When we looked into FIO performance with swiotlb enabled in VM, we found
swiotlb_bounce() is always called one more time than expected for each DMA
read request.
It turns out that the bounce buffer is copied to original DMA buffer twice
after the completion of a DMA request (one is done by in
dma_direct_sync_single_for_cpu(), the other by swiotlb_tbl_unmap_single()).
But the content in bounce buffer actually doesn't change between the two
rounds of copy. So, one round of copy is redundant.
Pass DMA_ATTR_SKIP_CPU_SYNC flag to swiotlb_tbl_unmap_single() to
skip the memory copy in it.
This fix increases FIO 64KB sequential read throughput in a guest with
swiotlb=force by 5.6%.
Fixes: 55897af630 ("dma-direct: merge swiotlb_dma_ops into the dma_direct code")
Reported-by: Wang Zhaoyang1 <zhaoyang1.wang@intel.com>
Reported-by: Gao Liang <liang.gao@intel.com>
Signed-off-by: Chao Gao <chao.gao@intel.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
We're moving sysctls out of kernel/sysctl.c as it is a mess. We
already moved all filesystem sysctls out. And with time the goal
is to move all sysctls out to their own subsystem/actual user.
kernel/sysctl.c has grown to an insane mess and its easy to run
into conflicts with it. The effort to move them out into various
subsystems is part of this.
Signed-off-by: Yan Zhu <zhuyan34@huawei.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Luis Chamberlain <mcgrof@kernel.org>
Link: https://lore.kernel.org/bpf/20220407070759.29506-1-zhuyan34@huawei.com
Rather than waiting until a CPU is first brought online, do the
initialisation of the cpuhp_cpu_state structure for each CPU during the
__init phase. This saves a (small) amount of non-__init memory and
avoids potential confusion about when the cpuhp_cpu_state struct is
valid.
Suggested-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Steven Price <steven.price@arm.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/r/20220411152233.474129-3-steven.price@arm.com
Currently the setting of the 'cpu' member of struct cpuhp_cpu_state in
cpuhp_create() is too late as it is used earlier in _cpu_up().
If kzalloc_node() in __smpboot_create_thread() fails then the rollback will
be done with st->cpu==0 causing CPU0 to be erroneously set to be dying,
causing the scheduler to get mightily confused and throw its toys out of
the pram.
However the cpu number is actually available directly, so simply remove
the 'cpu' member and avoid the problem in the first place.
Fixes: 2ea46c6fc9 ("cpumask/hotplug: Fix cpu_dying() state tracking")
Signed-off-by: Steven Price <steven.price@arm.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/r/20220411152233.474129-2-steven.price@arm.com
The check in flush_smp_call_function_queue() for callbacks that are sent
to offline CPUs currently checks whether the queue is empty.
However, flush_smp_call_function_queue() has just deleted all the
callbacks from the queue and moved all the entries into a local list.
This checks would only be positive if some callbacks were added in the
short time after llist_del_all() was called. This does not seem to be
the intention of this check.
Change the check to look at the local list to which the entries were
moved instead of the queue from which all the callbacks were just
removed.
Fixes: 8d056c48e4 ("CPU hotplug, smp: flush any pending IPI callbacks before CPU offline")
Signed-off-by: Nadav Amit <namit@vmware.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/r/20220319072015.1495036-1-namit@vmware.com
Change the comment to a normal (non-kernel-doc) comment to avoid
these kernel-doc warnings:
kernel/power/snapshot.c:335: warning: This comment starts with '/**', but
isn't a kernel-doc comment. Refer Documentation/doc-guide/kernel-doc.rst
* Data types related to memory bitmaps.
Signed-off-by: Haowen Bai <baihaowen@meizu.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Add parameter description in alloc_rtree_node() kernel-doc
comment and fix several inconsistent function name descriptions.
Remove some warnings found by running scripts/kernel-doc,
which is caused by using 'make W=1'.
kernel/power/snapshot.c:438: warning: Function parameter or member
'gfp_mask' not described in 'alloc_rtree_node'
kernel/power/snapshot.c:438: warning: Function parameter or member
'safe_needed' not described in 'alloc_rtree_node'
kernel/power/snapshot.c:438: warning: Function parameter or member 'ca'
not described in 'alloc_rtree_node'
kernel/power/snapshot.c:438: warning: Function parameter or member
'list' not described in 'alloc_rtree_node'
kernel/power/snapshot.c:916: warning: expecting prototype for
memory_bm_rtree_next_pfn(). Prototype was for memory_bm_next_pfn()
instead
kernel/power/snapshot.c:1947: warning: expecting prototype for
alloc_highmem_image_pages(). Prototype was for alloc_highmem_pages()
instead
kernel/power/snapshot.c:2230: warning: expecting prototype for load
header(). Prototype was for load_header() instead
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Signed-off-by: Yang Li <yang.lee@linux.alibaba.com>
[ rjw: Comment adjustments to avoid line breaks ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Currently pm_pr_dbg() is used to filter kernel pm debug messages based
on pm_debug_messages_on flag. The problem is if we enable/disable this
flag it will affect all pm_pr_dbg() calls at once, so we can't
individually control them.
This patch changes pm_pr_dbg() implementation as such:
- If pm_debug_messages_on is enabled, print the message.
- If pm_debug_messages_on is disabled and CONFIG_DYNAMIC_DEBUG is
enabled, only print the messages explicitly enabled on
/sys/kernel/debug/dynamic_debug/control.
- If pm_debug_messages_on is disabled and CONFIG_DYNAMIC_DEBUG is
disabled, don't print the message.
Signed-off-by: David Cohen <dacohen@pm.me>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The macro -DDEBUG is broadly enabled on kernel/power/ directory if
CONFIG_DYNAMIC_DEBUG is enabled. As side effect all debug messages using
pr_debug() and dev_dbg() are enabled by default on dynamic debug.
We're reworking pm_pr_dbg() to support dynamic debug, where pm_pr_dbg()
will print message if either pm_debug_messages_on flag is set or if it's
explicitly enabled on dynamic debug's control. That means if we let
-DDEBUG broadly set, the pm_debug_messages_on flag will be bypassed by
default on pm_pr_dbg() if dynamic debug is also enabled.
The files that directly use pr_debug() and dev_dbg() on kernel/power/ are:
- swap.c
- snapshot.c
- energy_model.c
And those files do not use pm_pr_dbg(). So if we limit -DDEBUG to them,
we keep the same functional behavior while allowing the pm_pr_dbg()
refactor.
Signed-off-by: David Cohen <dacohen@pm.me>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The Energy Model gets more bits used in 'flags'. Avoid adding another
debugfs file just to print what is the status of a new flag. Simply
remove old debugfs files and add one generic which prints all flags
as a hex value.
Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
Reviewed-by: Ionela Voinescu <ionela.voinescu@arm.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The .active_power() callback passes the device pointer when it's called.
Aligned with a convetion present in other subsystems and pass the 'dev'
as a first argument. It looks more cleaner.
Adjust all affected drivers which implement that API callback.
Suggested-by: Ionela Voinescu <ionela.voinescu@arm.com>
Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
Reviewed-by: Ionela Voinescu <ionela.voinescu@arm.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The Energy Model (EM) allows to provide the 'cost' values when the device
driver provides the .get_cost() optional callback. This removes
restriction which is in the EM calculation function of the 'cost'
for each performance state. Now, the driver is in charge of providing
the right values which are then used by Energy Aware Scheduler.
Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
Reviewed-by: Ionela Voinescu <ionela.voinescu@arm.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The Energy Model (EM) can be used on platforms which are missing real
power information. Those platforms would implement .get_cost() which
populates needed values for the Energy Aware Scheduler (EAS). The EAS
doesn't use 'power' fields from EM, but other frameworks might use them.
Thus, to avoid miss-usage of this specific type of EM, introduce a new
flags which can be checked by other frameworks.
Signed-off-by: Pierre Gondois <Pierre.Gondois@arm.com>
Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
Reviewed-by: Ionela Voinescu <ionela.voinescu@arm.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
A CPU will not show up in virtualized environment which includes an
Enclave. The VM splits its resources into a primary VM and a Enclave
VM. While the Enclave is active, the hypervisor will ignore all requests to
bring up a CPU and this CPU will remain in CPU_UP_PREPARE state.
The kernel will wait up to ten seconds for CPU to show up (do_boot_cpu())
and then rollback the hotplug state back to CPUHP_OFFLINE leaving the CPU
state in CPU_UP_PREPARE. The CPU state is set back to CPUHP_TEARDOWN_CPU
during the CPU_POST_DEAD stage.
After the Enclave VM terminates, the primary VM can bring up the CPU
again.
Allow to bring up the CPU if it is in the CPU_UP_PREPARE state.
[bigeasy: Rewrite commit description.]
Signed-off-by: Longpeng(Mike) <longpeng2@huawei.com>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Dongli Zhang <dongli.zhang@oracle.com>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Reviewed-by: Henry Wang <Henry.Wang@arm.com>
Link: https://lore.kernel.org/r/20220209080214.1439408-3-bigeasy@linutronix.de
A report of a 12-jiffy normal RCU CPU stall warning raises interesting
questions about the nature of time on the offending system. This commit
instruments rcu_sched_clock_irq(), which is RCU's hook into the
scheduling-clock interrupt, checking for the jiffies counter going
backwards.
Reported-by: Saravanan D <sarvanand@fb.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Suppose we have a kernel built with both CONFIG_RCU_STRICT_GRACE_PERIOD=y
and CONFIG_PREEMPT=y. Suppose further that an RCU reader from which RCU
core needs a quiescent state ends in rcu_preempt_deferred_qs_irqrestore().
This function will then invoke rcu_report_qs_rdp() in order to immediately
report that quiescent state. Unfortunately, it will not have cleared
that reader's CPU's rcu_data structure's ->cpu_no_qs.b.norm field.
As a result, rcu_report_qs_rdp() will take an early exit because it
will believe that this CPU has not yet encountered a quiescent state,
and there will be no reporting of the current quiescent state.
This commit therefore causes rcu_preempt_deferred_qs_irqrestore() to
clear the ->cpu_no_qs.b.norm field before invoking rcu_report_qs_rdp().
Kudos to Boqun Feng and Neeraj Upadhyay for helping with analysis of
this issue!
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The TASKS_RUDE_RCU does not select IRQ_WORK, which can result in build
failures for kernels that do not otherwise select IRQ_WORK. This commit
therefore causes the TASKS_RUDE_RCU Kconfig option to select IRQ_WORK.
Reported-by: Hyeonggon Yoo <42.hyeyoo@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The rcutorture module has an rcu_torture_writer task that repeatedly
performs writes, synchronizations, and deletes. There is a corner-case
check in rcu_torture_writer() wherein if nsynctypes is 0, a warning is
issued and the task waits to be stopped via a call to
torture_kthread_stopping() rather than performing any work.
There should be a return statement following this call to
torture_kthread_stopping(), as the intention with issuing the call to
torture_kthread_stopping() in the first place is to avoid the
rcu_torture_writer task from performing any work. Some of the work may even
be dangerous to perform, such as potentially causing a #DE due to
nsynctypes being used in a modulo operator when querying for sync updates
to issue.
This patch adds the missing return call. As a bonus, it also fixes a
checkpatch warning that was emitted due to the WARN_ONCE() call using the
name of the function rather than __func__.
Signed-off-by: David Vernet <void@manifault.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The rcutorture module is used to run torture tests that validate RCU.
rcutorture takes a variety of module parameters that configure the
functionality of the test. Amongst these parameters are the types of
synchronization mechanisms that the rcu_torture_writer and
rcu_torture_fakewriter tasks may use, and the torture_type of the run which
determines what read and sync operations are used by the various writer and
reader tasks that run throughout the test.
When the module is configured to only use sync types for which the
specified torture_type does not implement the necessary operations, we can
end up in a state where nsynctypes is 0. This is not an erroneous state,
but it currently crashes the kernel with a #DE due to nsynctypes being used
with a modulo operator in rcu_torture_fakewriter().
Here is an example of such a #DE:
$ insmod ./rcutorture.ko gp_cond=1 gp_cond_exp=0 gp_exp=0 gp_poll_exp=0
gp_normal=0 gp_poll=0 gp_poll_exp=0 verbose=9999 torture_type=trivial
...
[ 8536.525096] divide error: 0000 [#1] PREEMPT SMP PTI
[ 8536.525101] CPU: 30 PID: 392138 Comm: rcu_torture_fak Kdump: loaded Tainted: G S 5.17.0-rc1-00179-gc8c42c80febd #24
[ 8536.525105] Hardware name: Quanta Twin Lakes MP/Twin Lakes Passive MP, BIOS F09_3A23 12/08/2020
[ 8536.525106] RIP: 0010:rcu_torture_fakewriter+0xf1/0x2d0 [rcutorture]
[ 8536.525121] Code: 00 31 d2 8d 0c f5 00 00 00 00 48 63 c9 48 f7 f1 48 85 d2 0f 84 79 ff ff ff 48 89 e7 e8 78 78 01 00 48 63 0d 29 ca 00 00 31 d2 <48> f7 f1 8b 04 95 00 05 4e a0 83 f8 06 0f 84 ad 00 00 00 7f 1f 83
[ 8536.525124] RSP: 0018:ffffc9000777fef0 EFLAGS: 00010246
[ 8536.525127] RAX: 00000000223d006e RBX: cccccccccccccccd RCX: 0000000000000000
[ 8536.525130] RDX: 0000000000000000 RSI: ffffffff824315b9 RDI: ffffc9000777fef0
[ 8536.525132] RBP: ffffc9000487bb30 R08: 0000000000000002 R09: 000000000002a580
[ 8536.525134] R10: ffffffff82c5f920 R11: 0000000000000000 R12: ffff8881a2c35d00
[ 8536.525136] R13: ffff8881540c8d00 R14: ffffffffa04d39d0 R15: 0000000000000000
[ 8536.525137] FS: 0000000000000000(0000) GS:ffff88903ff80000(0000) knlGS:0000000000000000
[ 8536.525140] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ 8536.525142] CR2: 00007f839f022000 CR3: 0000000002c0a006 CR4: 00000000007706e0
[ 8536.525144] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[ 8536.525145] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[ 8536.525147] PKRU: 55555554
[ 8536.525148] Call Trace:
[ 8536.525150] <TASK>
[ 8536.525153] kthread+0xe8/0x110
[ 8536.525161] ? kthread_complete_and_exit+0x20/0x20
[ 8536.525167] ret_from_fork+0x22/0x30
[ 8536.525174] </TASK>
The solution is to gracefully handle the case of nsynctypes being 0 in
rcu_torture_fakewriter() by not performing any work. This is already being
done in rcu_torture_writer(), though there is a missing return on that path
which will be fixed in a subsequent patch.
Signed-off-by: David Vernet <void@manifault.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The scftorture test module's scf_handler() function is supposed to provide
three different distributions of short delays (including "no delay") and
one distribution of long delays, if specified by the scftorture.longwait
module parameter. However, the second of the two non-zero-wait short delays
is disabled due to the first such delay's "goto out" not being enclosed in
the "then" clause with the "udelay()".
This commit therefore adjusts the code to provide the intended set of
delays.
Fixes: e9d338a0b1 ("scftorture: Add smp_call_function() torture test")
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Tree RCU supports grace-period delays using the rcutree.gp_cleanup_delay,
rcutree.gp_init_delay, and rcutree.gp_preinit_delay kernel boot
parameters. These delays are strictly for debugging purposes, and have
proven quite effective at exposing bugs involving race with CPU-hotplug
operations. However, these delays can result in false positives when
used in conjunction with callback flooding, for example, those generated
by the rcutorture.fwd_progress kernel boot parameter.
This commit therefore suppresses grace-period delays while callback
flooding is in progress.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
If the cpu_possible_mask is sparse (for example, if bits are set only for
CPUs 0, 4, 8, ...), then rcu_tasks_invoke_cbs() will access per-CPU data
for a CPU not in cpu_possible_mask. It makes these accesses while doing
a workqueue-based binary search for non-empty callback lists. Although
this search must pass through CPUs not represented in cpu_possible_mask,
it has no need to check the callback list for such CPUs.
This commit therefore changes the rcu_tasks_invoke_cbs() function's
binary search so as to only check callback lists for CPUs present in
cpu_possible_mask.
Reported-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
If the rcupdate.rcu_task_enqueue_lim kernel boot parameter is set to
something greater than 1 and less than nr_cpu_ids, the code attempts to
use a subset of the CPU's RCU Tasks callback lists. This works, but only
if the cpu_possible_mask is contiguous. If there are "holes" in this
mask, the callback-enqueue code might attempt to access a non-existent
per-CPU ->rtcpu variable for a non-existent CPU. For example, if only
CPUs 0, 4, 8, 12, 16 and so on are in cpu_possible_mask, specifying
rcupdate.rcu_task_enqueue_lim=4 would cause the code to attempt to
use callback queues for non-existent CPUs 1, 2, and 3. Because such
systems have existed in the past and might still exist, the code needs
to gracefully handle this situation.
This commit therefore checks to see whether the desired CPU is present
in cpu_possible_mask, and, if not, searches for the next CPU. This means
that the systems administrator of a system with a sparse cpu_possible_mask
will need to account for this sparsity when specifying the value of
the rcupdate.rcu_task_enqueue_lim kernel boot parameter. For example,
setting this parameter to the value 4 will use only CPUs 0 and 4, which
CPU 4 getting three times the callback load of CPU 0.
This commit assumes that bit (nr_cpu_ids - 1) is always set in
cpu_possible_mask.
Link: https://lore.kernel.org/lkml/CANn89iKaNEwyNZ=L_PQnkH0LP_XjLYrr_dpyRKNNoDJaWKdrmg@mail.gmail.com/
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Currently, the show_rcu_tasks_generic_gp_kthread() function only looks
at CPU 0's callback lists. Although this is not fatal, it can confuse
debugging efforts in cases where any of the Tasks RCU flavors are in
per-CPU queueing mode. This commit therefore causes this function to
scan all CPUs' callback queues.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The use of hrtimers for RCU-tasks grace-period delays works well in
general, but can result in excessive grace-period delays for some
corner-case workloads. This commit therefore reverts to the use of
timers for non-RT kernels to mitigate those grace-period delays.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The synchronous RCU-tasks grace-period-wait primitives invoke
schedule_timeout_idle() to give readers a chance to exit their
read-side critical sections. Unfortunately, this fails during early
boot on PREEMPT_RT because PREEMPT_RT relies solely on ksoftirqd to run
timer handlers. Because ksoftirqd cannot operate until its kthreads
are spawned, there is a brief period of time following scheduler
initialization where PREEMPT_RT cannot run the timer handlers that
schedule_timeout_idle() relies on, resulting in a hang.
To avoid this boot-time hang, this commit replaces schedule_timeout_idle()
with schedule_hrtimeout(), so that the timer expires in hardirq context.
This is ensures that the timer fires even on PREEMPT_RT throughout the
irqs-enabled portions of boot as well as during runtime.
The timer is set to expire between fract and fract + HZ / 2 jiffies in
order to align with any other timers that might expire during that time,
thus reducing the number of wakeups.
Note that RCU-tasks grace periods are infrequent, so the use of hrtimer
should be fine. In contrast, in common-case code, user of hrtimer
could result in performance issues.
Cc: Martin KaFai Lau <kafai@fb.com>
Cc: Andrii Nakryiko <andrii@kernel.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The main Tasks RCU quiescent state is voluntary context switch. However,
userspace execution is also a valid quiescent state, and is a valuable one
for userspace applications that spin repeatedly executing light-weight
non-sleeping system calls. Currently, such an application can delay a
Tasks RCU grace period for many tens of seconds.
This commit therefore enlists the aid of the scheduler-clock interrupt to
provide a Tasks RCU quiescent state when it interrupted a task executing
in userspace.
[ paulmck: Apply feedback from kernel test robot. ]
Cc: Martin KaFai Lau <kafai@fb.com>
Cc: Neil Spring <ntspring@fb.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The waitqueue used by rcu_tasks_kthread() has always only one waiter.
With a guaranteed only one waiter, this can be replaced with rcuwait
which is smaller and simpler. With rcuwait based wake counterpart, the
irqwork function (call_rcu_tasks_iw_wakeup()) can be invoked hardirq
context because it is only a wake up and no sleeping locks are involved
(unlike the wait_queue_head).
As a side effect, this is also one piece of the puzzle to pass the RCU
selftest at early boot on PREEMPT_RT.
Replace wait_queue_head with rcuwait and let the irqwork run in hardirq
context on PREEMPT_RT.
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
RCU-tasks stall-warning messages are printed after the grace period is ten
minutes old. Unfortunately, most of us will have rebooted the system in
response to an apparently-hung command long before the ten minutes is up,
and will thus see what looks to be a silent hang.
This commit therefore adds pr_info() messages that are printed earlier.
These should avoid being classified as errors, but should give impatient
users a hint. These are controlled by new rcupdate.rcu_task_stall_info
and rcupdate.rcu_task_stall_info_mult kernel-boot parameters. The former
defines the initial delay in jiffies (defaulting to 10 seconds) and the
latter defines the multiplier (defaulting to 3). Thus, by default, the
first message will appear 10 seconds into the RCU-tasks grace period,
the second 40 seconds in, and the third 160 seconds in. There would be
a fourth at 640 seconds in, but the stall warning message appears 600
seconds in, and once a stall warning is printed for a given grace period,
no further informational messages are printed.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
While booting secondary CPUs, cpus_read_[lock/unlock] is not keeping
online cpumask stable. The transient online mask results in below
calltrace.
[ 0.324121] CPU1: Booted secondary processor 0x0000000001 [0x410fd083]
[ 0.346652] Detected PIPT I-cache on CPU2
[ 0.347212] CPU2: Booted secondary processor 0x0000000002 [0x410fd083]
[ 0.377255] Detected PIPT I-cache on CPU3
[ 0.377823] CPU3: Booted secondary processor 0x0000000003 [0x410fd083]
[ 0.379040] ------------[ cut here ]------------
[ 0.383662] WARNING: CPU: 0 PID: 10 at kernel/workqueue.c:3084 __flush_work+0x12c/0x138
[ 0.384850] Modules linked in:
[ 0.385403] CPU: 0 PID: 10 Comm: rcu_tasks_rude_ Not tainted 5.17.0-rc3-v8+ #13
[ 0.386473] Hardware name: Raspberry Pi 4 Model B Rev 1.4 (DT)
[ 0.387289] pstate: 20000005 (nzCv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
[ 0.388308] pc : __flush_work+0x12c/0x138
[ 0.388970] lr : __flush_work+0x80/0x138
[ 0.389620] sp : ffffffc00aaf3c60
[ 0.390139] x29: ffffffc00aaf3d20 x28: ffffffc009c16af0 x27: ffffff80f761df48
[ 0.391316] x26: 0000000000000004 x25: 0000000000000003 x24: 0000000000000100
[ 0.392493] x23: ffffffffffffffff x22: ffffffc009c16b10 x21: ffffffc009c16b28
[ 0.393668] x20: ffffffc009e53861 x19: ffffff80f77fbf40 x18: 00000000d744fcc9
[ 0.394842] x17: 000000000000000b x16: 00000000000001c2 x15: ffffffc009e57550
[ 0.396016] x14: 0000000000000000 x13: ffffffffffffffff x12: 0000000100000000
[ 0.397190] x11: 0000000000000462 x10: ffffff8040258008 x9 : 0000000100000000
[ 0.398364] x8 : 0000000000000000 x7 : ffffffc0093c8bf4 x6 : 0000000000000000
[ 0.399538] x5 : 0000000000000000 x4 : ffffffc00a976e40 x3 : ffffffc00810444c
[ 0.400711] x2 : 0000000000000004 x1 : 0000000000000000 x0 : 0000000000000000
[ 0.401886] Call trace:
[ 0.402309] __flush_work+0x12c/0x138
[ 0.402941] schedule_on_each_cpu+0x228/0x278
[ 0.403693] rcu_tasks_rude_wait_gp+0x130/0x144
[ 0.404502] rcu_tasks_kthread+0x220/0x254
[ 0.405264] kthread+0x174/0x1ac
[ 0.405837] ret_from_fork+0x10/0x20
[ 0.406456] irq event stamp: 102
[ 0.406966] hardirqs last enabled at (101): [<ffffffc0093c8468>] _raw_spin_unlock_irq+0x78/0xb4
[ 0.408304] hardirqs last disabled at (102): [<ffffffc0093b8270>] el1_dbg+0x24/0x5c
[ 0.409410] softirqs last enabled at (54): [<ffffffc0081b80c8>] local_bh_enable+0xc/0x2c
[ 0.410645] softirqs last disabled at (50): [<ffffffc0081b809c>] local_bh_disable+0xc/0x2c
[ 0.411890] ---[ end trace 0000000000000000 ]---
[ 0.413000] smp: Brought up 1 node, 4 CPUs
[ 0.413762] SMP: Total of 4 processors activated.
[ 0.414566] CPU features: detected: 32-bit EL0 Support
[ 0.415414] CPU features: detected: 32-bit EL1 Support
[ 0.416278] CPU features: detected: CRC32 instructions
[ 0.447021] Callback from call_rcu_tasks_rude() invoked.
[ 0.506693] Callback from call_rcu_tasks() invoked.
This commit therefore fixes this issue by applying a single-CPU
optimization to the RCU Tasks Rude grace-period process. The key point
here is that the purpose of this RCU flavor is to force a schedule on
each online CPU since some past event. But the rcu_tasks_rude_wait_gp()
function runs in the context of the RCU Tasks Rude's grace-period kthread,
so there must already have been a context switch on the current CPU since
the call to either synchronize_rcu_tasks_rude() or call_rcu_tasks_rude().
So if there is only a single CPU online, RCU Tasks Rude's grace-period
kthread does not need to anything at all.
It turns out that the rcu_tasks_rude_wait_gp() function's call to
schedule_on_each_cpu() causes problems during early boot. During that
time, there is only one online CPU, namely the boot CPU. Therefore,
applying this single-CPU optimization fixes early-boot instances of
this problem.
Link: https://lore.kernel.org/lkml/20220210184319.25009-1-treasure4paddy@gmail.com/T/
Suggested-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Padmanabha Srinivasaiah <treasure4paddy@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The rcu_spawn_gp_kthread() function is called as an early initcall,
which means that SMP initialization hasn't happened yet and only the
boot CPU is online. Therefore, create only the boost kthread for the
leaf node of the boot CPU.
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Cc: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Cc: Uladzislau Rezki <uladzislau.rezki@sony.com>
Cc: Joel Fernandes <joel@joelfernandes.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit instruments the acquisitions of the srcu_struct structure's
->lock, enabling the initiation of a transition from SRCU_SIZE_SMALL
to SRCU_SIZE_BIG when sufficient contention is experienced. The
instrumentation counts the number of trylock failures within the confines
of a single jiffy. If that number exceeds the value specified by the
srcutree.small_contention_lim kernel boot parameter (which defaults to
100), and if the value specified by the srcutree.convert_to_big kernel
boot parameter has the 0x10 bit set (defaults to 0), then a transition
will be automatically initiated.
By default, there will never be any transitions, so that none of the
srcu_struct structures ever gains an srcu_node array.
The useful values for srcutree.convert_to_big are:
0x00: Never convert.
0x01: Always convert at init_srcu_struct() time.
0x02: Convert when rcutorture prints its first round of statistics.
0x03: Decide conversion approach at boot given system size.
0x10: Convert if contention is encountered.
0x12: Convert if contention is encountered or when rcutorture prints
its first round of statistics, whichever comes first.
The value 0x11 acts the same as 0x01 because the conversion happens
before there is any chance of contention.
[ paulmck: Apply "static" feedback from kernel test robot. ]
Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Once there are contention-initiated size transitions, it will be
possible for rcutorture to initiate a transition at the same time
as a contention-initiated transition. This commit therefore creates
a concurrency-safe helper function named srcu_transition_to_big() to
safely initiate size transitions.
Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit adds a comment explaining why an unprotected call to
list_add() from srcu_funnel_gp_start() can be safe. TL;DR: It is only
called during very early boot when we don't have no steeking concurrency!
Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
When an srcu_struct structure is created (but not in a kernel module)
by DEFINE_SRCU() and friends, the per-CPU srcu_data structure is
statically allocated. In all other cases, that structure is obtained
from alloc_percpu(), in which case cleanup_srcu_struct() must invoke
free_percpu() on the resulting ->sda pointer in the srcu_struct pointer.
Which it does.
Except that it also invokes free_percpu() on the ->sda pointer
referencing the statically allocated per-CPU srcu_data structures.
Which free_percpu() is surprisingly OK with.
This commit nevertheless stops cleanup_srcu_struct() from freeing
statically allocated per-CPU srcu_data structures.
Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
You really shouldn't invoke srcu_torture_stats_print() after invoking
cleanup_srcu_struct(), but there is really no reason to get a
compiler-obfuscated per-CPU-variable NULL pointer dereference as the
diagnostic. This commit therefore checks for NULL ->sda and makes a
more polite console-message complaint in that case.
Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This commit adds an srcu_tree.convert_to_big kernel parameter that either
refuses to convert at all (0), converts immediately at init_srcu_struct()
time (1), or lets rcutorture convert it (2). An addition contention-based
dynamic conversion choice will be added, along with documentation.
[ paulmck: Apply callback-scanning feedback from Neeraj Upadhyay. ]
Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
For configurations where snp node tree is not initialized at
init time (added in subsequent commits), srcu_funnel_gp_start()
and srcu_funnel_exp_start() can potential traverse and observe
the snp nodes' transient (uninitialized) states. This can potentially
happen, when init_srcu_struct_nodes() initialization of sdp->mynode
races with srcu_funnel_gp_start() and srcu_funnel_exp_start()
Consider the case below where srcu_funnel_gp_start() observes
sdp->mynode to be not NULL and uses an uninitialized sdp->grpmask
P1 P2
init_srcu_struct_nodes() void srcu_funnel_gp_start(...)
{
for_each_possible_cpu(cpu) {
...
sdp->mynode = &snp_first[...];
for (snp = sdp->mynode;...) struct srcu_node *snp_leaf =
smp_load_acquire(&sdp->mynode)
... if (snp_leaf) {
for (snp = snp_leaf; ...)
...
if (snp == snp_leaf)
snp->srcu_data_have_cbs[idx] |=
sdp->grpmask;
sdp->grpmask =
1 << (cpu - sdp->mynode->grplo);
}
}
Similarly, init_srcu_struct_nodes() and srcu_funnel_exp_start() can
race, where srcu_funnel_exp_start() could observe state of snp lock
before spin_lock_init().
P1 P2
init_srcu_struct_nodes() void srcu_funnel_exp_start(...)
{
srcu_for_each_node_breadth_first(ssp, snp) { for (; ...) {
spin_lock_...(snp, )
spin_lock_init(&ACCESS_PRIVATE(snp, lock));
...
}
for_each_possible_cpu(cpu) {
...
sdp->mynode = &snp_first[...];
To avoid these issues, ensure that snp node tree initialization is
complete i.e. after SRCU_SIZE_WAIT_BARRIER srcu_size_state is reached,
before traversing the tree. Given that srcu_funnel_gp_start() and
srcu_funnel_exp_start() are called within SRCU read side critical
sections, this check is safe, in the sense that all callbacks are
enqueued on CPU0 srcu_cblist until SRCU_SIZE_WAIT_CALL is entered,
and these read side critical sections (containing srcu_funnel_gp_start()
and srcu_funnel_exp_start()) need to complete, before SRCU_SIZE_WAIT_CALL
is reached.
Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Currently, tree SRCU relies on the srcu_node structures being initialized
at the same time that the srcu_struct itself is initialized, and thus
use the initial grace-period sequence number as the initial value for
the srcu_node structure's ->srcu_have_cbs[] and ->srcu_gp_seq_needed_exp
fields. Although this has a high probability of also working when the
srcu_node array is allocated and initialized at some random later time,
it would be better to avoid leaving such things to chance.
This commit therefore initializes these fields with 0x2, which is a
recognizable invalid value. It then adds the required checks for this
invalid value in order to avoid confusion on long-running kernels
(especially those on 32-bit systems) that allocate and initialize
srcu_node arrays late in life.
Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Currently, srcu_funnel_gp_start() tests snp->srcu_have_cbs[idx] and then
separately assigns it to the snp_seq local variable. This commit does
the assignment earlier to simplify the code a bit. While in the area,
this commit also takes advantage of the 100-character line limit to put
the call to srcu_schedule_cbs_sdp() on a single line.
Co-developed-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
