My patch resized the pool size, but neglected to resize
the global table, which is obviously wrong since the global
table maps the pool's rxb to vid one to one. This results
in a panic in 9000 devices.
Add a build bug to avoid such a case in the future.

Fixes: 7b542436 ("iwlwifi: pcie: fine tune number of rxbs")
Reported-by: NHaim Dreyfuss <haim.dreyfuss@intel.com>
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

When trying to reach high Rx throughput of more than 500Mbps on
a device with a relatively weak CPU (Atom x5-Z8500), CPU utilization
may become a bottleneck. Analysis showed that we are looping in
iwl_pcie_rx_handle for very long periods which led to starvation
of other threads (iwl_pcie_rx_handle runs with _bh disabled).
We were handling Rx and allocating new buffers and the new buffers
were ready quickly enough to be available before we had finished
handling all the buffers available in the hardware. As a
consequence, we called iwl_pcie_rxq_restock to refill the hardware
with the new buffers, and start again handling new buffers without
exiting the function. Since we read the hardware pointer again when
we goto restart, new buffers were handled immediately instead of
exiting the function.

This patch avoids refilling RBs inside rx handling loop, unless an
emergency situation is reached. It also doesn't read the hardware
pointer again unless we are in an emergency (unlikely) case.
This significantly reduce the maximal time we spend in
iwl_pcie_rx_handle with _bh disabled.
Signed-off-by: NGregory Greenman <gregory.greenman@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

We kick the allocator when we have 2 RBDs that don't have
attached RBs, and the allocator allocates 8 RBs meaning
that it needs another 6 RBDs to attach the RBs to.
The design is that allocator should always have enough RBDs
to fulfill requests, so we give in advance 6 RBDs to the
allocator so that when it is kicked, it gets additional 2 RBDs
and has enough RBDs.
These RBDs were taken from the Rx queue itself, meaning
that each Rx queue didn't have the maximal number of
RBDs, but MAX - 6.
Change initial number of RBDs in the system to include both
queue size and allocator reserves.
Note the multi-queue is always 511 instead of 512 to avoid a
full queue since we cannot detect this state easily enough in
the 9000 arch.
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

128 byte chunk size is supported only on PCIe and not
on IOSF. For now, change it back to 64 byte.
Reported-by: NOren Givon <oren.givon@intel.com>
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Change the code to move rxbs directly from the allocator's
list to the queue's free list. This makes the code more
readable, saves the interim array and the double loop over
the free RBs.
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

In 9000 series A0 step the closed_rb_num is not wrapping around
properly. The queue is wrapping around as it should, so we can
W/A it by wrapping the closed_rb_num in the driver.
While at it, extend RX logging and add error handling of other
cases HW values may cause us to access invalid memory locations.
Add also a proper masking of vid value read from HW - this should
not have actual affect, but better to be on the safe side.
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Fine tune RFH registers further:
* Set default queue explicitly
* Set RFH to drop frames exceeding RB size
* Set the maximum rx transfer size to DRAM to 128 instead of 64
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Working with MSIX requires prior configuration.
This includes requesting interrupt vectors from the OS,
registering the vectors and mapping the optional causes to the
relevant interrupt. In addition add new interrupt handler
to handle MSIX interrupt.
Signed-off-by: NHaim Dreyfuss <haim.dreyfuss@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

When we load the firmware, we hold trans_pcie->mutex to
avoid nested flows. We also rely on the ISR to wake up the
thread when the DMA has finished copying a chunk. During
this flow, we enable the RF-Kill interrupt.

The problem is that the RF-Kill interrupt handler can take
the mutex and bring the device down. This means that if
we load the firmware while the RF-Kill switch is enabled
(which will happen when we load the INIT firmware to read
the device's capabilities and register to mac80211), we
may get an RF-Kill interrupt immediately and the ISR will
be waiting for the mutex held by the thread that is
currently loading the firmware. At this stage, the ISR
won't be able to service the DMA's interrupt needed to
wake up the thread that load the firmware. We are in a
deadlock situation which ends when the thread that loads
the firmware fails on timeout and releases the mutex.

To fix this, take the mutex later in the flow, disable
the interrupts and synchronize_irq() to give a chance to
the RF-Kill interrupt to run and complete.
After that, mask all the interrupts besides the DMA
interrupt and proceed with firmware load. Make sure to
check that there was no RF-Kill interrupt when the
interrupts were disabled.

This fixes https://bugzilla.kernel.org/show_bug.cgi?id=111361Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Previous patches enabled new 9000 hardware DMA for one queue
only.
Enable the actual multi-queue path and configuration now.
This requires also per-queue NAPI struct.
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

The 9000 series introduces several changes in the device
DMA operation.
As the device now supports multi-queue rx, several DMA channels
should be configured.
The flows of providing the device with the allocated RBDs now
changes as well - the device maintains a separate table of used
and free table.

The hardware may use the free table to feed RBDs to any queue.
This requires maintaing a shared table to map returned RBDs to
the original RXB - for that purpose the VID is introduced - an
internal identifier of the RB placed in the lower 12 bits and
returned by HW in the used data.

Another change is the support of 64 bit DMA address.
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

The 9000 series devices will support multi rx queues.
Current code has one static rx queue - change it to allocate
a number of queues per the device capability (pre-9000 devices
have the number of rx queues set to one).

Subsequent generalizations are:

Change the code to access an explicit numbered rx queue only
when the queue number is known - when handling interrupt, when
accessing the default queue and when iterating the queues.
The rest of the functions will receive the rx queue as a pointer.

Generalize the warning in allocation failure to consider the
allocator status instead of a single rx queue status.

Move the rx initial pool of memory buffers to be shared among
all the queues and allocated to the default queue on init.
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Host commands now have a group id, express this in printed messages.
Signed-off-by: NSharon Dvir <sharon.dvir@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

ilw@linux.intel.com is not available anymore.
linuxwifi@intel.com should be used instead.
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Since this pointer is not shown anywhere else, it's useless.
Remove it, just keeping the indexes.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

This message isn't very useful and presents a security risk
due to the use of %p - remove it.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

802.11ac allows A-MSDU that can be up to 12KB long. Since
an entire A-MSDU needs to fit into one single Receive
Buffer (RB), add support for big RBs.
Since this adds lots of pressure to the memory manager and
significantly increase the true_size of the RX buffers,
don't enable this by default.
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Part of reorganising wireless drivers directory and Kconfig.
Signed-off-by: NKalle Valo <kvalo@codeaurora.org>

The mac80211 patch to pass the NAPI struct only changed iwlwifi to
store the NAPI struct, but we can do better: pass it directly from
the lower transport layer to the opmode during RX, and then on to
mac80211 from there.

When we add multiple RX queues, we can then pass the appropriate
NAPI struct properly.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

As a preperation for multiple RX queues change the RBD
allocation model.

The new model includes a background allocator. The allocator is
called by the interrupt handler when there are two released
buffers by the queue, and the allocator starts allocating eight
pages per request.
When the queue has released 8 pages it tries claiming the
request. If the pages are not ready - it keeps claiming.
This new model should make sure that RBDs are always available
across the multiple queues.

The RBDs are transferred between the allocator and the queue.
The queue moves the free RBDs upon freeing them to the allocator.
The allocator moves them back to the queue's possession when the
request is claimed.
The allocator has an initial pool to make sure there are always RBDs
available for the request completion.
Release of the buffers at exit is done per pools - the allocator
frees its own initial pool and the queue frees its own pool.

Existing code refactor -
-Queue's initial pool is the size of the queue only as the allocation
of the new buffers no longer uses this pool.
-Removal of replenish background work, and replenish calls in the
interrupt handler and restock().
-The replenish() and the rxq used_list are used only during
initialization.
-Moved page allocation to a new function for code reuse.

New code -
Allocator code - new structure and functions.
Interrupt handler uses the allocator functions for replenishing buffers.
Reuse of the restock() method.
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Since the CSR_DRAM_INIT_TBL_WRITE_POINTER bit wasn't set
on ict reset, in some flows (like disable ict followed by
immediate reset ict) the driver and hardware went out
of sync (the driver cleared the ict_index, while the hw
kept it intact).

Fix it by setting the flag when resetting ict.
Signed-off-by: NEliad Peller <eliad@wizery.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

When the firmware crashes, we can't expect the Tx queues to
progress. Cancel their timer.
Reviewed-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

With the previous patch series, no opmode continues using the
command or handler_status (i.e. the return value from the RX)
so it can be removed now.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

This reverts commit 5f175703.

This patch introduced a high latency in buffer allocation
under extreme load. This latency caused a firmwre crash.
The same scenario works fine with this patch reverted.
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

This allows to ensure that we don't have races between them.
A user reported that stop_device was called twice upon
rfkill interrupt after suspend. When the interrupts are
enabled, and right after when we directly check the rfkill
state.
Reviewed-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

As a preperation for multiple RX queues change the RBD
allocation model.

The new model includes a background allocator. The allocator is
called by the interrupt handler when there are two released
buffers by the queue, and the allocator starts allocating eight
pages per request.
When the queue has released 8 pages it tries claiming the
request. If the pages are not ready - it keeps claiming.
This new model should make sure that RBDs are always available
across the multiple queues.

The RBDs are transferred between the allocator and the queue.
The queue moves the free RBDs upon freeing them to the allocator.
The allocator moves them back to the queue's possession when the
request is claimed.
The allocator has an initial pool to make sure there are always RBDs
available for the request completion.
Release of the buffers at exit is done per pools - the allocator
frees its own initial pool and the queue frees its own pool.

Existing code refactor -
-Queue's initial pool is the size of the queue only as the allocation
of the new buffers no longer uses this pool.
-Removal of replenish background work, and replenish calls in the
interrupt handler and restock().
-The replenish() and the rxq used_list are used only during
initialization.
-Moved page allocation to a new function for code reuse.

New code -
Allocator code - new structure and functions.
Interrupt handler uses the allocator functions for replenishing buffers.
Reuse of the restock() method.
Signed-off-by: NSara Sharon <sara.sharon@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Allow a cleaner way to access those hw-dependent registers,
instead of using the product family type etc.
Signed-off-by: NAvri Altman <avri.altman@intel.com>
Reviewed-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

For each RX packet until this patch there only was a debug
print of the HCMD and the offset. This adds also the
sequence number of the packet for easier matching between
what was sent, what came back / was received, and what
got stuck somewhere and was never responded by the FW.
Signed-off-by: NLiad Kaufman <liad.kaufman@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

When freeing the structures used for command data, clear their
memory as they may have contained key material at some point.
Also clear the duplicated buffer when freeing it to be safe;
currently key material is never put there but that may change.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Our legal structure changed at some point (see wikipedia), but
we forgot to immediately switch over to the new copyright
notice.

For files that we have modified in the time since the change,
add the proper copyright notice now.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

This variable always tracks a constant value (256) so there's
no need to have it. Removing it simplifies code generation,
reducing the .text size (by about 240 bytes on x86-64.)
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Besides the different allocation flags, they are really
the same. Pass the gfp_t flags as a parameter, and unify
them.
Reviewed-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Similar to the recent RX queue patch, this changes the need_update
handling for the TX queues to be clearer and only done when needed.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

When updating the write pointer, the TX queue should be locked
to get consistent state, fix that in the interrupt handler.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

When shadow registers are enabled, then need_update never needs
to be set, so move the need_update handling into the function
that really needs to do it (iwl_pcie_rxq_inc_wr_ptr) and also
separate the check when it woke up. While at it, convert it to
bool.

This also clarifies the locking and means the irq_lock needs to
no longer be held for any such updates.

The irq_lock also doesn't have to be held for restocking since
everything else locks the RX queue properly, so remove that and
finally disentangle the two locks entirely so there aren't any
dependencies between the two left.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Use the new NAPI infrastructure added to mac80211 to get
GRO. We don't really implement NAPI since we don't have
a real poll function and we never schedule a NAPI poll.
Instead of this, we collect all the packets we got from a
single interrupt and then call napi_gro_flush().

This allows us to benefit from GRO. In half duplex medium
like WiFi, its main advantage is that it reduces the number
of TCP Acks, hence improving the TCP Rx performance.

Since we call the Rx path with a spinlock held, remove
the might_sleep mention from the op_mode's API.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Reviewed-by: NIdo Yariv <ido@wizery.com>
[Squash different patches and rewrite the commit message]
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

When indicating RF-kill toggle to the higher layer, that
may in turn call back to the transport (for MVM at least)
to turn off the device quickly. Instead of that, allow it
to return whether or not the device should be turned off,
this gets rid of the call indirection and will help make
the API more consistent when we go back to non-threaded
interrupts again for PCIe.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

The various code blocks in iwl_pcie_[rt]xq_inc_wr_ptr
finally do the same things, so just merge them
all and make the functions cleaner.
Signed-off-by: NEliad Peller <eliadx.peller@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

In iwl_pcie_int_cause_non_ict, trans_pcie is used for lockdep
purposes only. Since this might not be enabled, trans_pcie
finds itself without user leading to a complaint from gcc.
Avoid using trans_pcie by inlining IWL_TRANS_GET_PCIE_TRANS.
Reviewed-by: NJohannes Berg <johannes.berg@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>

Add an inline helper function for getting an RX packet's
length or payload length and use it throughout the code
(most of which I did using an spatch.)

While at it, adjust some code, and remove a bogus comment
from the dvm calibration code.
Signed-off-by: NJohannes Berg <johannes.berg@intel.com>
Reviewed-by: NEran Harary <eran.harary@intel.com>
Signed-off-by: NEmmanuel Grumbach <emmanuel.grumbach@intel.com>