This reverts commit 9bd568a5.

That commit is shown to cause allocation failures during initialization
on some machines.

http://bugzilla.kernel.org/show_bug.cgi?id=14844Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

This rewrites the error handling policies in the TX status handler.
It tries to be error-tolerant as in "try hard to not crash the machine".
It won't recover from errors (that are bugs in the firmware or driver),
because that's impossible. However, it will return a more or less useful
error message and bail out. It also tries hard to use rate-limited messages
to not flood the syslog in case of a failure.
Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Enforce all device constraints on the descriptor memory region.

There are several constraints on the descriptor memory, as documented
in the specification. The current code does not enforce them and/or
incorrectly enforces them.

Those constraints are:
- The address limitations on 30/32bit engines, that also apply to
  the skbs.
- The 4k alignment requirement on 30/32bit engines.
- The 8k alignment requirement on 64bit engines.
Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

This removes the DMA/PIO queue locks. Locking is handled by
wl->mutex now.
Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Tested-by: NLarry Finger <Larry.Finger@lwfinger.net>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

In the old days we used one slot per frame. But when we changed that to 2,
we didn't raise the overall slot count. Which resulted in an effective
division of two to the number of slots.

Double the number of TX slots, so we have an effective hardware queue
of 128 frames per QoS queue.

Also optimize the TX header cache handling. We don't need a cached TX header
for slots that will never carry an actual header.
So we reduce the memory consumption of the cache by 50%.

So as a net result we end up with more or less the same memory usage before
and after this patch (except a few tiny meta structures), but have twice
the number of TX slots available.
Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

This fixes hidden bugs in the size handling of the DMA buffers.
This sets the RX buffer size to the theoretical max packet size and
fixes passing of the size values to the device (must not subtract the header offset).

These bugs are hidden and don't actually trigger due to the magic +100
offset for the buffer size.
Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

This patch converts mac80211 and all drivers to have transmit
information and status in skb->cb rather than allocating extra
memory for it and copying all the data around. To make it fit,
a union is used where only data that is necessary for all steps
is kept outside of the union.

A number of fixes were done by Ivo, as well as the rt2x00 part
of this patch.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohannes Berg <johannes@sipsolutions.net>
Acked-by: NDavid S. Miller <davem@davemloft.net>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

This adds PIO support back (D'oh!) for PCMCIA devices.
This is a complete rewrite of the old PIO code. It does actually work
and we get reasonable performance out of it on a modern machine.
On a PowerBook G4 I get a few MBit for TX and a few more for RX.
So it doesn't work as well as DMA (of course), but it's a _lot_ faster
than the old PIO code (only got a few kBit with that).

The limiting factor is the host CPU speed. So it will generate 100%
CPU usage when the network interface is heavily loaded. A voluntary preemption
point in the RX path makes sure Desktop Latency isn't hurt.

PIO is needed for 16bit PCMCIA devices, as we really don't want to poke with
the braindead DMA mechanisms on PCMCIA sockets. Additionally, not all
PCMCIA sockets do actually support DMA in 16bit mode (mine doesn't).
Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

This adds a few debugging counters, that are useful for debugging the
"card does not transmit" or "connection is unstable" kind of problems.
It's also useful for tuning an RC algorithm.
Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

This adds QOS support to the b43 driver.
QOS can be disabled on driver level with a module parameter for debugging purposes.
Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

This checks if the DMA address is bigger than what the controller can manage.
It will reallocate the buffers in the GFP_DMA zone in that case.
Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Remove b43 PIO support.
DMA works well on all supported devices. There's no reason to use PIO.
Additionally, new devices don't support PIO in hardware anymore.
b43 PIO support is dead and unused code.

After applying this patch please do
git rm drivers/net/wireless/b43/pio.h
git rm drivers/net/wireless/b43/pio.c
to remove the main PIO support code.
Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Signed-off-by: NMichael Buesch <mb@bu3sch.de>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>