Re: [Qemu-block] [PATCH RFC] block: add block-insert-node QMP command

[Manos Pitsidianakis]

On Fri, Oct 06, 2017 at 02:59:55PM +0200, Max Reitz wrote:

On 2017-10-04 23:04, Manos Pitsidianakis wrote:

On Wed, Oct 04, 2017 at 08:09:24PM +0200, Max Reitz wrote:

On 2017-10-04 19:05, Manos Pitsidianakis wrote:

On Wed, Oct 04, 2017 at 02:49:27PM +0200, Max Reitz wrote:

On 2017-08-15 09:45, Manos Pitsidianakis wrote:

block-insert-node and its pair command block-remove-node provide
runtime
insertion and removal of filter nodes.

block-insert-node takes a (parent, child) and (node, child) pair of
edges and unrefs the (parent, child) BdrvChild relationship and
creates
a new (parent, node) child with the same BdrvChildRole.

This is a different approach than x-blockdev-change which uses the
driver
methods bdrv_add_child() and bdrv_del_child(),


Why? :-)

Can't we reuse x-blockdev-change? As far as I'm concerned, this was one
of its roles, and at least I don't want to have both x-blockdev-change
and these new commands.

I think it would be a good idea just to change bdrv_add_child() and
bdrv_del_child(): If the driver has a bdrv_{add,del}_child() callback,
invoke that.  If it doesn't, then just attach the child.

Of course, it may turn out that x-blockdev-change is lacking something
(e.g. a way to specify as what kind of child a new node is to be
attached).  Then we should either extend it so that it covers what it's
lacking, or replace it completely by these new commands.  In the latter
case, however, they would need to cover the existing use case which is
reconfiguring the quorum driver.  (And that would mean it would have to
invoke bdrv_add_child()/bdrv_del_child() when the driver has them.)

Max



I think the two use cases are this:

a) Adding/removing a replication child to an existing quorum node
b) Insert a filter between two existing nodes.


For me both are the same: Adding/removing nodes into/from the graph.

The difference is how those children are added (or removed, but it's the
same in reverse): In case of quorum, you can simply attach a new child
because it supports a variable number of children.  Another case where
this would work are all block drivers which support backing files (you
can freely attach/detach those).


Doesn't blockdev-snapshot-sync cover this? (I may be missing something).


Not really, because it doesn't add a new child.  But it's still a good
point, because your block-insert-node command would make it obsolete,
actually.

blockdev-snapshot literally is a special-cased block-insert-node.  And
blockdev-snapshot-sync then is qemu-img create + blockdev-add +
blockdev-snapshot.


Now that we're on this topic, quorum might be a good candidate for
*_reopen when and if it lands on QMP: Reconfiguring the children could
be done by reopening the BDS with new options.


Hmmm...  I guess that would work, too.  But intuitively, that seems a
bit heavy-weight to me


In this series, it's not about adding or removing new children, but
instead "injecting" them into an edge: An existing child is replaced,
but it now serves as some child of the new one.

(I guess writing too much trying to get my point across, sorry...)

The gist is that for me it's not so much about "quorum" or "filter
nodes".  It's about adding a new child to a node vs. replacing an
existing one.


These are not directly compatible semantically, but as you said
x-blockdev-change can perform b) if bdrv_add_child()/bdrv_del_child()
are not implemented. Wouldn't that be unnecessary overloading?


Yes, I think it would be. :-)

So say we have these two trees in our graph:

[ Parent BDS -> Child BDS ]
[ Filter BDS -> Child BDS ]

So here's what you can do with x-blockdev-change (in theory; in practice
it can only do that for quorum):
- Remove a child, so
   [ Parent BDS -> Child BDS ]
 is split into two trees
   [ Parent BDS ] and
   [ Child BDS ]
- Add a child, so we can merge
   [ Parent BDS ] and
   [ Filter BDS -> Child BDS ]
 into
   [ Parent BDS -> Filter BDS -> Child BDS ]



Yes, of course this would have to be done in one transaction.


Would it?  If you want to put Filter BDS into the chain, you can just
create [ Filter BDS ] without any child, and then use a single
x-blockdev-change invocation to inject it.

The only thing that's necessary is that the filter BDS would have to be
able to handle having no child.


Yeah that was what I had in mind.


[...]


So after I've written all of this, I feel like the new insert-node and
remove-node commands are exactly what x-blockdev-change should do when
asked to replace a node.  The only difference is that x-blockdev-change
would allow you to replace any node with anything, without the
constraints that block-insert-node and block-remove-node exact.

(And node replacement with x-blockdev-change would work by specifying
all three parameters.)

Not sure if that makes sense, I hope it does. :-)



Hm, I can't think of a way to fit that into x-blockdev-change *and* keep
the bdrv_add_child/bdrv_del_child functionality into consideration
(since we'd have to keep both). This 

Re: [Qemu-block] [PATCH RFC] block: add block-insert-node QMP command

[Max Reitz]

On 2017-10-04 23:04, Manos Pitsidianakis wrote:
> On Wed, Oct 04, 2017 at 08:09:24PM +0200, Max Reitz wrote:
>> On 2017-10-04 19:05, Manos Pitsidianakis wrote:
>>> On Wed, Oct 04, 2017 at 02:49:27PM +0200, Max Reitz wrote:
 On 2017-08-15 09:45, Manos Pitsidianakis wrote:
> block-insert-node and its pair command block-remove-node provide
> runtime
> insertion and removal of filter nodes.
>
> block-insert-node takes a (parent, child) and (node, child) pair of
> edges and unrefs the (parent, child) BdrvChild relationship and
> creates
> a new (parent, node) child with the same BdrvChildRole.
>
> This is a different approach than x-blockdev-change which uses the
> driver
> methods bdrv_add_child() and bdrv_del_child(),

 Why? :-)

 Can't we reuse x-blockdev-change? As far as I'm concerned, this was one
 of its roles, and at least I don't want to have both x-blockdev-change
 and these new commands.

 I think it would be a good idea just to change bdrv_add_child() and
 bdrv_del_child(): If the driver has a bdrv_{add,del}_child() callback,
 invoke that.  If it doesn't, then just attach the child.

 Of course, it may turn out that x-blockdev-change is lacking something
 (e.g. a way to specify as what kind of child a new node is to be
 attached).  Then we should either extend it so that it covers what it's
 lacking, or replace it completely by these new commands.  In the latter
 case, however, they would need to cover the existing use case which is
 reconfiguring the quorum driver.  (And that would mean it would have to
 invoke bdrv_add_child()/bdrv_del_child() when the driver has them.)

 Max

>>>
>>> I think the two use cases are this:
>>>
>>> a) Adding/removing a replication child to an existing quorum node
>>> b) Insert a filter between two existing nodes.
>>
>> For me both are the same: Adding/removing nodes into/from the graph.
>>
>> The difference is how those children are added (or removed, but it's the
>> same in reverse): In case of quorum, you can simply attach a new child
>> because it supports a variable number of children.  Another case where
>> this would work are all block drivers which support backing files (you
>> can freely attach/detach those).
> 
> Doesn't blockdev-snapshot-sync cover this? (I may be missing something).

Not really, because it doesn't add a new child.  But it's still a good
point, because your block-insert-node command would make it obsolete,
actually.

blockdev-snapshot literally is a special-cased block-insert-node.  And
blockdev-snapshot-sync then is qemu-img create + blockdev-add +
blockdev-snapshot.

> Now that we're on this topic, quorum might be a good candidate for
> *_reopen when and if it lands on QMP: Reconfiguring the children could
> be done by reopening the BDS with new options.

Hmmm...  I guess that would work, too.  But intuitively, that seems a
bit heavy-weight to me

>> In this series, it's not about adding or removing new children, but
>> instead "injecting" them into an edge: An existing child is replaced,
>> but it now serves as some child of the new one.
>>
>> (I guess writing too much trying to get my point across, sorry...)
>>
>> The gist is that for me it's not so much about "quorum" or "filter
>> nodes".  It's about adding a new child to a node vs. replacing an
>> existing one.
>>
>>> These are not directly compatible semantically, but as you said
>>> x-blockdev-change can perform b) if bdrv_add_child()/bdrv_del_child()
>>> are not implemented. Wouldn't that be unnecessary overloading?
>>
>> Yes, I think it would be. :-)
>>
>> So say we have these two trees in our graph:
>>
>> [ Parent BDS -> Child BDS ]
>> [ Filter BDS -> Child BDS ]
>>
>> So here's what you can do with x-blockdev-change (in theory; in practice
>> it can only do that for quorum):
>> - Remove a child, so
>>    [ Parent BDS -> Child BDS ]
>>  is split into two trees
>>    [ Parent BDS ] and
>>    [ Child BDS ]
>> - Add a child, so we can merge
>>    [ Parent BDS ] and
>>    [ Filter BDS -> Child BDS ]
>>  into
>>    [ Parent BDS -> Filter BDS -> Child BDS ]
>>
> 
> Yes, of course this would have to be done in one transaction.

Would it?  If you want to put Filter BDS into the chain, you can just
create [ Filter BDS ] without any child, and then use a single
x-blockdev-change invocation to inject it.

The only thing that's necessary is that the filter BDS would have to be
able to handle having no child.

[...]

>> So after I've written all of this, I feel like the new insert-node and
>> remove-node commands are exactly what x-blockdev-change should do when
>> asked to replace a node.  The only difference is that x-blockdev-change
>> would allow you to replace any node with anything, without the
>> constraints that block-insert-node and block-remove-node exact.
>>
>> (And node replacement with x-blockdev-change would work by specifying
>> all three 

Re: [Qemu-block] [PATCH RFC] block: add block-insert-node QMP command

[Manos Pitsidianakis]

On Wed, Oct 04, 2017 at 08:09:24PM +0200, Max Reitz wrote:

On 2017-10-04 19:05, Manos Pitsidianakis wrote:

On Wed, Oct 04, 2017 at 02:49:27PM +0200, Max Reitz wrote:

On 2017-08-15 09:45, Manos Pitsidianakis wrote:

block-insert-node and its pair command block-remove-node provide runtime
insertion and removal of filter nodes.

block-insert-node takes a (parent, child) and (node, child) pair of
edges and unrefs the (parent, child) BdrvChild relationship and creates
a new (parent, node) child with the same BdrvChildRole.

This is a different approach than x-blockdev-change which uses the
driver
methods bdrv_add_child() and bdrv_del_child(),


Why? :-)

Can't we reuse x-blockdev-change? As far as I'm concerned, this was one
of its roles, and at least I don't want to have both x-blockdev-change
and these new commands.

I think it would be a good idea just to change bdrv_add_child() and
bdrv_del_child(): If the driver has a bdrv_{add,del}_child() callback,
invoke that.  If it doesn't, then just attach the child.

Of course, it may turn out that x-blockdev-change is lacking something
(e.g. a way to specify as what kind of child a new node is to be
attached).  Then we should either extend it so that it covers what it's
lacking, or replace it completely by these new commands.  In the latter
case, however, they would need to cover the existing use case which is
reconfiguring the quorum driver.  (And that would mean it would have to
invoke bdrv_add_child()/bdrv_del_child() when the driver has them.)

Max



I think the two use cases are this:

a) Adding/removing a replication child to an existing quorum node
b) Insert a filter between two existing nodes.


For me both are the same: Adding/removing nodes into/from the graph.

The difference is how those children are added (or removed, but it's the
same in reverse): In case of quorum, you can simply attach a new child
because it supports a variable number of children.  Another case where
this would work are all block drivers which support backing files (you
can freely attach/detach those).


Doesn't blockdev-snapshot-sync cover this? (I may be missing something).
Now that we're on this topic, quorum might be a good candidate for 
*_reopen when and if it lands on QMP: Reconfiguring the children could 
be done by reopening the BDS with new options.


In this series, it's not about adding or removing new children, but
instead "injecting" them into an edge: An existing child is replaced,
but it now serves as some child of the new one.

(I guess writing too much trying to get my point across, sorry...)

The gist is that for me it's not so much about "quorum" or "filter
nodes".  It's about adding a new child to a node vs. replacing an
existing one.


These are not directly compatible semantically, but as you said
x-blockdev-change can perform b) if bdrv_add_child()/bdrv_del_child()
are not implemented. Wouldn't that be unnecessary overloading?


Yes, I think it would be. :-)

So say we have these two trees in our graph:

[ Parent BDS -> Child BDS ]
[ Filter BDS -> Child BDS ]

So here's what you can do with x-blockdev-change (in theory; in practice
it can only do that for quorum):
- Remove a child, so
   [ Parent BDS -> Child BDS ]
 is split into two trees
   [ Parent BDS ] and
   [ Child BDS ]
- Add a child, so we can merge
   [ Parent BDS ] and
   [ Filter BDS -> Child BDS ]
 into
   [ Parent BDS -> Filter BDS -> Child BDS ]



Yes, of course this would have to be done in one transaction.


However, this is only possible with quorum because usually block drivers
don't support detaching children.

And here's what you can do with your commands (from what I can see):
- Replace a child (you call it insertion, but it really is just
 replacement of an existing child with the constraint that both nodes
 involved must have the same child):
   [ Parent BDS -> Child BDS ]
   [ Filter BDS -> Child BDS ]
 to
   [ Parent BDS -> Filter BDS -> Child BDS ]
- Replace a child (you call it removal, but it really is just
 replacement of a child with its child):
   [ Parent BDS -> Filter BDS -> Child BDS ]
 to
   [ Parent BDS -> Child BDS ]

This works on all BDSs because you don't change the number of children.


The interesting thing of course is that the "change" command can
actually add and remove children; where as the "insert" and "remove"
commands can only replace children.  So that's already a bit funny (one
command does two things; two commands do one thing).


That is true, but the replacing is more in terms of inserting and 
removing a node in a BDS chain.


And then of course you can simply modify x-blockdev-change so it can do
the same thing block-insert-node and block-remove-node can do: It just
needs another mode which can be used to replace a child (and its
description already states that it is supposed to be usable for that at
some point in the future).


So after I've written all of this, I feel like the new insert-node and
remove-node commands are exactly what 

Re: [Qemu-block] [PATCH RFC] block: add block-insert-node QMP command

[Max Reitz]

On 2017-10-04 19:05, Manos Pitsidianakis wrote:
> On Wed, Oct 04, 2017 at 02:49:27PM +0200, Max Reitz wrote:
>> On 2017-08-15 09:45, Manos Pitsidianakis wrote:
>>> block-insert-node and its pair command block-remove-node provide runtime
>>> insertion and removal of filter nodes.
>>>
>>> block-insert-node takes a (parent, child) and (node, child) pair of
>>> edges and unrefs the (parent, child) BdrvChild relationship and creates
>>> a new (parent, node) child with the same BdrvChildRole.
>>>
>>> This is a different approach than x-blockdev-change which uses the
>>> driver
>>> methods bdrv_add_child() and bdrv_del_child(),
>>
>> Why? :-)
>>
>> Can't we reuse x-blockdev-change? As far as I'm concerned, this was one
>> of its roles, and at least I don't want to have both x-blockdev-change
>> and these new commands.
>>
>> I think it would be a good idea just to change bdrv_add_child() and
>> bdrv_del_child(): If the driver has a bdrv_{add,del}_child() callback,
>> invoke that.  If it doesn't, then just attach the child.
>>
>> Of course, it may turn out that x-blockdev-change is lacking something
>> (e.g. a way to specify as what kind of child a new node is to be
>> attached).  Then we should either extend it so that it covers what it's
>> lacking, or replace it completely by these new commands.  In the latter
>> case, however, they would need to cover the existing use case which is
>> reconfiguring the quorum driver.  (And that would mean it would have to
>> invoke bdrv_add_child()/bdrv_del_child() when the driver has them.)
>>
>> Max
>>
> 
> I think the two use cases are this:
> 
> a) Adding/removing a replication child to an existing quorum node
> b) Insert a filter between two existing nodes.

For me both are the same: Adding/removing nodes into/from the graph.

The difference is how those children are added (or removed, but it's the
same in reverse): In case of quorum, you can simply attach a new child
because it supports a variable number of children.  Another case where
this would work are all block drivers which support backing files (you
can freely attach/detach those).

In this series, it's not about adding or removing new children, but
instead "injecting" them into an edge: An existing child is replaced,
but it now serves as some child of the new one.

(I guess writing too much trying to get my point across, sorry...)

The gist is that for me it's not so much about "quorum" or "filter
nodes".  It's about adding a new child to a node vs. replacing an
existing one.

> These are not directly compatible semantically, but as you said
> x-blockdev-change can perform b) if bdrv_add_child()/bdrv_del_child()
> are not implemented. Wouldn't that be unnecessary overloading?

Yes, I think it would be. :-)

So say we have these two trees in our graph:

[ Parent BDS -> Child BDS ]
[ Filter BDS -> Child BDS ]

So here's what you can do with x-blockdev-change (in theory; in practice
it can only do that for quorum):
- Remove a child, so
[ Parent BDS -> Child BDS ]
  is split into two trees
[ Parent BDS ] and
[ Child BDS ]
- Add a child, so we can merge
[ Parent BDS ] and
[ Filter BDS -> Child BDS ]
  into
[ Parent BDS -> Filter BDS -> Child BDS ]

However, this is only possible with quorum because usually block drivers
don't support detaching children.

And here's what you can do with your commands (from what I can see):
- Replace a child (you call it insertion, but it really is just
  replacement of an existing child with the constraint that both nodes
  involved must have the same child):
[ Parent BDS -> Child BDS ]
[ Filter BDS -> Child BDS ]
  to
[ Parent BDS -> Filter BDS -> Child BDS ]
- Replace a child (you call it removal, but it really is just
  replacement of a child with its child):
[ Parent BDS -> Filter BDS -> Child BDS ]
  to
[ Parent BDS -> Child BDS ]

This works on all BDSs because you don't change the number of children.


The interesting thing of course is that the "change" command can
actually add and remove children; where as the "insert" and "remove"
commands can only replace children.  So that's already a bit funny (one
command does two things; two commands do one thing).

And then of course you can simply modify x-blockdev-change so it can do
the same thing block-insert-node and block-remove-node can do: It just
needs another mode which can be used to replace a child (and its
description already states that it is supposed to be usable for that at
some point in the future).


So after I've written all of this, I feel like the new insert-node and
remove-node commands are exactly what x-blockdev-change should do when
asked to replace a node.  The only difference is that x-blockdev-change
would allow you to replace any node with anything, without the
constraints that block-insert-node and block-remove-node exact.

(And node replacement with x-blockdev-change would work by specifying
all three parameters.)

Not sure if that makes sense, I hope it does. 

Re: [Qemu-block] [PATCH RFC] block: add block-insert-node QMP command

[Manos Pitsidianakis]

On Wed, Oct 04, 2017 at 02:49:27PM +0200, Max Reitz wrote:

On 2017-08-15 09:45, Manos Pitsidianakis wrote:

block-insert-node and its pair command block-remove-node provide runtime
insertion and removal of filter nodes.

block-insert-node takes a (parent, child) and (node, child) pair of
edges and unrefs the (parent, child) BdrvChild relationship and 
creates

a new (parent, node) child with the same BdrvChildRole.

This is a different approach than x-blockdev-change which uses the driver
methods bdrv_add_child() and bdrv_del_child(),


Why? :-)

Can't we reuse x-blockdev-change? As far as I'm concerned, this was one
of its roles, and at least I don't want to have both x-blockdev-change
and these new commands.

I think it would be a good idea just to change bdrv_add_child() and
bdrv_del_child(): If the driver has a bdrv_{add,del}_child() callback,
invoke that.  If it doesn't, then just attach the child.

Of course, it may turn out that x-blockdev-change is lacking something
(e.g. a way to specify as what kind of child a new node is to be
attached).  Then we should either extend it so that it covers what it's
lacking, or replace it completely by these new commands.  In the latter
case, however, they would need to cover the existing use case which is
reconfiguring the quorum driver.  (And that would mean it would have to
invoke bdrv_add_child()/bdrv_del_child() when the driver has them.)

Max



I think the two use cases are this:

a) Adding/removing a replication child to an existing quorum node
b) Insert a filter between two existing nodes.

These are not directly compatible semantically, but as you said 
x-blockdev-change can perform b) if bdrv_add_child()/bdrv_del_child() 
are not implemented. Wouldn't that be unnecessary overloading?


signature.asc
Description: PGP signature

Re: [Qemu-block] [PATCH RFC] block: add block-insert-node QMP command

[Kevin Wolf]

Am 04.10.2017 um 14:23 hat Manos Pitsidianakis geschrieben:
> > > diff --git a/block.c b/block.c
> > > index 81bd51b670..f874aabbfb 100644
> > > --- a/block.c
> > > +++ b/block.c
> > > +/* insert 'node' as child bs of 'parent' node */
> > > +if (check_node_edge(parent, child, errp)) {
> > > +return;
> > > +}
> > > +parent_bs = bdrv_find_node(parent);
> > > +c = bdrv_find_child(parent_bs, child);
> > > +role = c->role;
> > > +assert(role == _file || role == _backing);
> > > +
> > > +bdrv_ref(node_bs);
> > > +
> > > +bdrv_drained_begin(parent_bs);
> > > +bdrv_unref_child(parent_bs, c);
> > > +if (role == _file) {
> > > +parent_bs->file = bdrv_attach_child(parent_bs, node_bs, "file",
> > > +_file, errp);
> > > +if (!parent_bs->file) {
> > > +parent_bs->file = bdrv_attach_child(parent_bs, child_bs, 
> > > "file",
> > > +_file, 
> > > _abort);
> > > +goto out;
> > > +}
> > > +} else if (role == _backing) {
> > > +parent_bs->backing = bdrv_attach_child(parent_bs, node_bs, 
> > > "backing",
> > > +   _backing, errp);
> > > +if (!parent_bs->backing) {
> > > +parent_bs->backing = bdrv_attach_child(parent_bs, child_bs,
> > > +   "backing", 
> > > _backing,
> > > +   _abort);
> > > +goto out;
> > > +}
> > > +}
> > 
> > I would prefer if we could find a solution to avoid requiring a specific
> > role. I'm not even sure that your assertion above is correct; can you
> > explain why c couldn't have any other role?
> > 
> > Instead of bdrv_unref_child/bdrv_attach_child, could we just change
> > where the child points to using bdrv_replace_child()? Then
> 
> bdrv_replace_child() uses bdrv_set_perm() and co. When I tried it at first I
> got errors like "Conflicts with use by ** as 'backing', which does not
> allow 'write' on disk". Presumably the permissions do not need to change but
> can we do bdrv_set_perm without bdrv_check_perm?

Which child is conflicting with which other child? Is c conflicting with
itself or something?

If unref_child/attach_child works without any other action in between,
there is no reason why replace_child shouldn't work, too. Maybe this is
a bug in bdrv_

> > parent_bs->file and parent_bs->backing (or whatever other variable
> > contains the BdrvChild pointer) can stay unchanged and just keep
> > working.
> > 
> > > +bdrv_refresh_filename(parent_bs);
> > > +bdrv_refresh_limits(parent_bs, NULL);
> > > +
> > > +out:
> > > +bdrv_drained_end(parent_bs);
> > > +}
> > 
> > > diff --git a/blockdev.c b/blockdev.c
> > > index 8e2fc6e64c..5195ec1b61 100644
> > > --- a/blockdev.c
> > > +++ b/blockdev.c
> > > @@ -4238,3 +4238,47 @@ QemuOptsList qemu_drive_opts = {
> > >  { /* end of list */ }
> > >  },
> > >  };
> > > +
> > > +void qmp_block_insert_node(const char *parent, const char *child,
> > > +   const char *node, Error **errp)
> > > +{
> > > +BlockDriverState *bs = bdrv_find_node(node);
> > > +if (!bs) {
> > > +error_setg(errp, "Node '%s' not found", node);
> > > +return;
> > > +}
> > > +if (!bs->monitor_list.tqe_prev) {
> > > +error_setg(errp, "Node '%s' is not owned by the monitor",
> > > +   bs->node_name);
> > > +return;
> > > +}
> > > +if (!bs->drv->is_filter) {
> > > +error_setg(errp, "Block format '%s' used by node '%s' does not 
> > > support"
> > > +   "insertion", bs->drv->format_name, bs->node_name);
> > > +return;
> > > +}
> > > +
> > > +bdrv_insert_node(parent, child, node, errp);
> > > +}
> > 
> > Do we need to acquire an AioContext lock somewhere?
> 
> the *_child() functions call drained_begin/end which I think might cover
> this case?

I don't think it's enough when you don't own the AioContext lock.

Kevin


signature.asc
Description: PGP signature

Re: [Qemu-block] [PATCH RFC] block: add block-insert-node QMP command

[Max Reitz]

On 2017-08-15 09:45, Manos Pitsidianakis wrote:
> block-insert-node and its pair command block-remove-node provide runtime
> insertion and removal of filter nodes.
> 
> block-insert-node takes a (parent, child) and (node, child) pair of
> edges and unrefs the (parent, child) BdrvChild relationship and creates
> a new (parent, node) child with the same BdrvChildRole.
> 
> This is a different approach than x-blockdev-change which uses the driver
> methods bdrv_add_child() and bdrv_del_child(),

Why? :-)

Can't we reuse x-blockdev-change? As far as I'm concerned, this was one
of its roles, and at least I don't want to have both x-blockdev-change
and these new commands.

I think it would be a good idea just to change bdrv_add_child() and
bdrv_del_child(): If the driver has a bdrv_{add,del}_child() callback,
invoke that.  If it doesn't, then just attach the child.

Of course, it may turn out that x-blockdev-change is lacking something
(e.g. a way to specify as what kind of child a new node is to be
attached).  Then we should either extend it so that it covers what it's
lacking, or replace it completely by these new commands.  In the latter
case, however, they would need to cover the existing use case which is
reconfiguring the quorum driver.  (And that would mean it would have to
invoke bdrv_add_child()/bdrv_del_child() when the driver has them.)

Max



signature.asc
Description: OpenPGP digital signature

Re: [Qemu-block] [PATCH RFC] block: add block-insert-node QMP command

[Manos Pitsidianakis]

On Fri, Sep 29, 2017 at 07:52:35PM +0200, Kevin Wolf wrote:

Am 15.08.2017 um 09:45 hat Manos Pitsidianakis geschrieben:

block-insert-node and its pair command block-remove-node provide runtime
insertion and removal of filter nodes.

block-insert-node takes a (parent, child) and (node, child) pair of
edges and unrefs the (parent, child) BdrvChild relationship and creates
a new (parent, node) child with the same BdrvChildRole.

This is a different approach than x-blockdev-change which uses the driver
methods bdrv_add_child() and bdrv_del_child(),

Signed-off-by: Manos Pitsidianakis 



diff --git a/qapi/block-core.json b/qapi/block-core.json
index 4d6ba1baef..16e19cb648 100644
--- a/qapi/block-core.json
+++ b/qapi/block-core.json
@@ -3947,3 +3947,63 @@
   'data' : { 'parent': 'str',
  '*child': 'str',
  '*node': 'str' } }
+
+##
+# @block-insert-node:
+#
+# Insert a filter node between a specific edge in the block driver state graph.
+# @parent:  the name of the parent node or device
+# @node:the name of the node to insert under parent
+# @child:   the name of the child of both node and parent
+#
+# Example:
+# Insert and remove a throttle filter on top of a device chain, between the
+# device 'ide0-hd0' and node 'node-A'
+#
+# -> {'execute': 'object-add',
+# "arguments": {
+#   "qom-type": "throttle-group",
+#   "id": "group0",
+#   "props" : { "limits": { "iops-total": 300 } }
+# }
+#}
+# <- { 'return': {} }
+# -> {'execute': 'blockdev-add',
+#   'arguments': {
+#   'driver': 'throttle',
+#   'node-name': 'throttle0',
+#   'throttle-group': 'group0',
+#   'file': 'node-A'
+#   }
+#}
+# <- { 'return': {} }
+# -> { 'execute': 'block-insert-node',
+#   'arguments': { 'parent': 'ide0-hd0', 'child': 'node-A', 'node': 
'throttle0' }
+#}
+# <- { 'return': {} }
+# -> { 'execute': 'block-remove-node',
+#   'arguments': { 'parent': 'ide0-hd0', 'child': 'node-A', 'node': 
'throttle0' }
+#}
+# <- { 'return': {} }
+# -> { 'execute': 'blockdev-del',
+#   'arguments': { 'node-name': 'throttle0' }
+#}
+# <- { 'return': {} }
+#
+##
+{ 'command': 'block-insert-node',
+  'data': { 'parent': 'str',
+ 'child': 'str',
+ 'node': 'str'} }


I would suggest a change to the meaning of @child: Instead of using the
node-name of the child BDS, I would use the name of the BdrvChild that
represents the link.

The reason for this is that the node-name could be ambiguous, if you
have two edges between the same two nodes.

The only use of the node-name of the child that I can remember was for
checking that the graph still looks like what the user expects. But I
think we came to the conclusion that there are no race conditions to
check for if we have manual block job deletion instead of automatic
completion which can involve surprise changes to the graph. So probably
we don't need the node-name even for this.


+##
+# @block-remove-node:
+#
+# Remove a filter node between two other nodes in the block driver state graph.
+# @parent:  the name of the parent node or device
+# @node:the name of the node to remove from parent
+# @child:   the name of the child of node which will go under parent
+##
+{ 'command': 'block-remove-node',
+  'data': { 'parent': 'str',
+ 'child': 'str',
+ 'node': 'str'} }


Same thing here.


diff --git a/block.c b/block.c
index 81bd51b670..f874aabbfb 100644
--- a/block.c
+++ b/block.c
+/* insert 'node' as child bs of 'parent' node */
+if (check_node_edge(parent, child, errp)) {
+return;
+}
+parent_bs = bdrv_find_node(parent);
+c = bdrv_find_child(parent_bs, child);
+role = c->role;
+assert(role == _file || role == _backing);
+
+bdrv_ref(node_bs);
+
+bdrv_drained_begin(parent_bs);
+bdrv_unref_child(parent_bs, c);
+if (role == _file) {
+parent_bs->file = bdrv_attach_child(parent_bs, node_bs, "file",
+_file, errp);
+if (!parent_bs->file) {
+parent_bs->file = bdrv_attach_child(parent_bs, child_bs, "file",
+_file, _abort);
+goto out;
+}
+} else if (role == _backing) {
+parent_bs->backing = bdrv_attach_child(parent_bs, node_bs, "backing",
+   _backing, errp);
+if (!parent_bs->backing) {
+parent_bs->backing = bdrv_attach_child(parent_bs, child_bs,
+   "backing", _backing,
+   _abort);
+goto out;
+}
+}


I would prefer if we could find a solution to avoid requiring a specific
role. I'm not even sure that your assertion above is correct; can you
explain why c couldn't have any other role?

Instead of bdrv_unref_child/bdrv_attach_child, could we just