* Jan 2, 2019
Rebase patches to make 9980 bisectable.
* Jan 2, 2019
Fix scheduling related assert when wal-peer is deleted with pending
tx buffers (bug 54, and others)
* Jan 7, 2019:
Fix specifying retransmits for AMPDU frames. It was previously ignored
since it is a 'software' retransmit instead of a hardware retransmit.
* Jan 9, 2019
Fix potential way to get zero rates selected (and then assert)
* Jan 18, 2019
pfsched has specific work-around to just return if we find invalid flags AND
if we are in an out-of-order situation. Maybe this is last of the pfsched
related issues (bug 54 and similar).
* Jan 24, 2019
The rcSibUpdate method can be called concurrently with IRQ tx-completion callback,
and that could potentially allow the tx-completion callback to see invalid state
and assert or otherwise mess up the rate-ctrl logic. So, disable IRQs in
rcSibUpdate to prevent this. Related to bug 58.
* Jan 28, 2019
Ensure that cached config is applied to ratectrl objects when fetched from
the cache. This should fix part of bug 58.
* Jan 28, 2019
Ensure that ratectrl objects from cachemgr are always initialized. This fixes
another part of bug 58.
* Jan 30, 2019
Better use of temporary rate-ctrl object. Make sure it is initialized, simplify
code path. This finishes up porting forward similar changes I made for wave-1
firmware long ago, and fixes another potential way to hit bug-58 issues.
* Jan 30, 2019
Cachemgr did not have a callback for when memory was logically freed. This means
that peers could keep stale references to rate-ctrl objects that were in process
of being DMA'd into to load a different peer's rate-ctrl state. This was causing
the bugcheck logic to fail early and often, and I suspect it might be a root cause
of bug 58 as well. The fix is to add a callback and set any 'deleted' memory references
to NULL so that we cannot access it accidentally. Thanks to excellent logs and patience
from the bug-58 reporter!
Signed-off-by: Koen Vandeputte <koen.vandeputte@ncentric.com>
5a8d03ceeb