Since the format of the HW descriptors is different between PPv2.1 and
PPv2.2, this commit introduces an intermediate union, with for now
only the PPv2.1 descriptors. The bulk of the driver code only
manipulates opaque mvpp2_tx_desc and mvpp2_rx_desc pointers, and the
descriptors can only be accessed and modified through the accessor
functions. A follow-up commit will add the descriptor definitions for
PPv2.2.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In preparation to the introduction for the support of PPv2.2 in the
mvpp2 driver, this commit adds a hw_version field to the struct
mvpp2, and uses the .data field of the DT match table to fill it in.

Having the MVPP21 and MVPP22 definitions available will allow to start
adding the necessary conditional code to support PPv2.2.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The PPv2.2 IP has a different TX and RX descriptor layout compared to
PPv2.1. In order to prepare for the introduction of PPv2.2 support in
mvpp2, this commit adds accessors for the different fields of the TX
and RX descriptors, and changes the code to use them.

For now, the mvpp2_port argument passed to the accessors is not used,
but it will be used in follow-up to update the descriptor according to
the version of the IP being used.

Apart from the mechanical changes to use the newly introduced
accessors, a few other changes, needed to use the accessors, are made:

 - The mvpp2_txq_inc_put() function now takes a mvpp2_port as first
   argument, as it is needed to use the accessors.

 - Similarly, the mvpp2_bm_cookie_build() gains a mvpp2_port first
   argument, for the same reason.

 - In mvpp2_rx_error(), instead of accessing the RX descriptor in each
   case of the switch, we introduce a local variable to store the
   packet size.

 - In mvpp2_tx_frag_process() and mvpp2_tx() instead of accessing the
   packet size from the TX descriptor, we use the actual value
   available in the function, which is used to set the TX descriptor
   packet size a few lines before.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The RX descriptors of the PPv2 hardware allow to store several
information, amongst which:

 - the DMA address of the buffer in which the data has been received
 - a "cookie" field, left to the use of the driver, and not used by the
   hardware

In the current implementation, the "cookie" field is used to store the
virtual address of the buffer, so that in the receive completion path,
we can easily get the virtual address of the buffer that corresponds to
a completed RX descriptors.

On PPv2.1, used on 32-bit platforms, those two fields are 32-bit wide,
which is enough to store a DMA address in the first field, and a virtual
address in the second field.

On PPv2.2, used on 64-bit platforms, these two fields have been extended
to 40 bits. While 40 bits is enough to store a DMA address (as long as
the DMA mask is 40 bits or lower), it is not enough to store a virtual
address. Therefore, the "cookie" field can no longer be used to store
the virtual address of the buffer.

However, as Russell King pointed out, the RX buffers are always
allocated in the kernel linear mapping, and therefore using
phys_to_virt() on the physical address of the RX buffer is possible and
correct.

Therefore, this commit changes the driver to use the "cookie" field to
store the physical address instead of the virtual
address. phys_to_virt() is used in the receive completion path to
retrieve the virtual address from the physical address.

It is obviously important to realize that the DMA address and physical
address are two different things, which is why we store both in the RX
descriptors. While those addresses may be identical in some situations,
it remains two distinct concepts, and both addresses should be handled
separately.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The mvpp2_txq_pend_desc_num_get() function only selects a TX queue, and
reads the number of pending descriptors. It is used in only one place,
in mvpp2_txq_clean(), where the TX queue has already been selected by a
write to MVPP2_TXQ_NUM_REG.

Therefore, this function is useless, and the caller can simply read the
value of the MVPP2_TXQ_PENDING_REG register instead.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This register is no longer used since commit edc660fa ("net: mvpp2:
replace TX coalescing interrupts with hrtimer").
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The "buffer header" functionality is a functionality used by the
hardware to split an incoming packets over multiple BM buffers if they
are not large enough. However, the mvpp2 driver guarantees that a pool
of BM buffers has buffers with a size large enough to store MTU-sized
packets. Therefore, this functionality is completely unused, and the
code can be removed, and we should never get a descriptor with bit
MVPP2_RXD_BUF_HDR set.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

As indicated by Russell King, the mvpp2 driver currently uses a lot
"phys" or "phys_addr" to store what really is a DMA address. This commit
clarifies this by using "dma" or "dma_addr" where appropriate.

This is especially important as we are going to introduce more changes
where the distinction between physical address and DMA address will be
key.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The mvpp2 is going to be extended to support the Marvell Armada 7K/8K
platform, which is ARM64. As a preparation to this work, this commit
enables building the mvpp2 driver on ARM64, by:

 - Adjusting the Kconfig dependency

 - Fixing the types used in the driver so that they are 32/64-bits
   compliant. We use dma_addr_t for DMA addresses, and unsigned long
   for virtual addresses.

It is worth mentioning that after this commit, the driver is for now
still only used on 32-bits platforms, and will only work on 32-bits
platforms.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This commit adapts the mvpp2 RX path to use the build_skb() method. Not
only build_skb() is now the recommended mechanism, but it also
simplifies the addition of support for the PPv2.2 variant.

Indeed, without build_skb(), we have to keep track for each RX
descriptor of the physical address of the packet buffer, and the virtual
address of the SKB. However, in PPv2.2 running on 64 bits platform,
there is not enough space in the descriptor to store the virtual address
of the SKB. So having to take care only of the address of the packet
buffer, and building the SKB upon reception helps in supporting PPv2.2.

The implementation is fairly straightforward:

 - mvpp2_skb_alloc() is renamed to mvpp2_buf_alloc() and no longer
   allocates a SKB. Instead, it allocates a buffer using the new
   mvpp2_frag_alloc() function, with enough space for the data and SKB.

 - The initialization of the RX buffers in mvpp2_bm_bufs_add() as well
   as the refill of the RX buffers in mvpp2_rx_refill() is adjusted
   accordingly.

 - Finally, the mvpp2_rx() is modified to use build_skb().
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Some of the MVPP2_PRS_RI_* definitions use the ~(value) syntax, which
doesn't compile nicely on 64-bit. Moreover, those definitions are in
fact unneeded, since they are always used in combination with a bit
mask that ensures only the appropriate bits are modified.

Therefore, such definitions should just be set to 0x0. In addition, as
suggested by Russell King, we change the _MASK definitions to also use
the BIT() macro so that it is clear they are related to the values
defined afterwards.

For example:

 #define MVPP2_PRS_RI_L2_CAST_MASK              0x600
 #define MVPP2_PRS_RI_L2_UCAST                  ~(BIT(9) | BIT(10))
 #define MVPP2_PRS_RI_L2_MCAST                  BIT(9)
 #define MVPP2_PRS_RI_L2_BCAST                  BIT(10)

becomes

 #define MVPP2_PRS_RI_L2_CAST_MASK              (BIT(9) | BIT(10))
 #define MVPP2_PRS_RI_L2_UCAST                  0x0
 #define MVPP2_PRS_RI_L2_MCAST                  BIT(9)
 #define MVPP2_PRS_RI_L2_BCAST                  BIT(10)

Because the values (MVPP2_PRS_RI_L2_UCAST, MVPP2_PRS_RI_L2_MCAST and
MVPP2_PRS_RI_L2_BCAST) are always applied with
MVPP2_PRS_RI_L2_CAST_MASK, and therefore there is no need for
MVPP2_PRS_RI_L2_UCAST to be defined as ~(BIT(9) | BIT(10)).

It fixes the following warnings when building the driver on a 64-bit
platform (which is not possible as of this commit, but will be enabled
in a follow-up commit):

drivers/net/ethernet/marvell/mvpp2.c: In function ‘mvpp2_prs_mac_promisc_set’:
drivers/net/ethernet/marvell/mvpp2.c:524:33: warning: large integer implicitly truncated to unsigned type [-Woverflow]
 #define MVPP2_PRS_RI_L2_UCAST   ~(BIT(9) | BIT(10))
                                 ^
drivers/net/ethernet/marvell/mvpp2.c:1459:33: note: in expansion of macro ‘MVPP2_PRS_RI_L2_UCAST’
   mvpp2_prs_sram_ri_update(&pe, MVPP2_PRS_RI_L2_UCAST,
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The mvpp2_bm_bufs_add() currently creates a fake cookie by calling
mvpp2_bm_cookie_pool_set(), just to be able to call
mvpp2_pool_refill(). But all what mvpp2_pool_refill() does is extract
the pool ID from the cookie, and call mvpp2_bm_pool_put() with this ID.

Instead of doing this convoluted thing, just call mvpp2_bm_pool_put()
directly, since we have the BM pool ID.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This commit drops dead code from the mvpp2 driver. The 'in_use' and
'in_use_thresh' fields of 'struct mvpp2_bm_pool' are
incremented/decremented/initialized in various places. But they are only
used in one place:

       if (is_recycle &&
           (atomic_read(&bm_pool->in_use) < bm_pool->in_use_thresh))
               return 0;

However 'is_recycle', passed as argument to mvpp2_rx_refill() is always
false. So in fact, this code is never reached, and the 'is_recycle'
argument is useless. So let's drop this code.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This commit remove a field of 'struct mvpp2_tx_queue' that is not used
anywhere.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The mvpp2_txq_bufs_free() function is called upon TX completion to DMA
unmap TX buffers, and free the corresponding SKBs. It gets the
references to the SKB to free and the DMA buffer to unmap from a per-CPU
txq_pcpu data structure.

However, the code currently increments the pointer to the next entry
before doing the DMA unmap and freeing the SKB. It does not cause any
visible problem because for a given SKB the TX completion is guaranteed
to take place on the CPU where the TX was started. However, it is much
more logical to increment the pointer to the next entry once the current
entry has been completely unmapped/released.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When configuring the MVPP2_ISR_RX_THRESHOLD_REG with the RX coalescing
time threshold, we do not check for the maximum allowed value supported
by the driver, which means we might overflow and use a bogus value. This
commit adds a check for this situation, and if a value higher than what
is supported by the hardware is provided, then we use the maximum value
supported by the hardware.

In order to achieve this in a way that avoids overflow and rounding
errors, we introduce two utility functions mvpp2_usec_to_cycles() and
cycles_to_usec(). Many thanks to Russell King for suggesting this
implementation.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Currently, mvpp2_rx_pkts_coal_set() does the following to avoid setting
a too large value for the RX coalescing by packet number:

  val = (pkts & MVPP2_OCCUPIED_THRESH_MASK);

This means that if you set a value that is slightly higher the the
maximum number of packets, you in fact get a very low value. It makes a
lot more sense to simply check if the value is too high, and if it's too
high, limit it to the maximum possible value.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

As noticed by Russell King, the last argument of
mvpp2_rx_{pkts,time}_coal_set() is useless, since the packet/time
coalescing value is already stored in the 'struct mvpp2_rx_queue *'
passed as argument to these functions. So passing the packet/time value
as an additional argument, and setting them again in the mvpp2_rx_queue
structure is useles.

This commit therefore gets rid of this additional argument, assuming the
caller has assigned the appropriate value to rxq->pkts_coal or
rxq->time_coal before calling the respective functions.
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Acked-by: NRussell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

When TX descriptors are filled in, the buffer DMA address is split
between the tx_desc->buf_phys_addr field (high-order bits) and
tx_desc->packet_offset field (5 low-order bits).

However, when we re-calculate the DMA address from the TX descriptor in
mvpp2_txq_inc_put(), we do not take tx_desc->packet_offset into
account. This means that when the DMA address is not aligned on a 32
bytes boundary, we end up calling dma_unmap_single() with a DMA address
that was not the one returned by dma_map_single().

This inconsistency is detected by the kernel when DMA_API_DEBUG is
enabled. We fix this problem by properly calculating the DMA address in
mvpp2_txq_inc_put().

Cc: <stable@vger.kernel.org>
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

if 'devm_kzalloc()' fails, we should release resources allocated so far,
just as done a few lines below.
Signed-off-by: NChristophe JAILLET <christophe.jaillet@wanadoo.fr>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

napi_complete_done() allows to opt-in for gro_flush_timeout,
added back in linux-3.19, commit 3b47d303
("net: gro: add a per device gro flush timer")

This allows for more efficient GRO aggregation without
sacrifying latencies.
Signed-off-by: NEric Dumazet <edumazet@google.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The network device operation for reading statistics is only called
in one place, and it ignores the return value. Having a structure
return value is potentially confusing because some future driver could
incorrectly assume that the return value was used.

Fix all drivers with ndo_get_stats64 to have a void function.
Signed-off-by: NStephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

ktime_set(S,N) was required for the timespec storage type and is still
useful for situations where a Seconds and Nanoseconds part of a time value
needs to be converted. For anything where the Seconds argument is 0, this
is pointless and can be replaced with a simple assignment.
Signed-off-by: NThomas Gleixner <tglx@linutronix.de>
Cc: Peter Zijlstra <peterz@infradead.org>

Since commit 71ce391d ("net: mvpp2: enable proper per-CPU TX
buffers unmapping"), we are not correctly DMA unmapping TX buffers for
fragments.

Indeed, the mvpp2_txq_inc_put() function only stores in the
txq_cpu->tx_buffs[] array the physical address of the buffer to be
DMA-unmapped when skb != NULL. In addition, when DMA-unmapping, we use
skb_headlen(skb) to get the size to be unmapped. Both of this works fine
for TX descriptors that are associated directly to a SKB, but not the
ones that are used for fragments, with a NULL pointer as skb:

 - We have a NULL physical address when calling DMA unmap
 - skb_headlen(skb) crashes because skb is NULL

This causes random crashes when fragments are used.

To solve this problem, we need to:

 - Store the physical address of the buffer to be unmapped
   unconditionally, regardless of whether it is tied to a SKB or not.

 - Store the length of the buffer to be unmapped, which requires a new
   field.

Instead of adding a third array to store the length of the buffer to be
unmapped, and as suggested by David Miller, this commit refactors the
tx_buffs[] and tx_skb[] arrays of 'struct mvpp2_txq_pcpu' into a
separate structure 'mvpp2_txq_pcpu_buf', to which a 'size' field is
added. Therefore, instead of having three arrays to allocate/free, we
have a single one, which also improve data locality, reducing the
impact on the CPU cache.

Fixes: 71ce391d ("net: mvpp2: enable proper per-CPU TX buffers unmapping")
Reported-by: NRaphael G <raphael.glon@corp.ovh.com>
Cc: Raphael G <raphael.glon@corp.ovh.com>
Cc: stable@vger.kernel.org
Signed-off-by: NThomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

gcc-7 detects a short memset in mvpp2, introduced in the original
merge of the driver:

drivers/net/ethernet/marvell/mvpp2.c: In function 'mvpp2_cls_init':
drivers/net/ethernet/marvell/mvpp2.c:3296:2: error: 'memset' used with length equal to number of elements without multiplication by element size [-Werror=memset-elt-size]

The result seems to be that we write uninitialized data into the
flow table registers, although we did not get any warning about
that uninitialized data usage.

Using sizeof() lets us initialize then entire array instead.

Fixes: 3f518509 ("ethernet: Add new driver for Marvell Armada 375 network unit")
Signed-off-by: NArnd Bergmann <arnd@arndb.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Implement ethtool::nway_reset using phy_ethtool_nway_reset. We are
already using dev->phydev all over the place so this comes for free.
Signed-off-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

mvneta: min_mtu 68, max_mtu 9676
- mtu validation routine mostly did range check, merge back into
  mvneta_change_mtu for simplicity

mvpp2: min_mtu 68, max_mtu 9676
- mtu validation routine mostly did range check, merge back into
  mvpp2_change_mtu for simplicity

pxa168_eth: min_mtu 68, max_mtu 9500

skge: min_mtu 60, max_mtu 9000

sky2: min_mtu 68, max_mtu 1500 or 9000, depending on hw

CC: netdev@vger.kernel.org
CC: Mirko Lindner <mlindner@marvell.com>
CC: Stephen Hemminger <stephen@networkplumber.org>
CC: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Signed-off-by: NJarod Wilson <jarod@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

of_node_put needs to be called when the device node which is got
from of_parse_phandle has finished using.
Signed-off-by: NPeter Chen <peter.chen@nxp.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

There are two generics functions phy_ethtool_{get|set}_link_ksettings,
so we can use them instead of defining the same code in the driver.
Signed-off-by: NPhilippe Reynes <tremyfr@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The private structure contain a pointer to phydev, but the structure
net_device already contain such pointer. So we can remove the pointer
phydev in the private structure, and update the driver to use the
one contained in struct net_device.
Signed-off-by: NPhilippe Reynes <tremyfr@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

L1_CACHE_BYTES may not be the real cacheline size, use cache_line_size
to determine the cacheline size in runtime.
Signed-off-by: NJisheng Zhang <jszhang@marvell.com>
Suggested-by: NMarcin Wojtas <mw@semihalf.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This is to fix the following maybe-uninitialized warning:

drivers/net/ethernet/marvell/mvpp2.c:6007:18: warning: 'err' may be
used uninitialized in this function [-Wmaybe-uninitialized]
Signed-off-by: NJisheng Zhang <jszhang@marvell.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The mvpp2 ip maybe used in SoCs which may have have 64bytes cacheline
size. Replace the MVPP2_CPU_D_CACHE_LINE_SIZE with L1_CACHE_BYTES.

And since dma_alloc_coherent() is always cacheline size aligned, so
remove the align checks.
Signed-off-by: NJisheng Zhang <jszhang@marvell.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The return value of kzalloc on failure of allocation of memory should
be -ENOMEM and not -1.

Found using Coccinelle. A simplified version of the semantic patch
used is:

//<smpl>
@@
expression *e;
position p,q;
@@

e@q = kzalloc(...);
if@p (e == NULL) {
...
return
- -1
+ -ENOMEM
;
}
//</smpl>

This function may also return -1 after calling mpp2_prs_tcam_port_map_get.
So that the function consistently returns meaningful error values on
failure, the -1 is changed to -EINVAL.
Signed-off-by: NAmitoj Kaur Chawla <amitoj1606@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In hitherto code in case of RX buffer allocation error during refill,
original buffer is pushed to the network stack, but the amount of
available buffer pointers in BM pool is decreased.

This commit fixes the situation by moving refill call before skb_put(),
and returning original buffer pointer to the pool in case of an error.
Signed-off-by: NMarcin Wojtas <mw@semihalf.com>

Fixes: 3f518509 ("ethernet: Add new driver for Marvell Armada 375
network unit")

Cc: <stable@vger.kernel.org> # v3.18+
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Each allocated buffer, whose pointer is put into BM pool is DMA-mapped.
Hence it should be properly unmapped after usage or when removing buffers
from pool.

This commit fixes DMA handling on RX path by adding dma_unmap_single() in
mvpp2_rx() and in mvpp2_bufs_free(). The latter function's argument number
had to be increased for this purpose.
Signed-off-by: NMarcin Wojtas <mw@semihalf.com>

Fixes: 3f518509 ("ethernet: Add new driver for Marvell Armada 375
network unit")

Cc: <stable@vger.kernel.org> # v3.18+
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The Tx descriptor release code currently calls dma_unmap_single() and
dev_kfree_skb_any() if the descriptor is associated with a non-NULL skb.
This condition is true only for the last fragment of the packet.

Since every descriptor's buffer is DMA-mapped it has to be properly
unmapped.
Signed-off-by: NMarcin Wojtas <mw@semihalf.com>

Fixes: 3f518509 ("ethernet: Add new driver for Marvell Armada 375
network unit")

Cc: <stable@vger.kernel.org> # v3.18+
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The PP2 controller is capable of per-CPU TX processing, which means there are
per-CPU banked register sets and queues. Current version of the driver supports
TX packet coalescing - once on given CPU sent packets amount reaches a threshold
value, an IRQ occurs. However, there is a single interrupt line responsible for
CPU0/1 TX and RX events (the latter is not per-CPU, the hardware does not
support RSS).

When the top-half executes the interrupt cause is not known. This is why in
NAPI poll function, along with RX processing, IRQ cause register on both
CPU's is accessed in order to determine on which of them the TX coalescing
threshold might have been reached. Thus the egress processing and releasing the
buffers is able to take place on the corresponding CPU. Hitherto approach lead
to an illegal usage of on_each_cpu function in softirq context.

The problem is solved by resigning from TX coalescing interrupts and separating
egress finalization from NAPI processing. For that purpose a method of using
hrtimer is introduced. In main transmit function (mvpp2_tx) buffers are released
once a software coalescing threshold is reached. In case not all the data is
processed a timer is set on this CPU - in its interrupt context a tasklet is
scheduled in which all queues are processed. At once only one timer per-CPU can
be running, which is controlled by a dedicated flag.

This commit removes TX processing from NAPI polling function, disables hardware
coalescing and enables hrtimer with tasklet, using new per-CPU port structure
(mvpp2_port_pcpu).
Signed-off-by: NMarcin Wojtas <mw@semihalf.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>