On Wed, Dec 16, 2020 at 01:35:14PM +0100, Kevin Wolf wrote:
Am 15.12.2020 um 18:23 hat Sergio Lopez geschrieben:
On Tue, Dec 15, 2020 at 04:01:19PM +0100, Kevin Wolf wrote:
Am 15.12.2020 um 14:15 hat Sergio Lopez geschrieben:
On Tue, Dec 15, 2020 at 01:12:33PM +0100, Kevin Wolf wrote:
Am 14.12.2020 um 18:05 hat Sergio Lopez geschrieben:
While processing the parents of a BDS, one of the parents may process
the child that's doing the tail recursion, which leads to a BDS being
processed twice. This is especially problematic for the aio_notifiers,
as they might attempt to work on both the old and the new AIO
contexts.
To avoid this, add the BDS pointer to the ignore list, and check the
child BDS pointer while iterating over the children.
Signed-off-by: Sergio Lopez <slp@redhat.com>
Ugh, so we get a mixed list of BdrvChild and BlockDriverState? :-/
I know, it's effective but quite ugly...
What is the specific scenario where you saw this breaking? Did you have
multiple BdrvChild connections between two nodes so that we would go to
the parent node through one and then come back to the child node through
the other?
I don't think this is a corner case. If the graph is walked top->down,
there's no problem since children are added to the ignore list before
getting processed, and siblings don't process each other. But, if the
graph is walked bottom->up, a BDS will start processing its parents
without adding itself to the ignore list, so there's nothing
preventing them from processing it again.
I don't understand. child is added to ignore before calling the parent
callback on it, so how can we come back through the same BdrvChild?
QLIST_FOREACH(child, &bs->parents, next_parent) {
if (g_slist_find(*ignore, child)) {
continue;
}
assert(child->klass->set_aio_ctx);
*ignore = g_slist_prepend(*ignore, child);
child->klass->set_aio_ctx(child, new_context, ignore);
}
Perhaps I'm missing something, but the way I understand it, that loop
is adding the BdrvChild pointer of each of its parents, but not the
BdrvChild pointer of the BDS that was passed as an argument to
b_s_a_c_i.
Generally, the caller has already done that.
In the theoretical case that it was the outermost call in the recursion
and it hasn't (I couldn't find any such case), I think we should still
call the callback for the passed BdrvChild like we currently do.
You didn't dump the BdrvChild here. I think that would add some
information on why we re-entered 0x555ee2fbf660. Maybe you can also add
bs->drv->format_name for each node to make the scenario less abstract?
I've generated another trace with more data:
bs=0x565505e48030 (backup-top) enter
bs=0x565505e48030 (backup-top) processing children
bs=0x565505e48030 (backup-top) calling bsaci child=0x565505e42090
(child->bs=0x565505e5d420)
bs=0x565505e5d420 (qcow2) enter
bs=0x565505e5d420 (qcow2) processing children
bs=0x565505e5d420 (qcow2) calling bsaci child=0x565505e41ea0
(child->bs=0x565505e52060)
bs=0x565505e52060 (file) enter
bs=0x565505e52060 (file) processing children
bs=0x565505e52060 (file) processing parents
bs=0x565505e52060 (file) processing itself
bs=0x565505e5d420 (qcow2) processing parents
bs=0x565505e5d420 (qcow2) calling set_aio_ctx child=0x5655066a34d0
bs=0x565505fbf660 (qcow2) enter
bs=0x565505fbf660 (qcow2) processing children
bs=0x565505fbf660 (qcow2) calling bsaci child=0x565505e41d20
(child->bs=0x565506bc0c00)
bs=0x565506bc0c00 (file) enter
bs=0x565506bc0c00 (file) processing children
bs=0x565506bc0c00 (file) processing parents
bs=0x565506bc0c00 (file) processing itself
bs=0x565505fbf660 (qcow2) processing parents
bs=0x565505fbf660 (qcow2) calling set_aio_ctx child=0x565505fc7aa0
bs=0x565505fbf660 (qcow2) calling set_aio_ctx child=0x5655068b8510
bs=0x565505e48030 (backup-top) enter
bs=0x565505e48030 (backup-top) processing children
bs=0x565505e48030 (backup-top) calling bsaci child=0x565505e3c450
(child->bs=0x565505fbf660)
bs=0x565505fbf660 (qcow2) enter
bs=0x565505fbf660 (qcow2) processing children
bs=0x565505fbf660 (qcow2) processing parents
bs=0x565505fbf660 (qcow2) processing itself
bs=0x565505e48030 (backup-top) processing parents
bs=0x565505e48030 (backup-top) calling set_aio_ctx child=0x565505e402d0
bs=0x565505e48030 (backup-top) processing itself
bs=0x565505fbf660 (qcow2) processing itself
Hm, is this complete? Is see no "processing itself" for
bs=0x565505e5d420. Or is this because it crashed before getting there?
Yes, it crashes there. I forgot to mention that, sorry.
Anyway, trying to reconstruct the block graph with BdrvChild pointers
annotated at the edges:
BlockBackend
|
v
backup-top ------------------------+
| | |
| +-----------------------+ |
| 0x5655068b8510 | | 0x565505e3c450
| | |
| 0x565505e42090 | |
v | |
qcow2 ---------------------+ | |
| | | |
| 0x565505e52060 | | | ??? [1]
| | | | |
v 0x5655066a34d0 | | | | 0x565505fc7aa0
file v v v v
qcow2 (backing)
|
| 0x565505e41d20
v
file
[1] This seems to be a BdrvChild with a non-BDS parent. Probably a
BdrvChild directly owned by the backup job.
So it seems this is happening:
backup-top (5e48030) <---------| (5)
| | |
| | (6) ------------> qcow2 (5fbf660)
| ^ |
| (3) | | (4)
|-> (1) qcow2 (5e5d420) ----- |-> file (6bc0c00)
|
|-> (2) file (5e52060)
backup-top (5e48030), the BDS that was passed as argument in the first
bdrv_set_aio_context_ignore() call, is re-entered when qcow2 (5fbf660)
is processing its parents, and the latter is also re-entered when the
first one starts processing its children again.
Yes, but look at the BdrvChild pointers, it is through different edges
that we come back to the same node. No BdrvChild is used twice.
If backup-top had added all of its children to the ignore list before
calling into the overlay qcow2, the backing qcow2 wouldn't eventually
have called back into backup-top.
I've tested a patch that first adds every child to the ignore list,
and then processes those that weren't there before, as you suggested
on a previous email. With that, the offending qcow2 is not re-entered,
so we avoid the crash, but backup-top is still entered twice: