Hi Steven,
Thanks for your explanation. Please see my comments below.
On 2019/3/21 0:29, Steven Price wrote:
On 19/03/2019 14:39, Heyi Guo wrote:
Hi Christoffer and Steve,
On 2019/3/15 16:59, Christoffer Dall wrote:
Hi Steve,
On Wed, Mar 13, 2019 at 10:11:30AM +0000, Steven Price wrote:
Personally I think what we need is:
* Either a patch like the one from Heyi Guo (save/restore CNTVCT_EL0) or
alternatively hooking up KVM_KVMCLOCK_CTRL to prevent the watchdog
firing when user space explicitly stops scheduling the guest for a
while.
If we save/restore CNTVCT_EL0 and the warning goes away, does the guest
wall clock timekeeping get all confused and does it figure this out
automagically somehow?
What's the source for guest wall clock timekeeping in current ARM64
implementation? Is it the value from CNTP_TVAL_EL0? Will guest OS keep
track of this? Or is the wall clock simply platform RTC?
I tested the RFC change and did not see anything unusual. Did I miss
someting?
Are you running ARM or ARM64? I'm assuming ARM64 here...
Yes, definitely :)
Unless I'm mistaken you should see the time reported by the guest not
progress when you have stopped it using the QEMU monitor console.
Running something like "while /bin/true; do date; sleep 1; done" should
allow you to see that without the patch time will jump in the guest
(i.e. the time reflects wall-clock time). With the patch I believe it
will not jump (i.e. the clock will be behind wall-clock time after the
pause).
I did the test and the result was exactly like you said.
However, this behaviour does depend on the exact system being emulated.
>From drivers/clocksource/arm_arch_timer.c:
static void __init arch_counter_register(unsigned type)
{
u64 start_count;
/* Register the CP15 based counter if we have one */
if (type & ARCH_TIMER_TYPE_CP15) {
if ((IS_ENABLED(CONFIG_ARM64) && !is_hyp_mode_available()) ||
arch_timer_uses_ppi == ARCH_TIMER_VIRT_PPI)
arch_timer_read_counter = arch_counter_get_cntvct;
else
arch_timer_read_counter = arch_counter_get_cntpct;
clocksource_counter.archdata.vdso_direct = vdso_default;
} else {
arch_timer_read_counter = arch_counter_get_cntvct_mem;
}
So we can see here that there are three functions for reading the
counter - there's the memory interface for CNTVCT, the "CP15" interface
also for CNTVCT, and an interface for CNTPCT. CNTPCT is only used for
!ARM64 or if hyp mode is available (not relevant until nested
virtualisation is added).
The upshot is that on arm64 we're using the virtual counter (CNTVCT).
Does KVM_KVMCLOCK_CTRL solve that problem?
It seems KVM_KVMCLOCK_CTRL is dedicated for guest pause/resume scenario,
but it relies on pvclock both in KVM and Guest and right now only X86
supports it :(
KVM_KVMCLOCK_CTRL simply provides a mechanism for the host to inform the
guest that a vCPU hasn't been scheduled for "a while". The guest can
then use that information to avoid triggering the watchdog timeout. As
you note it is only currently implemented for X86.
Could Steve provide more insights about the whole thing?
I'll try - see below.
Thanks,
Heyi
* KVM itself saving/restoring CNTVCT_EL0 during suspend/resume so the
guest doesn't see time pass during a suspend.
This smells like policy to me so I'd much prefer keeping as much
functionality in user space as possible. If we already have the APIs we
need from KVM, let's use them.
The problem with suspend/resume is that user space doesn't get a
look-in. There's no way (generically) for a user space application to
save/restore registers during the suspend. User space simply sees time
jump forwards - which is precisely what we're trying to hide from the
guest (as it otherwise gets upset that it appears a vCPU has deadlocked
for a while).
So while I agree this is "policy", it's something we need the kernel to
do on our behalf. Potentially we can put it behind an ioctl so that user
space can opt-in to it.
* Something equivalent to MSR_KVM_WALL_CLOCK_NEW for arm which allows
the guest to query the wall clock time from the host and provides an
offset between CNTVCT_EL0 to wall clock time which the KVM can update
during suspend/resume. This means that during a suspend/resume the guest
can observe that wall clock time has passed, without having to be
bothered about CNTVCT_EL0 jumping forwards.
Isn't the proper Arm architectural solution for this to read the
physical counter for wall clock time keeping ?
(Yes that will require a trap on physical counter reads after migration
on current systems, but migration sucks in terms of timekeeping
already.)
The physical counter is certainly tempting as it largely does what we
want as a secondary counter. However I'm wary of using it because it
starts to get "really interesting" when dealing with nested
virtualisation. Any by "really interesting" I of course mean horribly
broken :)
Basically the problem is that with nested virtualisation the offset
between the physical counter and the virtual counter is controlled by
the virtual-EL2. Because we want certain behaviours of the virtual
counter (pausing when the guest pauses) we have to replicate that onto
the physical counter to maintain the offset between the two that the
guest expects.
Given that it seems to be a non-starter when you introduce nested virt I
think we should just bite the bullet and implement a PV wall clock
mechanism for arm64 (similar to MSR_KVM_WALL_CLOCK_NEW).
If we decide to implement something like pvclock, shall we increase the
priority of this work? It seems there is much to do, including the spec, KVM,
qemu and guest kernel driver...
Thanks,
Heyi
This also has the advantage that it is going to be faster than the
physical counter when that has to be trapped (e.g. after migration).
Steve
.