A cleanup patch from my legacy timer series broke ia64 and led
to RCU stall errors and a fast system clock:
[ 909.360108] INFO: task systemd-sysv-ge:200 blocked for more than 127 seconds.
[ 909.360108] Not tainted 5.10.0+ #130
[ 909.360108] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 909.360108] task:systemd-sysv-ge state:D stack: 0 pid: 200 ppid: 189 flags:0x00000000
[ 909.364108]
[ 909.364108] Call Trace:
[ 909.364423] [<a00000010109b210>] __schedule+0x890/0x21e0
[ 909.364423] sp=e0000100487d7b70 bsp=e0000100487d1748
[ 909.368423] [<a00000010109cc00>] schedule+0xa0/0x240
[ 909.368423] sp=e0000100487d7b90 bsp=e0000100487d16e0
[ 909.368558] [<a00000010109ce70>] io_schedule+0x70/0xa0
[ 909.368558] sp=e0000100487d7b90 bsp=e0000100487d16c0
[ 909.372290] [<a00000010109e1c0>] bit_wait_io+0x20/0xe0
[ 909.372290] sp=e0000100487d7b90 bsp=e0000100487d1698
[ 909.374168] rcu: INFO: rcu_sched detected stalls on CPUs/tasks:
[ 909.376290] [<a00000010109d860>] __wait_on_bit+0xc0/0x1c0
[ 909.376290] sp=e0000100487d7b90 bsp=e0000100487d1648
[ 909.374168] rcu: 3-....: (2 ticks this GP) idle=19e/1/0x4000000000000002 softirq=1581/1581 fqs=2
[ 909.374168] (detected by 0, t=5661 jiffies, g=1089, q=3)
[ 909.376290] [<a00000010109da80>] out_of_line_wait_on_bit+0x120/0x140
[ 909.376290] sp=e0000100487d7b90 bsp=e0000100487d1610
[ 909.374168] Task dump for CPU 3:
[ 909.374168] task:khungtaskd state:R running task
Revert most of my patch to make this work again, including the extra
update_process_times()/profile_tick() and the local_irq_enable() in the
loop that I expected not to be needed here.
I have not found out exactly what goes wrong, and would suggest that
someone with hardware access tries to convert this code into a singleshot
clockevent driver, which should give better behavior in all cases.
Reported-by: John Paul Adrian Glaubitz <glaubitz@physik.fu-berlin.de>
Fixes: 2b49ddcef297 ("ia64: convert to legacy_timer_tick")
Cc: John Stultz <john.stultz@linaro.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Stephen Boyd <sboyd@kernel.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Split off from prev patch in the series that implements the syscall. Link: https://lkml.kernel.org/r/20201121144401.3727659-4-willemdebruijn.kernel@gmail.com Signed-off-by: Willem de Bruijn <willemb@google.com> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: Arnd Bergmann <arnd@arndb.de> Cc: Matthew Wilcox (Oracle) <willy@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull TIF_NOTIFY_SIGNAL updates from Jens Axboe: "This sits on top of of the core entry/exit and x86 entry branch from the tip tree, which contains the generic and x86 parts of this work. Here we convert the rest of the archs to support TIF_NOTIFY_SIGNAL. With that done, we can get rid of JOBCTL_TASK_WORK from task_work and signal.c, and also remove a deadlock work-around in io_uring around knowing that signal based task_work waking is invoked with the sighand wait queue head lock. The motivation for this work is to decouple signal notify based task_work, of which io_uring is a heavy user of, from sighand. The sighand lock becomes a huge contention point, particularly for threaded workloads where it's shared between threads. Even outside of threaded applications it's slower than it needs to be. Roman Gershman <romger@amazon.com> reported that his networked workload dropped from 1.6M QPS at 80% CPU to 1.0M QPS at 100% CPU after io_uring was changed to use TIF_NOTIFY_SIGNAL. The time was all spent hammering on the sighand lock, showing 57% of the CPU time there [1]. There are further cleanups possible on top of this. One example is TIF_PATCH_PENDING, where a patch already exists to use TIF_NOTIFY_SIGNAL instead. Hopefully this will also lead to more consolidation, but the work stands on its own as well" [1] https://github.com/axboe/liburing/issues/215 * tag 'tif-task_work.arch-2020-12-14' of git://git.kernel.dk/linux-block: (28 commits) io_uring: remove 'twa_signal_ok' deadlock work-around kernel: remove checking for TIF_NOTIFY_SIGNAL signal: kill JOBCTL_TASK_WORK io_uring: JOBCTL_TASK_WORK is no longer used by task_work task_work: remove legacy TWA_SIGNAL path sparc: add support for TIF_NOTIFY_SIGNAL riscv: add support for TIF_NOTIFY_SIGNAL nds32: add support for TIF_NOTIFY_SIGNAL ia64: add support for TIF_NOTIFY_SIGNAL h8300: add support for TIF_NOTIFY_SIGNAL c6x: add support for TIF_NOTIFY_SIGNAL alpha: add support for TIF_NOTIFY_SIGNAL xtensa: add support for TIF_NOTIFY_SIGNAL arm: add support for TIF_NOTIFY_SIGNAL microblaze: add support for TIF_NOTIFY_SIGNAL hexagon: add support for TIF_NOTIFY_SIGNAL csky: add support for TIF_NOTIFY_SIGNAL openrisc: add support for TIF_NOTIFY_SIGNAL sh: add support for TIF_NOTIFY_SIGNAL um: add support for TIF_NOTIFY_SIGNAL ...
Pull asm-generic cross-architecture timer cleanup from Arnd Bergmann: "This cleans up two ancient timer features that were never completed in the past, CONFIG_GENERIC_CLOCKEVENTS and CONFIG_ARCH_USES_GETTIMEOFFSET. There was only one user left for the ARCH_USES_GETTIMEOFFSET variant of clocksource implementations, the ARM EBSA110 platform. Rather than changing to use modern timekeeping, we remove the platform entirely as Russell no longer uses his machine and nobody else seems to have one any more. The conditional code for using arch_gettimeoffset() is removed as a result. For CONFIG_GENERIC_CLOCKEVENTS, there are still a couple of platforms not using clockevent drivers: parisc, ia64, most of m68k, and one Arm platform. These all do timer ticks slighly differently, and this gets cleaned up to the point they at least all call the same helper function. Instead of most platforms using 'select GENERIC_CLOCKEVENTS' in Kconfig, the polarity is now reversed, with the few remaining ones selecting LEGACY_TIMER_TICK instead" * tag 'asm-generic-timers-5.11' of git://git.kernel.org/pub/scm/linux/kernel/git/arnd/asm-generic: timekeeping: default GENERIC_CLOCKEVENTS to enabled timekeeping: remove xtime_update m68k: remove timer_interrupt() function m68k: change remaining timers to legacy_timer_tick m68k: m68328: use legacy_timer_tick() m68k: sun3/sun3c: use legacy_timer_tick m68k: split heartbeat out of timer function m68k: coldfire: use legacy_timer_tick() parisc: use legacy_timer_tick ARM: rpc: use legacy_timer_tick ia64: convert to legacy_timer_tick timekeeping: add CONFIG_LEGACY_TIMER_TICK timekeeping: remove arch_gettimeoffset net: remove am79c961a driver ARM: remove ebsa110 platform
Pull regset updates from Al Viro: "Dead code removal, mostly. The only exception is a bit of cleanups on itanic (getting rid of redundant stack unwinds - each access_uarea() call does it and we call that 7 times in a row in ptrace_[sg]etregs(), *after* having done it ourselves in the caller; location where the user registers have been spilled won't change under us, and we can bloody well just call access_elf_reg() directly, giving it the unw_frame_info we'd calculated for our own purposes)" * 'regset.followup' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs: c6x: kill ELF_CORE_COPY_FPREGS whack-a-mole: USE_ELF_CORE_DUMP [ia64] ptrace_[sg]etregs(): use access_elf_reg() instead of access_uarea() [ia64] missed cleanups from switch to regset coredumps arm: kill dump_task_regs()
Pull irqchip updates for 5.11 from Marc Zyngier: - Preliminary support for managed interrupts on platform devices - Correctly identify allocation of MSIs proxyied by another device - Remove the fasteoi IPI flow which has been proved useless - Generalise the Ocelot support to new SoCs - Improve GICv4.1 vcpu entry, matching the corresponding KVM optimisation - Work around spurious interrupts on Qualcomm PDC - Random fixes and cleanups Link: https://lore.kernel.org/r/20201212135626.1479884-1-maz@kernel.org
Wire up TIF_NOTIFY_SIGNAL handling for ia64. Cc: linux-ia64@vger.kernel.org [axboe: added fixes from Mike Rapoport <rppt@kernel.org>] Signed-off-by: Jens Axboe <axboe@kernel.dk>
The 3 architectures implementing CONFIG_VIRT_CPU_ACCOUNTING_NATIVE all have their own version of irq time accounting that dispatch the cputime to the appropriate index: hardirq, softirq, system, idle, guest... from an all-in-one function. Instead of having these ad-hoc versions, move the cputime destination dispatch decision to the core code and leave only the actual per-index cputime accounting to the architecture. Signed-off-by: Frederic Weisbecker <frederic@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Link: https://lore.kernel.org/r/20201202115732.27827-4-frederic@kernel.org
We call arch_cpu_idle() with RCU disabled, but then use
local_irq_{en,dis}able(), which invokes tracing, which relies on RCU.
Switch all arch_cpu_idle() implementations to use
raw_local_irq_{en,dis}able() and carefully manage the
lockdep,rcu,tracing state like we do in entry.
(XXX: we really should change arch_cpu_idle() to not return with
interrupts enabled)
Reported-by: Sven Schnelle <svens@linux.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Mark Rutland <mark.rutland@arm.com>
Tested-by: Mark Rutland <mark.rutland@arm.com>
Link: https://lkml.kernel.org/r/20201120114925.594122626@infradead.org
ia64 is the only architecture that calls xtime_update() in a loop,
once for each jiffie that has passed since the last event.
Before commit 3171a0305d62 ("[PATCH] simplify update_times (avoid
jiffies/jiffies_64 aliasing problem)") in 2006, it could not actually do
this any differently, but now it seems simpler to just pass the number
of jiffies that passed in the meantime.
While this loses the ability process interrupts in the middle of
the timer tick by calling local_irq_enable(), doing so is fairly
peculiar anyway and it seems better to just do what everyone
else does here.
Reviewed-by: Linus Walleij <linus.walleij@linaro.org>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>