[lkml]   [2019]   [Dec]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH RFC 04/15] KVM: Implement ring-based dirty memory tracking
On Tue, Dec 10, 2019 at 06:09:02PM +0100, Paolo Bonzini wrote:
> On 10/12/19 16:52, Peter Xu wrote:
> > On Tue, Dec 10, 2019 at 11:07:31AM +0100, Paolo Bonzini wrote:
> >>> I'm thinking whether I can start
> >>> to use this information in the next post on solving an issue I
> >>> encountered with the waitqueue.
> >>>
> >>> Current waitqueue is still problematic in that it could wait even with
> >>> the mmu lock held when with vcpu context.
> >>
> >> I think the idea of the soft limit is that the waiting just cannot
> >> happen. That is, the number of dirtied pages _outside_ the guest (guest
> >> accesses are taken care of by PML, and are subtracted from the soft
> >> limit) cannot exceed hard_limit - (soft_limit + pml_size).
> >
> > So the question go backs to, whether this is guaranteed somehow? Or
> > do you prefer us to keep the warn_on_once until it triggers then we
> > can analyze (which I doubt..)?
> Yes, I would like to keep the WARN_ON_ONCE just because you never know.
> Of course it would be much better to audit the calls to kvm_write_guest
> and figure out how many could trigger (e.g. two from the operands of an
> emulated instruction, 5 from a nested EPT walk, 1 from a page walk, etc.).

I would say we'd better either figure out all the caller's sites to
prove it will never overflow, or, I think we'll need the waitqueue at
least. The problem is if we release a kvm with WARN_ON_ONCE and at
last we found that it can be triggered and ring full can't be avoided,
then it means the interface and design is broken, and it could even be
too late to fix it after the interface is published.

(Actually I was not certain on previous clear_dirty interface where we
introduced a new capability for it. I'm not sure whether that can be
avoided because after all the initial version is not working at all,
and we fixed it up without changing the interface. However for this
one if at last we prove the design wrong, then we must introduce
another capability for it IMHO, and the interface is prone to change

So, with the hope that we could avoid the waitqueue, I checked all the
callers of mark_page_dirty_in_slot(). Since this initial work is only
for x86, I didn't look more into other archs, assuming that can be
done later when it is implemented for other archs (and this will for
sure also cover the common code):

mark_page_dirty_in_slot calls, per-vm (x86 only)
kvm_vm_ioctl_set_tss_addr [*]
vmx_create_vcpu [*]
kvm_arch_write_log_dirty [&]
kvm_guest_time_update [&]
nested_flush_cached_shadow_vmcs12 [&]
kvm_write_wall_clock [&]
kvm_pv_clock_pairing [&]
kvmgt_rw_gpa [?]
kvm_steal_time_set_preempted [&]
pv_eoi_put_user [&]
kvm_lapic_sync_to_vapic [&]
kvm_setup_pvclock_page [&]
record_steal_time [&]
apf_put_user [&]
init_rmode_tss [*] (see above)
init_rmode_identity_map [*] (see above)
kvm_hv_set_msr [&]
kvm_write_guest_offset_cached [&] (see above)
kvm_hv_set_msr_pw [&]

We should only need to look at the leaves of the traces because
they're where the dirty request starts. I'm marking all the leaves
with below criteria then it'll be easier to focus:

Cases with [*]: should not matter much
[&]: actually with a per-vcpu context in the upper layer
[?]: uncertain...

I'm a bit amazed after I took these notes, since I found that besides
those that could probbaly be ignored (marked as [*]), most of the rest
per-vm dirty requests are actually with a vcpu context.

Although now because we have kvm_get_running_vcpu() all cases for [&]
should be fine without changing anything, but I tend to add another
patch in the next post to convert all the [&] cases explicitly to pass
vcpu pointer instead of kvm pointer to be clear if no one disagrees,
then we verify that against kvm_get_running_vcpu().

So the only uncertainty now is kvmgt_rw_gpa() which is marked as [?].
Could this happen frequently? I would guess the answer is we don't
know (which means it can).

> > One thing to mention is that for with-vcpu cases, we probably can even
> > stop KVM_RUN immediately as long as either the per-vm or per-vcpu ring
> > reaches the softlimit, then for vcpu case it should be easier to
> > guarantee that. What I want to know is the rest of cases like ioctls
> > or even something not from the userspace (which I think I should read
> > more later..).
> Which ioctls? Most ioctls shouldn't dirty memory at all.

init_rmode_tss or init_rmode_identity_map. But I've marked them as
unimportant because they should only happen once at boot.

> >>> And if we see if the mark_page_dirty_in_slot() is not with a vcpu
> >>> context (e.g. kvm_mmu_page_fault) but with an ioctl context (those
> >>> cases we'll use per-vm dirty ring) then it's probably fine.
> >>>
> >>> My planned solution:
> >>>
> >>> - When kvm_get_running_vcpu() != NULL, we postpone the waitqueue waits
> >>> until we finished handling this page fault, probably in somewhere
> >>> around vcpu_enter_guest, so that we can do wait_event() after the
> >>> mmu lock released
> >>
> >> I think this can cause a race:
> >>
> >> vCPU 1 vCPU 2 host
> >> ---------------------------------------------------------------
> >> mark page dirty
> >> write to page
> >> treat page as not dirty
> >> add page to ring
> >>
> >> where vCPU 2 skips the clean-page slow path entirely.
> >
> > If we're still with the rule in userspace that we first do RESET then
> > collect and send the pages (just like what we've discussed before),
> > then IMHO it's fine to have vcpu2 to skip the slow path? Because
> > RESET happens at "treat page as not dirty", then if we are sure that
> > we only collect and send pages after that point, then the latest
> > "write to page" data from vcpu2 won't be lost even if vcpu2 is not
> > blocked by vcpu1's ring full?
> Good point, the race would become
> vCPU 1 vCPU 2 host
> ---------------------------------------------------------------
> mark page dirty
> write to page
> reset rings
> wait for mmu lock
> add page to ring
> release mmu lock
> reset...
> release mmu lock
> page is now dirty

Hmm, the page will be dirty after the reset, but is that an issue?

Or, could you help me to identify what I've missed?

> > Maybe we can also consider to let mark_page_dirty_in_slot() return a
> > value, then the upper layer could have a chance to skip the spte
> > update if mark_page_dirty_in_slot() fails to mark the dirty bit, so it
> > can return directly with RET_PF_RETRY.
> I don't think that's possible, most writes won't come from a page fault
> path and cannot retry.

Yep, maybe I should say it in the other way round: we only wait if
kvm_get_running_vcpu() == NULL. Then in somewhere near
vcpu_enter_guest(), we add a check to wait if per-vcpu ring is full.
Would that work?


Peter Xu

 \ /
  Last update: 2019-12-15 18:21    [W:0.133 / U:0.492 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site