Am 19.08.19 um 22:03 schrieb Johannes Berg:
On Mon, 2019-08-19 at 21:57 +0200, Alexander Wetzel wrote:
+
+ /* The new Tx API does not allow to pass the key or keyid of a MPDU to
+ * the hw, preventing us to control which key(id) to use per MPDU.
+ * Till that's fixed we can't use Extended Key ID for the newer cards.
Technically we still don't need per MPDU, we just need to switch which
one to use for TX after installing it for RX already.
The Extended Key ID API we finally merged in mac80211 is not notifying
the driver when to switch the key over to the other id.
Oh, right, good point.
The current API provides the key/keyid per MPDU and let's mac80211 have
the full control what's the correct key for each frame.
Yeah, but as you noticed we no longer have that control per MPDU with
the new TX API in iwlmvm.
That's especially critical for drivers setting
IEEE80211_KEY_FLAG_GENERATE_IV and/or supporting A-MPDU's. Allowing the
driver to override the mac80211 decision is only safe when the
driver/card generates the PNs itself and also handles the A-MPDU key
borders correctly.
Sure, the device does generate the PN itself now with the new TX API
too. It doesn't care about A-MPDU key borders, but it probably could
when taught to care about extended key ID.
While less desirable we still could get that working: The mvm driver
would have to detect the key borders and then tell the firmware to
switch over to the other key. But we would have to make sure to not
re-enable A-MPDU aggregation till the card really has switched.
I'm not entirely sure off the top of my head how it works, but it seems
possible that if we just assign a new PN to retransmits of the same
frame but in a new A-MPDU after key switching, it wouldn't actually
matter?
I was thinking about mac80211 re-enabling A-MPDU aggregation but the
driver still using the old key for some frames and then switching to the
new key within one A-MPDU, causing non-compliant A-MPDUs.
But then again maybe somewhere it's stated that we must use the same key
for all transmit attempts of a single frame? Not sure.
I know of course next to nothing how the hardware/firmware is handling
that, but I'm surprised this could be a problem. I assumed the card
would create a full MPDU (with the crypto headers) and use that for
retransmit and not create the MPDU new... After all you still have to
use the old PN and getting that right sounds tricky when you don't have
the full MPDU cached somehow. But I guess the high bandwidth and the
required buffer space forced you to extreme measures to conserve space...
I'm also not sure if we could actually assign a PN from the new key for
the retransmit, the hardware has to store those back into memory
normally.
So probably you're right, and we'd have to disable A-MPDUs until we have
no outstanding old-key-retransmits, but that seems manageable.
Extending the A-MPDU block mac80211 is using to enforce the A-MPDU key
borders should be not very hard. Figuring out the time when we safely
can restart it - when re-transmits indeed are also an issue - I'm less
sure about. I believe re-transmits are handled by the card and not the
driver but so far had no reason to dig deeper into that topic. But I
guess a call to switch over the keyid which blocks till that is done
should take care of it.
Alexander
Alexander
