The vanilla has_capability() function has been unused since 2018's
commit dcb569cf6a ("Smack: ptrace capability use fixes")
Remove it.
Fixup a comment in security/commoncap.c that referenced it.
Signed-off-by: Dr. David Alan Gilbert <linux@treblig.org>
Reviewed-by: Paul Moore <paul@paul-moore.com>
Signed-off-by: Serge Hallyn <sergeh@kernel.org>
Since we're going to approach integer overflow mitigation a type at a
time, we need to enable all of the associated sanitizers, and then opt
into types one at a time.
Rename the existing "signed wrap" sanitizer to just the entire topic area:
"integer wrap". Enable the implicit integer truncation sanitizers, with
required callbacks and tests.
Notably, this requires features (currently) only available in Clang,
so we can depend on the cc-option tests to determine availability
instead of doing version tests.
Link: https://lore.kernel.org/r/20250307041914.937329-1-kees@kernel.org
Signed-off-by: Kees Cook <kees@kernel.org>
Both KASAN and LOCKDEP are commonly enabled in building a debug kernel.
Each of them can significantly slow down the speed of a debug kernel.
Enabling KASAN instrumentation of the LOCKDEP code will further slow
things down.
Since LOCKDEP is a high overhead debugging tool, it will never get
enabled in a production kernel. The LOCKDEP code is also pretty mature
and is unlikely to get major changes. There is also a possibility of
recursion similar to KCSAN.
To evaluate the performance impact of disabling KASAN instrumentation
of lockdep.c, the time to do a parallel build of the Linux defconfig
kernel was used as the benchmark. Two x86-64 systems (Skylake & Zen 2)
and an arm64 system were used as test beds. Two sets of non-RT and RT
kernels with similar configurations except mainly CONFIG_PREEMPT_RT
were used for evaluation.
For the Skylake system:
Kernel Run time Sys time
------ -------- --------
Non-debug kernel (baseline) 0m47.642s 4m19.811s
[CONFIG_KASAN_INLINE=y]
Debug kernel 2m11.108s (x2.8) 38m20.467s (x8.9)
Debug kernel (patched) 1m49.602s (x2.3) 31m28.501s (x7.3)
Debug kernel
(patched + mitigations=off) 1m30.988s (x1.9) 26m41.993s (x6.2)
RT kernel (baseline) 0m54.871s 7m15.340s
[CONFIG_KASAN_INLINE=n]
RT debug kernel 6m07.151s (x6.7) 135m47.428s (x18.7)
RT debug kernel (patched) 3m42.434s (x4.1) 74m51.636s (x10.3)
RT debug kernel
(patched + mitigations=off) 2m40.383s (x2.9) 57m54.369s (x8.0)
[CONFIG_KASAN_INLINE=y]
RT debug kernel 3m22.155s (x3.7) 77m53.018s (x10.7)
RT debug kernel (patched) 2m36.700s (x2.9) 54m31.195s (x7.5)
RT debug kernel
(patched + mitigations=off) 2m06.110s (x2.3) 45m49.493s (x6.3)
For the Zen 2 system:
Kernel Run time Sys time
------ -------- --------
Non-debug kernel (baseline) 1m42.806s 39m48.714s
[CONFIG_KASAN_INLINE=y]
Debug kernel 4m04.524s (x2.4) 125m35.904s (x3.2)
Debug kernel (patched) 3m56.241s (x2.3) 127m22.378s (x3.2)
Debug kernel
(patched + mitigations=off) 2m38.157s (x1.5) 92m35.680s (x2.3)
RT kernel (baseline) 1m51.500s 14m56.322s
[CONFIG_KASAN_INLINE=n]
RT debug kernel 16m04.962s (x8.7) 244m36.463s (x16.4)
RT debug kernel (patched) 9m09.073s (x4.9) 129m28.439s (x8.7)
RT debug kernel
(patched + mitigations=off) 3m31.662s (x1.9) 51m01.391s (x3.4)
For the arm64 system:
Kernel Run time Sys time
------ -------- --------
Non-debug kernel (baseline) 1m56.844s 8m47.150s
Debug kernel 3m54.774s (x2.0) 92m30.098s (x10.5)
Debug kernel (patched) 3m32.429s (x1.8) 77m40.779s (x8.8)
RT kernel (baseline) 4m01.641s 18m16.777s
[CONFIG_KASAN_INLINE=n]
RT debug kernel 19m32.977s (x4.9) 304m23.965s (x16.7)
RT debug kernel (patched) 16m28.354s (x4.1) 234m18.149s (x12.8)
Turning the mitigations off doesn't seems to have any noticeable impact
on the performance of the arm64 system. So the mitigation=off entries
aren't included.
For the x86 CPUs, CPU mitigations has a much bigger
impact on performance, especially the RT debug kernel with
CONFIG_KASAN_INLINE=n. The SRSO mitigation in Zen 2 has an especially
big impact on the debug kernel. It is also the majority of the slowdown
with mitigations on. It is because the patched RET instruction slows
down function returns. A lot of helper functions that are normally
compiled out or inlined may become real function calls in the debug
kernel.
With !CONFIG_KASAN_INLINE, the KASAN instrumentation inserts a
lot of __asan_loadX*() and __kasan_check_read() function calls to memory
access portion of the code. The lockdep's __lock_acquire() function,
for instance, has 66 __asan_loadX*() and 6 __kasan_check_read() calls
added with KASAN instrumentation. Of course, the actual numbers may vary
depending on the compiler used and the exact version of the lockdep code.
With the Skylake test system, the parallel kernel build times reduction
of the RT debug kernel with this patch are:
CONFIG_KASAN_INLINE=n: -37%
CONFIG_KASAN_INLINE=y: -22%
The time reduction is less with CONFIG_KASAN_INLINE=y, but it is still
significant.
Setting CONFIG_KASAN_INLINE=y can result in a significant performance
improvement. The major drawback is a significant increase in the size
of kernel text. In the case of vmlinux, its text size increases from
45997948 to 67606807. That is a 47% size increase (about 21 Mbytes). The
size increase of other kernel modules should be similar.
With the newly added rtmutex and lockdep lock events, the relevant
event counts for the test runs with the Skylake system were:
Event type Debug kernel RT debug kernel
---------- ------------ ---------------
lockdep_acquire 1,968,663,277 5,425,313,953
rtlock_slowlock - 401,701,156
rtmutex_slowlock - 139,672
The __lock_acquire() calls in the RT debug kernel are x2.8 times of the
non-RT debug kernel with the same workload. Since the __lock_acquire()
function is a big hitter in term of performance slowdown, this makes
the RT debug kernel much slower than the non-RT one. The average lock
nesting depth is likely to be higher in the RT debug kernel too leading
to longer execution time in the __lock_acquire() function.
As the small advantage of enabling KASAN instrumentation to catch
potential memory access error in the lockdep debugging tool is probably
not worth the drawback of further slowing down a debug kernel, disable
KASAN instrumentation in the lockdep code to allow the debug kernels
to regain some performance back, especially for the RT debug kernels.
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Marco Elver <elver@google.com>
Reviewed-by: Andrey Konovalov <andreyknvl@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: https://lore.kernel.org/r/20250307232717.1759087-6-boqun.feng@gmail.com
A circular lock dependency splat has been seen involving down_trylock():
======================================================
WARNING: possible circular locking dependency detected
6.12.0-41.el10.s390x+debug
------------------------------------------------------
dd/32479 is trying to acquire lock:
0015a20accd0d4f8 ((console_sem).lock){-.-.}-{2:2}, at: down_trylock+0x26/0x90
but task is already holding lock:
000000017e461698 (&zone->lock){-.-.}-{2:2}, at: rmqueue_bulk+0xac/0x8f0
the existing dependency chain (in reverse order) is:
-> #4 (&zone->lock){-.-.}-{2:2}:
-> #3 (hrtimer_bases.lock){-.-.}-{2:2}:
-> #2 (&rq->__lock){-.-.}-{2:2}:
-> #1 (&p->pi_lock){-.-.}-{2:2}:
-> #0 ((console_sem).lock){-.-.}-{2:2}:
The console_sem -> pi_lock dependency is due to calling try_to_wake_up()
while holding the console_sem raw_spinlock. This dependency can be broken
by using wake_q to do the wakeup instead of calling try_to_wake_up()
under the console_sem lock. This will also make the semaphore's
raw_spinlock become a terminal lock without taking any further locks
underneath it.
The hrtimer_bases.lock is a raw_spinlock while zone->lock is a
spinlock. The hrtimer_bases.lock -> zone->lock dependency happens via
the debug_objects_fill_pool() helper function in the debugobjects code.
-> #4 (&zone->lock){-.-.}-{2:2}:
__lock_acquire+0xe86/0x1cc0
lock_acquire.part.0+0x258/0x630
lock_acquire+0xb8/0xe0
_raw_spin_lock_irqsave+0xb4/0x120
rmqueue_bulk+0xac/0x8f0
__rmqueue_pcplist+0x580/0x830
rmqueue_pcplist+0xfc/0x470
rmqueue.isra.0+0xdec/0x11b0
get_page_from_freelist+0x2ee/0xeb0
__alloc_pages_noprof+0x2c2/0x520
alloc_pages_mpol_noprof+0x1fc/0x4d0
alloc_pages_noprof+0x8c/0xe0
allocate_slab+0x320/0x460
___slab_alloc+0xa58/0x12b0
__slab_alloc.isra.0+0x42/0x60
kmem_cache_alloc_noprof+0x304/0x350
fill_pool+0xf6/0x450
debug_object_activate+0xfe/0x360
enqueue_hrtimer+0x34/0x190
__run_hrtimer+0x3c8/0x4c0
__hrtimer_run_queues+0x1b2/0x260
hrtimer_interrupt+0x316/0x760
do_IRQ+0x9a/0xe0
do_irq_async+0xf6/0x160
Normally a raw_spinlock to spinlock dependency is not legitimate
and will be warned if CONFIG_PROVE_RAW_LOCK_NESTING is enabled,
but debug_objects_fill_pool() is an exception as it explicitly
allows this dependency for non-PREEMPT_RT kernel without causing
PROVE_RAW_LOCK_NESTING lockdep splat. As a result, this dependency is
legitimate and not a bug.
Anyway, semaphore is the only locking primitive left that is still
using try_to_wake_up() to do wakeup inside critical section, all the
other locking primitives had been migrated to use wake_q to do wakeup
outside of the critical section. It is also possible that there are
other circular locking dependencies involving printk/console_sem or
other existing/new semaphores lurking somewhere which may show up in
the future. Let just do the migration now to wake_q to avoid headache
like this.
Reported-by: yzbot+ed801a886dfdbfe7136d@syzkaller.appspotmail.com
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: https://lore.kernel.org/r/20250307232717.1759087-3-boqun.feng@gmail.com
Pull perf event fixes from Ingo Molnar:
"Fix a race between PMU registration and event creation, and fix
pmus_lock vs. pmus_srcu lock ordering"
* tag 'perf-urgent-2025-03-07' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
perf/core: Fix perf_pmu_register() vs. perf_init_event()
perf/core: Fix pmus_lock vs. pmus_srcu ordering
Now not only CPUs can use energy efficiency models, but GPUs
can also use. On the other hand, even with only one CPU, we can also
use energy_model to align control in thermal.
So remove the dependence of SMP, and add the DEVFREQ.
Signed-off-by: Jeson Gao <jeson.gao@unisoc.com>
[Added missing SMP config option in DTPM_CPU dependency]
Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
Link: https://patch.msgid.link/20250307132649.4056210-1-lukasz.luba@arm.com
[ rjw: Subject edits ]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
The rseq_cs field is documented as being set to 0 by user-space prior to
registration, however this is not currently enforced by the kernel. This
can result in a segfault on return to user-space if the value stored in
the rseq_cs field doesn't point to a valid struct rseq_cs.
The correct solution to this would be to fail the rseq registration when
the rseq_cs field is non-zero. However, some older versions of glibc
will reuse the rseq area of previous threads without clearing the
rseq_cs field and will also terminate the process if the rseq
registration fails in a secondary thread. This wasn't caught in testing
because in this case the leftover rseq_cs does point to a valid struct
rseq_cs.
What we can do is clear the rseq_cs field on registration when it's
non-zero which will prevent segfaults on registration and won't break
the glibc versions that reuse rseq areas on thread creation.
Signed-off-by: Michael Jeanson <mjeanson@efficios.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Link: https://lore.kernel.org/r/20250306211223.109455-1-mjeanson@efficios.com
Cross-merge networking fixes after downstream PR (net-6.14-rc6).
Conflicts:
net/ethtool/cabletest.c
2bcf4772e4 ("net: ethtool: try to protect all callback with netdev instance lock")
637399bf7e ("net: ethtool: netlink: Allow NULL nlattrs when getting a phy_device")
No Adjacent changes.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Notice that em_dev_register_perf_domain() and the functions called by it
do not update objects pointed to by its cb and cpus parameters, so the
const modifier can be added to them.
This allows the return value of cpumask_of() or a pointer to a
struct em_data_callback declared as const to be passed to
em_dev_register_perf_domain() directly without explicit type
casting which is rather handy.
No intentional functional impact.
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Lukasz Luba <lukasz.luba@arm.com>
Link: https://patch.msgid.link/4648962.LvFx2qVVIh@rjwysocki.net
Pull vfs fixes from Christian Brauner:
- Fix spelling mistakes in idmappings.rst
- Fix RCU warnings in override_creds()/revert_creds()
- Create new pid namespaces with default limit now that pid_max is
namespaced
* tag 'vfs-6.14-rc6.fixes' of gitolite.kernel.org:pub/scm/linux/kernel/git/vfs/vfs:
pid: Do not set pid_max in new pid namespaces
doc: correcting two prefix errors in idmappings.rst
cred: Fix RCU warnings in override/revert_creds
During stress-testing, we found a kmemleak report for perf_event:
unreferenced object 0xff110001410a33e0 (size 1328):
comm "kworker/4:11", pid 288, jiffies 4294916004
hex dump (first 32 bytes):
b8 be c2 3b 02 00 11 ff 22 01 00 00 00 00 ad de ...;....".......
f0 33 0a 41 01 00 11 ff f0 33 0a 41 01 00 11 ff .3.A.....3.A....
backtrace (crc 24eb7b3a):
[<00000000e211b653>] kmem_cache_alloc_node_noprof+0x269/0x2e0
[<000000009d0985fa>] perf_event_alloc+0x5f/0xcf0
[<00000000084ad4a2>] perf_event_create_kernel_counter+0x38/0x1b0
[<00000000fde96401>] hardlockup_detector_event_create+0x50/0xe0
[<0000000051183158>] watchdog_hardlockup_enable+0x17/0x70
[<00000000ac89727f>] softlockup_start_fn+0x15/0x40
...
Our stress test includes CPU online and offline cycles, and updating the
watchdog configuration.
After reading the code, I found that there may be a race between cleaning up
perf_event after updating watchdog and disabling event when the CPU goes offline:
CPU0 CPU1 CPU2
(update watchdog) (hotplug offline CPU1)
... _cpu_down(CPU1)
cpus_read_lock() // waiting for cpu lock
softlockup_start_all
smp_call_on_cpu(CPU1)
softlockup_start_fn
...
watchdog_hardlockup_enable(CPU1)
perf create E1
watchdog_ev[CPU1] = E1
cpus_read_unlock()
cpus_write_lock()
cpuhp_kick_ap_work(CPU1)
cpuhp_thread_fun
...
watchdog_hardlockup_disable(CPU1)
watchdog_ev[CPU1] = NULL
dead_event[CPU1] = E1
__lockup_detector_cleanup
for each dead_events_mask
release each dead_event
/*
* CPU1 has not been added to
* dead_events_mask, then E1
* will not be released
*/
CPU1 -> dead_events_mask
cpumask_clear(&dead_events_mask)
// dead_events_mask is cleared, E1 is leaked
In this case, the leaked perf_event E1 matches the perf_event leak
reported by kmemleak. Due to the low probability of problem recurrence
(only reported once), I added some hack delays in the code:
static void __lockup_detector_reconfigure(void)
{
...
watchdog_hardlockup_start();
cpus_read_unlock();
+ mdelay(100);
/*
* Must be called outside the cpus locked section to prevent
* recursive locking in the perf code.
...
}
void watchdog_hardlockup_disable(unsigned int cpu)
{
...
perf_event_disable(event);
this_cpu_write(watchdog_ev, NULL);
this_cpu_write(dead_event, event);
+ mdelay(100);
cpumask_set_cpu(smp_processor_id(), &dead_events_mask);
atomic_dec(&watchdog_cpus);
...
}
void hardlockup_detector_perf_cleanup(void)
{
...
perf_event_release_kernel(event);
per_cpu(dead_event, cpu) = NULL;
}
+ mdelay(100);
cpumask_clear(&dead_events_mask);
}
Then, simultaneously performing CPU on/off and switching watchdog, it is
almost certain to reproduce this leak.
The problem here is that releasing perf_event is not within the CPU
hotplug read-write lock. Commit:
941154bd69 ("watchdog/hardlockup/perf: Prevent CPU hotplug deadlock")
introduced deferred release to solve the deadlock caused by calling
get_online_cpus() when releasing perf_event. Later, commit:
efe951d3de ("perf/x86: Fix perf,x86,cpuhp deadlock")
removed the get_online_cpus() call on the perf_event release path to solve
another deadlock problem.
Therefore, it is now possible to move the release of perf_event back
into the CPU hotplug read-write lock, and release the event immediately
after disabling it.
Fixes: 941154bd69 ("watchdog/hardlockup/perf: Prevent CPU hotplug deadlock")
Signed-off-by: Li Huafei <lihuafei1@huawei.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20241021193004.308303-1-lihuafei1@huawei.com
The ftrace selftest reported a failure because writing -1 to
sched_rt_runtime_us returns -EBUSY. This happens when the possible
CPUs are different from active CPUs.
Active CPUs are part of one root domain, while remaining CPUs are part
of def_root_domain. Since active cpumask is being used, this results in
cpus=0 when a non active CPUs is used in the loop.
Fix it by looping over the online CPUs instead for validating the
bandwidth calculations.
Signed-off-by: Shrikanth Hegde <sshegde@linux.ibm.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Juri Lelli <juri.lelli@redhat.com>
Link: https://lore.kernel.org/r/20250306052954.452005-2-sshegde@linux.ibm.com
It is already difficult for users to troubleshoot which of multiple pid
limits restricts their workload. The per-(hierarchical-)NS pid_max would
contribute to the confusion.
Also, the implementation copies the limit upon creation from
parent, this pattern showed cumbersome with some attributes in legacy
cgroup controllers -- it's subject to race condition between parent's
limit modification and children creation and once copied it must be
changed in the descendant.
Let's do what other places do (ucounts or cgroup limits) -- create new
pid namespaces without any limit at all. The global limit (actually any
ancestor's limit) is still effectively in place, we avoid the
set/unshare race and bumps of global (ancestral) limit have the desired
effect on pid namespace that do not care.
Link: https://lore.kernel.org/r/20240408145819.8787-1-mkoutny@suse.com/
Link: https://lore.kernel.org/r/20250221170249.890014-1-mkoutny@suse.com/
Fixes: 7863dcc72d ("pid: allow pid_max to be set per pid namespace")
Signed-off-by: Michal Koutný <mkoutny@suse.com>
Link: https://lore.kernel.org/r/20250305145849.55491-1-mkoutny@suse.com
Signed-off-by: Christian Brauner <brauner@kernel.org>
child_cfs_rq_on_list attempts to convert a 'prev' pointer to a cfs_rq.
This 'prev' pointer can originate from struct rq's leaf_cfs_rq_list,
making the conversion invalid and potentially leading to memory
corruption. Depending on the relative positions of leaf_cfs_rq_list and
the task group (tg) pointer within the struct, this can cause a memory
fault or access garbage data.
The issue arises in list_add_leaf_cfs_rq, where both
cfs_rq->leaf_cfs_rq_list and rq->leaf_cfs_rq_list are added to the same
leaf list. Also, rq->tmp_alone_branch can be set to rq->leaf_cfs_rq_list.
This adds a check `if (prev == &rq->leaf_cfs_rq_list)` after the main
conditional in child_cfs_rq_on_list. This ensures that the container_of
operation will convert a correct cfs_rq struct.
This check is sufficient because only cfs_rqs on the same CPU are added
to the list, so verifying the 'prev' pointer against the current rq's list
head is enough.
Fixes a potential memory corruption issue that due to current struct
layout might not be manifesting as a crash but could lead to unpredictable
behavior when the layout changes.
Fixes: fdaba61ef8 ("sched/fair: Ensure that the CFS parent is added after unthrottling")
Signed-off-by: Zecheng Li <zecheng@google.com>
Reviewed-and-tested-by: K Prateek Nayak <kprateek.nayak@amd.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Vincent Guittot <vincent.guittot@linaro.org>
Link: https://lore.kernel.org/r/20250304214031.2882646-1-zecheng@google.com
Many devices implement highly accurate clocks, which the kernel manages
as PTP Hardware Clocks (PHCs). Userspace applications rely on these
clocks to timestamp events, trace workload execution, correlate
timescales across devices, and keep various clocks in sync.
The kernel’s current implementation of PTP clocks does not enforce file
permissions checks for most device operations except for POSIX clock
operations, where file mode is verified in the POSIX layer before
forwarding the call to the PTP subsystem. Consequently, it is common
practice to not give unprivileged userspace applications any access to
PTP clocks whatsoever by giving the PTP chardevs 600 permissions. An
example of users running into this limitation is documented in [1].
Additionally, POSIX layer requires WRITE permission even for readonly
adjtime() calls which are used in PTP layer to return current frequency
offset applied to the PHC.
Add permission checks for functions that modify the state of a PTP
device. Continue enforcing permission checks for POSIX clock operations
(settime, adjtime) in the POSIX layer. Only require WRITE access for
dynamic clocks adjtime() if any flags are set in the modes field.
[1] https://lists.nwtime.org/sympa/arc/linuxptp-users/2024-01/msg00036.html
Changes in v4:
- Require FMODE_WRITE in ajtime() only for calls modifying the clock in
any way.
Acked-by: Richard Cochran <richardcochran@gmail.com>
Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Signed-off-by: Wojtek Wasko <wwasko@nvidia.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
File descriptor based pc_clock_*() operations of dynamic posix clocks
have access to the file pointer and implement permission checks in the
generic code before invoking the relevant dynamic clock callback.
Character device operations (open, read, poll, ioctl) do not implement a
generic permission control and the dynamic clock callbacks have no
access to the file pointer to implement them.
Extend struct posix_clock_context with a struct file pointer and
initialize it in posix_clock_open(), so that all dynamic clock callbacks
can access it.
Acked-by: Richard Cochran <richardcochran@gmail.com>
Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Wojtek Wasko <wwasko@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
PREEMPT_LAZY can be enabled stand-alone or alongside PREEMPT_DYNAMIC
which allows for dynamic switching of preemption models.
The choice of PREEMPT_RCU or not, however, is fixed at compile time.
Given that PREEMPT_RCU makes some trade-offs to optimize for latency
as opposed to throughput, configurations with limited preemption
might prefer the stronger forward-progress guarantees of PREEMPT_RCU=n.
Accordingly, explicitly limit PREEMPT_RCU=y to the latency oriented
preemption models: PREEMPT, PREEMPT_RT, and the runtime configurable
model PREEMPT_DYNAMIC.
This means the throughput oriented models, PREEMPT_NONE,
PREEMPT_VOLUNTARY, and PREEMPT_LAZY will run with PREEMPT_RCU=n.
Cc: Paul E. McKenney <paulmck@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Ankur Arora <ankur.a.arora@oracle.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
The rcutorture_one_extend_check() function's second last check assumes
that "preempt_count() & PREEMPT_MASK" is non-zero only if
RCUTORTURE_RDR_PREEMPT or RCUTORTURE_RDR_SCHED bit is set.
This works for preemptible RCU and for non-preemptible RCU running in
a non-preemptible kernel. But it fails for non-preemptible RCU running
in a preemptible kernel because then rcu_read_lock() is just
preempt_disable(), which increases preempt count.
This commit therefore adjusts this check to take into account the case
fo non-preemptible RCU running in a preemptible kernel.
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
The rcutorture_one_extend_check() function's last check assumes that
if cur_ops->readlock_nesting() returns greater than zero, either the
RCUTORTURE_RDR_RCU_1 or the RCUTORTURE_RDR_RCU_2 bit must be set, that
is, there must be at least one rcu_read_lock() in effect.
This works for preemptible RCU and for non-preemptible RCU running in
a non-preemptible kernel. But it fails for non-preemptible RCU running
in a preemptible kernel because then RCU's cur_ops->readlock_nesting()
function, which is rcu_torture_readlock_nesting(), will return
the PREEMPT_MASK mask bits from preempt_count(). The result will
be greater than zero if preemption is disabled, including by the
RCUTORTURE_RDR_PREEMPT and RCUTORTURE_RDR_SCHED bits.
This commit therefore adjusts this check to take into account the case
fo non-preemptible RCU running in a preemptible kernel.
[boqun: Fix the if condition and add comment]
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202502171415.8ec87c87-lkp@intel.com
Co-developed-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Co-developed-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Tested-by: kernel test robot <oliver.sang@intel.com>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
To reduce RCU noise for nohz_full configurations, osnoise depends
on cond_resched() providing quiescent states for PREEMPT_RCU=n
configurations. For PREEMPT_RCU=y configurations -- where
cond_resched() is a stub -- we do this by directly calling
rcu_momentary_eqs().
With (PREEMPT_LAZY=y, PREEMPT_DYNAMIC=n), however, we have a
configuration with (PREEMPTION=y, PREEMPT_RCU=n) where neither
of the above can help.
Handle that by providing an explicit quiescent state here for all
configurations.
As mentioned above this is not needed for non-stubbed cond_resched(),
but, providing a quiescent state here just pulls in one that a future
cond_resched() would provide, so doesn't cause any extra work for
this configuration.
Cc: Paul E. McKenney <paulmck@kernel.org>
Cc: Daniel Bristot de Oliveira <bristot@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Suggested-by: Paul E. McKenney <paulmck@kernel.org>
Acked-by: Daniel Bristot de Oliveira <bristot@kernel.org>
Signed-off-by: Ankur Arora <ankur.a.arora@oracle.com>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Acked-by: Steven Rostedt (Google) <rostedt@goodmis.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Switch for using of get_state_synchronize_rcu_full() and
poll_state_synchronize_rcu_full() pair to debug a normal
synchronize_rcu() call.
Just using "not" full APIs to identify if a grace period is
passed or not might lead to a false-positive kernel splat.
It can happen, because get_state_synchronize_rcu() compresses
both normal and expedited states into one single unsigned long
value, so a poll_state_synchronize_rcu() can miss GP-completion
when synchronize_rcu()/synchronize_rcu_expedited() concurrently
run.
To address this, switch to poll_state_synchronize_rcu_full() and
get_state_synchronize_rcu_full() APIs, which use separate variables
for expedited and normal states.
Reported-by: cheung wall <zzqq0103.hey@gmail.com>
Closes: https://lore.kernel.org/lkml/Z5ikQeVmVdsWQrdD@pc636/T/
Fixes: 988f569ae0 ("rcu: Reduce synchronize_rcu() latency")
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Link: https://lore.kernel.org/r/20250227131613.52683-3-urezki@gmail.com
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Currently "nfakewriters" parameter can be set to any value but
there is no possibility to adjust it automatically based on how
many CPUs a system has where a test is run on.
To address this, if the "nfakewriters" is set to negative it will
be adjusted to num_online_cpus() during torture initialization.
Reviewed-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Link: https://lore.kernel.org/r/20250227131613.52683-1-urezki@gmail.com
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
The WARN_ON_ONCE() in ct_kernel_exit_state() follows the call to
ct_state_inc(), which means that RCU is not watching this WARN_ON_ONCE().
This can (and does) result in extraneous lockdep warnings when this
WARN_ON_ONCE() triggers. These extraneous warnings are the opposite
of helpful.
Therefore, invert the WARN_ON_ONCE() condition and move it before the
call to ct_state_inc(). This does mean that the ct_state_inc() return
value can no longer be used in the WARN_ON_ONCE() condition, so discard
this return value and instead use a call to rcu_is_watching_curr_cpu().
This call is executed only in CONFIG_RCU_EQS_DEBUG=y kernels, so there
is no added overhead in production use.
[Boqun: Add the subsystem tag in the title]
Reported-by: Breno Leitao <leitao@debian.org>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Reviewed-by: Valentin Schneider <vschneid@redhat.com>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Link: https://lore.kernel.org/r/bd911cd9-1fe9-447c-85e0-ea811a1dc896@paulmck-laptop
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
Analysis of an rcutorture callback-based forward-progress test failure was
hampered by the lack of ->cblist segment lengths. This commit therefore
adds this information, so that what would have been ".W85620.N." (there
are some callbacks waiting for grace period sequence number 85620 and
some number more that have not yet been assigned to a grace period)
now prints as ".W2(85620).N6." (there are 2 callbacks waiting for grace
period 85620 and 6 not yet assigned to a grace period). Note that
"D" (done), "N" (next and not yet assigned to a grace period, and "B"
(bypass, also not yet assigned to a grace period) have just the number
of callbacks without the parenthesized grace-period sequence number.
In contrast, "W" (waiting for the current grace period) and "R" (ready
to wait for the next grace period to start) both have parenthesized
grace-period sequence numbers.
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
The timer and hrtimer softirq processing has moved to dedicated threads
for kernels built with CONFIG_IRQ_FORCED_THREADING=y. This results in
timers not expiring until later in early boot, which in turn causes the
RCU Tasks self-tests to hang in kernels built with CONFIG_PROVE_RCU=y,
which further causes the entire kernel to hang. One fix would be to
make timers work during this time, but there are no known users of RCU
Tasks grace periods during that time, so no justification for the added
complexity. Not yet, anyway.
This commit therefore moves the call to rcu_init_tasks_generic() from
kernel_init_freeable() to a core_initcall(). This works because the
timer and hrtimer kthreads are created at early_initcall() time.
Fixes: 49a1763950 ("softirq: Use a dedicated thread for timer wakeups on PREEMPT_RT.")
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Frederic Weisbecker <frederic@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Alexei Starovoitov <ast@kernel.org>
Cc: Andrii Nakryiko <andrii@kernel.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: <linux-trace-kernel@vger.kernel.org>
Tested-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
The get_state_synchronize_rcu_full() and poll_state_synchronize_rcu_full()
functions use the root rcu_node structure's ->gp_seq field to detect
the beginnings and ends of grace periods, respectively. This choice is
necessary for the poll_state_synchronize_rcu_full() function because
(give or take counter wrap), the following sequence is guaranteed not
to trigger:
get_state_synchronize_rcu_full(&rgos);
synchronize_rcu();
WARN_ON_ONCE(!poll_state_synchronize_rcu_full(&rgos));
The RCU callbacks that awaken synchronize_rcu() instances are
guaranteed not to be invoked before the root rcu_node structure's
->gp_seq field is updated to indicate the end of the grace period.
However, these callbacks might start being invoked immediately
thereafter, in particular, before rcu_state.gp_seq has been updated.
Therefore, poll_state_synchronize_rcu_full() must refer to the
root rcu_node structure's ->gp_seq field. Because this field is
updated under this structure's ->lock, any code following a call to
poll_state_synchronize_rcu_full() will be fully ordered after the
full grace-period computation, as is required by RCU's memory-ordering
semantics.
By symmetry, the get_state_synchronize_rcu_full() function should also
use this same root rcu_node structure's ->gp_seq field. But it turns out
that symmetry is profoundly (though extremely infrequently) destructive
in this case. To see this, consider the following sequence of events:
1. CPU 0 starts a new grace period, and updates rcu_state.gp_seq
accordingly.
2. As its first step of grace-period initialization, CPU 0 examines
the current CPU hotplug state and decides that it need not wait
for CPU 1, which is currently offline.
3. CPU 1 comes online, and updates its state. But this does not
affect the current grace period, but rather the one after that.
After all, CPU 1 was offline when the current grace period
started, so all pre-existing RCU readers on CPU 1 must have
completed or been preempted before it last went offline.
The current grace period therefore has nothing it needs to wait
for on CPU 1.
4. CPU 1 switches to an rcutorture kthread which is running
rcutorture's rcu_torture_reader() function, which starts a new
RCU reader.
5. CPU 2 is running rcutorture's rcu_torture_writer() function
and collects a new polled grace-period "cookie" using
get_state_synchronize_rcu_full(). Because the newly started
grace period has not completed initialization, the root rcu_node
structure's ->gp_seq field has not yet been updated to indicate
that this new grace period has already started.
This cookie is therefore set up for the end of the current grace
period (rather than the end of the following grace period).
6. CPU 0 finishes grace-period initialization.
7. If CPU 1’s rcutorture reader is preempted, it will be added to
the ->blkd_tasks list, but because CPU 1’s ->qsmask bit is not
set in CPU 1's leaf rcu_node structure, the ->gp_tasks pointer
will not be updated. Thus, this grace period will not wait on
it. Which is only fair, given that the CPU did not come online
until after the grace period officially started.
8. CPUs 0 and 2 then detect the new grace period and then report
a quiescent state to the RCU core.
9. Because CPU 1 was offline at the start of the current grace
period, CPUs 0 and 2 are the only CPUs that this grace period
needs to wait on. So the grace period ends and post-grace-period
cleanup starts. In particular, the root rcu_node structure's
->gp_seq field is updated to indicate that this grace period
has now ended.
10. CPU 2 continues running rcu_torture_writer() and sees that,
from the viewpoint of the root rcu_node structure consulted by
the poll_state_synchronize_rcu_full() function, the grace period
has ended. It therefore updates state accordingly.
11. CPU 1 is still running the same RCU reader, which notices this
update and thus complains about the too-short grace period.
The fix is for the get_state_synchronize_rcu_full() function to use
rcu_state.gp_seq instead of the root rcu_node structure's ->gp_seq field.
With this change in place, if step 5's cookie indicates that the grace
period has not yet started, then any prior code executed by CPU 2 must
have happened before CPU 1 came online. This will in turn prevent CPU
1's code in steps 3 and 11 from spanning CPU 2's grace-period wait,
thus preventing CPU 1 from being subjected to a too-short grace period.
This commit therefore makes this change. Note that there is no change to
the poll_state_synchronize_rcu_full() function, which as noted above,
must continue to use the root rcu_node structure's ->gp_seq field.
This is of course an asymmetry between these two functions, but is an
asymmetry that is absolutely required for correct operation. It is a
common human tendency to greatly value symmetry, and sometimes symmetry
is a wonderful thing. Other times, symmetry results in poor performance.
But in this case, symmetry is just plain wrong.
Nevertheless, the asymmetry does require an additional adjustment.
It is possible for get_state_synchronize_rcu_full() to see a given
grace period as having started, but for an immediately following
poll_state_synchronize_rcu_full() to see it as having not yet started.
Given the current rcu_seq_done_exact() implementation, this will
result in a false-positive indication that the grace period is done
from poll_state_synchronize_rcu_full(). This is dealt with by making
rcu_seq_done_exact() reach back three grace periods rather than just
two of them.
However, simply changing get_state_synchronize_rcu_full() function to
use rcu_state.gp_seq instead of the root rcu_node structure's ->gp_seq
field results in a theoretical bug in kernels booted with
rcutree.rcu_normal_wake_from_gp=1 due to the following sequence of
events:
o The rcu_gp_init() function invokes rcu_seq_start() to officially
start a new grace period.
o A new RCU reader begins, referencing X from some RCU-protected
list. The new grace period is not obligated to wait for this
reader.
o An updater removes X, then calls synchronize_rcu(), which queues
a wait element.
o The grace period ends, awakening the updater, which frees X
while the reader is still referencing it.
The reason that this is theoretical is that although the grace period
has officially started, none of the CPUs are officially aware of this,
and thus will have to assume that the RCU reader pre-dated the start of
the grace period. Detailed explanation can be found at [2] and [3].
Except for kernels built with CONFIG_PROVE_RCU=y, which use the polled
grace-period APIs, which can and do complain bitterly when this sequence
of events occurs. Not only that, there might be some future RCU
grace-period mechanism that pulls this sequence of events from theory
into practice. This commit therefore also pulls the call to
rcu_sr_normal_gp_init() to precede that to rcu_seq_start().
Although this fixes commit 91a967fd69 ("rcu: Add full-sized polling
for get_completed*() and poll_state*()"), it is not clear that it is
worth backporting this commit. First, it took me many weeks to convince
rcutorture to reproduce this more frequently than once per year.
Second, this cannot be reproduced at all without frequent CPU-hotplug
operations, as in waiting all of 50 milliseconds from the end of the
previous operation until starting the next one. Third, the TREE03.boot
settings cause multi-millisecond delays during RCU grace-period
initialization, which greatly increase the probability of the above
sequence of events. (Don't do this in production workloads!) Fourth,
the TREE03 rcutorture scenario was modified to use four-CPU guest OSes,
to have a single-rcu_node combining tree, no testing of RCU priority
boosting, and no random preemption, and these modifications were
necessary to reproduce this issue in a reasonable timeframe. Fifth,
extremely heavy use of get_state_synchronize_rcu_full() and/or
poll_state_synchronize_rcu_full() is required to reproduce this, and as
of v6.12, only kfree_rcu() uses it, and even then not particularly
heavily.
[boqun: Apply the fix [1], and add the comment before the moved
rcu_sr_normal_gp_init(). Additional links are added for explanation.]
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
Reviewed-by: Joel Fernandes (Google) <joel@joelfernandes.org>
Tested-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Link: https://lore.kernel.org/rcu/d90bd6d9-d15c-4b9b-8a69-95336e74e8f4@paulmck-laptop/ [1]
Link: https://lore.kernel.org/rcu/20250303001507.GA3994772@joelnvbox/ [2]
Link: https://lore.kernel.org/rcu/Z8bcUsZ9IpRi1QoP@pc636/ [3]
Reviewed-by: Joel Fernandes <joelagnelf@nvidia.com>
Signed-off-by: Boqun Feng <boqun.feng@gmail.com>
