On 18.07.23 17:57, Stefan Hajnoczi wrote:
On Wed, Jul 12, 2023 at 01:16:59PM +0200, Hanna Czenczek wrote:
For vhost-user devices, qemu can migrate the virtio state, but not the
back-end's internal state. To do so, we need to be able to transfer
this internal state between front-end (qemu) and back-end.
At this point, this new feature is added for the purpose of virtio-fs
migration. Because virtiofsd's internal state will not be too large, we
believe it is best to transfer it as a single binary blob after the
streaming phase.
These are the additions to the protocol:
- New vhost-user protocol feature VHOST_USER_PROTOCOL_F_MIGRATORY_STATE
It's not 100% clear whether "migratory" is related to live migration or
something else. I don't like the name :P.
The name "VHOST_USER_PROTOCOL_F_DEVICE_STATE" would be more obviously
associated with SET_DEVICE_STATE_FD and CHECK_DEVICE_STATE than
"MIGRATORY_STATE".
Sure, sure. Naming things is hard. :)
- SET_DEVICE_STATE_FD function: Front-end and back-end negotiate a pipe
over which to transfer the state.
Does it need to be a pipe or can it be another type of file (e.g. UNIX
domain socket)?
It’s difficult to say, honestly. It can be anything, but I’m not sure
how to describe that in this specification.
It must be any FD into which the state sender can write the state and
signal end of state by closing its FD; and from which the state receiver
can read the state, terminated by seeing an EOF. As you say, that
doesn’t mean that the sender has to write the state into the FD, nor
that the receiver has to read it (into memory), it’s just that either
side must ensure the other can do it.
In the future the fd may become bi-directional. Pipes are
uni-directional on Linux.
I suggest calling it a "file descriptor" and not mentioning "pipes"
explicitly.
Works here in the commit message, but in the document, we need to be
explicit about the requirements for this FD, i.e. the way in which
front-end and back-end can expect the FD to be usable. Calling it a
“pipe” was a simple way, but you’re right, it’s more general than that.
- CHECK_DEVICE_STATE: After the state has been transferred through the
pipe, the front-end invokes this function to verify success. There is
no in-band way (through the pipe) to indicate failure, so we need to
check explicitly.
Once the transfer pipe has been established via SET_DEVICE_STATE_FD
(which includes establishing the direction of transfer and migration
phase), the sending side writes its data into the pipe, and the reading
side reads it until it sees an EOF. Then, the front-end will check for
success via CHECK_DEVICE_STATE, which on the destination side includes
checking for integrity (i.e. errors during deserialization).
Suggested-by: Stefan Hajnoczi <stefa...@redhat.com>
Signed-off-by: Hanna Czenczek <hre...@redhat.com>
---
docs/interop/vhost-user.rst | 87 +++++++++++++++++++++++++++++++++++++
1 file changed, 87 insertions(+)
diff --git a/docs/interop/vhost-user.rst b/docs/interop/vhost-user.rst
index ac6be34c4c..c98dfeca25 100644
--- a/docs/interop/vhost-user.rst
+++ b/docs/interop/vhost-user.rst
@@ -334,6 +334,7 @@ in the ancillary data:
* ``VHOST_USER_SET_VRING_ERR``
* ``VHOST_USER_SET_BACKEND_REQ_FD`` (previous name
``VHOST_USER_SET_SLAVE_REQ_FD``)
* ``VHOST_USER_SET_INFLIGHT_FD`` (if ``VHOST_USER_PROTOCOL_F_INFLIGHT_SHMFD``)
+* ``VHOST_USER_SET_DEVICE_STATE_FD``
If *front-end* is unable to send the full message or receives a wrong
reply it will close the connection. An optional reconnection mechanism
@@ -497,6 +498,44 @@ it performs WAKE ioctl's on the userfaultfd to wake the
stalled
back-end. The front-end indicates support for this via the
``VHOST_USER_PROTOCOL_F_PAGEFAULT`` feature.
+.. _migrating_backend_state:
+
+Migrating back-end state
+^^^^^^^^^^^^^^^^^^^^^^^^
+
+If the back-end has internal state that is to be sent from source to
+destination,
Migration and the terms "source" and "destination" have not been
defined. Here is a suggestion for an introductory paragraph:
Migrating device state involves transferring the state from one
back-end, called the source, to another back-end, called the
destination. After migration, the destination transparently resumes
operation without requiring the driver to re-initialize the device at
the VIRTIO level. If the migration fails, then the source can
transparently resume operation until another migration attempt is
made.
You’re right, thanks! Maybe I’ll try to be even more verbose here, and
include what VM and guest do.
the front-end may be able to store and transfer it via an
+internal migration stream. Support for this is negotiated with the
+``VHOST_USER_PROTOCOL_F_MIGRATORY_STATE`` feature.
+
+First, a channel over which the state is transferred is established on
+the source side using the ``VHOST_USER_SET_DEVICE_STATE_FD`` message.
+This message has two parameters:
+
+* Direction of transfer: On the source, the data is saved, transferring
+ it from the back-end to the front-end. On the destination, the data
+ is loaded, transferring it from the front-end to the back-end.
+
+* Migration phase: Currently, only the period after memory transfer
"memory transfer" is vague. This sentence is referring to VM live
migration and guest RAM but it may be better to focus on just the device
perspective and not the VM:
The device perspective does include guest RAM, though, because the
back-end must log its memory modifications, so it is very much involved
in that process. I think it’s a good idea to note that the state
transfer will occur afterwards.
Migration is currently only supported while the device is suspended
and all of its rings are stopped. In the future, additional phases
might be support to allow iterative migration while the device is
running.
In any case, I’ll happily add this last sentence.
+ before switch-over is supported, in which the device is suspended and
+ all of its rings are stopped.
+
+Then, the writing end will write all the data it has, signalling the end
+of data by closing its end of the pipe. The reading end must read all
+of this data and process it:
+
+* If saving, the front-end will transfer this data to the destination,
To be extra clear:
...transfer this data to the destination through some
implementation-specific means.
Yep!
+ where it is loaded into the destination back-end.
+
+* If loading, the back-end must deserialize its internal state from the
+ transferred data and be set up to resume operation.
"and be set up to resume operation" is a little unclear to me. I guess
it means "in preparation for VHOST_USER_RESUME".
I don’t think the back-end on the destination will receive a RESUME. It
never gets a SUSPEND, after all. So this is about resuming operation
once the vrings are kicked, and resuming it like it was left on the
source when the back-end was SUSPEND-ed there.
+
+After the front-end has seen all data transferred (saving: seen an EOF
+on the pipe; loading: closed its end of the pipe), it sends the
+``VHOST_USER_CHECK_DEVICE_STATE`` message to verify that data transfer
+was successful in the back-end, too. The back-end responds once it
+knows whether the tranfer and processing was successful or not.
+
Memory access
-------------
@@ -891,6 +930,7 @@ Protocol features
#define VHOST_USER_PROTOCOL_F_STATUS 16
#define VHOST_USER_PROTOCOL_F_XEN_MMAP 17
#define VHOST_USER_PROTOCOL_F_SUSPEND 18
+ #define VHOST_USER_PROTOCOL_F_MIGRATORY_STATE 19
Front-end message types
-----------------------
@@ -1471,6 +1511,53 @@ Front-end message types
before. The back-end must again begin processing rings that are not
stopped, and it may resume background operations.
+``VHOST_USER_SET_DEVICE_STATE_FD``
+ :id: 43
+ :equivalent ioctl: N/A
+ :request payload: device state transfer parameters
+ :reply payload: ``u64``
+
+ The front-end negotiates a pipe over which to transfer the back-end’s
+ internal state during migration. For this purpose, this message is
+ accompanied by a file descriptor that is to be the back-end’s end of
+ the pipe. If the back-end can provide a more efficient pipe (i.e.
+ because it internally already has a pipe into/from which to
+ put/receive state), it can ignore this and reply with a different file
+ descriptor to serve as the front-end’s end.
+
+ The request payload contains parameters for the subsequent data
+ transfer, as described in the :ref:`Migrating back-end state
+ <migrating_backend_state>` section. That section also explains the
+ data transfer itself.
+
+ The value returned is both an indication for success, and whether a
+ new pipe file descriptor is returned: Bits 0–7 are 0 on success, and
+ non-zero on error. Bit 8 is the invalid FD flag; this flag is set
+ when there is no file descriptor returned. When this flag is not set,
+ the front-end must use the returned file descriptor as its end of the
+ pipe. The back-end must not both indicate an error and return a file
+ descriptor.
Is the invalid FD flag necessary? The front-end can check whether or not
an fd was passed along with the result, so I'm not sure why the result
also needs to communicate this.
If the front-end can check this, shouldn’t the back-end also generally
be able to check whether the front-end has passed an FD in the ancillary
data? We do have this flag in messages sent by the front-end that can
optionally provide an FD (SET_VRING_KICK, SET_VRING_CALL), so I thought
it would be good for symmetry to keep this convention every time an FD
is optional in communication between front-end and back-end, in either
direction.
Hanna
_______________________________________________
Virtio-fs mailing list
Virtio-fs@redhat.com
https://listman.redhat.com/mailman/listinfo/virtio-fs
