James
Thanks for taking the time to review this and the pointers.
On 11/11/20 10:11 AM, James Morse wrote:
Hi Henry,
On 06/11/2020 23:25, Henry Willard wrote:
machine_shutdown() is called by kernel_kexec() to shutdown
the non-boot CPUs prior to starting the new kernel. The
implementation of machine_shutdown() varies by architecture.
Many make an interprocessor call, such as smp_send_stop(),
to stop the non-boot CPUs. On some architectures the CPUs make
some sort of firmware call to stop the CPU. On some architectures
without the necessary firmware support to stop the CPU, the CPUs
go into a disabled loop, which is not suitable for supporting
kexec. On Arm64 systems that support PSCI, CPUs can be stopped
with a PSCI CPU_OFF call.
All this variation is because we want to to get the CPU back in a sane state,
as if we'd
just come from cold boot. Without the platform firmware doing its
initialisation, the only
way we have of doing this is to run the cpuhp callbacks to take the CPU offline
cleanly.
If it is unsafe to call cpu_ops.cpu_die (or cpu_die) on Arm except from
cpuhp shouldn't something detect that?
Arm64 machine_shutdown() uses the CPU hotplug infrastructure via
smp_shutdown_nonboot_cpus() to stop each CPU. This is relatively
slow and takes a best case of .02 to .03 seconds per CPU which are
stopped sequentially.
Hmmm, looks like cpuhp doesn't have a way to run the callbacks in parallel...
This can take the better part of a second for
all the CPUs to be stopped depending on how many CPUs are present.
If for some reason the CPUs are busy at the time of the kexec reboot,
it can take several seconds to shut them all down.
Busy doing what?
Executing user code
I assume the problem is CPUs starting work on behalf of user-space, which is now
pointless, which prevents them from scheduling into the cpuhp work quickly.
Does hoisting kexec's conditional call to freeze_processes() above the #ifdef -
so that
user-space threads are no longer schedule-able improve things here?
It might help the worst cases, but even on an idle system it takes a while.
Each CPU shuts itself down by calling PSCI CPU_OFF.
In some applications such as embedded systems, which need a very
fast reboot (less than a second), this may be too slow.
Where does this requirement come from? Surely kexec is part of a software
update, not
regular operation.
The requirement comes from the owner of the larger environment of which
this embedded system is a part. So, yes, this is part of software
maintenance of a component during regular operation.
This patch reverts to using smp_send_stop() to signal all
CPUs to stop immediately. Currently smp_send_stop() causes each cpu
to call local_cpu_stop(), which goes into a disabled loop. This patch
modifies local_cpu_stop() to call cpu_die() when kexec_in_progress
is true, so that the CPU calls PSCI CPU_OFF just as in the case of
smp_shutdown_nonboot_cpus().
This is appropriate for panic(), as we accept it may fail.
For Kexec(), the CPU must go offline, otherwise we can't overwrite the code it
was
running. The arch code can't just call CPU_OFF in any context. See 5.5 CPU_OFF'
of
https://developer.arm.com/documentation/den0022/d
5.5.2 describes what the OS must do first, in particular interrupts must be
migrated away
from the CPU calling CPU_OFF. Currently the cpuhp notifiers do this, which
after this
patch, no longer run.
I believe this is done by irq_migrate_all_off_this_cpu(), which is
called by take_cpu_down() scheduled on the processor to be shutdown by
stop_machine_cpuslocked(). I missed that. While take_cpu_down() is
running on the target processor the boot CPU is waiting, which is part
of the reason for the latency.
You're going to need some duct-tape here, but I recall the proposed
'ARCH_OFFLINE_CPUS_ON_REBOOT', which would help, but isn't a complete thing.
From the
discussion:
https://lore.kernel.org/lkml/87h80vwta7....@nanos.tec.linutronix.de/
https://lore.kernel.org/lkml/alpine.deb.2.21.1908201321200.2...@nanos.tec.linutronix.de/
using cpuhp to offline these CPUs is the right thing to do.
If the problem is its too slow, can we tackled that instead?
I think it is relatively slow because the CPUs are shutdown
sequentially. Besides ia64, most of the architectures that support
kexec, appear to kill them all at once. I will try again. Thanks for the
pointers.
Using smp_send_stop() instead of
smp_shutdown_nonboot_cpus() reduces the shutdown time for 23 CPUs
from about .65 seconds on an idle system to less than 5 msecs. On a
busy system smp_shutdown_nonboot_cpus() may take several seconds,
while smp_send_stop() needs only the 5 msecs.
diff --git a/arch/arm64/kernel/process.c b/arch/arm64/kernel/process.c
index 4784011cecac..2568452a2417 100644
--- a/arch/arm64/kernel/process.c
+++ b/arch/arm64/kernel/process.c
@@ -142,12 +143,22 @@ void arch_cpu_idle_dead(void)
* This must completely disable all secondary CPUs; simply causing those CPUs
* to execute e.g. a RAM-based pin loop is not sufficient. This allows the
* kexec'd kernel to use any and all RAM as it sees fit, without having to
- * avoid any code or data used by any SW CPU pin loop. The CPU hotplug
- * functionality embodied in smpt_shutdown_nonboot_cpus() to achieve this.
+ * avoid any code or data used by any SW CPU pin loop. The target stop function
+ * will call cpu_die() if kexec_in_progress is set.
*/
void machine_shutdown(void)
{
- smp_shutdown_nonboot_cpus(reboot_cpu);
+ unsigned long timeout;
+
+ /*
+ * Don't wait forever, but no longer than a second
+ */
For kexec we must wait for the CPU to exit the current kernel. If it doesn't we
can't
overwrite the current memory image with the kexec payload.
If all the CPUs haven't exited (num_online_cpus() > 1), machine_kexec()
will panic if it isn't a crash kernel. Preferably we want to find out
sooner rather than later if it isn't going to finish.
Features like CNP allow CPUs to share TLB entries. If a CPU is left behind in
the older
kernel, the code its is executing will be overwritten and its behaviour stops
being
predictable. It may start allocating junk TLB entries, that CNP allows CPUs in
the new
kernel to use, resulting in hard to debug crashes.
For kdump we avoid this problem by ensuring the old and new kernels never
overlap. The old
kernel doesn't even have the kdump carveout mapped.
+ timeout = USEC_PER_SEC;
+
+ smp_send_stop();
+ while (num_online_cpus() > 1 && timeout--)
+ udelay(1);
+ return;
}
/*
Thanks,
James
Thanks,
Henry