Felix Fietkau [Sun, 28 Aug 2011 19:11:01 +0000 (21:11 +0200)]
mac80211: retry sending failed BAR frames later instead of tearing down aggr
Unfortunately failed BAR tx attempts happen more frequently than I
expected, and the resulting aggregation teardowns cause performance
issues, as the aggregation session does not always get re-established
properly.
Instead of tearing down the entire aggr session, we can simply store the
SSN of the last failed BAR tx attempt, wait for the first successful
tx status event, and then send another BAR with the same SSN.
Felix Fietkau [Sat, 27 Aug 2011 23:52:10 +0000 (01:52 +0200)]
ath9k_hw: do not apply the 2.4 ghz ack timeout workaround to cts
It is only used to workaround interoperability issues related to longer
delays in receiving the block ack, so it is not necessary to apply it
to the CTS exchange.
Should improve throughput slightly, especially when there are lots
of retransmissions.
Javier Cardona [Tue, 6 Sep 2011 19:10:43 +0000 (12:10 -0700)]
mac80211: Defer tranmission of mesh path errors
Under failure conditions, the mesh stack sends PERR messages to the
previous sender of the failed frame. This happens in the tx feedback
path, in which the transmission queue lock may be taken. Avoid a
deadlock by sending the path error via the pending queue.
Felix Fietkau [Sat, 27 Aug 2011 22:32:25 +0000 (00:32 +0200)]
ath9k: defer ath_tx_setup_buffer setup to the first tx attempt during aggr
With sequence number and buffer allocation deferred to when they're needed
for the first time, it becomes much easier to start dropping packets from
the tid queue if necessary, e.g. when latency suddenly increases. This can
lead to some future improvements in buffer management for better latency.
Felix Fietkau [Sat, 27 Aug 2011 22:32:23 +0000 (00:32 +0200)]
ath9k: move the sequence number from ath_frame_info to ath_buf
It is only necessary for BAW tracking and moving it to the ath_buf
makes it easier to add further improvements, such as deferring
seqno allocation in the aggregation path.
Felix Fietkau [Sat, 27 Aug 2011 22:32:22 +0000 (00:32 +0200)]
ath9k: clean up the aggregation tid queue
Use a sk_buff_head instead containing skbs instead of a list_head
containing ath_bufs. This makes it easier to decouple the aggregation
code from the ath_buf struct
John W. Linville [Mon, 29 Aug 2011 19:47:58 +0000 (15:47 -0400)]
wl12xx/sdio_test.c: fix build breakage from WL127X_FW_NAME change
Commit c302b2c959164622558474871ae942da0e484a38 ("wl12xx: Use a single
fw for both STA and AP roles") changed the name of the firmware name
definition, breaking the build of wl12xx/sdio_test.c.
After analysis, this is not a bug, but a false warning. Nonetheless,
a cleanup is in order to prevent some future janitor proposing
the wrong fix, as I did in my original patch.
ath9k: Add debugfs support for mac/baseband samples
This patch keep track of number of samples that includes
DMA debugs registers, PCU observe, CR, channel noise,
cycle conters, noisefloor history buffer and last N number
of tx and rx descriptor status. These samples are grouped
in table manner which dumping in debgufs.
Felix Fietkau [Sat, 27 Aug 2011 08:25:27 +0000 (10:25 +0200)]
ath9k: fix regression in sending aggregated packets
The recent commit "ath9k: Send legacy rated frames as unaggregated"
introduced a check to ensure that packets with non-MCS rates set in
the rate series will not be aggregated. However, it failed to check
if the rate series is valid before testing the flags, thus breaking
aggregation for normal MCS-only packets if the last series is unset.
For AR9287 v1.3+ chips, MAC runs at 117MHz. But the initvals
IFS parameters are loaded based on 44/88MHz clockrate. So
eifs/usec from ini should not be used for AR9287 v1.3+.
The mentioned values are tested on 2 chain HT40 mode.
Larry Finger [Fri, 26 Aug 2011 18:41:38 +0000 (20:41 +0200)]
b43: Relax requirement for descriptors to be in the DMA zone
When 64-bit DMA was first used, there were problems with the
BCM4311 (14e4:4311). The problem was "fixed" by using the GFP_DMA
flag in the allocation of coherent ring descriptor memory.
The original problem is now believed to have been due to bugs in
the 64-bit DMA implementation in the rest of the kernel, and that
those bugs have been fixed. Accordingly, the requirement for the
descriptors to be in the DMA zone is relaxed.
Johannes Berg [Fri, 26 Aug 2011 06:13:56 +0000 (23:13 -0700)]
iwlagn: handle GO powersave
In order to implement support for GO powersave on
the P2P client side, the ucode needs to know what
GO we're trying to authenticate/associate with,
it needs to have a station entry and the BSSID in
the RXON set.
Implement the new mac80211 callbacks to give this
data to the device.
Since this is also useful for the device when a
normal connection is established, also program it
with the information in that case.
iwlagn: move the stop / wake queue logic to transport layer
priv->mac80211_registered and priv->hw needed to move to shared.
stop_queue API was added in order to allow the upper layer to stop
the SW queues for regulatory purposes.
iwlagn: move the disable agg logic to transport layer
Since all the check_empty logic is now in the transport layer,
the upper layer doesn't need to know anything about tx queues.
The disable aggregation flow was the last to know what a tx queue
is, so move it too.
iwlagn: move the check_empty logic to the transport layer
This logic is responsible to tell mac80211 when the HW queues are
empty and the BA session can be started / torn down.
Fix a bug on the way:
When the the Tx BA session is stopped and the HW queues aren't empty,
we stop the SW queue to drain the HW queue and then switch to the
legacy HW queue. This is the IWL_EMPTYING_HW_QUEUE_DELBA state.
While in this state, we never wake the SW queue, even when the HW
queue is almost empty, since we need to drain it completely. Look
at iwl_trans_pcie_reclaim regarding this.
Once the HW queue is really empty, we must wake the SW queue in order
to get traffic to the legacy queue.
This step was missing leading to an odd situation were the traffic
would just stall after we tore down a Tx BA session while the HW
queue was not empty.
iwlagn: move the mapping ac to queue / fifo to transport
This mapping is transport related.
This allows us to remove the notion of tx queue from the tx path in
the upper layer.
iwl_wake_any_queue moved to transport layer since it needs to access
these mappings.
The TX API is nicer now:
iwlagn: upper layer uses slabs to allocate tx cmds
In a near future, the upper layer won't be aware of the tx queues.
This allows to remove one place where the upper layer needed to
provide the tx queue index to the transport layer.
This also saves around 1.5MB.
the mapping tx_queue -> fifo is really transport related. The upper
layer should be involved in such things.
Note that upon agg_disable, the queue is always mapped to fifo 0, but
this doesn't matter since when the queue will be setup again for a
new BA session, it will be configured to the good fifo anyway.
iwlagn: upper layer stores iwl_rxon_context in skb's CB
This removes the need for iwl_tx_info.
Each tx queue holds an array of skbs, the transport layer doesn't
need to know anything about the context in which a specific skb is
sent.
When we remove the module, we free all the tx and rx resources.
Before doing that, we'd better stop the tx / rx activity. Calling
iwl_trans_stop_device in iwl_remove helps also to remove a few API
functions:
* rx_free: happens in iwl_trans_free
* tx_free: happens in iwl_trans_free
* disable_sync_irq: happens in iwl_trans_stop_device
iwlagn: remove dereferences of priv from transport
There are still quite a few, but much less.
A few fields have been moved /copied to hw_params which sits in the
shared area:
* priv->cfg->base_params->num_of_ampdu_queues
* priv->cfg->base_params->shadow_reg_enable
* priv->cfg->sku
* priv->ucode_owner
The rate scaling and the transport need to access the data in
iwl_tid_data, hence the move.
Note that the only component in the upper layer that needs this data
is the rate scaling. Refactoring the rate scaling may help to move
iwl_tid_data from the shared area to the transport area.
Move all the iwlXXX_abgn_cfg forward declaration to a separate file
so that iwl-pci.c doesn't need to include iwl-agn.h that includes
all iwl-dev.h
This allows to provide real encapsulation. Dereferencing iwl_priv
in the bus layer will now lead to a compilation error.
Daniel Halperin [Fri, 26 Aug 2011 06:11:15 +0000 (23:11 -0700)]
iwlagn: fix compile warnings when CONFIG_PM_SLEEP is not set
CC [M] drivers/net/wireless/iwlwifi/iwl-pci.o
drivers/net/wireless/iwlwifi/iwl-pci.c:506: warning: \91iwl_pci_suspend\92 defined but not used
drivers/net/wireless/iwlwifi/iwl-pci.c:519: warning: \91iwl_pci_resume\92 defined but not used
These are only used if CONFIG_PM_SLEEP is enabled. CONFIG_PM depends
(CONFIG_PM_SLEEP || CONFIG_PM_RUNTIME), so it can be set without
CONFIG_PM_SLEEP selected.
Now that the reclaim flow has been moved to the transport layer, a lot of
functions can be made static or don't need to be exported outside the transport
layer.
The reclaim flow is really transport related. Define a simple API to allow the
upper layer to request from the transport layer to reclaim packets until an
index written in the Tx response / BA notification.
The transport layer prepares a list of the packets that are being freed and
passes this list to the upper layer.
Between the two layers, the CB of the skb is used to pass a pointer to the
context (BSS / PAN) in which the skb was sent.
It is accessed by the transport layer only, hence the move.
The debugfs handlers that accessed it moved to the transport layer too.
The rx_handlers part of it stayed in the upper layer and a special debugfs
has been added for it
Since the ISR is entirely in the transport layer, its data should be in the pcie
specific region.
Change sync_irq to first disable and then synchronize the IRQ.
iwl_isr and iwl_isr_ict now receive iwl_trans.
Remove iwl_transport_register which was a W/A. The bus layer knows what
transport to use. So now, the bus layer gives the upper layer a pointer to the
iwl_trans_ops struct that it wants to use. The upper layer then, allocates the
desired transport layer using iwl_trans_ops->alloc function.
As a result of this, priv->trans, no longer exists, priv holds a pointer to
iwl_shared, which holds a pointer to iwl_trans. This required to change all the
calls to the transport layer from upper layer. While we were at it, trans_X
inlines have been renamed to iwl_trans_X to avoid confusions, which of course
required to rename the functions inside the transport layer because of
conflicts in names. So the static API functions inside the transport layer
implementation have been renamed to iwl_trans_pcie_X.
Until now, the IRQ / Tasklet were initialized in iwl_transport_layer. This is
confusing since the registration doesn't mean to request IRQ, so I added a
handler for that.
iwlagn: modify the debug macro to be usable by all the layers
Since all the layers need to print debug message, the debug macro cannot suppose
that they will be given iwl_priv as a parameter and then dereference it.
Use iwl_shared instead.
iwlagn: introduce struct iwl-shared - known by all layers
This struct will hold pointers to all the layers, so that every layer will find
the pointers it needs when calling another layer.
Note that the drv_data set to struct device is now a pointer to
struct iwl_shared.
This solves of bug that I introduced in
iwlagn: simplify the bus architecture
Bug description:
sysfs gets the the driver data from struct device. Till the aforementioned
patch, dev_get_drvdata would return iwl_priv. After the patch, dev_get_drvdata
return iwl_bus which is buggy since the sysfs handlers rely on this value, and
sysfs handlers need iwl_priv.
Now, dev_get_drvdata return iwl-shared. Since we have pointers to all the
layers in iwl_shared, every layer will be able to get the pointer it needs:
bus layer will gets iwl_bus from the PCI suspend callbacks, and the sysfs
handlers will get the iwl_priv they need.
In order to keep good encapsulation, we need to avoid to dereference iwl_priv
from a different layer. This is why instead of including iwl-dev.h from
iwl-shared.h, I added a forward declaration to iwl_priv. Moreover we keep type
safety while providing encapsulation.
It will hold declaration of functions and forward declaration of struct that
are used by several layers.
This will allow modules not to include iwl_priv. iwl_bus and iwl_trans are
still visible to all.
All the layers share the module parameters, move the struct to iwl-shared.h.
Also add all module parameters to iwl_mod_params instead of having them as
global static. This includes
iwlagn: use iwl_get_debug_level instead of iwl_debug_level
The latter may return incomplete information.
For example, if one switched IWL_DL_TX on through sysfs, IWL_DL_TX bit would
have been set in priv->debug_level, but since iwl_alloc_traffic_mem looked at
iwl_debug_level only, it wouldn't have allocated the tx_traffic buffer.
Peter Huewe [Tue, 7 Jun 2011 20:36:14 +0000 (22:36 +0200)]
net/mac80211/debugfs: Convert to kstrou8_from_user
This patch replaces the code for getting an number from a
userspace buffer by a simple call to kstrou8_from_user.
This makes it easier to read and less error prone.
Since the old buffer was only 10 bytes long and the value is masked by a
nibble-mask anyway, we don't need to use kstrtoul but rather kstrtou8.
carl9170: Use do_div for 64-bit division to fix 32-bit kernels
Use the do_div macro for 64-bit division. Otherwise, the module will
reference __udivdi3 under 32-bit kernels, which is not allowed in
kernel space.
drivers/built-in.o: In function `carl9170_collect_tally':
cmd.c:191: undefined reference to `__udivdi3'
cmd.c:192: undefined reference to `__udivdi3'
cmd.c:193: undefined reference to `__udivdi3'
Here are the AR9580 1.0 initvals checksums using the
Atheros initvals-tools [1]. This is useful for when
we udate the initvals again with other values. It ensures
that we match the same initvals used internally. The
tool is documented on the wiki [2].
projects / linux.git / log