Pull timer fix from Borislav Petkov:
- Prevent a tick storm when a dedicated timekeeper CPU in nohz_full
mode runs for prolonged periods with interrupts disabled and ends up
programming the next tick in the past, leading to that storm
* tag 'timers_urgent_for_v5.16_rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
timers/nohz: Last resort update jiffies on nohz_full IRQ entry
Pull scheduler fixes from Borislav Petkov:
- Properly init uclamp_flags of a runqueue, on first enqueuing
- Fix preempt= callback return values
- Correct utime/stime resource usage reporting on nohz_full to return
the proper times instead of shorter ones
* tag 'sched_urgent_for_v5.16_rc4' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/uclamp: Fix rq->uclamp_max not set on first enqueue
preempt/dynamic: Fix setup_preempt_mode() return value
sched/cputime: Fix getrusage(RUSAGE_THREAD) with nohz_full
Commit d81ae8aac8 ("sched/uclamp: Fix initialization of struct
uclamp_rq") introduced a bug where uclamp_max of the rq is not reset to
match the woken up task's uclamp_max when the rq is idle.
The code was relying on rq->uclamp_max initialized to zero, so on first
enqueue
static inline void uclamp_rq_inc_id(struct rq *rq, struct task_struct *p,
enum uclamp_id clamp_id)
{
...
if (uc_se->value > READ_ONCE(uc_rq->value))
WRITE_ONCE(uc_rq->value, uc_se->value);
}
was actually resetting it. But since commit d81ae8aac8 changed the
default to 1024, this no longer works. And since rq->uclamp_flags is
also initialized to 0, neither above code path nor uclamp_idle_reset()
update the rq->uclamp_max on first wake up from idle.
This is only visible from first wake up(s) until the first dequeue to
idle after enabling the static key. And it only matters if the
uclamp_max of this task is < 1024 since only then its uclamp_max will be
effectively ignored.
Fix it by properly initializing rq->uclamp_flags = UCLAMP_FLAG_IDLE to
ensure uclamp_idle_reset() is called which then will update the rq
uclamp_max value as expected.
Fixes: d81ae8aac8 ("sched/uclamp: Fix initialization of struct uclamp_rq")
Signed-off-by: Qais Yousef <qais.yousef@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <Valentin.Schneider@arm.com>
Tested-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Link: https://lkml.kernel.org/r/20211202112033.1705279-1-qais.yousef@arm.com
getrusage(RUSAGE_THREAD) with nohz_full may return shorter utime/stime
than the actual time.
task_cputime_adjusted() snapshots utime and stime and then adjust their
sum to match the scheduler maintained cputime.sum_exec_runtime.
Unfortunately in nohz_full, sum_exec_runtime is only updated once per
second in the worst case, causing a discrepancy against utime and stime
that can be updated anytime by the reader using vtime.
To fix this situation, perform an update of cputime.sum_exec_runtime
when the cputime snapshot reports the task as actually running while
the tick is disabled. The related overhead is then contained within the
relevant situations.
Reported-by: Hasegawa Hitomi <hasegawa-hitomi@fujitsu.com>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Hasegawa Hitomi <hasegawa-hitomi@fujitsu.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Masayoshi Mizuma <m.mizuma@jp.fujitsu.com>
Acked-by: Phil Auld <pauld@redhat.com>
Link: https://lore.kernel.org/r/20211026141055.57358-3-frederic@kernel.org
When at least one CPU runs in nohz_full mode, a dedicated timekeeper CPU
is guaranteed to stay online and to never stop its tick.
Meanwhile on some rare case, the dedicated timekeeper may be running
with interrupts disabled for a while, such as in stop_machine.
If jiffies stop being updated, a nohz_full CPU may end up endlessly
programming the next tick in the past, taking the last jiffies update
monotonic timestamp as a stale base, resulting in an tick storm.
Here is a scenario where it matters:
0) CPU 0 is the timekeeper and CPU 1 a nohz_full CPU.
1) A stop machine callback is queued to execute somewhere.
2) CPU 0 reaches MULTI_STOP_DISABLE_IRQ while CPU 1 is still in
MULTI_STOP_PREPARE. Hence CPU 0 can't do its timekeeping duty. CPU 1
can still take IRQs.
3) CPU 1 receives an IRQ which queues a timer callback one jiffy forward.
4) On IRQ exit, CPU 1 schedules the tick one jiffy forward, taking
last_jiffies_update as a base. But last_jiffies_update hasn't been
updated for 2 jiffies since the timekeeper has interrupts disabled.
5) clockevents_program_event(), which relies on ktime_get(), observes
that the expiration is in the past and therefore programs the min
delta event on the clock.
6) The tick fires immediately, goto 3)
7) Tick storm, the nohz_full CPU is drown and takes ages to reach
MULTI_STOP_DISABLE_IRQ, which is the only way out of this situation.
Solve this with unconditionally updating jiffies if the value is stale
on nohz_full IRQ entry. IRQs and other disturbances are expected to be
rare enough on nohz_full for the unconditional call to ktime_get() to
actually matter.
Reported-by: Paul E. McKenney <paulmck@kernel.org>
Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Paul E. McKenney <paulmck@kernel.org>
Link: https://lore.kernel.org/r/20211026141055.57358-2-frederic@kernel.org
The 'kprobe::data_size' is unsigned, thus it can not be negative. But if
user sets it enough big number (e.g. (size_t)-8), the result of 'data_size
+ sizeof(struct kretprobe_instance)' becomes smaller than sizeof(struct
kretprobe_instance) or zero. In result, the kretprobe_instance are
allocated without enough memory, and kretprobe accesses outside of
allocated memory.
To avoid this issue, introduce a max limitation of the
kretprobe::data_size. 4KB per instance should be OK.
Link: https://lkml.kernel.org/r/163836995040.432120.10322772773821182925.stgit@devnote2
Cc: stable@vger.kernel.org
Fixes: f47cd9b553 ("kprobes: kretprobe user entry-handler")
Reported-by: zhangyue <zhangyue1@kylinos.cn>
Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Pull scheduler fix from Thomas Gleixner:
"A single scheduler fix to ensure that there is no stale KASAN shadow
state left on the idle task's stack when a CPU is brought up after it
was brought down before"
* tag 'sched-urgent-2021-11-28' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
sched/scs: Reset task stack state in bringup_cpu()
Pull perf fix from Thomas Gleixner:
"A single fix for perf to prevent it from sending SIGTRAP to another
task from a trace point event as it's not possible to deliver a
synchronous signal to a different task from there"
* tag 'perf-urgent-2021-11-28' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
perf: Ignore sigtrap for tracepoints destined for other tasks
Pull locking fixes from Thomas Gleixner:
"Two regression fixes for reader writer semaphores:
- Plug a race in the lock handoff which is caused by inconsistency of
the reader and writer path and can lead to corruption of the
underlying counter.
- down_read_trylock() is suboptimal when the lock is contended and
multiple readers trylock concurrently. That's due to the initial
value being read non-atomically which results in at least two
compare exchange loops. Making the initial readout atomic reduces
this significantly. Whith 40 readers by 11% in a benchmark which
enforces contention on mmap_sem"
* tag 'locking-urgent-2021-11-28' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip:
locking/rwsem: Optimize down_read_trylock() under highly contended case
locking/rwsem: Make handoff bit handling more consistent
Pull another tracing fix from Steven Rostedt:
"Fix the fix of pid filtering
The setting of the pid filtering flag tested the "trace only this pid"
case twice, and ignored the "trace everything but this pid" case.
The 5.15 kernel does things a little differently due to the new sparse
pid mask introduced in 5.16, and as the bug was discovered running the
5.15 kernel, and the first fix was initially done for that kernel,
that fix handled both cases (only pid and all but pid), but the
forward port to 5.16 created this bug"
* tag 'trace-v5.16-rc2-3' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace:
tracing: Test the 'Do not trace this pid' case in create event
When creating a new event (via a module, kprobe, eprobe, etc), the
descriptors that are created must add flags for pid filtering if an
instance has pid filtering enabled, as the flags are used at the time the
event is executed to know if pid filtering should be done or not.
The "Only trace this pid" case was added, but a cut and paste error made
that case checked twice, instead of checking the "Trace all but this pid"
case.
Link: https://lore.kernel.org/all/202111280401.qC0z99JB-lkp@intel.com/
Fixes: 6cb206508b ("tracing: Check pid filtering when creating events")
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Pull tracing fixes from Steven Rostedt:
"Two fixes to event pid filtering:
- Make sure newly created events reflect the current state of pid
filtering
- Take pid filtering into account when recording trigger events.
(Also clean up the if statement to be cleaner)"
* tag 'trace-v5.16-rc2-2' of git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-trace:
tracing: Fix pid filtering when triggers are attached
tracing: Check pid filtering when creating events
If a event is filtered by pid and a trigger that requires processing of
the event to happen is a attached to the event, the discard portion does
not take the pid filtering into account, and the event will then be
recorded when it should not have been.
Cc: stable@vger.kernel.org
Fixes: 3fdaf80f4a ("tracing: Implement event pid filtering")
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
Pull power management fixes from Rafael Wysocki:
"These address three issues in the intel_pstate driver and fix two
problems related to hibernation.
Specifics:
- Make intel_pstate work correctly on Ice Lake server systems with
out-of-band performance control enabled (Adamos Ttofari).
- Fix EPP handling in intel_pstate during CPU offline and online in
the active mode (Rafael Wysocki).
- Make intel_pstate support ITMT on asymmetric systems with
overclocking enabled (Srinivas Pandruvada).
- Fix hibernation image saving when using the user space interface
based on the snapshot special device file (Evan Green).
- Make the hibernation code release the snapshot block device using
the same mode that was used when acquiring it (Thomas Zeitlhofer)"
* tag 'pm-5.16-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
PM: hibernate: Fix snapshot partial write lengths
PM: hibernate: use correct mode for swsusp_close()
cpufreq: intel_pstate: ITMT support for overclocked system
cpufreq: intel_pstate: Fix active mode offline/online EPP handling
cpufreq: intel_pstate: Add Ice Lake server to out-of-band IDs
When pid filtering is activated in an instance, all of the events trace
files for that instance has the PID_FILTER flag set. This determines
whether or not pid filtering needs to be done on the event, otherwise the
event is executed as normal.
If pid filtering is enabled when an event is created (via a dynamic event
or modules), its flag is not updated to reflect the current state, and the
events are not filtered properly.
Cc: stable@vger.kernel.org
Fixes: 3fdaf80f4a ("tracing: Implement event pid filtering")
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
snapshot_write() is inappropriately limiting the amount of data that can
be written in cases where a partial page has already been written. For
example, one would expect to be able to write 1 byte, then 4095 bytes to
the snapshot device, and have both of those complete fully (since now
we're aligned to a page again). But what ends up happening is we write 1
byte, then 4094/4095 bytes complete successfully.
The reason is that simple_write_to_buffer()'s second argument is the
total size of the buffer, not the size of the buffer minus the offset.
Since simple_write_to_buffer() accounts for the offset in its
implementation, snapshot_write() can just pass the full page size
directly down.
Signed-off-by: Evan Green <evgreen@chromium.org>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Commit 39fbef4b0f ("PM: hibernate: Get block device exclusively in
swsusp_check()") changed the opening mode of the block device to
(FMODE_READ | FMODE_EXCL).
In the corresponding calls to swsusp_close(), the mode is still just
FMODE_READ which triggers the warning in blkdev_flush_mapping() on
resume from hibernate.
So, use the mode (FMODE_READ | FMODE_EXCL) also when closing the
device.
Fixes: 39fbef4b0f ("PM: hibernate: Get block device exclusively in swsusp_check()")
Signed-off-by: Thomas Zeitlhofer <thomas.zeitlhofer+lkml@ze-it.at>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
To hot unplug a CPU, the idle task on that CPU calls a few layers of C
code before finally leaving the kernel. When KASAN is in use, poisoned
shadow is left around for each of the active stack frames, and when
shadow call stacks are in use. When shadow call stacks (SCS) are in use
the task's saved SCS SP is left pointing at an arbitrary point within
the task's shadow call stack.
When a CPU is offlined than onlined back into the kernel, this stale
state can adversely affect execution. Stale KASAN shadow can alias new
stackframes and result in bogus KASAN warnings. A stale SCS SP is
effectively a memory leak, and prevents a portion of the shadow call
stack being used. Across a number of hotplug cycles the idle task's
entire shadow call stack can become unusable.
We previously fixed the KASAN issue in commit:
e1b77c9298 ("sched/kasan: remove stale KASAN poison after hotplug")
... by removing any stale KASAN stack poison immediately prior to
onlining a CPU.
Subsequently in commit:
f1a0a376ca ("sched/core: Initialize the idle task with preemption disabled")
... the refactoring left the KASAN and SCS cleanup in one-time idle
thread initialization code rather than something invoked prior to each
CPU being onlined, breaking both as above.
We fixed SCS (but not KASAN) in commit:
63acd42c0d ("sched/scs: Reset the shadow stack when idle_task_exit")
... but as this runs in the context of the idle task being offlined it's
potentially fragile.
To fix these consistently and more robustly, reset the SCS SP and KASAN
shadow of a CPU's idle task immediately before we online that CPU in
bringup_cpu(). This ensures the idle task always has a consistent state
when it is running, and removes the need to so so when exiting an idle
task.
Whenever any thread is created, dup_task_struct() will give the task a
stack which is free of KASAN shadow, and initialize the task's SCS SP,
so there's no need to specially initialize either for idle thread within
init_idle(), as this was only necessary to handle hotplug cycles.
I've tested this on arm64 with:
* gcc 11.1.0, defconfig +KASAN_INLINE, KASAN_STACK
* clang 12.0.0, defconfig +KASAN_INLINE, KASAN_STACK, SHADOW_CALL_STACK
... offlining and onlining CPUS with:
| while true; do
| for C in /sys/devices/system/cpu/cpu*/online; do
| echo 0 > $C;
| echo 1 > $C;
| done
| done
Fixes: f1a0a376ca ("sched/core: Initialize the idle task with preemption disabled")
Reported-by: Qian Cai <quic_qiancai@quicinc.com>
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Valentin Schneider <valentin.schneider@arm.com>
Tested-by: Qian Cai <quic_qiancai@quicinc.com>
Link: https://lore.kernel.org/lkml/20211115113310.35693-1-mark.rutland@arm.com/
syzbot reported that the warning in perf_sigtrap() fires, saying that
the event's task does not match current:
| WARNING: CPU: 0 PID: 9090 at kernel/events/core.c:6446 perf_pending_event+0x40d/0x4b0 kernel/events/core.c:6513
| Modules linked in:
| CPU: 0 PID: 9090 Comm: syz-executor.1 Not tainted 5.15.0-syzkaller #0
| Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
| RIP: 0010:perf_sigtrap kernel/events/core.c:6446 [inline]
| RIP: 0010:perf_pending_event_disable kernel/events/core.c:6470 [inline]
| RIP: 0010:perf_pending_event+0x40d/0x4b0 kernel/events/core.c:6513
| ...
| Call Trace:
| <IRQ>
| irq_work_single+0x106/0x220 kernel/irq_work.c:211
| irq_work_run_list+0x6a/0x90 kernel/irq_work.c:242
| irq_work_run+0x4f/0xd0 kernel/irq_work.c:251
| __sysvec_irq_work+0x95/0x3d0 arch/x86/kernel/irq_work.c:22
| sysvec_irq_work+0x8e/0xc0 arch/x86/kernel/irq_work.c:17
| </IRQ>
| <TASK>
| asm_sysvec_irq_work+0x12/0x20 arch/x86/include/asm/idtentry.h:664
| RIP: 0010:__raw_spin_unlock_irqrestore include/linux/spinlock_api_smp.h:152 [inline]
| RIP: 0010:_raw_spin_unlock_irqrestore+0x38/0x70 kernel/locking/spinlock.c:194
| ...
| coredump_task_exit kernel/exit.c:371 [inline]
| do_exit+0x1865/0x25c0 kernel/exit.c:771
| do_group_exit+0xe7/0x290 kernel/exit.c:929
| get_signal+0x3b0/0x1ce0 kernel/signal.c:2820
| arch_do_signal_or_restart+0x2a9/0x1c40 arch/x86/kernel/signal.c:868
| handle_signal_work kernel/entry/common.c:148 [inline]
| exit_to_user_mode_loop kernel/entry/common.c:172 [inline]
| exit_to_user_mode_prepare+0x17d/0x290 kernel/entry/common.c:207
| __syscall_exit_to_user_mode_work kernel/entry/common.c:289 [inline]
| syscall_exit_to_user_mode+0x19/0x60 kernel/entry/common.c:300
| do_syscall_64+0x42/0xb0 arch/x86/entry/common.c:86
| entry_SYSCALL_64_after_hwframe+0x44/0xae
On x86 this shouldn't happen, which has arch_irq_work_raise().
The test program sets up a perf event with sigtrap set to fire on the
'sched_wakeup' tracepoint, which fired in ttwu_do_wakeup().
This happened because the 'sched_wakeup' tracepoint also takes a task
argument passed on to perf_tp_event(), which is used to deliver the
event to that other task.
Since we cannot deliver synchronous signals to other tasks, skip an event if
perf_tp_event() is targeted at another task and perf_event_attr::sigtrap is
set, which will avoid ever entering perf_sigtrap() for such events.
Fixes: 97ba62b278 ("perf: Add support for SIGTRAP on perf events")
Reported-by: syzbot+663359e32ce6f1a305ad@syzkaller.appspotmail.com
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/YYpoCOBmC/kJWfmI@elver.google.com
We found that a process with 10 thousnads threads has been encountered
a regression problem from Linux-v4.14 to Linux-v5.4. It is a kind of
workload which will concurrently allocate lots of memory in different
threads sometimes. In this case, we will see the down_read_trylock()
with a high hotspot. Therefore, we suppose that rwsem has a regression
at least since Linux-v5.4. In order to easily debug this problem, we
write a simply benchmark to create the similar situation lile the
following.
```c++
#include <sys/mman.h>
#include <sys/time.h>
#include <sys/resource.h>
#include <sched.h>
#include <cstdio>
#include <cassert>
#include <thread>
#include <vector>
#include <chrono>
volatile int mutex;
void trigger(int cpu, char* ptr, std::size_t sz)
{
cpu_set_t set;
CPU_ZERO(&set);
CPU_SET(cpu, &set);
assert(pthread_setaffinity_np(pthread_self(), sizeof(set), &set) == 0);
while (mutex);
for (std::size_t i = 0; i < sz; i += 4096) {
*ptr = '\0';
ptr += 4096;
}
}
int main(int argc, char* argv[])
{
std::size_t sz = 100;
if (argc > 1)
sz = atoi(argv[1]);
auto nproc = std::thread::hardware_concurrency();
std::vector<std::thread> thr;
sz <<= 30;
auto* ptr = mmap(nullptr, sz, PROT_READ | PROT_WRITE, MAP_ANON |
MAP_PRIVATE, -1, 0);
assert(ptr != MAP_FAILED);
char* cptr = static_cast<char*>(ptr);
auto run = sz / nproc;
run = (run >> 12) << 12;
mutex = 1;
for (auto i = 0U; i < nproc; ++i) {
thr.emplace_back(std::thread([i, cptr, run]() { trigger(i, cptr, run); }));
cptr += run;
}
rusage usage_start;
getrusage(RUSAGE_SELF, &usage_start);
auto start = std::chrono::system_clock::now();
mutex = 0;
for (auto& t : thr)
t.join();
rusage usage_end;
getrusage(RUSAGE_SELF, &usage_end);
auto end = std::chrono::system_clock::now();
timeval utime;
timeval stime;
timersub(&usage_end.ru_utime, &usage_start.ru_utime, &utime);
timersub(&usage_end.ru_stime, &usage_start.ru_stime, &stime);
printf("usr: %ld.%06ld\n", utime.tv_sec, utime.tv_usec);
printf("sys: %ld.%06ld\n", stime.tv_sec, stime.tv_usec);
printf("real: %lu\n",
std::chrono::duration_cast<std::chrono::milliseconds>(end -
start).count());
return 0;
}
```
The functionality of above program is simply which creates `nproc`
threads and each of them are trying to touch memory (trigger page
fault) on different CPU. Then we will see the similar profile by
`perf top`.
25.55% [kernel] [k] down_read_trylock
14.78% [kernel] [k] handle_mm_fault
13.45% [kernel] [k] up_read
8.61% [kernel] [k] clear_page_erms
3.89% [kernel] [k] __do_page_fault
The highest hot instruction, which accounts for about 92%, in
down_read_trylock() is cmpxchg like the following.
91.89 │ lock cmpxchg %rdx,(%rdi)
Sice the problem is found by migrating from Linux-v4.14 to Linux-v5.4,
so we easily found that the commit ddb20d1d3a ("locking/rwsem: Optimize
down_read_trylock()") caused the regression. The reason is that the
commit assumes the rwsem is not contended at all. But it is not always
true for mmap lock which could be contended with thousands threads.
So most threads almost need to run at least 2 times of "cmpxchg" to
acquire the lock. The overhead of atomic operation is higher than
non-atomic instructions, which caused the regression.
By using the above benchmark, the real executing time on a x86-64 system
before and after the patch were:
Before Patch After Patch
# of Threads real real reduced by
------------ ------ ------ ----------
1 65,373 65,206 ~0.0%
4 15,467 15,378 ~0.5%
40 6,214 5,528 ~11.0%
For the uncontended case, the new down_read_trylock() is the same as
before. For the contended cases, the new down_read_trylock() is faster
than before. The more contended, the more fast.
Signed-off-by: Muchun Song <songmuchun@bytedance.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Waiman Long <longman@redhat.com>
Link: https://lore.kernel.org/r/20211118094455.9068-1-songmuchun@bytedance.com
There are some inconsistency in the way that the handoff bit is being
handled in readers and writers that lead to a race condition.
Firstly, when a queue head writer set the handoff bit, it will clear
it when the writer is being killed or interrupted on its way out
without acquiring the lock. That is not the case for a queue head
reader. The handoff bit will simply be inherited by the next waiter.
Secondly, in the out_nolock path of rwsem_down_read_slowpath(), both
the waiter and handoff bits are cleared if the wait queue becomes
empty. For rwsem_down_write_slowpath(), however, the handoff bit is
not checked and cleared if the wait queue is empty. This can
potentially make the handoff bit set with empty wait queue.
Worse, the situation in rwsem_down_write_slowpath() relies on wstate,
a variable set outside of the critical section containing the ->count
manipulation, this leads to race condition where RWSEM_FLAG_HANDOFF
can be double subtracted, corrupting ->count.
To make the handoff bit handling more consistent and robust, extract
out handoff bit clearing code into the new rwsem_del_waiter() helper
function. Also, completely eradicate wstate; always evaluate
everything inside the same critical section.
The common function will only use atomic_long_andnot() to clear bits
when the wait queue is empty to avoid possible race condition. If the
first waiter with handoff bit set is killed or interrupted to exit the
slowpath without acquiring the lock, the next waiter will inherit the
handoff bit.
While at it, simplify the trylock for loop in
rwsem_down_write_slowpath() to make it easier to read.
Fixes: 4f23dbc1e6 ("locking/rwsem: Implement lock handoff to prevent lock starvation")
Reported-by: Zhenhua Ma <mazhenhua@xiaomi.com>
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Waiman Long <longman@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20211116012912.723980-1-longman@redhat.com
