Re: [PATCH v4 08/25] block: introduce assert_bdrv_graph_writable

[Emanuele Giuseppe Esposito <eesposit@redhat.com>]

On 18/11/2021 16:17, Hanna Reitz wrote:
> On 18.11.21 10:55, Emanuele Giuseppe Esposito wrote:
>>
>> On 12/11/2021 15:40, Hanna Reitz wrote:
>>> On 25.10.21 12:17, Emanuele Giuseppe Esposito wrote:
>>>> We want to be sure that the functions that write the child and
>>>> parent list of a bs are under BQL and drain.
>>>>
>>>> BQL prevents from concurrent writings from the GS API, while
>>>> drains protect from I/O.
>>>>
>>>> TODO: drains are missing in some functions using this assert.
>>>> Therefore a proper assertion will fail. Because adding drains
>>>> requires additional discussions, they will be added in future
>>>> series.
>>>>
>>>> Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
>>>> Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
>>>> Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
>>>> ---
>>>>   block.c                                |  5 +++++
>>>>   block/io.c                             | 11 +++++++++++
>>>>   include/block/block_int-global-state.h | 10 +++++++++-
>>>>   3 files changed, 25 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/block.c b/block.c
>>>> index 41c5883c5c..94bff5c757 100644
>>>> --- a/block.c
>>>> +++ b/block.c
>>>> @@ -2734,12 +2734,14 @@ static void 
>>>> bdrv_replace_child_noperm(BdrvChild *child,
>>>>           if (child->klass->detach) {
>>>>               child->klass->detach(child);
>>>>           }
>>>> +        assert_bdrv_graph_writable(old_bs);
>>>>           QLIST_REMOVE(child, next_parent);
>>>
>>> I think this belongs above the .detach() call (and the QLIST_REMOVE() 
>>> belongs into the .detach() implementation, as done in 
>>> https://lists.nongnu.org/archive/html/qemu-block/2021-11/msg00240.html, 
>>> which has been merged to Kevin’s block branch).
>>
>> Yes, I rebased on kwolf/block branch. Thank you for pointing that out.
>>>
>>>>       }
>>>>       child->bs = new_bs;
>>>>       if (new_bs) {
>>>> +        assert_bdrv_graph_writable(new_bs);
>>>>           QLIST_INSERT_HEAD(&new_bs->parents, child, next_parent);
>>>
>>> In both these places it’s a bit strange that the assertion is done on 
>>> the child nodes.  The subgraph starting from them isn’t modified 
>>> after all, so their subgraph technically doesn’t need to be 
>>> writable.  I think a single assertion on the parent node would be 
>>> preferable.
>>>
>>> I presume the problem with that is that we don’t have the parent node 
>>> here?  Do we need a new BdrvChildClass method that performs this 
>>> assertion on the parent node?
>>>
>>
>> Uhm I am not sure what you mean here.
>>
>> Just to recap on how I see this: the assertion 
>> assert_bdrv_graph_writable(bs) is basically used to make sure we are 
>> protecting the write on some fields (childrens and parents lists in 
>> this patch) of a given @bs.
> 
> Oh, OK.  I understood it to mean that the subgraph starting at `bs` is 
> mutable, i.e. including all of its recursive children.

Yes, sorry for the confusion. We want to drain the subgraph starting at 
`bs`, also because we modify both parent's .children list and children's 
.parent list.

> 
> And yes, you’re right, the child BDSs are indeed modified, too, so we 
> should check them, too.
> 
>> It should work like a rwlock: reading is allowed to be concurrent, but 
>> a write should stop all readers to prevent concurrency issues. This is 
>> achieved by draining.
> 
> Draining works on a subgraph, so I suppose it does mean that the whole 
> subgraph will be mutable.  But no matter, at least new_bs is not in the 
> parent’s subgraph, so it wouldn’t be included in the check if we only 
> checked the parent.
> 
>> Let's use the first case that you point out, old_bs (it's specular for 
>> new_bs):
>>
>> >> +        assert_bdrv_graph_writable(old_bs);
>> >>           QLIST_REMOVE(child, next_parent);
>>
>> So old_bs should be the child "son" (child->bs), meaning 
>> old_bs->parents contains the child. Therefore when a child is removed 
>> by old_bs, we need to be sure we are doing it safely.
>>
>> So we should check that if old_bs exists, old_bs should be drained, to 
>> prevent any other iothread from reading the ->parents list that is 
>> being updated.
>>
>> The only thing to keep in mind in this case is that just wrapping a 
>> drain around that won't be enough, because then the child won't be 
>> included in the drain_end(old_bs). Therefore the right way to cover 
>> this drain-wise once the assertion also checks for drains is:
>>
>> drain_begin(old_bs)
>> assert_bdrv_graph_writable(old_bs)
>> QLIST_REMOVE(child, next_parent)
>> /* old_bs will be under drain_end, but not the child */
>> bdrv_parent_drained_end_single(child);
>> bdrv_drained_end(old_bs);
>>
>> I think you agree on this so far.
>>
>> Now I think your concern is related to the child "parent", namely 
>> child->opaque. The problem is that in the .detach and .attach 
>> callbacks we are firstly adding/removing the child from the list, and 
>> then calling drain on the subtree.
> 
> It was my impression that you’d want bdrv_replace_child_noperm() to 
> always be called in a section where the subgraph starting from the 
> parent BDS is drained, not just the child BDSs that are swapped (old_bs 
> and new_bs).
> 
> My abstract concern is that bdrv_replace_child_noperm() does not modify 
> only old_bs and new_bs, but actually modifies the whole subgraph 
> starting at the parent node.  A concrete example for this is that we 
> modify not only the children’s parent lists, but also the parent’s 
> children list.
> 
> That’s why I’m asking why we only check that the graph is writable for 
> the children, when actually I feel like we’re modifying the parent, too.

Ok I think I understood what you mean, and I think I addressed this in 
the new series version but was not addressed here.

Maybe what I do is redundant, but I:
1) drain the childrens when we swap them
2) modify .attach and .detach to drain child->opaque (parent) too.
More precisely, I think it should be enough to change 
bdrv_child_cb_attach/detach in the way I showed below, since we are 
touching the parent's .children list only there.
blk_root_attach/detach seem not to deal with its children, so a drain 
there is not necessary.

This should cover everything we need.

> 
>> We would ideally need to do the opposite:
>>
>> assert_bdrv_graph_writable(bs);
>> QLIST_REMOVE(child, next);
>> bdrv_unapply_subtree_drain(child, bs);
>>
>> In this case I think this would actually work, because removing/adding 
>> the child from the ->children list beforehand just prevents an 
>> additional recursion call (I think, and the fact that tests are 
>> passing seems to confirm my theory).
>>
>> Of course you know this stuff better than me, so let me know if 
>> something here is wrong.
> 
> Well.  I’m mostly wondering why you’re discussing how to do the drain 
> right when I was mostly curious about why we check the children and not 
> the parent for whether the graph is mutable at their respective 
> position. O:)
> 
> It was my impression that so far we mostly wrapped graph change 
> operations in drained sections (starting at the parent) and not leave it 
> to bdrv_replace_child_noperm() to do so.  That function only deals with 
> drain stuff because it balances the drain counters on old_bs and new_bs 
> to match the counter on the parent’s subgraph.

"mostly" is the problem, because if you try to use the full assertion 
introduced in this patch (ie assert also for drains), it will fail.

So I figured I would (will) put the drains as near as possible to the 
section where they are needed, since bdrv_replace_child_noperm is being 
called in many different ways. But this is a discussion for a future series.

Emanuele