advancement when the leaf rcu_node ->lock is not immediately available.
(In this case, the no-CBs grace-period kthread will eventually do any
needed callback advancement.)
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Currently, __call_rcu_nocb_wake() conditionally acquires the leaf rcu_node
structure's ->lock, and only afterwards does rcu_advance_cbs_nowake()
check to see if it is possible to advance callbacks without potentially
needing to awaken the grace-period kthread. Given that the no-awaken
check can be done locklessly, this commit reverses the order, so that
rcu_advance_cbs_nowake() is invoked without holding the leaf rcu_node
structure's ->lock and rcu_advance_cbs_nowake() checks the grace-period
state before conditionally acquiring that lock, thus reducing the number
of needless acquistions of the leaf rcu_node structure's ->lock.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Currently, when the square root of the number of CPUs is rounded down
by int_sqrt(), this round-down is applied to the number of callback
kthreads per grace-period kthreads. This makes almost no difference
for large systems, but results in oddities such as three no-CBs
grace-period kthreads for a five-CPU system, which is a bit excessive.
This commit therefore causes the round-down to apply to the number of
no-CBs grace-period kthreads, so that systems with from four to eight
CPUs have only two no-CBs grace period kthreads.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
A given rcu_data structure's ->nocb_lock can be acquired very frequently
by the corresponding CPU and occasionally by the corresponding no-CBs
grace-period and callbacks kthreads. In particular, these two kthreads
will have frequent gaps between ->nocb_lock acquisitions that are roughly
a grace period in duration. This means that any excessive ->nocb_lock
contention will be due to the CPU's acquisitions, and this in turn
enables a very naive contention-avoidance strategy to be quite effective.
This commit therefore modifies rcu_nocb_lock() to first
attempt a raw_spin_trylock(), and to atomically increment a
separate ->nocb_lock_contended across a raw_spin_lock(). This new
->nocb_lock_contended field is checked in __call_rcu_nocb_wake() when
interrupts are enabled, with a spin-wait for contending acquisitions
to complete, thus allowing the kthreads a chance to acquire the lock.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Currently, the code provides an extra wakeup for the no-CBs grace-period
kthread if one of its CPUs is generating excessive numbers of callbacks.
But satisfying though it is to wake something up when things are going
south, unless the thing being awakened can actually help solve the
problem, that extra wakeup does nothing but consume additional CPU time,
which is exactly what you don't want during a call_rcu() flood.
This commit therefore avoids doing anything if the corresponding
no-CBs callback kthread is going full tilt. Otherwise, if advancing
callbacks immediately might help and if the leaf rcu_node structure's
lock is immediately available, this commit invokes a new variant of
rcu_advance_cbs() that advances callbacks only if doing so won't require
awakening the grace-period kthread (not to be confused with any of the
no-CBs grace-period kthreads).
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
It might be hard to imagine having more than two billion callbacks
queued on a single CPU's ->cblist, but someone will do it sometime.
This commit therefore makes __call_rcu_nocb_wake() handle this situation
by upgrading local variable "len" from "int" to "long".
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Currently, wake_nocb_gp_defer() simply stores whatever waketype was
passed in, which can result in a RCU_NOCB_WAKE_FORCE being downgraded
to RCU_NOCB_WAKE, which could in turn delay callback processing.
This commit therefore adds a check so that wake_nocb_gp_defer() only
updates ->nocb_defer_wakeup when the update increases the forcefulness,
thus avoiding downgrades.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
The __call_rcu_nocb_wake() function and its predecessors set
->qlen_last_fqs_check to zero for the first callback and to LONG_MAX / 2
for forced reawakenings. The former can result in a too-quick reawakening
when there are many callbacks ready to invoke and the latter prevents a
second reawakening. This commit therefore sets ->qlen_last_fqs_check
to the current number of callbacks in both cases. While in the area,
this commit also moves both assignments under ->nocb_lock.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Historically, no-CBs CPUs allowed the scheduler-clock tick to be
unconditionally disabled on any transition to idle or nohz_full userspace
execution (see the rcu_needs_cpu() implementations). Unfortunately,
the checks used by rcu_needs_cpu() are defeated now that no-CBs CPUs
use ->cblist, which might make users of battery-powered devices rather
unhappy. This commit therefore adds explicit rcu_segcblist_is_offloaded()
checks to return to the historical energy-efficient semantics.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Some compilers complain that wait_gp_seq might be used uninitialized
in nocb_gp_wait(). This cannot actually happen because when wait_gp_seq
is uninitialized, needwait_gp must be false, which prevents wait_gp_seq
from being used. But this analysis is apparently beyond some compilers,
so this commit adds a bogus initialization of wait_gp_seq for the sole
purpose of suppressing the false-positive warning.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit removes the obsolete nocb_q_count and nocb_q_count_lazy
fields, also removing rcu_get_n_cbs_nocb_cpu(), adjusting
rcu_get_n_cbs_cpu(), and making rcutree_migrate_callbacks() once again
disable the ->cblist fields of offline CPUs.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Currently the RCU callbacks for no-CBs CPUs are queued on a series of
ad-hoc linked lists, which means that these callbacks cannot benefit
from "drive-by" grace periods, thus suffering needless delays prior
to invocation. In addition, the no-CBs grace-period kthreads first
wait for callbacks to appear and later wait for a new grace period,
which means that callbacks appearing during a grace-period wait can
be delayed. These delays increase memory footprint, and could even
result in an out-of-memory condition.
This commit therefore enqueues RCU callbacks from no-CBs CPUs on the
rcu_segcblist structure that is already used by non-no-CBs CPUs. It also
restructures the no-CBs grace-period kthread to be checking for incoming
callbacks while waiting for grace periods. Also, instead of waiting
for a new grace period, it waits for the closest grace period that will
cause some of the callbacks to be safe to invoke. All of these changes
reduce callback latency and thus the number of outstanding callbacks,
in turn reducing the probability of an out-of-memory condition.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
As a first step towards making no-CBs CPUs use the ->cblist, this commit
leaves the ->cblist enabled for these CPUs. The main reason to make
no-CBs CPUs use ->cblist is to take advantage of callback numbering,
which will reduce the effects of missed grace periods which in turn will
reduce forward-progress problems for no-CBs CPUs.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
The idea behind the checks for extended quiescent states at the end of
__call_rcu_nocb() is to handle cases where call_rcu() is invoked directly
from within an extended quiescent state, for example, from the idle loop.
However, this will result in a timer-mediated deferred wakeup, which
will cause the needed wakeup to happen within a jiffy or thereabouts.
There should be no forward-progress concerns, and if there are, the proper
response is to exit the extended quiescent state while executing the
endless blast of call_rcu() invocations, for example, using RCU_NONIDLE().
Given the more realistic case of an isolated call_rcu() invocation, there
should be no problem.
This commit therefore removes the checks for invoking call_rcu() within
an extended quiescent state for on no-CBs CPUs.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
RCU callback processing currently uses rcu_is_nocb_cpu() to determine
whether or not the current CPU's callbacks are to be offloaded.
This works, but it is not so good for cache locality. Plus use of
->cblist for offloaded callbacks will greatly increase the frequency
of these checks. This commit therefore adds a ->offloaded flag to the
rcu_segcblist structure to provide a more flexible and cache-friendly
means of checking for callback offloading.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
NULLing the RCU_NEXT_TAIL pointer was a clever way to save a byte, but
forward-progress considerations would require that this pointer be both
NULL and non-NULL, which, absent a quantum-computer port of the Linux
kernel, simply won't happen. This commit therefore creates as separate
->enabled flag to replace the current NULL checks.
[ paulmck: Add include files per 0day test robot and -next. ]
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit causes the no-CBs grace-period/callback hierarchy to be
printed to the console when the dump_tree kernel boot parameter is set.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit changes the name of the rcu_nocb_leader_stride kernel
boot parameter to rcu_nocb_gp_stride in order to account for the new
distinction between callback and grace-period no-CBs kthreads.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
The nocb_cb_wait() function traces a "FollowerSleep" trace_rcu_nocb_wake()
event, which never was documented and is now misleading. This commit
therefore changes "FollowerSleep" to "CBSleep", documents this, and
updates the documentation for "Sleep" as well.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit renames rdp_leader to rdp_gp in order to account for the
new distinction between callback and grace-period no-CBs kthreads.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit adjusts naming to account for the new distinction between
callback and grace-period no-CBs kthreads.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit adjusts naming to account for the new distinction between
callback and grace-period no-CBs kthreads. While in the area, it also
updates local variables.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit adjusts naming to account for the new distinction between
callback and grace-period no-CBs kthreads.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit adjusts naming to account for the new distinction between
callback and grace-period no-CBs kthreads.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Currently, there is one no-CBs rcuo kthread per CPU, and these kthreads
are divided into groups. The first rcuo kthread to come online in a
given group is that group's leader, and the leader both waits for grace
periods and invokes its CPU's callbacks. The non-leader rcuo kthreads
only invoke callbacks.
This works well in the real-time/embedded environments for which it was
intended because such environments tend not to generate all that many
callbacks. However, given huge floods of callbacks, it is possible for
the leader kthread to be stuck invoking callbacks while its followers
wait helplessly while their callbacks pile up. This is a good recipe
for an OOM, and rcutorture's new callback-flood capability does generate
such OOMs.
One strategy would be to wait until such OOMs start happening in
production, but similar OOMs have in fact happened starting in 2018.
It would therefore be wise to take a more proactive approach.
This commit therefore features per-CPU rcuo kthreads that do nothing
but invoke callbacks. Instead of having one of these kthreads act as
leader, each group has a separate rcog kthread that handles grace periods
for its group. Because these rcuog kthreads do not invoke callbacks,
callback floods on one CPU no longer block callbacks from reaching the
rcuc callback-invocation kthreads on other CPUs.
This change does introduce additional kthreads, however:
1. The number of additional kthreads is about the square root of
the number of CPUs, so that a 4096-CPU system would have only
about 64 additional kthreads. Note that recent changes
decreased the number of rcuo kthreads by a factor of two
(CONFIG_PREEMPT=n) or even three (CONFIG_PREEMPT=y), so
this still represents a significant improvement on most systems.
2. The leading "rcuo" of the rcuog kthreads should allow existing
scripting to affinity these additional kthreads as needed, the
same as for the rcuop and rcuos kthreads. (There are no longer
any rcuob kthreads.)
3. A state-machine approach was considered and rejected. Although
this would allow the rcuo kthreads to continue their dual
leader/follower roles, it complicates callback invocation
and makes it more difficult to consolidate rcuo callback
invocation with existing softirq callback invocation.
The introduction of rcuog kthreads should thus be acceptable.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit simply rewords comments to prepare for leader nocb kthreads
doing only grace-period work and callback shuffling. This will mean
the addition of replacement kthreads to invoke callbacks. The "leader"
and "follower" thus become less meaningful, so the commit changes no-CB
comments with these strings to "GP" and "CB", respectively. (Give or
take the usual grammatical transformations.)
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
This commit simply renames rcu_data fields to prepare for leader
nocb kthreads doing only grace-period work and callback shuffling.
This will mean the addition of replacement kthreads to invoke callbacks.
The "leader" and "follower" thus become less meaningful, so the commit
changes no-CB fields with these strings to "gp" and "cb", respectively.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
The return value of rcu_spawn_one_boost_kthread() is not used any longer.
This commit therefore changes its return type from int to void, and
removes the cast to void from its callers.
Signed-off-by: Byungchul Park <byungchul.park@lge.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Commit bb73c52bad ("rcu: Don't disable preemption for Tiny and Tree
RCU readers") removed the barrier() calls from rcu_read_lock() and
rcu_write_lock() in CONFIG_PREEMPT=n&&CONFIG_PREEMPT_COUNT=n kernels.
Within RCU, this commit was OK, but it failed to account for things like
get_user() that can pagefault and that can be reordered by the compiler.
Lack of the barrier() calls in rcu_read_lock() and rcu_read_unlock()
can cause these page faults to migrate into RCU read-side critical
sections, which in CONFIG_PREEMPT=n kernels could result in too-short
grace periods and arbitrary misbehavior. Please see commit 386afc9114
("spinlocks and preemption points need to be at least compiler barriers")
and Linus's commit 66be4e66a7 ("rcu: locking and unlocking need to
always be at least barriers"), this last of which restores the barrier()
call to both rcu_read_lock() and rcu_read_unlock().
This commit removes barrier() calls that are no longer needed given that
the addition of them in Linus's commit noted above. The combination of
this commit and Linus's commit effectively reverts commit bb73c52bad
("rcu: Don't disable preemption for Tiny and Tree RCU readers").
Reported-by: Herbert Xu <herbert@gondor.apana.org.au>
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
[ paulmck: Fix embarrassing typo located by Alan Stern. ]
Because __rcu_read_unlock() can be preempted just before the call to
rcu_read_unlock_special(), it is possible for a task to be preempted just
before it would have fully exited its RCU read-side critical section.
This would result in a needless extension of that critical section until
that task was resumed, which might in turn result in a needlessly
long grace period, needless RCU priority boosting, and needless
force-quiescent-state actions. Therefore, rcu_note_context_switch()
invokes __rcu_read_unlock() followed by rcu_preempt_deferred_qs() when
it detects this situation. This action by rcu_note_context_switch()
ends the RCU read-side critical section immediately.
Of course, once the task resumes, it will invoke rcu_read_unlock_special()
redundantly. This is harmless because the fact that a preemption
happened means that interrupts, preemption, and softirqs cannot
have been disabled, so there would be no deferred quiescent state.
While ->rcu_read_lock_nesting remains less than zero, none of the
->rcu_read_unlock_special.b bits can be set, and they were all zeroed by
the call to rcu_note_context_switch() at task-preemption time. Therefore,
setting ->rcu_read_unlock_special.b.exp_hint to false has no effect.
Therefore, the extra call to rcu_preempt_deferred_qs_irqrestore()
would return immediately. With one possible exception, which is
if an expedited grace period started just as the task was being
resumed, which could leave ->exp_deferred_qs set. This will cause
rcu_preempt_deferred_qs_irqrestore() to invoke rcu_report_exp_rdp(),
reporting the quiescent state, just as it should. (Such an expedited
grace period won't affect the preemption code path due to interrupts
having already been disabled.)
But when rcu_note_context_switch() invokes __rcu_read_unlock(), it
is doing so with preemption disabled, hence __rcu_read_unlock() will
unconditionally defer the quiescent state, only to immediately invoke
rcu_preempt_deferred_qs(), thus immediately reporting the deferred
quiescent state. It turns out to be safe (and faster) to instead
just invoke rcu_preempt_deferred_qs() without the __rcu_read_unlock()
middleman.
Because this is the invocation during the preemption (as opposed to
the invocation just after the resume), at least one of the bits in
->rcu_read_unlock_special.b must be set and ->rcu_read_lock_nesting
must be negative. This means that rcu_preempt_need_deferred_qs() must
return true, avoiding the early exit from rcu_preempt_deferred_qs().
Thus, rcu_preempt_deferred_qs_irqrestore() will be invoked immediately,
as required.
This commit therefore simplifies the CONFIG_PREEMPT=y version of
rcu_note_context_switch() by removing the "else if" branch of its
"if" statement. This change means that all callers that would have
invoked rcu_read_unlock_special() followed by rcu_preempt_deferred_qs()
will now simply invoke rcu_preempt_deferred_qs(), thus avoiding the
rcu_read_unlock_special() middleman when __rcu_read_unlock() is preempted.
Cc: rcu@vger.kernel.org
Cc: kernel-team@android.com
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Threaded interrupts provide additional interesting interactions between
RCU and raise_softirq() that can result in self-deadlocks in v5.0-2 of
the Linux kernel. These self-deadlocks can be provoked in susceptible
kernels within a few minutes using the following rcutorture command on
an 8-CPU system:
tools/testing/selftests/rcutorture/bin/kvm.sh --duration 5 --configs "TREE03" --bootargs "threadirqs"
Although post-v5.2 RCU commits have at least greatly reduced the
probability of these self-deadlocks, this was entirely by accident.
Although this sort of accident should be rowdily celebrated on those
rare occasions when it does occur, such celebrations should be quickly
followed by a principled patch, which is what this patch purports to be.
The key point behind this patch is that when in_interrupt() returns
true, __raise_softirq_irqoff() will never attempt a wakeup. Therefore,
if in_interrupt(), calls to raise_softirq*() are both safe and
extremely cheap.
This commit therefore replaces the in_irq() calls in the "if" statement
in rcu_read_unlock_special() with in_interrupt() and simplifies the
"if" condition to the following:
if (irqs_were_disabled && use_softirq &&
(in_interrupt() ||
(exp && !t->rcu_read_unlock_special.b.deferred_qs))) {
raise_softirq_irqoff(RCU_SOFTIRQ);
} else {
/* Appeal to the scheduler. */
}
The rationale behind the "if" condition is as follows:
1. irqs_were_disabled: If interrupts are enabled, we should
instead appeal to the scheduler so as to let the upcoming
irq_enable()/local_bh_enable() do the rescheduling for us.
2. use_softirq: If this kernel isn't using softirq, then
raise_softirq_irqoff() will be unhelpful.
3. a. in_interrupt(): If this returns true, the subsequent
call to raise_softirq_irqoff() is guaranteed not to
do a wakeup, so that call will be both very cheap and
quite safe.
b. Otherwise, if !in_interrupt() the raise_softirq_irqoff()
might do a wakeup, which is expensive and, in some
contexts, unsafe.
i. The "exp" (an expedited RCU grace period is being
blocked) says that the wakeup is worthwhile, and:
ii. The !.deferred_qs says that scheduler locks
cannot be held, so the wakeup will be safe.
Backporting this requires considerable care, so no auto-backport, please!
Fixes: 05f415715c ("rcu: Speed up expedited GPs when interrupting RCU reader")
Reported-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
In !use_softirq runs, we clearly cannot rely on raise_softirq() and
its lightweight bit setting, so we must instead do some form of wakeup.
In the absence of a self-IPI when interrupts are disabled, these wakeups
can be delayed until the next interrupt occurs. This means that calling
invoke_rcu_core() doesn't actually do any expediting.
In this case, it is better to take the "else" clause, which sets the
current CPU's resched bits and, if there is an expedited grace period
in flight, uses IRQ-work to force the needed self-IPI. This commit
therefore removes the "else if" clause that calls invoke_rcu_core().
Reported-by: Scott Wood <swood@redhat.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
The dump_blkd_tasks() function dumps at most 10 blocked tasks, ignoring
the value of the ncheck parameter. This commit therefore substitutes
the value of ncheck for the hard-coded value of 10. Because all callers
currently pass 10 as the number, this patch does not change behavior,
but it is clearly an accident waiting to happen.
Signed-off-by: Neeraj Upadhyay <neeraju@codeaurora.org>
Reviewed-by: Mukesh Ojha <mojha@codeaurora.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
The rcu_data structure's ->deferred_qs field is used to indicate that the
current CPU is blocking an expedited grace period (perhaps a future one).
Given that it is used only for expedited grace periods, its current name
is misleading, so this commit renames it to ->exp_deferred_qs.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
When rcu_read_unlock_special() is invoked with interrupts disabled, is
either not in an interrupt handler or is not using RCU_SOFTIRQ, is not
the first RCU read-side critical section in the chain, and either there
is an expedited grace period in flight or this is a NO_HZ_FULL kernel,
the end of the grace period can be unduly delayed. The reason for this
is that it is not safe to do wakeups in this situation.
This commit fixes this problem by using the irq_work subsystem to
force a later interrupt handler in a clean environment. Because
set_tsk_need_resched(current) and set_preempt_need_resched() are
invoked prior to this, the scheduler will force a context switch
upon return from this interrupt (though perhaps at the end of any
interrupted preempt-disable or BH-disable region of code), which will
invoke rcu_note_context_switch() (again in a clean environment), which
will in turn give RCU the chance to report the deferred quiescent state.
Of course, by then this task might be within another RCU read-side
critical section. But that will be detected at that time and reporting
will be further deferred to the outermost rcu_read_unlock(). See
rcu_preempt_need_deferred_qs() and rcu_preempt_deferred_qs() for more
details on the checking.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
When running in an interrupt handler, raise_softirq() and
raise_softirq_irqoff() have extremely low overhead: They simply set a
bit in a per-CPU mask, which is checked upon exit from that interrupt
handler. Therefore, if rcu_read_unlock_special() is invoked within an
interrupt handler and RCU_SOFTIRQ is in use, this commit make use of
raise_softirq_irqoff() even if there is no expedited grace period in
flight and even if this is not a nohz_full CPU.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Currently, rcu_read_unlock_special() will do wakeups whenever it is safe
to do so. However, wakeups are expensive, and they are only really
needed when the just-ended RCU read-side critical section is blocking
an expedited grace period (in which case speed is of the essence)
or on a nohz_full CPU (where it might be a good long time before an
interrupt arrives). This commit therefore checks for these conditions,
and does the expensive wakeups only if doing so would be useful.
Note it can be rather expensive to determine whether or not the current
task (as opposed to the current CPU) is blocking the current expedited
grace period. Doing so requires traversing the ->blkd_tasks list, which
can be quite long. This commit therefore cheats: If the current task
is on a given ->blkd_tasks list, and some task on that list is blocking
the current expedited grace period, the code assumes that the current
task is blocking that expedited grace period.
Reported-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
When RCU core processing is offloaded from RCU_SOFTIRQ to the rcuc
kthreads, a full and unconditional wakeup is required to initiate RCU
core processing. In contrast, when RCU core processing is carried
out by RCU_SOFTIRQ, a raise_softirq() suffices. Of course, there are
situations where raise_softirq() does a full wakeup, but these do not
occur with normal usage of rcu_read_unlock().
The reason that full wakeups can be problematic is that the scheduler
sometimes invokes rcu_read_unlock() with its pi or rq locks held,
which can of course result in deadlock in CONFIG_PREEMPT=y kernels when
rcu_read_unlock() invokes the scheduler. Scheduler invocations can happen
in the following situations: (1) The just-ended reader has been subjected
to RCU priority boosting, in which case rcu_read_unlock() must deboost,
(2) Interrupts were disabled across the call to rcu_read_unlock(), so
the quiescent state must be deferred, requiring a wakeup of the rcuc
kthread corresponding to the current CPU.
Now, the scheduler may hold one of its locks across rcu_read_unlock()
only if preemption has been disabled across the entire RCU read-side
critical section, which in the days prior to RCU flavor consolidation
meant that rcu_read_unlock() never needed to do wakeups. However, this
is no longer the case for any but the first rcu_read_unlock() following a
condition (e.g., preempted RCU reader) requiring special rcu_read_unlock()
attention. For example, an RCU read-side critical section might be
preempted, but preemption might be disabled across the rcu_read_unlock().
The rcu_read_unlock() must defer the quiescent state, and therefore
leaves the task queued on its leaf rcu_node structure. If a scheduler
interrupt occurs, the scheduler might well invoke rcu_read_unlock() with
one of its locks held. However, the preempted task is still queued, so
rcu_read_unlock() will attempt to defer the quiescent state once more.
When RCU core processing is carried out by RCU_SOFTIRQ, this works just
fine: The raise_softirq() function simply sets a bit in a per-CPU mask
and the RCU core processing will be undertaken upon return from interrupt.
Not so when RCU core processing is carried out by the rcuc kthread: In this
case, the required wakeup can result in deadlock.
The initial solution to this problem was to use set_tsk_need_resched() and
set_preempt_need_resched() to force a future context switch, which allows
rcu_preempt_note_context_switch() to report the deferred quiescent state
to RCU's core processing. Unfortunately for expedited grace periods,
there can be a significant delay between the call for a context switch
and the actual context switch.
This commit therefore introduces a ->deferred_qs flag to the task_struct
structure's rcu_special structure. This flag is initially false, and
is set to true by the first call to rcu_read_unlock() requiring special
attention, then finally reset back to false when the quiescent state is
finally reported. Then rcu_read_unlock() attempts full wakeups only when
->deferred_qs is false, that is, on the first rcu_read_unlock() requiring
special attention. Note that a chain of RCU readers linked by some other
sort of reader may find that a later rcu_read_unlock() is once again able
to do a full wakeup, courtesy of an intervening preemption:
rcu_read_lock();
/* preempted */
local_irq_disable();
rcu_read_unlock(); /* Can do full wakeup, sets ->deferred_qs. */
rcu_read_lock();
local_irq_enable();
preempt_disable()
rcu_read_unlock(); /* Cannot do full wakeup, ->deferred_qs set. */
rcu_read_lock();
preempt_enable();
/* preempted, >deferred_qs reset. */
local_irq_disable();
rcu_read_unlock(); /* Can again do full wakeup, sets ->deferred_qs. */
Such linked RCU readers do not yet seem to appear in the Linux kernel, and
it is probably best if they don't. However, RCU needs to handle them, and
some variations on this theme could make even raise_softirq() unsafe due to
the possibility of its doing a full wakeup. This commit therefore also
avoids invoking raise_softirq() when the ->deferred_qs set flag is set.
Signed-off-by: Paul E. McKenney <paulmck@linux.ibm.com>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
