On 10.10.23 12:36, Alex Bennée wrote:
Hanna Czenczek <hre...@redhat.com> writes:
On 10.10.23 06:00, Yajun Wu wrote:
On 10/9/2023 5:13 PM, Hanna Czenczek wrote:
External email: Use caution opening links or attachments
On 09.10.23 11:07, Hanna Czenczek wrote:
On 09.10.23 10:21, Hanna Czenczek wrote:
On 07.10.23 04:22, Yajun Wu wrote:
[...]
<snip>
So as far as I understand, the feature is supposed to rely on
implementation-specific behavior between specifically qemu as a
front-end and dpdk as a back-end, nothing else. Honestly, that to me
is a very good reason to deprecate it. That would make it clear that
any implementation that implements it does so because it relies on
implementation-specific behavior from other implementations.
Option 2 is to fix it. It is not right to use this broadly defined
feature with its clear protocol as given in the virtio specification
just to set and clear a single bit (DRIVER_OK). The vhost-user
specification points to that virtio protocol. We must adhere to the
protocol. And note that we must not reset devices just because the VM
is paused/resumed. (That is why I wanted to deprecate SET_STATUS, so
that Stefan’s series would introduce RESET_DEVICE where we need it,
and we can (for now) ignore the SET_STATUS 0 in vhost_dev_stop().)
Option 3 would be to just be honest in the specification, and limit
the scope of F_STATUS to say the only bit that matters is DRIVER_OK.
I would say this is not really different from deprecating, though it
wouldn’t affect your case. However, I understand Alex relies on a
full status byte. I’m still interested to know why that is.
For an F_TRANSPORT backend (or whatever the final name ends up being) we
need the backend to have full control of the status byte because all the
handling of VirtIO is deferred to it. Therefor it has to handle all the
feature negotiation and indicate when the device needs resetting.
(side note: feature negotiation is another slippery area when QEMU gets
involved in gating which feature bits may or may not be exposed to the
backend. The only one it should ever mask is F_UNUSED which is used
(sic) to trigger the vhost protocol negotiation)
That’s the thing, feature negotiation is done with GET_FEATURES and
SET_FEATURES. Configuring F_REPLY_ACK lets SET_FEATURES return errors.
Indicating that the device needs reset is a good point, there is no
other feature to do that. (And something qemu currently ignores, just
like any value the device returns through GET_STATUS, but that’s besides
the point.)
Option 4 is of course not to do anything, and leave everything as-is,
waiting for the next person to stir the hornet’s nest.
Cc-ing Alex on this mail, because to me, this seems like an important
detail when he plans on using the byte in the future. If we need a
virtio status byte, I can’t see how we could use the existing F_STATUS
for it.
What would we use instead of F_STATUS to query the Device Status field?
We would emulate it in the front-end, just like we need to do for
back-ends without F_STATUS. We can’t emulate the DEVICE_NEEDS_RESET
bit, though, that’s correct.
Given that qemu currently ignores DEVICE_NEEDS_RESET, I’m not 100 %
convinced that your use case has a hard dependency on F_STATUS. However,
this still does make a fair point in general that it would be useful to
keep it.
That still leaves us with the situation that currently, the only
implementations with F_STATUS support are qemu and dpdk, which both
handle it incorrectly. Furthermore, the specification leaves much to be
desired, specifically in how F_STATUS interacts with other vhost-user
commands (which is something I cited as a reason for my original patch),
i.e. whether RESET_DEVICE and SET_STATUS 0 are equivalent, and whether
failures in feature negotiation must result in both SET_FEATURES
returning an error (with F_REPLY_ACK), and FEATURES_OK being reset in
the status byte, or whether either is sufficient. What happens when
DEVICE_NEEDS_RESET is set, i.e. do we just need RESET_DEVICE /
SET_STATUS 0, or do we also need to reset some protocol state? (This is
also connected to the fact that what happens on RESET_DEVICE is largely
undefined, which I said on Stefan’s series.)
In general, because we have our own transport, we should make a note how
it interacts with the status negotiation phases, i.e. that GET_FEATURES
must not be called before S_ACKNOWLEDGE | S_DRIVER are set, that
FEATURES_OK must be set after the SET_FEATURES call, and that DRIVER_OK
must not be set without FEATURES_OK set / SET_FEATURES having returned
success. Here we would also answer the question about the interaction
of F_REPLY_ACK+SET_FEATURES with F_STATUS, specifically whether an
implementation with F_REPLY_ACK even needs to read back the status byte
after setting FEATURES_OK because it could have got the feature
negotiation result already as a result of the SET_FEATURES call.
After migration, can you just set all flags immediately or do we need to
follow this step-by-step protocol? I think we do need to do it
step-by-step, mostly for simplicity in the back-end, i.e. that it just
sees a normal device start-up.
We should also clarify whether SET_STATUS can fail, i.e. whether setting
an invalid status (is setting FEATURES_OK when the device doesn’t think
so invalid?) has SET_STATUS fail (with F_REPLY_ACK) and/or immediately
gets the device into DEVICE_NEEDS_RESET.
We should clarify whether SET_STATUS can block. The current use of
DRIVER_OK seems to indicate to me that dpdk does do time-consuming
operations when it sees DRIVER_OK (code looks like it, too) and only
returns when that’s done, but naïvely, I would expect SET_STATUS to be
just setting some value and doing whatever needs to be done in the
background, not actually launching and blocking on an operation.
I think it is dangerous to just push ahead with using F_STATUS without
acknowledging that its implementation is broken right now, and that it
is so *on purpose* because the DRIVER_OK bit is the only thing that it
was supposed to be used for. Using it for its purported original use
(actually the virtio status byte) is contradictory to that. It’s
probably fixable, but I think it requires taking a step back and seeing
what needs to be done to remedy the conflict. If you rip out all the
existing STATUS code and replace it such that qemu will let the guest
have full control over the status byte (except for migration, where we
restore it on the destination, which will result in DRIVER_OK being set
at the end, fulfilling that requirement), that will fix the
implementation in qemu. I think. But the specification should be
amended to think about all these corner cases, not least because I think
they will also affect your implementation.
(The answers to many of the questions I raise for documentation may be
obvious to you, based on “in virtio, it’s just an MMIO byte that’s
written and read, so the rest follows from there”. But evidently the
implementation we have kind of ignores that e.g. SET_STATUS 0 is a reset
(6f8be29ec17d44496b9ed67599bceaaba72d1864 is a work-around, not much
more) or that there is actually a protocol to setting the status flags
and you can’t just set them all at once, so I don’t think the answers
are immediately obvious, and should be documented.)
As for me and the original patch: I claimed nobody really needs
F_STATUS, you say you do, so plainly, I assumed wrong and will naturally
take my hands off of F_STATUS and just ensure not to implement it in any
back-end until you’ve fixed it, as Yajun has advised. I’d still prefer
mentioning this advice in the documentation until it’s fixed, but, you
know, I wouldn’t be the first one to say “I now know about the quirk, so
I can work around it, no need to tell anyone else as long as my stuff
works”.
Hanna
