We only need 4 bytes of headroom for alignment
purposes in the RX frame. It was previously higher
for optimization purposes which are no longer
possible due to DMA mappings.

v2: Fix patch error
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When rt2x00queue_alloc_rxskbs() fails rt2x00queue_unitialize()
will be called which will free all rxskb. So we don't need
to do this in the rt2x00queue_alloc_rxskb() function as well.

rt2x00queue_free_skb() unmaps the DMA but doesn't clear the
allocation flag. Since the code is copied from rt2x00queue_unmap_skb()
anyway (and that function does clear the flag) we might as well
use that function directly.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

With a bit of code moving to rt2x00lib within the
TX and RX paths we can now remove a lot of EXPORT_SYMBOL_GPL()
statements. This cleans up the interface between rt2x00lib
and the drivers and has the additional benefit that rt2x00pci
and rt2x00usb are trimmed down in size as well since they
have less to do.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

The current PCI drivers require a lot of pre-allocated DMA buffers. Reduce this
by using dynamically mapped skb's (using pci_map_single) instead of the pre-
allocated DMA buffers that are allocated at device start-up time.

At the same time move common RX path code into rt2x00lib from rt2x00pci and
rt2x00usb, as the RX paths now are now almost the same.
Signed-off-by: NGertjan van Wingerde <gwingerde@kpnplanet.nl>
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

In preparation of replacing the statically allocated DMA buffers with
dynamically mapped skbs, centralize the allocation of RX skbs to rt2x00queue.c
and let rt2x00pci already use them.
Signed-off-by: NGertjan van Wingerde <gwingerde@kpnplanet.nl>
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

With the introduction of the ieee80211 fc handlers
we can now remove the rt2x00.h versions to use the
global versions.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Reviewed-by: NHarvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Replaced by the new helper ieee80211_has_morefrags which is
more consistent with the intent of the function.
Signed-off-by: NHarvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When rt2x00pci will be switched over to dynamically mapped skb's
instead of statically allocated DMA buffers, it no longer can handle
alignment of RX packets in a copy step, and needs to implement the
same scheme as rt2x00usb does.

In order to make the patch on dynamically mapped skb's smaller,
already centralize the alignment handling into rt2x00lib. This allows
us to move more code in rt2x00lib, and thus remove code duplication
between rt2x00usb and rt2x00pci.
Signed-off-by: NGertjan van Wingerde <gwingerde@kpnplanet.nl>
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

TX queues shouldn't be kicked after each frame that is put into the
queue.  This could cause problems during RTS and CTS-to-self as well
as with fragmentation. In all those cases you want all frames to be
send out in a single burst. Off course we shouldn't let the queue fill
up entirely, thus we introduce a 10% threshold which, when reached,
will force the frames to be send out regardless of the frame.

In addition we should prevent queues to become full in such a way
that the tx() handler can fail. Instead of stopping the queue when
it is full, we should stop it when it is below the threshold.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

The write_tx_data functions in rt2x00pci and rt2x00usb have
a lot in common. This moves that duplicate code into
rt2x00queue_write_tx_frame().
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

qid should be initialized to QID_BEACON and QID_ATIM
for the beacon and atim quue. This makes checking for
a particular queue much saner, and it shouldn't harm,
because the only places where the value is send to
the hardware, we are allowed to send any value we
want since it is only used as argument in the
TX done register.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
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 patch modifies struct ieee80211_tx_control to give band
info and the rate index (instead of rate pointers) to drivers.
This mostly serves to reduce the TX control structure size to
make it fit into skb->cb so that the fragmentation code can
put it there and we can think about passing it to drivers that
way in the future.

The rt2x00 driver update was done by Ivo, thanks.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohannes Berg <johannes@sipsolutions.net>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Split rt2x00lib_write_tx_desc() up into a TX descriptor initializor
and TX descriptor writer.

This split is required to properly allow mac80211 to move its
tx_control structure into the skb->cb array.
The rt2x00queue_create_tx_descriptor() function will read all tx control
information and convert it into a rt2x00 TX descriptor information structure.
After that function is complete, we have all information we needed from the
tx control structure and are free to start writing into the skb->cb array
for our own purposes.
rt2x00queue_write_tx_descriptor() will be in charge of really sending
the TX descriptor to the hardware and kicking the TX queue.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

With the integration of the mac80211 multiqueue patches it has become possible that the
mac80211 layer modifies the number of TX queues that is stored inside the ieee80211_hw
structure, especially when multi-queue is not selected.

The rt2x00 drivers are not well suited to handle that situation, as they allocate the
queue structures before mac80211 has modified the number of queues it is going to use,
and also expect the number of allocated queues to match the hardware implementation.

Hence, ensure that rt2x00 maintains by itself the number of queues that the hardware
supports, and, at the same time, making is not dependent on the preservation of contents
inside a mac80211 structure.
Signed-off-by: NGertjan van Wingerde <gwingerde@kpnplanet.nl>
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

Use the rt2x00 queue enumeration as much as possible,
removing the usage of the mac80211 queue numbering
wherever it is possible.

This makes it easier for mac80211 to change it queue
identification scheme without having to deal with
big changes in the rt2x00 code.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

The queue->lock could be grabbed from interrupt context,
which could lead to lockdep panic like this:

kernel: ======================================================
kernel: [ INFO: soft-safe -> soft-unsafe lock order detected ]
kernel: 2.6.25-0.95.rc4.fc9 #1
kernel: ------------------------------------------------------
kernel: rt2500pci/1251 [HC0[0]:SC0[1]:HE1:SE0] is trying to acquire:
kernel:  (&queue->lock){--..}, at: [<ffffffff88213339>] rt2x00queue_get_entry+0x5a/0x81 [rt2x00lib]
kernel:
kernel: and this task is already holding:
kernel:  (_xmit_IEEE80211){-...}, at: [<ffffffff8122e9a3>] __qdisc_run+0x84/0x1a9
kernel: which would create a new lock dependency:
kernel:  (_xmit_IEEE80211){-...} -> (&queue->lock){--..}
kernel:
kernel: but this new dependency connects a soft-irq-safe lock:
kernel:  (_xmit_ETHER){-+..}
kernel: ... which became soft-irq-safe at:
kernel:   [<ffffffffffffffff>] 0xffffffffffffffff
kernel:
kernel: to a soft-irq-unsafe lock:
kernel:  (&queue->lock){--..}
kernel: ... which became soft-irq-unsafe at:
kernel: ...  [<ffffffff810545a2>] __lock_acquire+0x62d/0xd63
kernel:   [<ffffffff81054d36>] lock_acquire+0x5e/0x78
kernel:   [<ffffffff812a1497>] _spin_lock+0x26/0x53
kernel:   [<ffffffff88212f98>] rt2x00queue_reset+0x16/0x40 [rt2x00lib]
kernel:   [<ffffffff88212fd4>] rt2x00queue_alloc_entries+0x12/0xab [rt2x00lib]
kernel:   [<ffffffff88213091>] rt2x00queue_initialize+0x24/0xf2 [rt2x00lib]
kernel:   [<ffffffff88212036>] rt2x00lib_start+0x3b/0xd4 [rt2x00lib]
kernel:   [<ffffffff88212609>] rt2x00mac_start+0x18/0x1a [rt2x00lib]
kernel:   [<ffffffff881b9a4b>] ieee80211_open+0x1f3/0x46d [mac80211]
kernel:   [<ffffffff8121d980>] dev_open+0x4d/0x8b
kernel:   [<ffffffff8121d41e>] dev_change_flags+0xaf/0x172
kernel:   [<ffffffff81224fc2>] do_setlink+0x276/0x338
kernel:   [<ffffffff81225198>] rtnl_setlink+0x114/0x116
kernel:   [<ffffffff812262fc>] rtnetlink_rcv_msg+0x1d8/0x1f9
kernel:   [<ffffffff8123649a>] netlink_rcv_skb+0x3e/0xac
kernel:   [<ffffffff8122611a>] rtnetlink_rcv+0x29/0x33
kernel:   [<ffffffff81235eed>] netlink_unicast+0x1fe/0x26b
kernel:   [<ffffffff81236224>] netlink_sendmsg+0x2ca/0x2dd
kernel:   [<ffffffff812103b3>] sock_sendmsg+0xfd/0x120
kernel:   [<ffffffff812105a8>] sys_sendmsg+0x1d2/0x23c
kernel:   [<ffffffff8100c1c7>] tracesys+0xdc/0xe1
kernel:   [<ffffffffffffffff>] 0xffffffffffffffff

This can be fixed by using the irqsave/irqrestore versions
during the queue->lock handling.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

As Adam Baker reported the queue->qid was not initialized
correctly. The QID_AC_BE was assigned to the RX ring.

This will move the queue initialization into a seperate function
and makes sure that all queues are initialized directly with the
correct qids.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When calculating the offset to add to the queue entry base to get the
individual entry's private data area the base address must be treated
as a char * not a struct queue_entry so we can do byte oriented
pointer arithmetic with it.
Signed-off-by: NAdam Baker <linux@baker-net.org.uk>
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

When Q_INDEX is increased, the queue->length should be increased,
only when Q_INDEX_DONE is increased should queue_length be descreased.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

As Johannes Berg indicated the BEACON and AFTER_BEACON
queue indeces in mac80211 should be removed because they
are too hardware specific. This patch adds the queue index
defines into rt2x00queue.h and removes the dependency of
the defines inside mac80211.h.

Also move rt2x00pci_beacon_update() into rt2400pci and
rt2500pci individually since it is no longer a generic
function since rt61 and rt2800 no longer use that.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>

This introduces a big queue handling overhaul, this also
renames "ring" to "queues".

Move queue handling into rt2x00queue.c and the matching header,
use Kerneldoc to improve rt2x00 library documentation.

Access to the queues is now protected under a spinlock, this
to prevent race conditions which could corrupt the indexing
system of the queue.

Each queue entry allocates x bytes for driver/device specific data,
this cleans up the queue structure significantly and improves
code readability.

rt2500usb no longer needs 2 entries in the beacon queue to correctly
send out the guardian byte. This is now handled in the entry specific
structure.

rt61 and rt73 now use the correct descriptor size for beacon frames,
since this data is written into the registers not the entire TXD
descriptor was used but instead of a subset of it named TXINFO.

Finally this also fixes numerous other bugs related to incorrect
beacon handling or beacon related code.
Signed-off-by: NIvo van Doorn <IvDoorn@gmail.com>
Signed-off-by: NJohn W. Linville <linville@tuxdriver.com>