Conversation
Signed-off-by: László Várady <[email protected]>
Signed-off-by: László Várady <[email protected]>
… msgs Signed-off-by: László Várady <[email protected]>
Signed-off-by: László Várady <[email protected]>
|
Manually tested. Can not reproduce original crash. Verified from my side |
bazsi
left a comment
bazsi
left a comment
There was a problem hiding this comment.
This looks sane. I don't have a complete understanding on how the dynamic window allocation and release operates, so this approve is more like "I trust @MrAnno as he was behind the original implementation and this looks legit", and since this is a crashing bug, let's get this merged, even if I don't understand the bug and the fix.
|
I didn't write this part of the code, but the issue here is the following: The current implementation is not what we originally wanted: we just try to divide the dynamic window between ACTIVE connections fairly. Sometimes it is not possible because a previous connection already used up a big portion of the window, so a reclaim request is made in that case to return the missing window as soon as possible. This reclaim request needs to be processed before we free the connection (we do that periodically, but that's not enough because the scheduling of that timer is not guaranteed to be executed before the connection closes). |
Signed-off-by: László Várady <[email protected]>
3 participants
No description provided.