allocator: Fix panic when allocations happen at init time #1651
Conversation
|
I think we should also make sure this scenario is covered in unit tests. I'm not very familiar with the tests for this part of the code, so it would be great if someone could point me in the right direction or submit a separate PR for test coverage. |
|
LGTM |
| @@ -81,6 +81,7 @@ func (a *Allocator) doNetworkInit(ctx context.Context) error { | |||
| unallocatedNetworks: make(map[string]*api.Network), | |||
| ingressNetwork: newIngressNetwork(), | |||
| } | |||
| a.netCtx = nc | |||
There was a problem hiding this comment.
The reason a.netCtx is initialized at the end is to make sure in case of failures we don't return from doNetworkInit with a netCtx which shouldn't be there in the Allocator. This is because the Allocator has a longer life time than what happens in doNetworkInit itself.
So in all of doNetworkInit I've just passed the netCtx as an argument to functions that need it. May be we should do the same for taskCreateNetworkAttachments i.e add netCtx as an argument?
There was a problem hiding this comment.
Sounds kind of dangerous TBH. What about a deferred closure in doNetworkInit that clears a.netCtx if an error is being returned?
There was a problem hiding this comment.
Ok, I think a deferred closure for error handling should satisfy that requirement as well. I am good with that.
Current coverage is 56.54% (diff: 69.04%)@@ master #1651 diff @@
==========================================
Files 90 90
Lines 14551 14552 +1
Methods 0 0
Messages 0 0
Branches 0 0
==========================================
- Hits 8248 8229 -19
- Misses 5214 5233 +19
- Partials 1089 1090 +1
|
a.netCtx is initialized too late, so if allocations happen as part of doNetworkInit, a nil pointer dereference will cause a panic. Initialize a.netCtx earlier and use a.netCtx directly in member functions instead of passing the network context separately, so there is no confusion about which to use. Also change allocator.go to have separate entries in the waitgroup for initialization and actually running the allocator, and defer `Done` for both. This should prevent a panic like this from leading to a deadlock, since the deferred `Done` will be reached. Signed-off-by: Aaron Lehmann <aaron.lehmann@docker.com>
aaronlehmann
force-pushed
the
allocator-crash
branch
from
dab9090
to
a8066c1
Compare
|
@mrjana: Updated to add a closure that clears |
|
@aaronlehmann Adding a test case to cover this scenario would be nice. I will try to add one. |
|
@aaronlehmann @mrjana Please add a test case in a separate PR |
To identify the issue in allocator. Signed-off-by: Madhu Venugopal <madhu@docker.com>
Also Cherry-pick moby/swarmkit#1651 to identify the issue in allocator. Signed-off-by: Madhu Venugopal <madhu@docker.com>
a.netCtxis initialized too late, so if allocations happen as part ofdoNetworkInit, a nil pointer dereference will cause a panic.Initialize
a.netCtxearlier and usea.netCtxdirectly in memberfunctions instead of passing the network context separately, so there is
no confusion about which to use.
Also change allocator.go to have separate entries in the waitgroup for
initialization and actually running the allocator, and defer
Doneforboth. This should prevent a panic like this from leading to a deadlock,
since the deferred
Donewill be reached.See moby/moby#25432
cc @mrjana @LK4D4 @tonistiigi