qemu-devel
[Top][All Lists]
Advanced
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Reducing vdpa migration downtime because of memory pin / maps

From: Si-Wei Liu
Subject: Re: Reducing vdpa migration downtime because of memory pin / maps
Date: Mon, 17 Jul 2023 12:56:58 -0700
User-agent: Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:102.0) Gecko/20100101 Thunderbird/102.12.0 
Hey,
I am now back from the break. Sorry for the delayed response, please see in line. 

On 7/9/2023 11:04 PM, Eugenio Perez Martin wrote:

On Sat, Jul 8, 2023 at 11:14 AM Si-Wei Liu <si-wei.liu@oracle.com> wrote:



On 7/5/2023 10:46 PM, Eugenio Perez Martin wrote:

On Thu, Jul 6, 2023 at 2:13 AM Si-Wei Liu <si-wei.liu@oracle.com> wrote:


On 7/5/2023 11:03 AM, Eugenio Perez Martin wrote:

On Tue, Jun 27, 2023 at 8:36 AM Si-Wei Liu <si-wei.liu@oracle.com> wrote:

On 6/9/2023 7:32 AM, Eugenio Perez Martin wrote:

On Fri, Jun 9, 2023 at 12:39 AM Si-Wei Liu <si-wei.liu@oracle.com> wrote:

On 6/7/23 01:08, Eugenio Perez Martin wrote:

On Wed, Jun 7, 2023 at 12:43 AM Si-Wei Liu <si-wei.liu@oracle.com> wrote:

Sorry for reviving this old thread, I lost the best timing to follow up
on this while I was on vacation. I have been working on this and found
out some discrepancy, please see below.

On 4/5/23 04:37, Eugenio Perez Martin wrote:

Hi!

As mentioned in the last upstream virtio-networking meeting, one of
the factors that adds more downtime to migration is the handling of
the guest memory (pin, map, etc). At this moment this handling is
bound to the virtio life cycle (DRIVER_OK, RESET). In that sense, the
destination device waits until all the guest memory / state is
migrated to start pinning all the memory.

The proposal is to bind it to the char device life cycle (open vs
close),

Hmmm, really? If it's the life cycle for char device, the next guest /
qemu launch on the same vhost-vdpa device node won't make it work.


Maybe my sentence was not accurate, but I think we're on the same page here.

Two qemu instances opening the same char device at the same time are
not allowed, and vhost_vdpa_release clean all the maps. So the next
qemu that opens the char device should see a clean device anyway.

I mean the pin can't be done at the time of char device open, where the
user address space is not known/bound yet. The earliest point possible
for pinning would be until the vhost_attach_mm() call from SET_OWNER is
done.

Maybe we are deviating, let me start again.

Using QEMU code, what I'm proposing is to modify the lifecycle of the
.listener member of struct vhost_vdpa.

At this moment, the memory listener is registered at
vhost_vdpa_dev_start(dev, started=true) call for the last vhost_dev,
and is unregistered in both vhost_vdpa_reset_status and
vhost_vdpa_cleanup.

My original proposal was just to move the memory listener registration
to the last vhost_vdpa_init, and remove the unregister from
vhost_vdpa_reset_status. The calls to vhost_vdpa_dma_map/unmap would
be the same, the device should not realize this change.

This can address LM downtime latency for sure, but it won't help
downtime during dynamic SVQ switch - which still needs to go through the
full unmap/map cycle (that includes the slow part for pinning) from
passthrough to SVQ mode. Be noted not every device could work with a
separate ASID for SVQ descriptors. The fix should expect to work on
normal vDPA vendor devices without a separate descriptor ASID, with
platform IOMMU underneath or with on-chip IOMMU.


At this moment the SVQ switch is very inefficient mapping-wise, as it
unmap all the GPA->HVA maps and overrides it. In particular, SVQ is
allocated in low regions of the iova space, and then the guest memory
is allocated in this new IOVA region incrementally.

Yep. The key to build this fast path for SVQ switching I think is to
maintain the identity mapping for the passthrough queues so that QEMU
can reuse the old mappings for guest memory (e.g. GIOVA identity mapped
to GPA) while incrementally adding new mappings for SVQ vrings.


We can optimize that if we place SVQ in a free GPA area instead.

Here's a question though: it might not be hard to find a free GPA range
for the non-vIOMMU case (allocate iova from beyond the 48bit or 52bit
ranges), but I'm not sure if easy to find a free GIOVA range for the
vIOMMU case - particularly this has to work in the same entire 64bit
IOVA address ranges that (for now) QEMU won't be able to "reserve" a
specific IOVA ranges for SVQ from the vIOMMU. Do you foresee this can be
done for every QEMU emulated vIOMMU (intel-iommu amd-iommu, arm smmu and
virito-iommu) so that we can call it out as a generic means for SVQ
switching optimization?


In the case vIOMMU allocates a new block we will use the same algorithm as now:
* Find a new free IOVA chunk of the same size
* Map this new SVQ IOVA, that may or may not be the same as SVQ

Since we must go through the translation phase to sanitize guest's
available descriptors anyway, it has zero added cost.

Not sure I followed, this can work but doesn't seem able to reuse the
old host kernel mappings for guest memory, hence still requires remap of
the entire host IOVA ranges when SVQ IOVA comes along. I think by
maintaining 1:1 identity map on guest memory, we don't have to bother
tearing down existing HVA->HPA mappings in kernel thus save the
expensive pinning calls at large. I don't clearly see under this scheme,
how the new SVQ IOVA may work with potential conflict on IOVA space from
hotplugged memory - in this case the 1:1 IOVA->GPA identity guest memory
mapping can't be kept.


There is no need to maintain the 1:1 for memory mapped after the
pinning. The bigger reason to maintain them is to reduce the downtime
because of pinning.
Yes, if guest users don't care about SVQ switching downtime there's no need to maintain 1:1, and SVQ translation like today should work for them. However most live migration users who care about downtime during live migration also care about the downtime for SVQ switching - you don't want to say that brownout time like 300ms or so in the mid of live migration at the cost of complete service loss of 4 to 5 seconds at the start of migration is a win to them. What I said above has the presumption that we both are looking at (the possibility of) a generic/software way to optimize/reduce pinning overhead on the downtime - say what can be done at QEMU level or host kernel to avoid pinning (at SVQ switch) rather than put burden on every hardware vendor to implement a separate ASID for SVQ. 


After that, we can reuse the method we're using at
this moment, looking for a new free hole for the new map. ew only need
to pin the new map.

Consider this sequence:
- initially host device uses original GPA for guest memory mapping
- live migration starts off, QEMU's iova tree maps guest memory GPA 1:1 to IOVA - SVQ new maps allocated from a free hole on iova tree that happen to fall just above the IOVA range for guest memory GPA - new memory hot plugged to guest while migration is still going on in source host. although this hot plugged memory region sits right above the guest memory blocks in guest memory space (GPA), it will get new map in a separate range (not 1:1 mapped to GPA any more!) from the QEMU iova tree. Since QEMU mediates and translates virtqueue access to memory via SVQ (all guest memory maps to the same IOVA for SVQ), so far so good - for some reason live migration fails before VM is able to be migrated to the destination host, VM has to resume from the source host immediately - since live migration had failed, QEMU will unmap SVQ vrings from the IOVA tree and then stop shadowing virtqueue access - the host vDPA device now has incorrect passthrough GPA mapping for the part of newly plugged guest memory, as that belongs to the IOVA range where SVQ translation is in use
Although it might not be easy for SVQ vrings and plugged memory to collide on the same GPA/IOVA range, this is something we should prevent from happening in the first place I'd assume. You can say we only look from higher IOVA ranges to map SVQ so that the lower range can be reserved for latent hot plug memory, but this still needs complicated implementation to deal with IOVA range fragmentation and special collision prevention from breaking the 1:1 map to guest memory space. 

Regards,
-Siwei




Another option would be to move the SVQ vring to a new region, but I
don't see any advantage on maintaining 1:1 mapping at that point.

See above. For pinning avoidance point of view (i.e. QEMU software
optimization on SVQ switching downtime), I think it's crucial to
maintain 1:1 mapping and move SVQ vring to a specific region. But I'm
not sure how complex it will be for QEMU to get it implemented in a
clean way.

Thanks,
-Siwei


If this QEMU/vIOMMU "hack" is not universally feasible, I would rather
build a fast path in the kernel via a new vhost IOTLB command, say
INVALIDATE_AND_UPDATE_ALL, to atomically flush all existing
(passthrough) mappings and update to use the SVQ ones in a single batch,
while keeping the pages for guest memory always pinned (the kernel will
make this decision). This doesn't expose pinning to userspace, and can
also fix downtime issue.


    All
of the "translations" still need to be done, to ensure the guest
doesn't have access to SVQ vring. That way, qemu will not send all the
unmaps & maps, only the new ones. And vhost/vdpa does not need to call
unpin_user_page / pin_user_pages for all the guest memory.

More optimizations include the batching of the SVQ vrings.

Nods.


One of the concerns was that it could delay VM initialization, and I
didn't profile it but I think that may be the case.

Yes, that's the concern here - we should not introduce regression to
normal VM boot process/time. In case of large VM it's very easy to see
the side effect if we go this way.


     I'm not sure about
the right fix but I think the change is easy to profile. If that is
the case, we could:
* use a flag (listener->address_space ?) and only register the
listener in _init if waiting for a migration, do it in _start
otherwise.

Just doing this alone won't help SVQ mode switch downtime, see the
reason stated above.


* something like io_uring, where the map can be done in parallel and
we can fail _start if some of them fails.

This can alleviate the problem somehow, but still sub-optimal and not
scalable with larger size. I'd like zero or least pinning to be done at
the SVQ switch or migration time.


To reduce even further the pinning at SVQ time we would need to
preallocate SVQ vrings before suspending the device.

Yep. Preallocate SVQ vrings at the start of migration, but before
suspending the device. This will work under the assumption that we can
reserve or "steal" some ranges from the GPA or GIOVA space...

Thanks,
-Siwei


Actually I think the counterpart vhost_detach_mm() only gets
handled in vhost_vdpa_release() at device close time is a resulting
artifact and amiss of today's vhost protocol - the opposite RESET_OWNER
call is not made mandatory hence only seen implemented in vhost-net
device today. One qemu instance could well exec(3) another new qemu
instance to live upgrade itself while keeping all emulated devices and
guest alive. The current vhost design simply prohibits this from happening.


Ok, I was not aware of this. Thanks for explaining it!


       so all the guest memory can be pinned for all the guest / qemu
lifecycle.

I think to tie pinning to guest / qemu process life cycle makes more
sense. Essentially this pinning part needs to be decoupled from the
iotlb mapping abstraction layer, and can / should work as a standalone
uAPI. Such that QEMU at the destination may launch and pin all guest's
memory as needed without having to start the device, while awaiting any
incoming migration request. Though problem is, there's no existing vhost
uAPI that could properly serve as the vehicle for that. SET_OWNER /
SET_MEM_TABLE / RESET_OWNER seems a remote fit.. Any objection against
introducing a new but clean vhost uAPI for pinning guest pages, subject
to guest's life cycle?


I think that to pin or not pin memory maps should be a kernel
decision, not to be driven by qemu.

It's kernel decision for sure. I am with this part.


I'm not against it if needed, but
let me know if the current "clean at close" address your concerns.

To better facilitate QEMU exec (live update) case, I propose we add new
vhost uAPI pair for explicit pinning request - which would live with
user mm's, or more precisely qemu instance's lifecycle.


Ok I see your problem better now, but I think it should be solved at
kernel level. Does that live update need to forcefully unpin and pin
the memory again,

No, it should avoid the unpin&pin cycle, otherwise it'd defeat the
downtime expectation. The exec(3)'d process should inherit the page
pinning and/or mlock accounting from the original QEMU process, while
keeping original page pinning intact. Physical page mappings for DMA can
be kept as is to avoid the need of reprogramming device, though in this
case the existing vhost iotlb entries should be updated to reflect the
new HVA in the exec(3)'d QEMU process.


     or that is just a consequence of how it works the
memory listener right now?

Why not extend the RESET_OWNER to the rest of devices? It seems the
most natural way to me.

Not sure, I think RESET_OWNER might be too heavy weighted to implement
live update, and people are not clear what the exact semantics are by
using it (which part of the device state is being reset, and how
much)... In addition, people working on iommufd intended to make this a
"one-shot" ioctl e.g. CHANGE_OWNER instead of RESET_OWNER+SET_OWNER:

https://lore.kernel.org/kvm/Y5Ibvv9PNMifi0NF@ziepe.ca/

New uAPI to just change ownership of mm seems a better fit to me...


I'm not sure about the right solution here, but there are other
proposals to batch ioctls. But maybe creating a new one fits better.

Thanks!


Thanks,
-Siwei


Thanks!



Another concern is the use_va stuff, originally it tags to the device
level and is made static at the time of device instantiation, which is
fine. But others to come just find a new home at per-group level or
per-vq level struct. Hard to tell whether or not pinning is actually
needed for the latter use_va friends, as they are essentially tied to
the virtio life cycle or feature negotiation. While guest / Qemu starts
way earlier than that. Perhaps just ignore those sub-device level use_va
usages? Presumably !use_va at the device level is sufficient to infer
the need of pinning for device?


I don't follow this. But I have the feeling that the subject of my
original mail is way more accurate if I would have said just "memory
maps".

I think the iotlb layer in vhost-vdpa just provides the abstraction for
mapping, not pinning. Although in some case mapping implicitly relies on
pinning for DMA purpose, it doesn't have to tie to that in uAPI
semantics. We can do explicit on-demand pinning for cases for e.g.
warming up device at live migration destination.



I still consider the way to fix it is to actually delegate that to the
kernel vdpa, so it can choose if a particular ASID needs the pin or
not. But let me know if I missed something.

You can disregard this for now. I will discuss that further with you
guys while bind_mm and per-group use_va stuffs are landed.

Thanks!
-Siwei




Thanks!


Regards,
-Siwei



This has two main problems:
* At this moment the reset semantics forces the vdpa device to unmap
all the memory. So this change needs a vhost vdpa feature flag.
* This may increase the initialization time. Maybe we can delay it if
qemu is not the destination of a LM. Anyway I think this should be
done as an optimization on top.

Any ideas or comments in this regard?

Thanks!
reply via email to
[Prev in Thread] Current Thread [Next in Thread]