Re: [PATCH 2/4] vhost-vdpa: reset vendor specific mapping to initial state in .release

From: Si-Wei Liu
Date: Thu Oct 19 2023 - 18:29:08 EST




On 10/19/2023 7:39 AM, Eugenio Perez Martin wrote:
On Thu, Oct 19, 2023 at 10:27 AM Jason Wang <jasowang@xxxxxxxxxx> wrote:
On Thu, Oct 19, 2023 at 2:47 PM Si-Wei Liu <si-wei.liu@xxxxxxxxxx> wrote:


On 10/18/2023 7:53 PM, Jason Wang wrote:
On Wed, Oct 18, 2023 at 4:49 PM Si-Wei Liu <si-wei.liu@xxxxxxxxxx> wrote:

On 10/18/2023 12:00 AM, Jason Wang wrote:
Unfortunately, it's a must to stick to ABI. I agree it's a mess but we
don't have a better choice. Or we can fail the probe if userspace
doesn't ack this feature.
Antoher idea we can just do the following in vhost_vdpa reset?

config->reset()
if (IOTLB_PERSIST is not set) {
config->reset_map()
}

Then we don't have the burden to maintain them in the parent?

Thanks
Please see my earlier response in the other email, thanks.

----------------%<----------------%<----------------

First, the ideal fix would be to leave this reset_vendor_mappings()
emulation code on the individual driver itself, which already has the
broken behavior.
So the point is, not about whether the existing behavior is "broken"
or not.
Hold on, I thought earlier we all agreed upon that the existing behavior
of vendor driver self-clearing maps during .reset violates the vhost
iotlb abstraction and also breaks the .set_map/.dma_map API. This is
100% buggy driver implementation itself that we should discourage or
eliminate as much as possible (that's part of the goal for this series),
I'm not saying it's not an issue, what I'm saying is, if the fix
breaks another userspace, it's a new bug in the kernel. See what Linus
said in [1]

"If a change results in user programs breaking, it's a bug in the kernel."

but here you seem to go existentialism and suggests the very opposite
that every .set_map/.dma_map driver implementation, regardless being the
current or the new/upcoming, should unconditionally try to emulate the
broken reset behavior for the sake of not breaking older userspace.
Such "emulation" is not done at the parent level. New parents just
need to implement reset_map() or not. everything could be done inside
vhost-vDPA as pseudo code that is shown above.

Set
aside the criteria and definition for how userspace can be broken, can
we step back to the original question why we think it's broken, and what
we can do to promote good driver implementation instead of discuss the
implementation details?
I'm not sure I get the point of this question. I'm not saying we don't
need to fix, what I am saying is that such a fix must be done in a
negotiable way. And it's better if parents won't get any burden. It
can just decide to implement reset_map() or not.

Reading the below response I found my major
points are not heard even if written for quite a few times.
I try my best to not ignore any important things, but I can't promise
I will not miss any. I hope the above clarifies my points.

It's not
that I don't understand the importance of not breaking old userspace, I
appreciate your questions and extra patience, however I do feel the
"broken" part is very relevant to our discussion here.
If it's broken (in the sense of vhost IOTLB API) that you agree, I think
we should at least allow good driver implementations; and when you think
about the possibility of those valid good driver cases
(.set_map/.dma_map implementations that do not clear maps in .reset),
you might be able to see why it's coded the way as it is now.

It's about whether we could stick to the old behaviour without
too much cost. And I believe we could.

And just to clarify here, reset_vendor_mappings() = config->reset_map()

But today there's no backend feature negotiation
between vhost-vdpa and the parent driver. Do we want to send down the
acked_backend_features to parent drivers?
There's no need to do that with the above code, or anything I missed here?

config->reset()
if (IOTLB_PERSIST is not set) {
config->reset_map()
}
Implementation issue: this implies reset_map() has to be there for every
.set_map implementations, but vendor driver implementation for custom
IOMMU could well implement DMA ops by itself instead of .reset_map. This
won't work for every set_map driver (think about the vduse case).
Well let me do it once again, reset_map() is not mandated:

config->reset()
if (IOTLB_PERSIST is not set) {
if (config->reset_map)
config->reset_map()
To avoid new parent drivers
I am afraid it's not just new parent drivers, but any well behaved driver today may well break userspace if go with this forced emulation code, if they have to implement reset_map for some reason (e.g. restored to 1:1 passthrough mapping or other default state in mapping). For new userspace and user driver we can guard against it using the IOTLB_PERSIST flag, but the above code would get a big chance to break setup with good driver and older userspace in practice.

And .reset_map implementation doesn't necessarily need to clear maps. For e.g. IOMMU API compliant driver that only needs simple DMA model for passthrough, all .reset_map has to do is toggle to 1:1 mapping mode to the default/initial state without taking care of maps, as vhost_vdpa_unmap(0, -1ULL) earlier should have done the map cleaning job already.


to have this behavior if they need to
implement reset_map,

What if we add a new callback like "config->buggy_virtio_reset_map",
different from regular reset_map callback at cleanup?
Right, separating out the need for old behavior emulation from .reset_map is much cleaner, and this is what individual broken driver has to maintain without penalizing other good drivers. Good to see what I said earlier is heard.

Only mlx5 and
vdpa_sim need to implement it, with a big warning, and new parent
drivers can trust they'll never have the old bad behavior.
Let's see what Jason will say about it and try to converge on this first, I think he seemed to imply that this is part of ABI that every driver has to make compromise for. I'd better get it ack'ed before proceeding to the rest.

Thanks,
-Siwei


}

Did you see any issue with VDUSE in this case?

But this is not the the point I was making. I think if you agree this is
purely buggy driver implementation of its own, we should try to isolate
this buggy behavior to individual driver rather than overload vhost-vdpa
or vdpa core's role to help implement the emulation of broken driver
behavior.
As I pointed out, if it is not noticeable in the userspace, that's
fine but it's not.

I don't get why .reset is special here, the abuse of .reset to
manipulate mapping could also happen in other IOMMU unrelated driver
entries like in .suspend, or in queue_reset.
Who can abuse reset here? It is totally under the control of
vhost-vDPA and it's not visible to uAPI. And we can fully control the
behaviour of vhost-vDPA.

If someday userspace is
found coded around similar buggy driver implementation in other driver
ops, do we want to follow and duplicate the same emulation in vdpa core
as the precedent is already set here around .reset?
I think so, have you seen the links I give you? If you want to go
through the one from Linus thread[1], you can see the one that unbreak
virtio-IOMMU[2]:

1) Someday, we spot invalidate with size 0 is a bug
2) We fix this bug by not allowing this
3) But virtio-IOMMU userspace find that size 0 actually clean all the
IOTLB so it depends on the behaviour
4) So the virtio-IOMMU userspace find it can't work after 2)
5) Then we recover the behaviour before 2) via [2]

Another example is the IOTLB_MSG_V2, V1 suffers from in-stable ABI in
32bit archs, most of the userspace survives since it never runs on
32bit archs. The fix is to introduce a V2 but we will stick to V1 by
default if V2 is not acknowledged by the userspace.

I think the above 2 examples are sufficient for us to understand the
case. If not, I can help to clarify more since I'm involved in those 2
fixes.

The buggy driver can fail in a lot of other ways indefinitely during
reset, if there's a buggy driver that's already broken the way as how it
is and happens to survive with all userspace apps, we just don't care
and let it be.
Without IOTLB_PRESIST it doesn't break. With IOTLB_PERSIST and if the
reset_map() is done unconditionally, it can break. That's my point.

There's no way we can enumerate all those buggy behaviors
in .reset_map itself, it's overloading that driver API too much.
If it is not noticeable by userspace, we can do any fix at will. But
it is not, we don't have another choice. Especially considering the
cost is rather low.

Second, IOTLB_PERSIST is needed but not sufficient. Due to lack of
backend feature negotiation in parent driver, if vhost-vdpa has to
provide the old-behaviour emulation for compatibility on driver's
behalf, it needs to be done per-driver basis. There could be good
on-chip or vendor IOMMU implementation which doesn't clear the IOTLB in
.reset, and vendor specific IOMMU doesn't have to provide .reset_map,
Then we just don't offer IOTLB_PRESIST, isn't this by design?
Think about the vduse case, it can work with DMA ops directly so doesn't
have to implement .reset_map, unless for some specific good reason.
Because it's a conforming and valid/good driver implementation, we may
still allow it to advertise IOTLB_PERSIST to userspace.
I would like to know why this can't work in this case:

config->reset()
if (IOTLB_PERSIST is not set) {
if (config->reset_map)
config->reset_map()
}

Which belongs to
the 3rd bullet below:

https://lore.kernel.org/virtualization/1696928580-7520-4-git-send-email-si-wei.liu@xxxxxxxxxx/

There are 3 cases that backend may claim this feature bit on:

- parent device that has to work with platform IOMMU
- parent device with on-chip IOMMU that has the expected
.reset_map support in driver
- parent device with vendor specific IOMMU implementation
that explicitly declares the specific backend feature

we
should allow these good driver implementations rather than
unconditionally stick to some specific problematic behavior for every
other good driver.
Then you can force reset_map() with set_map() that is what I suggest
in another thread, no?
This is exactly what I was afraid of that broken behavior emulation may
become a dangerous slippery slope - in principle we should encourage
good driver implementation, as they can work totally fine with older
userspace. Why do they have to bother emulating broken behavior just
because some other driver's misbehaving?
Please read the link [1], Linus has explained it.

And what's the boundary for
this hack, do drivers backed by platform IOMMU even have to emulate (if
not why not, and is there substantial difference in between)?
The boundary is whether the behaviour change could be noticed but
userspace. And I've shown you it's not a burden with the pseudo codes.
If not, please explain why.

After
getting through all of this, do you still believe everything is just as
easy and simple as what thought to be?
The truth is that bugs exist everywhere. We can't promise there's no
bug when developing an uAPI or subsystem. For kernel code, the bug
that touches uAPI might be fixed in a way that doesn't break existing
userspace. If you look at how downstream to maintain kABI, you will be
supersized furtherly.

Btw, I thought I was expecting but still haven't got the clear answers
to what was the goal to do all this, we spent a lot of time trying to
unbreak userspace,
The code is pretty simple. But yes, the time spent on justifying it
might take some time. That's the community. People need time to
understand each other's points.

but looks to me as if we were trying every possible
way to break userspace
How could my suggestions break a userspace?

or try to approximate to the same brokenness
mlx5_vdpa may have caused to the userspace. What we will get eventually
from these lengthy discussions?
Siwei, I'd really suggest you read the link I gave you. You may get
the answer. What's more, It doesn't cost too much then we know for
sure there would not be any issue, why not choose the hard way?

On the other hand, if you think it from
vhost-vdpa user perspective, you'll clearly see there's just a couple of
ways to unbreak userspace from the internal broken map which is out of
sync with vhost-vdpa iotlb after device reset.
Patches are more than welcomed.

If this brokenness was
something universally done from the vhost-vdpa layer itself, I'd feel
it's more of a shared problem, but this is not the case I see it here.
While the long standing mlx5_vdpa/vdpa_sim issue is 100% misuse of
.reset op in a wrong way per IOMMU API definition. Why leaving this
discrepancy to the individual driver is not even an option, I'm still
not sure?
Sorry? I start with a switch in the driver, and then I try to avoid
that. And it seems you don't want a burden on the driver as well.
Where did you see I say we can't do that in the driver? What I
disagree with is to use a module parameter.

Even if I fail, it doesn't mean we can't do that in the driver code.
If you read the link[1] you can see the offending commit is a change
in uvcvideo driver.

Thanks


Thanks,
-Siwei

Then we need a set of device flags (backend_features
bit again?) to indicate the specific driver needs upper layer's help on
old-behaviour emulation.

Last but not least, I'm not sure how to properly emulate
reset_vendor_mappings() from vhost-vdpa layer. If a vendor driver has no
.reset_map op implemented, or if .reset_map has a slightly different
implementation than what it used to reset the iotlb in the .reset op,
See above, for reset_vendor_mappings() I meant config->reset_map() exactly.

Thanks

then this either becomes effectively dead code if no one ends up using,
or the vhost-vdpa emulation is helpless and limited in scope, unable to
cover all the cases.

----------------%<----------------%<----------------