[Xen-devel] [PATCH v4 0/2] docs: Migration design documents

[Xen-devel] [PATCH v4 0/2] docs: Migration design documents

[Paul Durrant]

Paul Durrant (2):
  docs/designs: Add a design document for non-cooperative live migration
  docs/designs: Add a design document for migration of xenstore data

 docs/designs/non-cooperative-migration.md | 272 ++++++++++++++++++++++
 docs/designs/xenstore-migration.md        | 121 ++++++++++
 2 files changed, 393 insertions(+)
 create mode 100644 docs/designs/non-cooperative-migration.md
 create mode 100644 docs/designs/xenstore-migration.md

-- 
2.20.1


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

[Xen-devel] [PATCH v4 1/2] docs/designs: Add a design document for non-cooperative live migration

[Paul Durrant]

It has become apparent to some large cloud providers that the current
model of cooperative migration of guests under Xen is not usable as it
relies on software running inside the guest, which is likely beyond the
provider's control.
This patch introduces a proposal for non-cooperative live migration,
designed not to rely on any guest-side software.

Signed-off-by: Paul Durrant <pdurrant@amazon.com>
---
Cc: Andrew Cooper <andrew.cooper3@citrix.com>
Cc: George Dunlap <George.Dunlap@eu.citrix.com>
Cc: Ian Jackson <ian.jackson@eu.citrix.com>
Cc: Jan Beulich <jbeulich@suse.com>
Cc: Julien Grall <julien@xen.org>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Stefano Stabellini <sstabellini@kernel.org>
Cc: Wei Liu <wl@xen.org>

v4:
 - Fix issues raised by Wei

v2:
 - Use the term 'non-cooperative' instead of 'transparent'
 - Replace 'trust in' with 'reliance on' when referring to guest-side
   software
---
 docs/designs/non-cooperative-migration.md | 272 ++++++++++++++++++++++
 1 file changed, 272 insertions(+)
 create mode 100644 docs/designs/non-cooperative-migration.md

diff --git a/docs/designs/non-cooperative-migration.md b/docs/designs/non-cooperative-migration.md
new file mode 100644
index 0000000000..5db3939db5
--- /dev/null
+++ b/docs/designs/non-cooperative-migration.md
@@ -0,0 +1,272 @@
+# Non-Cooperative Migration of Guests on Xen
+
+## Background
+
+The normal model of migration in Xen is driven by the guest because it was
+originally implemented for PV guests, where the guest must be aware it is
+running under Xen and is hence expected to co-operate. This model dates from
+an era when it was assumed that the host administrator had control of at least
+the privileged software running in the guest (i.e. the guest kernel) which may
+still be true in an enterprise deployment but is not generally true in a cloud
+environment. The aim of this design is to provide a model which is purely host
+driven, requiring no co-operation from the software running in the
+guest, and is thus suitable for cloud scenarios.
+
+PV guests are out of scope for this project because, as is outlined above, they
+have a symbiotic relationship with the hypervisor and therefore a certain level
+of co-operation can be assumed.
+
+HVM guests can already be migrated on Xen without guest co-operation but only
+if they don’t have PV drivers installed[1] or are in power state S3. The
+reason for not expecting co-operation if the guest is in S3 is obvious, but the
+reason co-operation is expected if PV drivers are installed is due to the
+nature of PV protocols.
+
+## Xenstore Nodes and Domain ID
+
+The PV driver model consists of a *frontend* and a *backend*. The frontend runs
+inside the guest domain and the backend runs inside a *service domain* which
+may or may not be domain 0. The frontend and backend typically pass data via
+memory pages which are shared between the two domains, but this channel of
+communication is generally established using xenstore (the store protocol
+itself being an exception to this for obvious chicken-and-egg reasons).
+
+Typical protocol establishment is based on use of two separate xenstore
+*areas*. If we consider PV drivers for the *netif* protocol (i.e. class vif)
+and assume the guest has domid X, the service domain has domid Y, and the vif
+has index Z then the frontend area will reside under the parent node:
+
+`/local/domain/Y/device/vif/Z`
+
+All backends, by convention, typically reside under parent node:
+
+`/local/domain/X/backend`
+
+and the normal backend area for vif Z would be:
+
+`/local/domain/X/backend/vif/Y/Z`
+
+but this should not be assumed.
+
+The toolstack will place two nodes in the frontend area to explicitly locate
+the backend:
+
+    * `backend`: the fully qualified xenstore path of the backend area
+    * `backend-id`: the domid of the service domain
+
+and similarly two nodes in the backend area to locate the frontend area:
+
+    * `frontend`: the fully qualified xenstore path of the frontend area
+    * `frontend-id`: the domid of the guest domain
+
+
+The guest domain only has write permission to the frontend area and similarly
+the service domain only has write permission to the backend area, but both ends
+have read permission to both areas.
+
+Under both frontend and backend areas is a node called *state*. This is key to
+protocol establishment. Upon PV device creation the toolstack will set the
+value of both state nodes to 1 (XenbusStateInitialising[2]). This should cause
+enumeration of appropriate devices in both the guest and service domains. The
+backend device, once it has written any necessary protocol specific information
+into the xenstore backend area (to be read by the frontend driver) will update
+the backend state node to 2 (XenbusStateInitWait). From this point on PV
+protocols differ slightly; the following illustration is true of the netif
+protocol.
+
+Upon seeing a backend state value of 2, the frontend driver will then read the
+protocol specific information, write details of grant references (for shared
+pages) and event channel ports (for signalling) that it has created, and set
+the state node in the frontend area to 4 (XenbusStateConnected). Upon see this
+frontend state, the backend driver will then read the grant references (mapping
+the shared pages) and event channel ports (opening its end of them) and set the
+state node in the backend area to 4. Protocol establishment is now complete and
+the frontend and backend start to pass data.
+
+Because the domid of both ends of a PV protocol forms a key part of negotiating
+the data plane for that protocol (because it is encoded into both xenstore
+nodes and node paths), and because guest’s own domid and the domid of the
+service domain are visible to the guest in xenstore (and hence may cached
+internally), and neither are necessarily preserved during migration, it is
+hence necessary to have the co-operation of the frontend in re-negotiating the
+protocol using the new domid after migration.
+
+Moreover the backend-id value will be used by the frontend driver in setting up
+grant table entries and event channels to communicate with the service domain,
+so the co-operation of the guest is required to re-establish these in the new
+host environment after migration.
+
+Thus if we are to change the model and support migration of a guest with PV
+drivers, without the co-operation of the frontend driver code, the paths and
+values in both the frontend and backend xenstore areas must remain unchanged
+and valid in the new host environment, and the grant table entries and event
+channels must be preserved (and remain operational once guest execution is
+resumed).
+
+Because the service domain’s domid is used directly by the guest in setting
+up grant entries and event channels, the backend drivers in the new host
+environment must be provided by service domain with the same domid. Also,
+because the guest can sample its own domid from the frontend area and use it in
+hypercalls (e.g. HVMOP_set_param) rather than DOMID_SELF, the guest domid must
+also be preserved to maintain the ABI.
+
+Furthermore, it will necessary to modify backend drivers to re-establish
+communication with frontend drivers without perturbing the content of the
+backend area or requiring any changes to the values of the xenstore state nodes.
+
+## Other Para-Virtual State
+
+### Shared Rings
+
+Because the console and store protocol shared pages are actually part of the
+guest memory image (in an E820 reserved region just below 4G) then the content
+will get migrated as part of the guest memory image. Hence no additional code
+is require to prevent any guest visible change in the content.
+
+### Shared Info
+
+There is already a record defined in *libxenctrl Domain Image Format* [3]
+called `SHARED_INFO` which simply contains a complete copy of the domain’s
+shared info page. It is not currently incuded in an HVM (type `0x0002`)
+migration stream. It may be feasible to include it as an optional record
+but it is not clear that the content of the shared info page ever needs
+to be preserved for an HVM guest.
+
+For a PV guest the `arch_shared_info` sub-structure contains important
+information about the guest’s P2M, but this information is not relevant for
+an HVM guest where the P2M is not directly manipulated via the guest. The other
+state contained in the `shared_info` structure relates the domain wall-clock
+(the state of which should already be transferred by the `RTC` HVM context
+information which contained in the `HVM_CONTEXT` save record) and some event
+channel state (particularly if using the *2l* protocol). Event channel state
+will need to be fully transferred if we are not going to require the guest
+co-operation to re-open the channels and so it should be possible to re-build a
+shared info page for an HVM guest from such other state.
+
+Note that the shared info page also contains an array of `XEN_LEGACY_MAX_VCPUS`
+(32) `vcpu_info` structures. A domain may nominate a different guest physical
+address to use for the vcpu info. This is mandatory for if a domain wants to
+use more than 32 vCPUs and optional for legacy vCPUs. This mapping is not
+currently transferred in the migration state so this will either need to be
+added into an existing save record, or an additional type of save record will
+be needed.
+
+### Xenstore Watches
+
+As mentioned above, no domain Xenstore state is currently transferred in the
+migration stream. There is a record defined in *libxenlight Domain Image
+Format* [4] called `EMULATOR_XENSTORE_DATA` for transferring Xenstore nodes
+relating to emulators but no record type is defined for nodes relating to the
+domain itself, nor for registered *watches*. A XenStore watch is a mechanism
+used by PV frontend and backend drivers to request a notification if the value
+of a particular node (e.g. the other end’s state node) changes, so it is
+important that watches continue to function after a migration. One or more new
+save records will therefore be required to transfer Xenstore state. It will
+also be necessary to extend the *store* protocol[5] with mechanisms to allow
+the toolstack to acquire the list of watches that the guest has registered and
+for the toolstack to register a watch on behalf of a domain.
+
+### Event channels
+
+Event channels are essentially the para-virtual equivalent of interrupts. They
+are an important part of post PV protocols. Normally a frontend driver creates
+an *inter-domain* event channel between its own domain and the domain running
+the backend, which it discovers using the `backend-id` node in Xenstore (see
+above), by making a `EVTCHNOP_alloc_unbound` hypercall. This hypercall
+allocates an event channel object in the hypervisor and assigns a *local port*
+number which is then written into the frontend area in Xenstore. The backend
+driver then reads this port number and *binds* to the event channel by
+specifying it, and the value of `frontend-id`, as *remote domain* and *remote
+port* (respectively) to a `EVTCHNOP_bind_interdomain` hypercall. Once
+connection is established in this fashion frontend and backend drivers can use
+the event channel as a *mailbox* to notify each other when a shared ring has
+been updated with new requests or response structures.
+
+Currently no event channel state is preserved on migration, requiring frontend
+and backend drivers to create and bind a complete new set of event channels in
+order to re-establish a protocol connection. Hence, one or more new save
+records will be required to transfer event channel state in order to avoid the
+need for explicit action by frontend drivers running in the guest. Note that
+the local port numbers need to preserved in this state as they are the only
+context the guest has to refer to the hypervisor event channel objects.
+ Note also that the PV *store* (Xenstore access) and *console* protocols also
+rely on event channels which are set up by the toolstack. Normally, early in
+migration, the toolstack running on the remote host would set up a new pair of
+event channels for these protocols in the destination domain. These may not be
+assigned the same local port numbers as the protocols running in the source
+domain. For non-cooperative migration these channels must either be created with
+fixed port numbers, or their creation must be avoided and instead be included
+in the general event channel state record(s).
+
+### Grant table
+
+The grant table is essentially the para-virtual equivalent of an IOMMU. For
+example, the shared rings of a PV protocol are *granted* by a frontend driver
+to the backend driver by allocating *grant entries* in the guest’s table,
+filling in details of the memory pages and then writing the *grant references*
+(the index values of the grant entries) into Xenstore. The grant references of
+the protocol buffers themselves are typically written directly into the request
+structures passed via a shared ring.
+
+The guest is responsible for managing its own grant table. No hypercall is
+required to grant a memory page to another domain. It is sufficient to find an
+unused grant entry and set bits in the entry to give read and/or write access
+to a remote domain also specified in the entry along with the page frame
+number. Thus the layout and content of the grant table logically forms part of
+the guest state.
+
+Currently no grant table state is migrated, requiring a guest to separately
+maintain any state that it wishes to persist elsewhere in its memory image and
+then restore it after migration. Thus to avoid the need for such explicit
+action by the guest, one or more new save records will be required to migrate
+the contents of the grant table.
+
+# Outline Proposal
+
+* PV backend drivers will be modified to unilaterally re-establish connection
+to a frontend if the backend state node is restored with value 4
+(XenbusStateConnected)[6].
+
+* The toolstack should be modified to allow domid to be randomized on initial
+creation or default migration, but make it identical to the source domain on
+non-cooperative migration. Non-Cooperative migration will have to be denied if the
+domid is unavailable on the target host, but randomization of domid on creation
+should hopefully minimize the likelihood of this. Non-Cooperative migration to
+localhost will clearly not be possible. Patches have already been sent to
+`xen-devel` to make this change[7].
+
+* `xenstored` should be modified to implement the new mechanisms needed. See
+*Other Para-Virtual State* above. A further design document will propose
+additional protocol messages.
+
+* Within the migration stream extra save records will be defined as required.
+See *Other Para-Virtual State* above. A further design document will propose
+modifications to the libxenlight and libxenctrl Domain Image Formats.
+
+* An option should be added to the toolstack to initiate a non-cooperative
+migration, instead of the (default) potentially co-operative migration.
+Essentially this should skip the check to see if PV drivers and migrate as if
+there are none present, but also enabling the extra save records. Note that at
+least some of the extra records should only form part of a non-cooperative
+migration stream. For example, migrating event channel state would be counter
+productive in a normal migration as this will essentially leak event channel
+objects at the receiving end. Others, such as grant table state, could
+potentially harmlessly form part of a normal migration stream.
+
+* * *
+[1] PV drivers are deemed to be installed if the HVM parameter
+*HVM_PARAM_CALLBACK_IRQ* has been set to a non-zero value.
+
+[2] See https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=xen/include/public/io/xenbus.h
+
+[3] See https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=docs/specs/libxc-migration-stream.pandoc
+
+[4] See https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=docs/specs/libxl-migration-stream.pandoc
+
+[5] See https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=docs/misc/xenstore.txt
+
+[6] `xen-blkback` and `xen-netback` have already been modified in Linux to do
+this.
+
+[7] See https://lists.xenproject.org/archives/html/xen-devel/2020-01/msg00632.html
+
-- 
2.20.1


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH v4 1/2] docs/designs: Add a design document for non-cooperative live migration

[Andrew Cooper]

On 29/01/2020 14:47, Paul Durrant wrote:
> diff --git a/docs/designs/non-cooperative-migration.md b/docs/designs/non-cooperative-migration.md
> new file mode 100644
> index 0000000000..5db3939db5
> --- /dev/null
> +++ b/docs/designs/non-cooperative-migration.md
> @@ -0,0 +1,272 @@
> +# Non-Cooperative Migration of Guests on Xen
> +
> +## Background
> +
> +The normal model of migration in Xen is driven by the guest because it was
> +originally implemented for PV guests, where the guest must be aware it is
> +running under Xen and is hence expected to co-operate. 

For PV guests, is more than "expected to co-operate".

Migrating a PV guest involves rewriting every pagetable entry with a
different MFN, so even before you consider things like the PV protocols,
there is no way this could be done without the cooperation of the guest.

Sadly, this fact was depended upon for migration of the PV protocols,
and has migrated (excuse the pun) into the HVM world as well.

> This model dates from
> +an era when it was assumed that the host administrator had control of at least
> +the privileged software running in the guest (i.e. the guest kernel) which may
> +still be true in an enterprise deployment but is not generally true in a cloud
> +environment.

I haven't seen it discussed elsewhere, but even enterprise environments
have problems.

Having host admin == guest admin doesn't mean that guest drivers aren't
buggy, or that the VM doesn't explode on migrate.

The simple fact is that involving the guest kernel adds unnecessary
moving parts which can (and do with a non-zero probability) go wrong.

>  The aim of this design is to provide a model which is purely host
> +driven, requiring no co-operation from the software running in the
> +guest, and is thus suitable for cloud scenarios.
> +
> +PV guests are out of scope for this project because, as is outlined above, they
> +have a symbiotic relationship with the hypervisor and therefore a certain level
> +of co-operation can be assumed.

If nothing else, I'd at least suggest s/can be assumed/is necessary/.

> +Because the service domain’s domid is used directly by the guest in setting
> +up grant entries and event channels, the backend drivers in the new host
> +environment must be provided by service domain with the same domid. Also,
> +because the guest can sample its own domid from the frontend area and use it in
> +hypercalls (e.g. HVMOP_set_param) rather than DOMID_SELF, the guest domid must
> +also be preserved to maintain the ABI.

Has this been true since forever?  The grant and event APIs took some
care to avoid the guest needing to know its own domid.

> +
> +Furthermore, it will necessary to modify backend drivers to re-establish
> +communication with frontend drivers without perturbing the content of the
> +backend area or requiring any changes to the values of the xenstore state nodes.
> +
> +## Other Para-Virtual State
> +
> +### Shared Rings
> +
> +Because the console and store protocol shared pages are actually part of the
> +guest memory image (in an E820 reserved region just below 4G) 

Typically*.

Their exact location is entirely up to the domain builder, and tend not
to be there for PVH guests which aren't trying to fit the two frames
into a BAR.

> then the content
> +will get migrated as part of the guest memory image. Hence no additional code
> +is require to prevent any guest visible change in the content.

I do agree with this conclusion however.

> +### Shared Info
> +
> +There is already a record defined in *libxenctrl Domain Image Format* [3]
> +called `SHARED_INFO` which simply contains a complete copy of the domain’s
> +shared info page. It is not currently incuded in an HVM (type `0x0002`)
> +migration stream. It may be feasible to include it as an optional record
> +but it is not clear that the content of the shared info page ever needs
> +to be preserved for an HVM guest.
> +
> +For a PV guest the `arch_shared_info` sub-structure contains important
> +information about the guest’s P2M, but this information is not relevant for
> +an HVM guest where the P2M is not directly manipulated via the guest. The other
> +state contained in the `shared_info` structure relates the domain wall-clock
> +(the state of which should already be transferred by the `RTC` HVM context
> +information which contained in the `HVM_CONTEXT` save record) and some event
> +channel state (particularly if using the *2l* protocol). Event channel state
> +will need to be fully transferred if we are not going to require the guest
> +co-operation to re-open the channels and so it should be possible to re-build a
> +shared info page for an HVM guest from such other state.
> +
> +Note that the shared info page also contains an array of `XEN_LEGACY_MAX_VCPUS`
> +(32) `vcpu_info` structures. A domain may nominate a different guest physical
> +address to use for the vcpu info. This is mandatory for if a domain wants to
> +use more than 32 vCPUs and optional for legacy vCPUs. This mapping is not
> +currently transferred in the migration state so this will either need to be
> +added into an existing save record, or an additional type of save record will
> +be needed.

For non-cooperative migration in the current ABI, a minimum is to know
where the shared info frame is mapped, so it can be re-mapped on behalf
of the guest on the destination side.

The rest of this section will be very good evidence in the "new guest
ABI" design.

> +### Grant table
> +
> +The grant table is essentially the para-virtual equivalent of an IOMMU.

TBH, I think "shared memory" is a much better analogy than an IOMMU. 
OTOH, perhaps that doesn't cope with the grant copy aspect quite as well
as I'd like.

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH v4 1/2] docs/designs: Add a design document for non-cooperative live migration

[Durrant, Paul]

> -----Original Message-----
> From: Andrew Cooper <andrew.cooper3@citrix.com>
> Sent: 29 January 2020 19:47
> To: Durrant, Paul <pdurrant@amazon.co.uk>; xen-devel@lists.xenproject.org
> Cc: George Dunlap <George.Dunlap@eu.citrix.com>; Ian Jackson
> <ian.jackson@eu.citrix.com>; Jan Beulich <jbeulich@suse.com>; Julien Grall
> <julien@xen.org>; Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>; Stefano
> Stabellini <sstabellini@kernel.org>; Wei Liu <wl@xen.org>
> Subject: Re: [PATCH v4 1/2] docs/designs: Add a design document for non-
> cooperative live migration
> 
> On 29/01/2020 14:47, Paul Durrant wrote:
> > diff --git a/docs/designs/non-cooperative-migration.md
> b/docs/designs/non-cooperative-migration.md
> > new file mode 100644
> > index 0000000000..5db3939db5
> > --- /dev/null
> > +++ b/docs/designs/non-cooperative-migration.md
> > @@ -0,0 +1,272 @@
> > +# Non-Cooperative Migration of Guests on Xen
> > +
> > +## Background
> > +
> > +The normal model of migration in Xen is driven by the guest because it
> was
> > +originally implemented for PV guests, where the guest must be aware it
> is
> > +running under Xen and is hence expected to co-operate.
> 
> For PV guests, is more than "expected to co-operate".
> 
> Migrating a PV guest involves rewriting every pagetable entry with a
> different MFN, so even before you consider things like the PV protocols,
> there is no way this could be done without the cooperation of the guest.

Yes, the P2M will change and this is visible to the guest, but does a PV guest need to take action when this occurs? I'm not sure.

> 
> Sadly, this fact was depended upon for migration of the PV protocols,
> and has migrated (excuse the pun) into the HVM world as well.
> 

Alas yes.

> > This model dates from
> > +an era when it was assumed that the host administrator had control of
> at least
> > +the privileged software running in the guest (i.e. the guest kernel)
> which may
> > +still be true in an enterprise deployment but is not generally true in
> a cloud
> > +environment.
> 
> I haven't seen it discussed elsewhere, but even enterprise environments
> have problems.
> 
> Having host admin == guest admin doesn't mean that guest drivers aren't
> buggy, or that the VM doesn't explode on migrate.

No, but at least the host admin has a chance to test and update guest software to be 'reasonably' confident that migration will work before employing it en masse.

> 
> The simple fact is that involving the guest kernel adds unnecessary
> moving parts which can (and do with a non-zero probability) go wrong.
> 

Yes, having written the frontend side of migration in the Windows drivers it is *very* hard to get right, particularly in Windows where one has to deal with the complex and asynchronous PnP subsystem colliding with a migration. The network driver also requires a multi-reader/single-writer lock with odd semantics (w.r.t. to IRQL) which I had to code myself (https://xenbits.xen.org/gitweb/?p=pvdrivers/win/xenvif.git;a=blob;f=src/xenvif/mrsw.h). It took years of fixing subtle races (in that and elsewhere) to get to the (AFAIK) reliable code we have now. 
Avoiding execution of code like this (in all OS) certainly avoids the opportunity for subtle bugs to manifest themselves.

> >  The aim of this design is to provide a model which is purely host
> > +driven, requiring no co-operation from the software running in the
> > +guest, and is thus suitable for cloud scenarios.
> > +
> > +PV guests are out of scope for this project because, as is outlined
> above, they
> > +have a symbiotic relationship with the hypervisor and therefore a
> certain level
> > +of co-operation can be assumed.
> 
> If nothing else, I'd at least suggest s/can be assumed/is necessary/.

Ok. I'll make that modification.

> 
> > +Because the service domain’s domid is used directly by the guest in
> setting
> > +up grant entries and event channels, the backend drivers in the new
> host
> > +environment must be provided by service domain with the same domid.
> Also,
> > +because the guest can sample its own domid from the frontend area and
> use it in
> > +hypercalls (e.g. HVMOP_set_param) rather than DOMID_SELF, the guest
> domid must
> > +also be preserved to maintain the ABI.
> 
> Has this been true since forever?  The grant and event APIs took some
> care to avoid the guest needing to know its own domid.
> 

The guest doesn't need to know its domid; DOMID_SELF will work, but the guest *can* use its own domid in this case (whereas I think grant and event ops will insist on DOMID_SELF unless referring to another domain). As far as I know this has been the case since forever and so I don't think it is something we can change now unless we move to a new ABI.

> > +
> > +Furthermore, it will necessary to modify backend drivers to re-
> establish
> > +communication with frontend drivers without perturbing the content of
> the
> > +backend area or requiring any changes to the values of the xenstore
> state nodes.
> > +
> > +## Other Para-Virtual State
> > +
> > +### Shared Rings
> > +
> > +Because the console and store protocol shared pages are actually part
> of the
> > +guest memory image (in an E820 reserved region just below 4G)
> 
> Typically*.
> 
> Their exact location is entirely up to the domain builder, and tend not
> to be there for PVH guests which aren't trying to fit the two frames
> into a BAR.

Ok, I'll add the 'typically' in there. The exact detail is not that important.

> 
> > then the content
> > +will get migrated as part of the guest memory image. Hence no
> additional code
> > +is require to prevent any guest visible change in the content.
> 
> I do agree with this conclusion however.
> 

Good :-)

> > +### Shared Info
> > +
> > +There is already a record defined in *libxenctrl Domain Image Format*
> [3]
> > +called `SHARED_INFO` which simply contains a complete copy of the
> domain’s
> > +shared info page. It is not currently incuded in an HVM (type `0x0002`)
> > +migration stream. It may be feasible to include it as an optional
> record
> > +but it is not clear that the content of the shared info page ever needs
> > +to be preserved for an HVM guest.
> > +
> > +For a PV guest the `arch_shared_info` sub-structure contains important
> > +information about the guest’s P2M, but this information is not relevant
> for
> > +an HVM guest where the P2M is not directly manipulated via the guest.
> The other
> > +state contained in the `shared_info` structure relates the domain wall-
> clock
> > +(the state of which should already be transferred by the `RTC` HVM
> context
> > +information which contained in the `HVM_CONTEXT` save record) and some
> event
> > +channel state (particularly if using the *2l* protocol). Event channel
> state
> > +will need to be fully transferred if we are not going to require the
> guest
> > +co-operation to re-open the channels and so it should be possible to
> re-build a
> > +shared info page for an HVM guest from such other state.
> > +
> > +Note that the shared info page also contains an array of
> `XEN_LEGACY_MAX_VCPUS`
> > +(32) `vcpu_info` structures. A domain may nominate a different guest
> physical
> > +address to use for the vcpu info. This is mandatory for if a domain
> wants to
> > +use more than 32 vCPUs and optional for legacy vCPUs. This mapping is
> not
> > +currently transferred in the migration state so this will either need
> to be
> > +added into an existing save record, or an additional type of save
> record will
> > +be needed.
> 
> For non-cooperative migration in the current ABI, a minimum is to know
> where the shared info frame is mapped, so it can be re-mapped on behalf
> of the guest on the destination side.
> 

True, and the same for the grant tables, although it occurs to me that by turning these into domheap pages (as part of getting rid of shared xenheap pages... for other reasons) means the content should be migrated anyway, so we'll only need save records for the GFNs themselves.

> The rest of this section will be very good evidence in the "new guest
> ABI" design.
> 
> > +### Grant table
> > +
> > +The grant table is essentially the para-virtual equivalent of an IOMMU.
> 
> TBH, I think "shared memory" is a much better analogy than an IOMMU.
> OTOH, perhaps that doesn't cope with the grant copy aspect quite as well
> as I'd like.
> 

Well the table allows the guest to create a 'mapping' into a grant ref address space, and then those addresses are passed to 'PV devices', so the IOMMU analogy seemed most appropriate.

  Paul

> ~Andrew
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH v4 1/2] docs/designs: Add a design document for non-cooperative live migration

[Andrew Cooper]

On 30/01/2020 09:15, Durrant, Paul wrote:
>>> diff --git a/docs/designs/non-cooperative-migration.md
>> b/docs/designs/non-cooperative-migration.md
>>> new file mode 100644
>>> index 0000000000..5db3939db5
>>> --- /dev/null
>>> +++ b/docs/designs/non-cooperative-migration.md
>>> @@ -0,0 +1,272 @@
>>> +# Non-Cooperative Migration of Guests on Xen
>>> +
>>> +## Background
>>> +
>>> +The normal model of migration in Xen is driven by the guest because it
>> was
>>> +originally implemented for PV guests, where the guest must be aware it
>> is
>>> +running under Xen and is hence expected to co-operate.
>> For PV guests, is more than "expected to co-operate".
>>
>> Migrating a PV guest involves rewriting every pagetable entry with a
>> different MFN, so even before you consider things like the PV protocols,
>> there is no way this could be done without the cooperation of the guest.
> Yes, the P2M will change and this is visible to the guest, but does a PV guest need to take action when this occurs? I'm not sure.

Consider a non-cooperative migration happening to occur between e.g.
reading a PTE and looking it up in the M2P.

I'm sure there are other cases where chaos would reign.

>> The simple fact is that involving the guest kernel adds unnecessary
>> moving parts which can (and do with a non-zero probability) go wrong.
>>
> Yes, having written the frontend side of migration in the Windows drivers it is *very* hard to get right, particularly in Windows where one has to deal with the complex and asynchronous PnP subsystem colliding with a migration. The network driver also requires a multi-reader/single-writer lock with odd semantics (w.r.t. to IRQL) which I had to code myself (https://xenbits.xen.org/gitweb/?p=pvdrivers/win/xenvif.git;a=blob;f=src/xenvif/mrsw.h). It took years of fixing subtle races (in that and elsewhere) to get to the (AFAIK) reliable code we have now. 
> Avoiding execution of code like this (in all OS) certainly avoids the opportunity for subtle bugs to manifest themselves.

There are a number of things in the guest API/ABI which are particularly
poor in retrospect.  They all need to go.

>
>>> +Because the service domain’s domid is used directly by the guest in
>> setting
>>> +up grant entries and event channels, the backend drivers in the new
>> host
>>> +environment must be provided by service domain with the same domid.
>> Also,
>>> +because the guest can sample its own domid from the frontend area and
>> use it in
>>> +hypercalls (e.g. HVMOP_set_param) rather than DOMID_SELF, the guest
>> domid must
>>> +also be preserved to maintain the ABI.
>> Has this been true since forever?  The grant and event APIs took some
>> care to avoid the guest needing to know its own domid.
>>
> The guest doesn't need to know its domid; DOMID_SELF will work, but the guest *can* use its own domid in this case (whereas I think grant and event ops will insist on DOMID_SELF unless referring to another domain). As far as I know this has been the case since forever and so I don't think it is something we can change now unless we move to a new ABI.

That smells rather like stub-qemu reasons.

~Andrew

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

[Xen-devel] [PATCH v4 2/2] docs/designs: Add a design document for migration of xenstore data

[Paul Durrant]

This patch details proposes extra migration data and xenstore protocol
extensions to support non-cooperative live migration of guests.

Signed-off-by: Paul Durrant <pdurrant@amazon.com>
---
Cc: Andrew Cooper <andrew.cooper3@citrix.com>
Cc: George Dunlap <George.Dunlap@eu.citrix.com>
Cc: Ian Jackson <ian.jackson@eu.citrix.com>
Cc: Jan Beulich <jbeulich@suse.com>
Cc: Julien Grall <julien@xen.org>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Stefano Stabellini <sstabellini@kernel.org>
Cc: Wei Liu <wl@xen.org>

v4:
 - Drop the restrictions on special paths

v3:
 - New in v3
---
 docs/designs/xenstore-migration.md | 121 +++++++++++++++++++++++++++++
 1 file changed, 121 insertions(+)
 create mode 100644 docs/designs/xenstore-migration.md

diff --git a/docs/designs/xenstore-migration.md b/docs/designs/xenstore-migration.md
new file mode 100644
index 0000000000..991236e201
--- /dev/null
+++ b/docs/designs/xenstore-migration.md
@@ -0,0 +1,121 @@
+# Xenstore Migration
+
+## Background
+
+The design for *Non-Cooperative Migration of Guests*[1] explains that extra
+save records are required in the migrations stream to allow a guest running
+PV drivers to be migrated without its co-operation. Moreover the save
+records must include details of registered xenstore watches as well as
+content; information that cannot currently be recovered from `xenstored`,
+and hence some extension to the xenstore protocol[2] will also be required.
+
+The *libxenlight Domain Image Format* specification[3] already defines a
+record type `EMULATOR_XENSTORE_DATA` but this is not suitable for
+transferring xenstore data pertaining to the domain directly as it is
+specified such that keys are relative to the path
+`/local/domain/$dm_domid/device-model/$domid`. Thus it is necessary to
+define at least one new save record type.
+
+## Proposal
+
+### New Save Record
+
+A new mandatory record type should be defined within the libxenlight Domain
+Image Format:
+
+`0x00000007: DOMAIN_XENSTORE_DATA`
+
+The format of each of these new records should be as follows:
+
+
+```
+0     1     2     3     4     5     6     7 octet
++------------------------+------------------------+
+| type                   | record specific data   |
++------------------------+                        |
+...
++-------------------------------------------------+
+```
+
+
+| Field | Description |
+|---|---|
+| `type` | 0x00000000: invalid |
+|        | 0x00000001: node data |
+|        | 0x00000002: watch data |
+|        | 0x00000003 - 0xFFFFFFFF: reserved for future use |
+
+
+where data is always in the form of a NUL separated and terminated tuple
+as follows
+
+
+**node data**
+
+
+`<path>|<value>|<perm-as-string>|`
+
+
+`<path>` is considered relative to the domain path `/local/domain/$domid`
+and hence must not begin with `/`.
+`<path>` and `<value>` should be suitable to formulate a `WRITE` operation
+to the receiving xenstore and `<perm-as-string>` should be similarly suitable
+to formulate a subsequent `SET_PERMS` operation.
+
+**watch data**
+
+
+`<path>|<token>|`
+
+`<path>` again is considered relative and, together with `<token>`, should
+be suitable to formulate an `ADD_DOMAIN_WATCHES` operation (see below).
+
+
+### Protocol Extension
+
+The `WATCH` operation does not allow specification of a `<domid>`; it is
+assumed that the watch pertains to the domain that owns the shared ring
+over which the operation is passed. Hence, for the tool-stack to be able
+to register a watch on behalf of a domain a new operation is needed:
+
+```
+ADD_DOMAIN_WATCHES      <domid>|<watch>|+
+
+Adds watches on behalf of the specified domain.
+
+<watch> is a NUL separated tuple of <path>|<token>. The semantics of this
+operation are identical to the domain issuing WATCH <path>|<token>| for
+each <watch>.
+```
+
+The watch information for a domain also needs to be extracted from the
+sending xenstored so the following operation is also needed:
+
+```
+GET_DOMAIN_WATCHES      <domid>|<index>   <gencnt>|<watch>|* 
+
+Gets the list of watches that are currently registered for the domain.
+
+<watch> is a NUL separated tuple of <path>|<token>. The sub-list returned
+will start at <index> into the the overall list of watches and may be
+truncated such that the returned data fits within XENSTORE_PAYLOAD_MAX.
+If <index> is beyond the end of the overall list then the returned sub-
+list will be empty. If the value of <gencnt> changes then it indicates
+that the overall watch list has changed and thus it may be necessary
+to re-issue the operation for previous values of <index>.
+```
+
+It may also be desirable to state in the protocol specification that
+the `INTRODUCE` operation should not clear the `<mfn>` specified such that
+a `RELEASE` operation followed by an `INTRODUCE` operation form an
+idempotent pair. The current implementation of *C xentored* does this
+(in the `domain_conn_reset()` function) but this could be dropped as this
+behaviour is not currently specified and the page will always be zeroed
+for a newly created domain.
+
+
+* * *
+
+[1] See https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=docs/designs/non-cooperative-migration.md
+[2] See https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=docs/misc/xenstore.txt
+[3] See https://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=docs/specs/libxl-migration-stream.pandoc
-- 
2.20.1


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH v4 2/2] docs/designs: Add a design document for migration of xenstore data

[Marek Marczykowski-Górecki]

[-- Attachment #1.1: Type: text/plain, Size: 2866 bytes --]

On Wed, Jan 29, 2020 at 02:47:02PM +0000, Paul Durrant wrote:

<snip>

> +**node data**
> +
> +
> +`<path>|<value>|<perm-as-string>|`
> +
> +
> +`<path>` is considered relative to the domain path `/local/domain/$domid`
> +and hence must not begin with `/`.

How backend settings are going to be serialized? For example vif backend
has a bunch of feature-* entries, which should not change under the
guest feet during non-cooperative migration.

> +`<path>` and `<value>` should be suitable to formulate a `WRITE` operation
> +to the receiving xenstore and `<perm-as-string>` should be similarly suitable
> +to formulate a subsequent `SET_PERMS` operation.
> +
> +**watch data**
> +
> +
> +`<path>|<token>|`
> +
> +`<path>` again is considered relative and, together with `<token>`, should
> +be suitable to formulate an `ADD_DOMAIN_WATCHES` operation (see below).
> +
> +
> +### Protocol Extension
> +
> +The `WATCH` operation does not allow specification of a `<domid>`; it is
> +assumed that the watch pertains to the domain that owns the shared ring
> +over which the operation is passed. Hence, for the tool-stack to be able
> +to register a watch on behalf of a domain a new operation is needed:
> +
> +```
> +ADD_DOMAIN_WATCHES      <domid>|<watch>|+
> +
> +Adds watches on behalf of the specified domain.
> +
> +<watch> is a NUL separated tuple of <path>|<token>. The semantics of this
> +operation are identical to the domain issuing WATCH <path>|<token>| for
> +each <watch>.
> +```

Normal WATCH operation triggers an event immediately. Is it intended in
this case too? On the other hand, guest should cope with spurious watch
events, so probably not an issue.

> +The watch information for a domain also needs to be extracted from the
> +sending xenstored so the following operation is also needed:
> +
> +```
> +GET_DOMAIN_WATCHES      <domid>|<index>   <gencnt>|<watch>|* 
> +
> +Gets the list of watches that are currently registered for the domain.
> +
> +<watch> is a NUL separated tuple of <path>|<token>. The sub-list returned
> +will start at <index> into the the overall list of watches and may be
> +truncated such that the returned data fits within XENSTORE_PAYLOAD_MAX.
> +If <index> is beyond the end of the overall list then the returned sub-
> +list will be empty. If the value of <gencnt> changes then it indicates
> +that the overall watch list has changed and thus it may be necessary
> +to re-issue the operation for previous values of <index>.
> +```

In what units <index> is expressed? bytes? entries?
Can the response be truncated at arbitrary place, or only between
records?

-- 
Best Regards,
Marek Marczykowski-Górecki
Invisible Things Lab
A: Because it messes up the order in which people normally read text.
Q: Why is top-posting such a bad thing?

[-- Attachment #1.2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH v4 2/2] docs/designs: Add a design document for migration of xenstore data

[Durrant, Paul]

> -----Original Message-----
> From: Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
> Sent: 29 January 2020 21:23
> To: Durrant, Paul <pdurrant@amazon.co.uk>
> Cc: xen-devel@lists.xenproject.org; Stefano Stabellini
> <sstabellini@kernel.org>; Julien Grall <julien@xen.org>; Wei Liu
> <wl@xen.org>; Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>; George
> Dunlap <George.Dunlap@eu.citrix.com>; Andrew Cooper
> <andrew.cooper3@citrix.com>; Ian Jackson <ian.jackson@eu.citrix.com>
> Subject: Re: [Xen-devel] [PATCH v4 2/2] docs/designs: Add a design
> document for migration of xenstore data
> 
> On Wed, Jan 29, 2020 at 02:47:02PM +0000, Paul Durrant wrote:
> 
> <snip>
> 
> > +**node data**
> > +
> > +
> > +`<path>|<value>|<perm-as-string>|`
> > +
> > +
> > +`<path>` is considered relative to the domain path
> `/local/domain/$domid`
> > +and hence must not begin with `/`.
> 
> How backend settings are going to be serialized?

They're not going to be. The toolstack will construct new backends; co-operation will be required from them in that they must support re-binding (which the latest versions of blkback and netback do). We will need to white-list backends that are known to do this and refuse non-cooperative migration if any other backend is use.

> For example vif backend
> has a bunch of feature-* entries, which should not change under the
> guest feet during non-cooperative migration.
> 

The frontend will normally come back in 'connected' state, in which case a change in any feature flags will be irrelevant until the next (if any) re-connection (since the protocols only sample them at that point). If the frontend is not connected then it will sample the feature flags in the usual way. If the frontend/backend are in transition then the locking in the backend should prevent the 'unbind' from completing until some level of stability has been reached.

> > +`<path>` and `<value>` should be suitable to formulate a `WRITE`
> operation
> > +to the receiving xenstore and `<perm-as-string>` should be similarly
> suitable
> > +to formulate a subsequent `SET_PERMS` operation.
> > +
> > +**watch data**
> > +
> > +
> > +`<path>|<token>|`
> > +
> > +`<path>` again is considered relative and, together with `<token>`,
> should
> > +be suitable to formulate an `ADD_DOMAIN_WATCHES` operation (see below).
> > +
> > +
> > +### Protocol Extension
> > +
> > +The `WATCH` operation does not allow specification of a `<domid>`; it
> is
> > +assumed that the watch pertains to the domain that owns the shared ring
> > +over which the operation is passed. Hence, for the tool-stack to be
> able
> > +to register a watch on behalf of a domain a new operation is needed:
> > +
> > +```
> > +ADD_DOMAIN_WATCHES      <domid>|<watch>|+
> > +
> > +Adds watches on behalf of the specified domain.
> > +
> > +<watch> is a NUL separated tuple of <path>|<token>. The semantics of
> this
> > +operation are identical to the domain issuing WATCH <path>|<token>| for
> > +each <watch>.
> > +```
> 
> Normal WATCH operation triggers an event immediately. Is it intended in
> this case too? On the other hand, guest should cope with spurious watch
> events, so probably not an issue.

I don't think it matters if one is triggered or not. As you say, the watch protocol allows them to spuriously fire so the guest should cope either way.

> 
> > +The watch information for a domain also needs to be extracted from the
> > +sending xenstored so the following operation is also needed:
> > +
> > +```
> > +GET_DOMAIN_WATCHES      <domid>|<index>   <gencnt>|<watch>|*
> > +
> > +Gets the list of watches that are currently registered for the domain.
> > +
> > +<watch> is a NUL separated tuple of <path>|<token>. The sub-list
> returned
> > +will start at <index> into the the overall list of watches and may be
> > +truncated such that the returned data fits within XENSTORE_PAYLOAD_MAX.
> > +If <index> is beyond the end of the overall list then the returned sub-
> > +list will be empty. If the value of <gencnt> changes then it indicates
> > +that the overall watch list has changed and thus it may be necessary
> > +to re-issue the operation for previous values of <index>.
> > +```
> 
> In what units <index> is expressed? bytes? entries?

I thought entries was reasonably well implied since I refer to a list, but I can make it explicit.

> Can the response be truncated at arbitrary place, or only between
> records?
> 

Only between records. Again I'll add some words to make that clear.

  Paul

> --
> Best Regards,
> Marek Marczykowski-Górecki
> Invisible Things Lab
> A: Because it messes up the order in which people normally read text.
> Q: Why is top-posting such a bad thing?
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel

Re: [Xen-devel] [PATCH v4 2/2] docs/designs: Add a design document for migration of xenstore data

[Marek Marczykowski-Górecki]

[-- Attachment #1.1: Type: text/plain, Size: 2667 bytes --]

On Thu, Jan 30, 2020 at 08:45:49AM +0000, Durrant, Paul wrote:
> > -----Original Message-----
> > From: Marek Marczykowski-Górecki <marmarek@invisiblethingslab.com>
> > Sent: 29 January 2020 21:23
> > To: Durrant, Paul <pdurrant@amazon.co.uk>
> > Cc: xen-devel@lists.xenproject.org; Stefano Stabellini
> > <sstabellini@kernel.org>; Julien Grall <julien@xen.org>; Wei Liu
> > <wl@xen.org>; Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>; George
> > Dunlap <George.Dunlap@eu.citrix.com>; Andrew Cooper
> > <andrew.cooper3@citrix.com>; Ian Jackson <ian.jackson@eu.citrix.com>
> > Subject: Re: [Xen-devel] [PATCH v4 2/2] docs/designs: Add a design
> > document for migration of xenstore data
> > 
> > On Wed, Jan 29, 2020 at 02:47:02PM +0000, Paul Durrant wrote:
> > 
> > <snip>
> > 
> > > +**node data**
> > > +
> > > +
> > > +`<path>|<value>|<perm-as-string>|`
> > > +
> > > +
> > > +`<path>` is considered relative to the domain path
> > `/local/domain/$domid`
> > > +and hence must not begin with `/`.
> > 
> > How backend settings are going to be serialized?
> 
> They're not going to be. The toolstack will construct new backends; co-operation will be required from them in that they must support re-binding (which the latest versions of blkback and netback do). We will need to white-list backends that are known to do this and refuse non-cooperative migration if any other backend is use.
> 
> > For example vif backend
> > has a bunch of feature-* entries, which should not change under the
> > guest feet during non-cooperative migration.
> > 
> 
> The frontend will normally come back in 'connected' state, in which case a change in any feature flags will be irrelevant until the next (if any) re-connection (since the protocols only sample them at that point). If the frontend is not connected then it will sample the feature flags in the usual way. If the frontend/backend are in transition then the locking in the backend should prevent the 'unbind' from completing until some level of stability has been reached.

Yes, but regardless of xenstore entries, those do represent what backend
can and is willing to do. If some feature disappears (in actual backend,
not only its xenstore representation) during non-cooperative migration,
I guess frontent won't be happy about that.

How are they related to feature-* entries in frontent? Are frontend
features separate set, or a confirmation what backend's features will be
used?

-- 
Best Regards,
Marek Marczykowski-Górecki
Invisible Things Lab
A: Because it messes up the order in which people normally read text.
Q: Why is top-posting such a bad thing?

[-- Attachment #1.2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

[-- Attachment #2: Type: text/plain, Size: 157 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel