Re: [PATCH RFC 3/4] vl: Sync dirty bits for system resets during precopy

[Peter Xu <peterx@redhat.com>]

On Wed, May 06, 2020 at 11:53:40AM +0100, Dr. David Alan Gilbert wrote:
> * Peter Xu (peterx@redhat.com) wrote:
> > System resets will also reset system memory layout.  Although the memory layout
> > after the reset should probably the same as before the reset, we still need to
> > do frequent memory section removals and additions during the reset process.
> > Those operations could accidentally lose per-mem-section information like KVM
> > memslot dirty bitmaps.
> > 
> > Previously we keep those dirty bitmaps by sync it during memory removal.
> > However that's hard to make it right after all [1].  Instead, we sync dirty
> > pages before system reset if we know we're during a precopy migration.  This
> > should solve the same problem explicitly.
> > 
> > [1] https://lore.kernel.org/qemu-devel/20200327150425.GJ422390@xz-x1/
> 
> I think the problem is knowing whether this is sufficient or whether you
> definitely need to do it for other cases of removal/readd.

Right that's the tricky part.  I planned to look into the other potential cases
after I get a feedback from Paolo or anyone that think this approach is
acceptable, but of course I can even start to look into them now.

When I said this should mostly be for reset, it comes from this commit message:

    commit 4cc856fabae1447d53890e707c70f257a7691174
    Author: zhanghailiang <zhang.zhanghailiang@huawei.com>
    Date:   Thu Apr 2 19:26:31 2015 +0000

    kvm-all: Sync dirty-bitmap from kvm before kvm destroy the corresponding dirty_bitmap
    
    Sometimes, we destroy the dirty_bitmap in kvm_memory_slot before any sync action
    occur, this bit in dirty_bitmap will be missed, and which will lead the corresponding
    dirty pages to be missed in migration.
    
    This usually happens when do migration during VM's Start-up or Reboot.

That's where the get-dirty-log was missing and Hailiang reported an issue with
that probably on system reboot (but, again, I know nothing about the use case
behind this...).

Today I even tried to dig deeper into the initial place that this is
introduced, and it's:

    commit 3fbffb628c001bd540dc9c1805bdf7aa8591da4d
    Author: Avi Kivity <avi@redhat.com>
    Date:   Sun Jan 15 16:13:59 2012 +0200

    kvm: flush the dirty log when unregistering a slot
    
    Otherwise, the dirty log information is lost in the kernel forever.
    
    Fixes opensuse-12.1 boot screen, which changes the vga windows rapidly.
    
    Signed-off-by: Avi Kivity <avi@redhat.com>

So it's Avi's 8 years ago patch probably trying to fix a display issue since
VGA tracks dirty bit without migration.  IIUC this should belong to the case
you raised the concern about "removing a device MMIO region with RAM backed".
So I'll definitely look into that too, which I planned to, after all.

> 
> However, assuming we need to do it during reset, how do we know this is
> the right point to do it, and not something further inside the reset
> process?  Is this just on the basis that vcpus are stopped?

"Stopped vcpus" is not the hint for "this is the right place", but definitely a
good thing because we don't even need to consider multiple layers of dirty bit
caches when they're stopped (for which the last layer is the PML, which is
unluckily per-vcpu rather than per-memslot).  It's just something nice and
something extra we got.  Even if vcpus are not stopped at this point (I believe
the above VGA MMIO region removal could be the case where I need to take care
of later, that we will have to collect dirty bits during the vcpus running),
however as I mentioned in the other thread it should be a stack high enough to
do some global dirty sync (because both dirty log and dirty ring need some
global sync, and I believe that's one of the reasons why current code is still
racy - we always need to flush the PML on every vcpu even for dirty log!).

Sorry I went too far... should probably go back to the topic on "the right
place" - I think this should be the right place for reset because AFAIK all the
memslot add/remove is done inside device reset hooks, qemu_devices_reset() (or
MachineClass.reset(), which could also change stuff, but finally calls
qemu_devices_reset() too).  I tried to verify this by enable the tracepoint at
trace_kvm_set_user_memory() and also dump something at qemu_system_reset()
before qemu_devices_reset(), I think it proves my understanding that all
trace_kvm_set_user_memory() triggers after that point.

Thanks,

-- 
Peter Xu