Pull scheduler updates from Thomas Gleixner:
- Cleanup and improvement of NUMA balancing
- Refactoring and improvements to the PELT (Per Entity Load Tracking)
code
- Watchdog simplification and related cleanups
- The usual pile of small incremental fixes and improvements
* 'sched-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (41 commits)
watchdog: Reduce message verbosity
stop_machine: Reflow cpu_stop_queue_two_works()
sched/numa: Move task_numa_placement() closer to numa_migrate_preferred()
sched/numa: Use group_weights to identify if migration degrades locality
sched/numa: Update the scan period without holding the numa_group lock
sched/numa: Remove numa_has_capacity()
sched/numa: Modify migrate_swap() to accept additional parameters
sched/numa: Remove unused task_capacity from 'struct numa_stats'
sched/numa: Skip nodes that are at 'hoplimit'
sched/debug: Reverse the order of printing faults
sched/numa: Use task faults only if numa_group is not yet set up
sched/numa: Set preferred_node based on best_cpu
sched/numa: Simplify load_too_imbalanced()
sched/numa: Evaluate move once per node
sched/numa: Remove redundant field
sched/debug: Show the sum wait time of a task group
sched/fair: Remove #ifdefs from scale_rt_capacity()
sched/core: Remove get_cpu() from sched_fork()
sched/cpufreq: Clarify sugov_get_util()
sched/sysctl: Remove unused sched_time_avg_ms sysctl
...
The rcutorture test module currently increments both successes and error
for the barrier test upon error, which results in misleading statistics
being printed. This commit therefore changes the code to increment the
success counter only when the test actually passes.
This change was tested by by returning from the barrier callback without
incrementing the callback counter, thus introducing what appeared to
rcutorture to be rcu_barrier() failures.
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
When rcutorture is built in to the kernel, an earlier patch detects
that and raises the priority of RCU's kthreads to allow rcutorture's
RCU priority boosting tests to succeed.
However, if rcutorture is built as a module, those priorities must be
raised manually via the rcutree.kthread_prio kernel boot parameter.
If this manual step is not taken, rcutorture's RCU priority boosting
tests will fail due to kthread starvation. One approach would be to
raise the default priority, but that risks breaking existing users.
Another approach would be to allow runtime adjustment of RCU's kthread
priorities, but that introduces numerous "interesting" race conditions.
This patch therefore instead detects too-low priorities, and prints a
message and disables the RCU priority boosting tests in that case.
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The get_seconds() call is deprecated because it overflows on 32-bit
architectures. The algorithm in rcu_torture_stall() can deal with
the overflow, but another problem here is that using a CLOCK_REALTIME
stamp can lead to a false-positive stall warning when a settimeofday()
happens concurrently.
Using ktime_get_seconds() instead avoids those issues and will never
overflow. The added cast to 'unsigned long' however is necessary to
make ULONG_CMP_LT() work correctly.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Currently, with RCU_BOOST disabled, I get no failures when forcing
rcutorture to test RCU boost priority inversion. The reason seems to be
that we don't check for failures if the callback never ran at all for
the duration of the boost-test loop.
Further, the 'rtb' and 'rtbf' counters seem to be used inconsistently.
'rtb' is incremented at the start of each test and 'rtbf' is incremented
per-cpu on each failure of call_rcu. So its possible 'rtbf' > 'rtb'.
To test the boost with rcutorture, I did following on a 4-CPU x86 machine:
modprobe rcutorture test_boost=2
sleep 20
rmmod rcutorture
With patch:
rtbf: 8 rtb: 12
Without patch:
rtbf: 0 rtb: 2
In summary this patch:
- Increments failed and total test counters once per boost-test.
- Checks for failure cases correctly.
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Currently rcutorture is not able to torture RCU boosting properly. This
is because the rcutorture's boost threads which are doing the torturing
may be throttled due to RT throttling.
This patch makes rcutorture use the right torture technique (unthrottled
rcutorture boost tasks) for torturing RCU so that the test fails
correctly when no boost is available.
Currently this requires accessing sysctl_sched_rt_runtime directly, but
that should be Ok since rcutorture is test code. Such direct access is
also only possible if rcutorture is used as a built-in so make it
conditional on that.
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
For RCU implementations supporting multiple types of reader protection,
rcutorture currently randomly selects the combinations of types of
protection for each phase of each reader. The problem with this,
for example, given the four kinds of protection for RCU-sched
(local_irq_disable(), local_bh_disable(), preempt_disable(), and
rcu_read_lock_sched()), the reader will be protected by a single
mechanism only 25% of the time. We really heavier testing of single
read-side mechanisms.
This commit therefore uses only a single mechanism about 60% of the time,
half of the time explicitly and one-eighth of the time by chance.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This commit enables rcutorture to test whether RCU properly aggregates
different types of read-side critical sections into a larger section
covering the set. It does this by extending an initial read-side
critical section randomly for a random number of extensions. There is
a new rcu_torture_ops field ->extendable that specifies what extensions
are permitted for a given flavor of RCU (for example, SRCU does not
permit any extensions, while RCU-sched permits all types). Note that
if a given operation (for example, local_bh_disable()) extends an RCU
read-side critical section, then rcutorture feels free to also start
and end the critical section with that operation's type of disabling.
Disabling operations include local_bh_disable(), local_irq_disable(),
and preempt_disable(). This commit also adds a new "busted_srcud"
torture type, which verifies rcutorture's ability to detect extensions
of RCU read-side critical sections that are not handled. Gotta test
the test, after all!
Note that it is not legal to invoke local_bh_disable() with interrupts
disabled, and this transition is avoided by overriding the random-number
generator when it wants to call local_bh_disable() while interrupts
are disabled. The code instead leaves both interrupts and bh/softirq
disabled in this case.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This commit saves a few lines of code by making rcu_torture_timer()
invoke rcu_torture_one_read(), thus completing the consolidation of
code between rcu_torture_timer() and rcu_torture_reader().
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Currently, the rcu_torture_timer() function uses a single global
torture_random_state structure protected by a single global lock.
This conflicts to some extent with performance and scalability,
but even more with the goal of consolidating read-side testing
with rcu_torture_reader(). This commit therefore creates a per-CPU
torture_random_state structure for use by rcu_torture_timer() and
eliminates the lock.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
[ paulmck: Make rcu_torture_timer_rand static, per 0day Test Robot report. ]
Currently, rcu_torture_timer() relies on a lock to guard updates to
n_rcu_torture_timers. Unfortunately, consolidating code with
rcu_torture_reader() will dispense with this lock. This commit
therefore makes n_rcu_torture_timers be an atomic_long_t and uses
atomic_long_inc() to carry out the update.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This commit extracts the code executed on each pass through the loop
in rcu_torture_reader() into a new rcu_torture_one_read() function.
This new function will also be used by rcu_torture_timer().
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Back when RCU had a debugfs interface, there was a test version and
sequence number that allowed associating debugfs data with a particular
test run, where the test run started with modprobe and ended with rmmod,
which was how tests were run back on the old ABAT system within IBM.
But rcutorture testing no longer runs on ABAT, and there is no longer an
RCU debugfs interface, so there is no longer any need for test versions
and sequence numbers.
This commit therefore removes the rcutorture_record_test_transition()
and rcutorture_record_progress() functions, and along with them the
rcutorture_testseq and rcutorture_vernum variables that they update.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Some RCU bugs have been sensitive to the frequency of CPU-hotplug
operations, which have been gradually increased over time. But this
frequency is now at the one-second lower limit that can be specified using
the rcutorture.onoff_interval kernel parameter. This commit therefore
changes the units of rcutorture.onoff_interval from seconds to jiffies,
and also sets the value specified for this kernel parameter in the TREE03
rcutorture scenario to 200, which is 200 milliseconds for HZ=1000.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The rcutorture RCU priority boosting tests fail even with CONFIG_RCU_BOOST
set because rcutorture's threads run at the same priority as the default
RCU kthreads (RT class with priority of 1).
This patch checks if RCU torture is built into the kernel and if so,
assigns RT priority 1 to the RCU threads, allowing the rcutorture boost
tests to pass.
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This commit adds the SRCU grace-period number to the rcutorture statistics
printout, which allows it to be compared to the rcutorture "Writer stall
state" message.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The ->dynticks_nmi_nesting field records the nesting depth of both
interrupt and NMI handlers. Because the kernel can enter interrupts
and never leave them (and vice versa) and because NMIs can interrupt
manipulation of the ->dynticks_nmi_nesting field, the values in this
field must be both chosen and maniupated very carefully. As a result,
although the value is zero when the corresponding CPU is executing
neither an interrupt nor an NMI handler, it is 4,611,686,018,427,387,906
on 64-bit systems when there is a single level of interrupt/NMI handling
in progress.
This number is difficult to remember and interpret, so this commit
switches the output to hexadecimal, resulting in the much nicer
0x4000000000000002.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The current implementatation of rcu_seq_diff() follows tradition in
providing a rough-and-ready approximation of the number of elapsed grace
periods between the two rcu_seq values. However, this difference is
used to flag RCU-failure "near misses", which can be a valuable debugging
aid, so more exactitude would be an improvement. This commit therefore
improves the accuracy of rcu_seq_diff().
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Currently, the range of jiffies_till_{first,next}_fqs are checked and
adjusted on and on in the loop of rcu_gp_kthread on runtime.
However, it's enough to check them only when setting them, not every
time in the loop. So make them handled on a setting time via sysfs.
Signed-off-by: Byungchul Park <byungchul.park@lge.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This commit adds any in-the-future ->gp_seq_needed fields to the
diagnostics for an rcutorture writer stall warning message.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
At the end of rcu_tasks_kthread() there's a lonely
schedule_timeout_uninterruptible() call with no apparent rationale for
its existence. But there is. It is to keep the thread from going into
a tight loop if there's some anomaly. That really needs a comment.
Link: http://lkml.kernel.org/r/20180524223839.GU3803@linux.vnet.ibm.com
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Joel Fernandes found that the synchronize_rcu_tasks() was taking a
significant amount of time. He demonstrated it with the following test:
# cd /sys/kernel/tracing
# while [ 1 ]; do x=1; done &
# echo '__schedule_bug:traceon' > set_ftrace_filter
# time echo '!__schedule_bug:traceon' > set_ftrace_filter;
real 0m1.064s
user 0m0.000s
sys 0m0.004s
Where it takes a little over a second to perform the synchronize,
because there's a loop that waits 1 second at a time for tasks to get
through their quiescent points when there's a task that must be waited
for.
After discussion we came up with a simple way to wait for holdouts but
increase the time for each iteration of the loop but no more than a
full second.
With the new patch we have:
# time echo '!__schedule_bug:traceon' > set_ftrace_filter;
real 0m0.131s
user 0m0.000s
sys 0m0.004s
Which drops it down to 13% of what the original wait time was.
Link: http://lkml.kernel.org/r/20180523063815.198302-2-joel@joelfernandes.org
Reported-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Suggested-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
rcu_seq_snap may be tricky to decipher. Lets document how it works with
an example to make it easier.
Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
[ paulmck: Shrink comment as suggested by Peter Zijlstra. ]
Currently, rcu_check_gp_start_stall() waits for one second after the first
request before complaining that a grace period has not yet started. This
was desirable while testing the conversion from ->future_gp_needed[] to
->gp_seq_needed, but it is a bit on the hair-trigger side for production
use under heavy load. This commit therefore makes this wait time be
exactly that of the RCU CPU stall warning, allowing easy adjustment of
both timeouts to suit the distribution or installation at hand.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The rcu_cpu_has_callbacks() function is now used in all configurations,
so this commit removes the __maybe_unused.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This function is in rcuperf.c, which is not an include file, so there
is no problem dropping the "inline", especially given that this function
is invoked only twice per rcuperf run. This commit therefore delegates
the inlining decision to the compiler by dropping the "inline".
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This function is in rcutorture.c, which is not an include file, so there
is no problem dropping the "inline", especially given that this function
is invoked only twice per rcutorture run. This commit therefore delegates
the inlining decision to the compiler by dropping the "inline".
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
These functions are in kernel/rcu/tree.c, which is not an include file,
so there is no problem dropping the "inline", especially given that these
functions are nowhere near a fastpath. This commit therefore delegates
the inlining decision to the compiler by dropping the "inline".
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
One danger of using __maybe_unused is that the compiler doesn't yell
at you when you remove the last reference, witness rcu_bind_gp_kthread()
and its local variable "cpu". This commit removes this local variable.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The rcu_kick_nohz_cpu() function is no longer used, and the functionality
it used to provide is now provided by a call to resched_cpu() in the
force-quiescent-state function rcu_implicit_dynticks_qs(). This commit
therefore removes rcu_kick_nohz_cpu().
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The rcu_preempt_qs() function only applies to the CPU, not the task.
A task really is allowed to invoke this function while in an RCU-preempt
read-side critical section, but only if it has first added itself to
some leaf rcu_node structure's ->blkd_tasks list.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The rcu_dynticks_momentary_idle() function is invoked only from
rcu_momentary_dyntick_idle(), and neither function is particularly
large. This commit therefore saves a few lines by inlining
rcu_dynticks_momentary_idle() into rcu_momentary_dyntick_idle().
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
If any scheduling-clock interrupt interrupts an RCU-preempt read-side
critical section, the interrupted task's ->rcu_read_unlock_special.b.need_qs
field is set. This causes the outermost rcu_read_unlock() to incur the
extra overhead of calling into rcu_read_unlock_special(). This commit
reduces that overhead by setting ->rcu_read_unlock_special.b.need_qs only
if the grace period has been in effect for more than one second.
Why one second? Because this is comfortably smaller than the minimum
RCU CPU stall-warning timeout of three seconds, but long enough that the
.need_qs marking should happen quite rarely. And if your RCU read-side
critical section has run on-CPU for a full second, it is not unreasonable
to invest some CPU time in ending the grace period quickly.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The naming and comments associated with some RCU-tasks code make
the faulty assumption that context switches due to cond_resched()
are voluntary. As several people pointed out, this is not the case.
This commit therefore updates function names and comments to better
reflect current reality.
Reported-by: Byungchul Park <byungchul.park@lge.com>
Reported-by: Joel Fernandes <joel@joelfernandes.org>
Reported-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This commit also adjusts some whitespace while in the area.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
[ paulmck: Revert string-breaking %s as requested by Andy Shevchenko. ]
We expect a quiescent state of TASKS_RCU when cond_resched_tasks_rcu_qs()
is called, no matter whether it actually be scheduled or not. However,
it currently doesn't report the quiescent state when the task enters
into __schedule() as it's called with preempt = true. So make it report
the quiescent state unconditionally when cond_resched_tasks_rcu_qs() is
called.
And in TINY_RCU, even though the quiescent state of rcu_bh also should
be reported when the tick interrupt comes from user, it doesn't. So make
it reported.
Lastly in TREE_RCU, rcu_note_voluntary_context_switch() should be
reported when the tick interrupt comes from not only user but also idle,
as an extended quiescent state.
Signed-off-by: Byungchul Park <byungchul.park@lge.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
[ paulmck: Simplify rcutiny portion given no RCU-tasks for !PREEMPT. ]
Because rcu_read_unlock_special() is no longer used outside of
kernel/rcu/tree_plugin.h, this commit makes it static.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
CPUs are expected to report quiescent states when coming online and
when going offline, and grace-period initialization is supposed to
handle any race conditions where a CPU's ->qsmask bit is set just after
it goes offline. This commit adds diagnostics for the case where an
offline CPU nevertheless has a grace period waiting on it.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Grace-period initialization first processes any recent CPU-hotplug
operations, and then initializes state for the new grace period. These
two phases of initialization are currently not distinguished in debug
prints, but the distinction is valuable in a number of debug situations.
This commit therefore introduces two new values for ->gp_state,
RCU_GP_ONOFF and RCU_GP_INIT, in order to make this distinction.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Interactions between CPU-hotplug operations and grace-period
initialization can result in dump_blkd_tasks(). One of the first
debugging actions in this case is to search back in dmesg to work
out which of the affected rcu_node structure's CPUs are online and to
determine the last CPU-hotplug operation affecting any of those CPUs.
This can be laborious and error-prone, especially when console output
is lost.
This commit therefore causes dump_blkd_tasks() to dump the state of
the affected rcu_node structure's CPUs and the last grace period during
which the last offline and online operation affected each of these CPUs.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
This commit updates dump_blkd_tasks() to print out quiescent-state
bitmasks for the rcu_node structures further up the tree. This
information helps debugging of interactions between CPU-hotplug
operations and RCU grace-period initialization.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Now that quiescent states for newly offlined CPUs are reported either
when that CPU goes offline or at the end of grace-period initialization,
the CPU-hotplug failsafe in the force-quiescent-state code path is no
longer needed.
This commit therefore removes this failsafe.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Now that quiescent-state reporting is fully event-driven, this commit
removes the check for a lost quiescent state from force_qs_rnp().
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The main race with the early part of grace-period initialization appears
to be with CPU hotplug. To more fully open this race window, this commit
moves the rcu_gp_slow() from the beginning of the early initialization
loop to follow that loop, thus widening the race window, especially for
the rcu_node structures that are initialized last. This commit also
expands rcutree.gp_preinit_delay from 3 to 12, giving the same overall
delay in the grace period, but concentrated in the spot where it will
do the most good.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
RCU should only be waiting on CPUs that were online at the time that the
current grace period started. Failure to abide by this rule can result
in confusing splats during grace-period cleanup and initialization.
This commit therefore adds a check to RCU-preempt's preempted-task
queuing that checks for waiting on newly onlined CPUs.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Without special fail-safe quiescent-state-propagation checks, grace-period
hangs can result from the following scenario:
1. CPU 1 goes offline.
2. Because CPU 1 is the only CPU in the system blocking the current
grace period, the grace period ends as soon as
rcu_cleanup_dying_idle_cpu()'s call to rcu_report_qs_rnp()
returns.
3. At this point, the leaf rcu_node structure's ->lock is no longer
held: rcu_report_qs_rnp() has released it, as it must in order
to awaken the RCU grace-period kthread.
4. At this point, that same leaf rcu_node structure's ->qsmaskinitnext
field still records CPU 1 as being online. This is absolutely
necessary because the scheduler uses RCU (in this case on the
wake-up path while awakening RCU's grace-period kthread), and
->qsmaskinitnext contains RCU's idea as to which CPUs are online.
Therefore, invoking rcu_report_qs_rnp() after clearing CPU 1's
bit from ->qsmaskinitnext would result in a lockdep-RCU splat
due to RCU being used from an offline CPU.
5. RCU's grace-period kthread awakens, sees that the old grace period
has completed and that a new one is needed. It therefore starts
a new grace period, but because CPU 1's leaf rcu_node structure's
->qsmaskinitnext field still shows CPU 1 as being online, this new
grace period is initialized to wait for a quiescent state from the
now-offline CPU 1.
6. Without the fail-safe force-quiescent-state checks, there would
be no quiescent state from the now-offline CPU 1, which would
eventually result in RCU CPU stall warnings and memory exhaustion.
It would be good to get rid of the special fail-safe quiescent-state
propagation checks, and thus it would be good to fix things so that
the above scenario cannot happen. This commit therefore adds a new
->ofl_lock to the rcu_state structure. This lock is held by rcu_gp_init()
across the applying of buffered online and offline operations to the
rcu_node tree, and it is also held by rcu_cleanup_dying_idle_cpu()
when buffering a new offline operation. This prevents rcu_gp_init()
from acquiring the leaf rcu_node structure's lock during the interval
between when rcu_cleanup_dying_idle_cpu() invokes rcu_report_qs_rnp(),
which releases ->lock and the re-acquisition of that same lock.
This in turn prevents the failure scenario outlined above, and will
hopefully eventually allow removal of the offline-CPU checks from the
force-quiescent-state code path.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Without special fail-safe quiescent-state-propagation checks, grace-period
hangs can result from the following scenario:
1. A task running on a given CPU is preempted in its RCU read-side
critical section.
2. That CPU goes offline, and there are now no online CPUs
corresponding to that CPU's leaf rcu_node structure.
3. The rcu_gp_init() function does the first phase of grace-period
initialization, and sets the aforementioned leaf rcu_node
structure's ->qsmaskinit field to all zeroes. Because there
is a blocked task, it does not propagate the zeroing of either
->qsmaskinit or ->qsmaskinitnext up the rcu_node tree.
4. The task resumes on some other CPU and exits its critical section.
There is no grace period in progress, so the resulting quiescent
state is not reported up the tree.
5. The rcu_gp_init() function does the second phase of grace-period
initialization, which results in the leaf rcu_node structure
being initialized to expect no further quiescent states, but
with that structure's parent expecting a quiescent-state report.
The parent will never receive a quiescent state from this leaf
rcu_node structure, so the grace period will hang, resulting in
RCU CPU stall warnings.
It would be good to get rid of the special fail-safe quiescent-state
propagation checks. This commit therefore checks the leaf rcu_node
structure's ->wait_blkd_tasks field during grace-period initialization.
If this flag is set, the rcu_report_qs_rnp() is invoked to immediately
report the possible quiescent state. While in the neighborhood, this
commit also report quiescent states for any CPUs that went offline between
the two phases of grace-period initialization, thus reducing grace-period
delays and hopefully eventually allowing removal of offline-CPU checks
from the force-quiescent-state code path.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Consider the following sequence of events in a PREEMPT=y kernel:
1. All CPUs corresponding to a given leaf rcu_node structure are
offline.
2. The first phase of the rcu_gp_init() function's grace-period
initialization runs, and sets that rcu_node structure's
->qsmaskinit to zero, as it should.
3. One of the CPUs corresponding to that rcu_node structure comes
back online. Note that because this CPU came online after the
grace period started, this grace period can safely ignore this
newly onlined CPU.
4. A task running on the newly onlined CPU enters an RCU-preempt
read-side critical section, and is then preempted. Because
the corresponding rcu_node structure's ->qsmask is zero,
rcu_preempt_ctxt_queue() leaves the rcu_node structure's
->gp_tasks field NULL, as it should.
5. The rcu_gp_init() function continues running the second phase of
grace-period initialization. The ->qsmask field of the parent of
the aforementioned leaf rcu_node structure is set to not expect
a quiescent state from the leaf, as is only right and proper.
However, when rcu_gp_init() reaches the leaf, it invokes
rcu_preempt_check_blocked_tasks(), which sees that the leaf's
->blkd_tasks list is non-empty, and therefore sets the leaf's
->gp_tasks field to reference the first task on that list.
6. The grace period ends before the preempted task resumes, which
is perfectly fine, given that this grace period was under no
obligation to wait for that task to exit its late-starting
RCU-preempt read-side critical section. Unfortunately, the
leaf's ->gp_tasks field is non-NULL, so rcu_gp_cleanup() splats.
After all, it appears to rcu_gp_cleanup() that the grace period
failed to wait for a task that was supposed to be blocking that
grace period.
This commit avoids this false-positive splat by adding a check of both
->qsmaskinit and ->wait_blkd_tasks to rcu_preempt_check_blocked_tasks().
If both ->qsmaskinit and ->wait_blkd_tasks are zero, then the task must
have entered its RCU-preempt read-side critical section late (after all,
the CPU that it is running on was not online at that time), which means
that the upper-level rcu_node structure won't be waiting for anything
on the leaf anyway.
If ->wait_blkd_tasks is non-zero, then there is at least one task on
ths rcu_node structure's ->blkd_tasks list whose RCU read-side
critical section predates the current grace period. If ->qsmaskinit
is non-zero, there is at least one CPU that was online at the start
of the current grace period. Thus, if both are zero, there is nothing
to wait for.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Linus Torvalds
Paul E. McKenney
Joel Fernandes (Google)
Arnd Bergmann
Byungchul Park
Steven Rostedt (VMware)
Joe Perches