Linu Cherian reported a WARN in cleanup_srcu_struct() when shutting
down a guest running iperf on a VFIO assigned device. This happens
because irqfd_wakeup() calls srcu_read_lock(&kvm->irq_srcu) in interrupt
context, while a worker thread does the same inside kvm_set_irq(). If the
interrupt happens while the worker thread is executing __srcu_read_lock(),
updates to the Classic SRCU ->lock_count[] field or the Tree SRCU
->srcu_lock_count[] field can be lost.
The docs say you are not supposed to call srcu_read_lock() and
srcu_read_unlock() from irq context, but KVM interrupt injection happens
from (host) interrupt context and it would be nice if SRCU supported the
use case. KVM is using SRCU here not really for the "sleepable" part,
but rather due to its IPI-free fast detection of grace periods. It is
therefore not desirable to switch back to RCU, which would effectively
revert commit 719d93cd5f ("kvm/irqchip: Speed up KVM_SET_GSI_ROUTING",
2014-01-16).
However, the docs are overly conservative. You can have an SRCU instance
only has users in irq context, and you can mix process and irq context
as long as process context users disable interrupts. In addition,
__srcu_read_unlock() actually uses this_cpu_dec() on both Tree SRCU and
Classic SRCU. For those two implementations, only srcu_read_lock()
is unsafe.
When Classic SRCU's __srcu_read_unlock() was changed to use this_cpu_dec(),
in commit 5a41344a3d ("srcu: Simplify __srcu_read_unlock() via
this_cpu_dec()", 2012-11-29), __srcu_read_lock() did two increments.
Therefore it kept __this_cpu_inc(), with preempt_disable/enable in
the caller. Tree SRCU however only does one increment, so on most
architectures it is more efficient for __srcu_read_lock() to use
this_cpu_inc(), and any performance differences appear to be down in
the noise.
Unlike Classic and Tree SRCU, Tiny SRCU does increments and decrements on
a single variable. Therefore, as Peter Zijlstra pointed out, Tiny SRCU's
implementation already supports mixed-context use of srcu_read_lock()
and srcu_read_unlock(), at least as long as uses of srcu_read_lock()
and srcu_read_unlock() in each handler are nested and paired properly.
In other words, it is still illegal to (say) invoke srcu_read_lock()
in an interrupt handler and to invoke the matching srcu_read_unlock()
in a softirq handler. Therefore, the only change required for Tiny SRCU
is to its comments.
Fixes: 719d93cd5f ("kvm/irqchip: Speed up KVM_SET_GSI_ROUTING")
Reported-by: Linu Cherian <linuc.decode@gmail.com>
Suggested-by: Linu Cherian <linuc.decode@gmail.com>
Cc: kvm@vger.kernel.org
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Paolo Bonzini <pbonzini@redhat.com>
Linus noticed that the <linux/rcu_segcblist.h> has huge inline functions
which should not be inline at all.
As a first step in cleaning this up, move them all to kernel/rcu/ and
only keep an absolute minimum of data type defines in the header:
before: -rw-r--r-- 1 mingo mingo 22284 May 2 10:25 include/linux/rcu_segcblist.h
after: -rw-r--r-- 1 mingo mingo 3180 May 2 10:22 include/linux/rcu_segcblist.h
More can be done, such as uninlining the large functions, which inlining
is unjustified even if it's an RCU internal matter.
Reported-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
The default value for the kernel boot parameter srcutree.exp_holdoff
is 50 microseconds, which is too long for good Tree SRCU performance
(compared to Classic SRCU) on the workloads tested by Mike Galbraith.
This commit therefore sets the default value to 25 microseconds, which
shows excellent results in Mike's testing.
Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Mike Galbraith <efault@gmx.de>
On small systems, in the absence of readers, expedited SRCU grace
periods can complete in less than a microsecond. This means that an
eight-CPU system can have all CPUs doing synchronize_srcu() in a tight
loop and almost always expedite. This might actually be desirable in
some situations, but in general it is a good way to needlessly burn
CPU cycles. And in those situations where it is desirable, your friend
is the function synchronize_srcu_expedited().
For other situations, this commit adds a kernel parameter that specifies
a holdoff between completing the last SRCU grace period and auto-expediting
the next. If the next grace period starts before the holdoff expires,
auto-expediting is disabled. The holdoff is 50 microseconds by default,
and can be tuned to the desired number of nanoseconds. A value of zero
disables auto-expediting.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Mike Galbraith <efault@gmx.de>
Classic SRCU in effect expedites the first synchronize_srcu() when SRCU
is idle, and Mike Galbraith demonstrated that some use cases do in fact
rely on this behavior. In particular, Mike showed that Steven Rostedt's
hotplug stress script takes 55 seconds with Classic SRCU and more than
16 -minutes- when running Tree SRCU. Assuming that each Tree SRCU's call
to synchronize_srcu() takes four milliseconds, this implies that Steven's
test invokes synchronize_srcu() in isolation, but more than once per
200 microseconds. Mike used ftrace to demonstrate that the time between
successive calls to synchronize_srcu() ranged from 118 to 342 microseconds,
with one outlier at 80 milliseconds. This data clearly indicates that
Tree SRCU needs to expedite the first invocation of synchronize_srcu()
during an SRCU idle period.
This commit therefor introduces a srcu_might_be_idle() function that
probabilistically checks whether or not SRCU is idle. This function is
used by synchronize_rcu() as an additional criterion in deciding whether
or not to expedite.
(Hat trick to Peter Zijlstra for his earlier suggestion that this might
in fact be a problem. Which for all I know might have motivated Mike to
look into it.)
Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Mike Galbraith <efault@gmx.de>
Commit f60d231a87 ("srcu: Crude control of expedited grace periods")
introduced a per-srcu_struct atomic counter to track outstanding
requests for grace periods. This works, but represents a memory-contention
bottleneck. This commit therefore uses the srcu_node combining tree
to remove this bottleneck.
This commit adds new ->srcu_gp_seq_needed_exp fields to the
srcu_data, srcu_node, and srcu_struct structures, which track the
farthest-in-the-future grace period that must be expedited, which in
turn requires that all nearer-term grace periods also be expedited.
Requests for expediting start with the srcu_data structure, run up
through the srcu_node tree, and end at the srcu_struct structure.
Note that it may be necessary to expedite a grace period that just
now started, and this is handled by a new srcu_funnel_exp_start()
function, which is invoked when the grace period itself is already
in its way, but when that grace period was not marked as expedited.
A new srcu_get_delay() function returns zero if there is at least one
expedited SRCU grace period in flight, or SRCU_INTERVAL otherwise.
This function is used to calculate delays: Normal grace periods
are allowed to extend in order to cover more requests with a given
grace-period computation, which decreases per-request overhead.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Mike Galbraith <efault@gmx.de>
In the past, SRCU was simple enough that there was little point in
making the rcutorture writer stall messages print the SRCU grace-period
number state. With the advent of Tree SRCU, this has changed. This
commit therefore makes Classic, Tiny, and Tree SRCU report this state
to rcutorture as needed.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Mike Galbraith <efault@gmx.de>
The current Tree SRCU implementation schedules a workqueue for every
srcu_data covered by a given leaf srcu_node structure having callbacks,
even if only one of those srcu_data structures actually contains
callbacks. This is clearly inefficient for workloads that don't feature
callbacks everywhere all the time. This commit therefore adds an array
of masks that are used by the leaf srcu_node structures to track exactly
which srcu_data structures contain callbacks.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Mike Galbraith <efault@gmx.de>
Although Tree SRCU does reduce delays when there is at least one
synchronize_srcu_expedited() invocation pending, srcu_schedule_cbs_snp()
still waits for SRCU_INTERVAL before invoking callbacks. Since
synchronize_srcu_expedited() now posts a callback and waits for
that callback to do a wakeup, this destroys the expedited nature of
synchronize_srcu_expedited(). This destruction became apparent to
Marc Zyngier in the guise of a guest-OS bootup slowdown from five
seconds to no fewer than forty seconds.
This commit therefore invokes callbacks immediately at the end of the
grace period when there is at least one synchronize_srcu_expedited()
invocation pending. This brought Marc's guest-OS bootup times back
into the realm of reason.
Reported-by: Marc Zyngier <marc.zyngier@arm.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Tested-by: Marc Zyngier <marc.zyngier@arm.com>
Peter Zijlstra proposed using SRCU to reduce mmap_sem contention [1,2],
however, there are workloads that could result in a high volume of
concurrent invocations of call_srcu(), which with current SRCU would
result in excessive lock contention on the srcu_struct structure's
->queue_lock, which protects SRCU's callback lists. This commit therefore
moves SRCU to per-CPU callback lists, thus greatly reducing contention.
Because a given SRCU instance no longer has a single centralized callback
list, starting grace periods and invoking callbacks are both more complex
than in the single-list Classic SRCU implementation. Starting grace
periods and handling callbacks are now handled using an srcu_node tree
that is in some ways similar to the rcu_node trees used by RCU-bh,
RCU-preempt, and RCU-sched (for example, the srcu_node tree shape is
controlled by exactly the same Kconfig options and boot parameters that
control the shape of the rcu_node tree).
In addition, the old per-CPU srcu_array structure is now named srcu_data
and contains an rcu_segcblist structure named ->srcu_cblist for its
callbacks (and a spinlock to protect this). The srcu_struct gets
an srcu_gp_seq that is used to associate callback segments with the
corresponding completion-time grace-period number. These completion-time
grace-period numbers are propagated up the srcu_node tree so that the
grace-period workqueue handler can determine whether additional grace
periods are needed on the one hand and where to look for callbacks that
are ready to be invoked.
The srcu_barrier() function must now wait on all instances of the per-CPU
->srcu_cblist. Because each ->srcu_cblist is protected by ->lock,
srcu_barrier() can remotely add the needed callbacks. In theory,
it could also remotely start grace periods, but in practice doing so
is complex and racy. And interestingly enough, it is never necessary
for srcu_barrier() to start a grace period because srcu_barrier() only
enqueues a callback when a callback is already present--and it turns out
that a grace period has to have already been started for this pre-existing
callback. Furthermore, it is only the callback that srcu_barrier()
needs to wait on, not any particular grace period. Therefore, a new
rcu_segcblist_entrain() function enqueues the srcu_barrier() function's
callback into the same segment occupied by the last pre-existing callback
in the list. The special case where all the pre-existing callbacks are
on a different list (because they are in the process of being invoked)
is handled by enqueuing srcu_barrier()'s callback into the RCU_DONE_TAIL
segment, relying on the done-callbacks check that takes place after all
callbacks are inovked.
Note that the readers use the same algorithm as before. Note that there
is a separate srcu_idx that tells the readers what counter to increment.
This unfortunately cannot be combined with srcu_gp_seq because they
need to be incremented at different times.
This commit introduces some ugly #ifdefs in rcutorture. These will go
away when I feel good enough about Tree SRCU to ditch Classic SRCU.
Some crude performance comparisons, courtesy of a quickly hacked rcuperf
asynchronous-grace-period capability:
Callback Queuing Overhead
-------------------------
# CPUS Classic SRCU Tree SRCU
------ ------------ ---------
2 0.349 us 0.342 us
16 31.66 us 0.4 us
41 --------- 0.417 us
The times are the 90th percentiles, a statistic that was chosen to reject
the overheads of the occasional srcu_barrier() call needed to avoid OOMing
the test machine. The rcuperf test hangs when running Classic SRCU at 41
CPUs, hence the line of dashes. Despite the hacks to both the rcuperf code
and that statistics, this is a convincing demonstration of Tree SRCU's
performance and scalability advantages.
[1] https://lwn.net/Articles/309030/
[2] https://patchwork.kernel.org/patch/5108281/
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
[ paulmck: Fix initialization if synchronize_srcu_expedited() called first. ]
The TREE_SRCU rewrite is large and a bit on the non-simple side, so
this commit helps reduce risk by allowing the old v4.11 SRCU algorithm
to be selected using a new CLASSIC_SRCU Kconfig option that depends
on RCU_EXPERT. The default is to use the new TREE_SRCU and TINY_SRCU
algorithms, in order to help get these the testing that they need.
However, if your users do not require the update-side scalability that
is to be provided by TREE_SRCU, select RCU_EXPERT and then CLASSIC_SRCU
to revert back to the old classic SRCU algorithm.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
