On 02/27/13 11:27, Dan Williams wrote:
On Mon, 2013-02-25 at 14:07 +0100, Felix Fietkau wrote:
On 2013-02-25 11:25 AM, Johannes Berg wrote:
On Mon, 2013-02-25 at 06:54 +0100, Felix Fietkau wrote:
Most devices have some kind of connection manager that has a high-level
perspective of when it's fully connected (which includes DHCP/bootp).
Why not just let that connection manager set a sane maximum network
latency value via pm_qos network_latency and derive btcoex weight
changing and multi-channel settings from that?
Frankly, I don't think that's going to work well. We tried using the
pm_qos framework once and nothing ever used it. Android isn't going to
change to it, so we'd be stuck with hacks like setting pm_qos in
wpa_supplicant which is just as awkward.
If only the connection manager gets changed to use it, that would
already be enough. It doesn't have to be pushed into dhcp clients and
other applications.
Also, what you mostly want isn't really so much a weight but rather a
time-based approach to give it high priority until the connection
handshake completes (we already do for auth/assoc/... until authorized)
so I think using the pm_qos framework to give priority wouldn't work
very well since there'd also be no way to tell when it was "done"
Just release the latency requirement in the connection manager once the
handshake is done. It knows...
We also don't know what IP configuration method will get used; whether
it will be IPv6 RA, DHCPv4 or DHCPv6, IPv4 autoconf, or static. Only
the connection manager knows that. Only the connection manager/DHCP
client know when they expect a lease renew operation to start too.
wpa_supplicant doesn't know any of these things either since it doesn't
do anything IP related.
I think the best approach here is to allow the higher layers to hint to
the driver that some operations that are about to start must be "more
reliable". That includes EAPOL, DHCP, IP autoconfiguration, etc. Then
when the higher layers know the operation is finished, they can indicate
the operations are done and the driver can go do whatever it wants.
Indeed the RFC approach was explicit about the scope of this interface
being BT coex or actually BT coex override. Johannes proposes one
dedicated to DHCP as a similar interface in Android is used for that
right now. Abstracting it to "more reliable" mainly avoids renaming it
when someone comes up with a use-case other than BT coex or DHCP.
The driver/stack may wish to do any of [set 1Mb rate, block rate
control, change BT coex, turn on microwave protection, whatever] and
that's great, the upper layers don't care about what the driver does,
just that the reliability of the operation is preserved.
It is actually the reliability of the connection, but it may depend on
what you mean by "operation" here. I think from user-space perpective
this API is at most a notification to the driver, which *may* result in
a more reliable protocol exchange in terms of reliability and/or latency.
Regards,
Arend
--
To unsubscribe from this list: send the line "unsubscribe linux-wireless" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
