On 24.07.23 19:55, Stefan Hajnoczi wrote:
On Wed, Jul 19, 2023 at 03:59:32PM +0200, Hanna Czenczek wrote:
On 18.07.23 16:25, Stefan Hajnoczi wrote:
On Tue, Jul 11, 2023 at 05:52:23PM +0200, Hanna Czenczek wrote:
When stopping the VM, qemu wants all devices to fully cease any
operation, too. Currently, we can only have vhost-user back-ends stop
processing vrings, but not background operations. Add the SUSPEND and
RESUME commands from vDPA, which allow the front-end (qemu) to tell
back-ends to cease all operations, including those running in the
background.
qemu's current work-around for this is to reset the back-end instead of
suspending it, which will not work for back-ends that have internal
state that must be preserved across e.g. stop/cont.
Note that the given specification requires the back-end to delay
processing kicks (i.e. starting vrings) until the device is resumed,
instead of requiring the front-end not to send kicks while suspended.
qemu starts devices (and would just resume them) only when the VM is in
a running state, so it would be difficult to have qemu delay kicks until
the device is resumed, which is why this patch specifies handling of
kicks as it does.
Signed-off-by: Hanna Czenczek <hre...@redhat.com>
---
docs/interop/vhost-user.rst | 35 +++++++++++++++++++++++++++++++++--
1 file changed, 33 insertions(+), 2 deletions(-)
diff --git a/docs/interop/vhost-user.rst b/docs/interop/vhost-user.rst
index 5a070adbc1..ac6be34c4c 100644
--- a/docs/interop/vhost-user.rst
+++ b/docs/interop/vhost-user.rst
@@ -381,6 +381,10 @@ readable) on the descriptor specified by
``VHOST_USER_SET_VRING_KICK``
or receiving the in-band message ``VHOST_USER_VRING_KICK`` if negotiated,
and stop ring upon receiving ``VHOST_USER_GET_VRING_BASE``.
+While the back-end is suspended (via ``VHOST_USER_SUSPEND``), it must
+never process rings, and thus also delay handling kicks until the
If you respin this series, I suggest replacing "never" with "not" to
emphasize that ring processing is only skipped while the device is
suspended (rather than forever). "Never" feels too strong to use when
describing a temporary state.
Sure.
+back-end is resumed again.
+
Rings can be enabled or disabled by ``VHOST_USER_SET_VRING_ENABLE``.
If ``VHOST_USER_F_PROTOCOL_FEATURES`` has not been negotiated, the
@@ -479,8 +483,9 @@ supplied by ``VhostUserLog``.
ancillary data, it may be used to inform the front-end that the log has
been modified.
-Once the source has finished migration, rings will be stopped by the
-source. No further update must be done before rings are restarted.
+Once the source has finished migration, the device will be suspended and
+its rings will be stopped by the source. No further update must be done
+before the device and its rings are resumed.
This paragraph is abstract and doesn't directly identify the mechanisms
or who does what:
- "the device will be suspended" via VHOST_USER_SUSPEND (or reset when
VHOST_USER_SUSPEND is not supported?) or automatically by the device
itself or some other mechanism?
OK, I’ll add VHOST_USER_SUSPEND.
So far I hadn’t considered making a note of resetting as a fallback right in
the specification. I don’t think I would want it in the specification, but
on the other hand, it is probably important for back-end authors to know
that if they do not implement SUSPEND, their device is going to be reset by
qemu.
Can we make that a ”may”, i.e.:
```
Once the source has finished migration, the device will be suspended via
VHOST_USER_SUSPEND and its rings will be stopped by the source.
I'm not sure what "its rings will be stopped by the source" means
exactly. Is it summarizing the effect of VHOST_USER_SUSPEND or is there
an additional action after VHOST_USER_SUSPEND that stops rings? And I'm
not sure whether "by the source" means by the front-end or back-end on
the source machine?
This is pre-existing text and I assumed it (with not doubt) to refer to
GET_VRING_BASE, because that is how rings are stopped.
I can improve the existing documentation and add the reference to
GET_VRING_BASE, and say that it clearly must come from the front-end.
No further
update must be done before the device and its rings are resumed.
"Update" to what? Guest RAM? Device state? Rings?
I feel like this text is too vague for a spec. People may interpret it
differently. Can you make rephrase this to more concrete?
Honestly, no. This is pre-existing, and I have the same questions as
you do.
I cannot “rephrase” this to make it more concrete. I can try to
actually specify that was was left unspecified, but that would be a
change in specification that would require its own patch, separate from
this series.
Personally, I’ve generally taken this sentence to be fluff. If the
rings are stopped, clearly, they should not be accessed at all. Probably
the back-end should also refrain from writing to guest memory, because
that is a logical conclusion from having the rings stopped. But now
it’s even clearer: The back-end ideally is suspended, which directly
means not to modify guest memory, and not to perform “background
operations”. Updating device state of course is possible through
vhost-user commands, because those must always be executed.
So basically it’s just “Device and rings are stopped (RESUME and
GET_VRING_BASE, resp.), so you know, adhere to that.”
Or maybe I’m completely wrong and “Once the source has finished
migration, rings will be stopped by the source. No further update must
be done before rings are restarted.” is to be taken together and the
second sentence just refers to the rings, i.e. the front-end stops the
rings, and the back-end must not update them. Or it means that the
front-end must not send any commands to the back-end until restarting
the rings, but that feels practically impossible.
Again, because this sentence currently doesn’t specify anything, really,
changing it to carry any meaning would be to add to the specification,
not just clarify it.
If and only
if the back-end does not support VHOST_USER_SUSPEND, the front-end may reset
it instead (via VHOST_USER_SET_STATUS, VHOST_USER_RESET_DEVICE, or
VHOST_USER_RESET_OWNER).
```
I’m unsure about the “If and only if” – older qemu versions will break this,
but I feel like we must promise back-end writers that if they implement
SUSPEND, their back-end is not going to be reset; if it is, and something
breaks because of it, it’s the front-end that must be updated to match the
specification.
I this the trick is to say "if and only if VHOST_USER_F_SUSPEND has not
been negotiated". That way really only new front-ends that support
VHOST_USER_SUSPEND are required to use suspend instead of reset and
older versions of QEMU will not violate this statement.
Ah, right, thanks!
- "before the device and its rings are resumed" via VHOST_USER_RESUME?
And is this referring to the source device?
Yes, via VHOST_USER_RESUME, and restarting the rings by starting them (i.e.
a kick).
Whether this is referring to the source device… Well, the text as it was
before begs the exact same question, so honestly, I don’t know for sure.
“Restarting” only makes sense if the rings were stopped before, so I assume
it’s referring to the source, e.g. for the case of a failed migration.
RESUME at least definitely will only happen after a prior SUSPEND, so this
one will definitely only apply on the source side.
Please rephrase the paragraph to identify the vhost-user messages
involved.
In postcopy migration the back-end is started before all the memory has
been received from the source host, and care must be taken to avoid
@@ -885,6 +890,7 @@ Protocol features
#define VHOST_USER_PROTOCOL_F_CONFIGURE_MEM_SLOTS 15
#define VHOST_USER_PROTOCOL_F_STATUS 16
#define VHOST_USER_PROTOCOL_F_XEN_MMAP 17
+ #define VHOST_USER_PROTOCOL_F_SUSPEND 18
Front-end message types
-----------------------
@@ -1440,6 +1446,31 @@ Front-end message types
query the back-end for its device status as defined in the Virtio
specification.
+``VHOST_USER_SUSPEND``
+ :id: 41
+ :equivalent ioctl: VHOST_VDPA_SUSPEND
+ :request payload: N/A
+ :reply payload: N/A
+
+ When the ``VHOST_USER_PROTOCOL_F_SUSPEND`` protocol feature has been
+ successfully negotiated, this message is submitted by the front-end to
+ have the back-end cease all operations except for handling vhost-user
+ requests. The back-end must stop processing all virt queues, and it
+ must not perform any background operations. It may not resume until a
"background operations" are not defined. What does it mean:
- Anything that writes to memory slots
- Anything that changes the visible state of the device
- Anything that changes the non-visible internal state of the device
- etc
?
My best answer (honestly) is: You tell me. This series is introducing
SUSPEND/RESUME because qemu wants to reset devices to make them stop
“background operations”, and this would break virtiofsd if any form of reset
were actually implemented. The implementation of SUSPEND/RESUME in
virtiofsd on the other hand is supposed to basically be a no-op (besides
delaying ring processing until a RESUME, but even if we processed them
before, i.e. really make SUSPEND/RESUME no-ops, it would most likely work
out fine), so I have no idea what kind of background operations we are even
talking about, or whether any such actually exist in practice.
I don’t know what anyone had in mind when introducing the RESET. It comes
from vDPA (c3716f260bf moved it from vdpa into the common code), and exists
since the code was originally added (108a64818e6), so there’s no comment
explaining why it exists. I can’t explain what the back-end is supposed to
stop doing, because so far it isn’t explained anywhere in qemu, its git log,
or in any documentation (there basically is no vdpa documentation).
I can only say what I just completely naïvely assumed it to mean so far:
That the back-end basically should stop doing anything besides handling and
replying to vhost-user messages. If such a message requires changing any
state, visible or not, then this state change is permissible.
Okay, I suggest the following instead of "background operations":
- Changes to the device state produced by SET_DEVICE_STATE_FD.
This is definitely something that I would absolutely allow after
SUSPEND. If the front-end does not wish the back-end to do this, it can
just not send this command while the back-end is suspended.
- Writing to memory regions.
- Signalling that buffers have been used.
This feels both too tight and not concrete enough. The only buffers I
can think of are buffers supplied through virt queues, which I intended
to already be included in “stop processing all virt queues”. (I took
this to mean the used-buffer ring, too, but I can of course be more
explicit about this, e.g. “stop processing all virt queues, including
returning buffers to the driver”.) “Signalling” sounds like triggering
the callfd, but that also seems clear; if you can’t process virt queues,
including returning buffers, you can never trigger the callfd (or send
VHOST_USER_BACKEND_VRING_CALL), because there can never be a new buffer
returned to the driver.
So too tight because it feels like a subset of virt queue processing,
but not concrete enough because “buffers” makes me feel like I’m
overlooking something besides virt queues.
- Signalling that the configuration space has changed.
Maybe this could be more general, i.e. the back-end must not send any
vhost-user messages to the front-end?
The goal is to ensure the device state and memory regions are stable and
that back-end doesn't trigger activity in the front-end.
If the goal is to ensure that the device state is stable, I feel like we
must then specify precisely this, and not just to say it must not be
mutable through SET_DEVICE_STATE_FD: In general, the state is more
likely to be changed by other factors after all.
On the other hand, I also don’t see why that’s a goal. For migration,
it might seem desirable, but I don’t actually think so: The back-end is
required to send a consistent device state anyway, so it is up to the
back-end to ensure the state is consistent, we don’t need to make it a
requirement for SUSPEND. It is implicitly clear from the migration
model that if the device state were to change after the back-end has
sent it to the front-end, those change will be lost on the destination
side, so it is clear that the back-end must anticipate this and work
around it.
Other than migration, qemu doesn’t see the device state at all. I don’t
see why internal state should not change between stop/cont. If a device
experience some signal that (for some reason) it can’t pause to receive
only after the subsequent RESUME, it might need to make note of that
signal to act on it after RESUME. I would consider that a change in
internal state, and I don’t immediately see a problem with it. (It may
be problematic when migrating, because receiving such signals on the
source side after transferring the state would mean they’re lost, but
again, this is something the device clearly has to solve, e.g. by
redirecting the signals to the destination starting with the
SET_TRANSFER_STATE_FD call.)
Hanna
