In PF_PACKET's packet mmap(), we can avoid using one atomic_inc()
and one atomic_dec() call in skb destructor and use a percpu
reference count instead in order to determine if packets are
still pending to be sent out. Micro-benchmark with [1] that has
been slightly modified (that is, protcol = 0 in socket(2) and
bind(2)), example on a rather crappy testing machine; I expect
it to scale and have even better results on bigger machines:

./packet_mm_tx -s7000 -m7200 -z700000 em1, avg over 2500 runs:

With patch:    4,022,015 cyc
Without patch: 4,812,994 cyc

time ./packet_mm_tx -s64 -c10000000 em1 > /dev/null, stable:

With patch:
  real         1m32.241s
  user         0m0.287s
  sys          1m29.316s

Without patch:
  real         1m38.386s
  user         0m0.265s
  sys          1m35.572s

In function tpacket_snd(), it is okay to use packet_read_pending()
since in fast-path we short-circuit the condition already with
ph != NULL, since we have next frames to process. In case we have
MSG_DONTWAIT, we also do not execute this path as need_wait is
false here anyway, and in case of _no_ MSG_DONTWAIT flag, it is
okay to call a packet_read_pending(), because when we ever reach
that path, we're done processing outgoing frames anyway and only
look if there are skbs still outstanding to be orphaned. We can
stay lockless in this percpu counter since it's acceptable when we
reach this path for the sum to be imprecise first, but we'll level
out at 0 after all pending frames have reached the skb destructor
eventually through tx reclaim. When people pin a tx process to
particular CPUs, we expect overflows to happen in the reference
counter as on one CPU we expect heavy increase; and distributed
through ksoftirqd on all CPUs a decrease, for example. As
David Laight points out, since the C language doesn't define the
result of signed int overflow (i.e. rather than wrap, it is
allowed to saturate as a possible outcome), we have to use
unsigned int as reference count. The sum over all CPUs when tx
is complete will result in 0 again.

The BUG_ON() in tpacket_destruct_skb() we can remove as well. It
can _only_ be set from inside tpacket_snd() path and we made sure
to increase tx_ring.pending in any case before we called po->xmit(skb).
So testing for tx_ring.pending == 0 is not too useful. Instead, it
would rather have been useful to test if lower layers didn't orphan
the skb so that we're missing ring slots being put back to
TP_STATUS_AVAILABLE. But such a bug will be caught in user space
already as we end up realizing that we do not have any
TP_STATUS_AVAILABLE slots left anymore. Therefore, we're all set.

Btw, in case of RX_RING path, we do not make use of the pending
member, therefore we also don't need to use up any percpu memory
here. Also note that __alloc_percpu() already returns a zero-filled
percpu area, so initialization is done already.

  [1] http://wiki.ipxwarzone.com/index.php5?title=Linux_packet_mmapSigned-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

In tpacket_snd(), when we've discovered a first frame that is
not in status TP_STATUS_SEND_REQUEST, and return a NULL buffer,
we exit the send routine in case of MSG_DONTWAIT, since we've
finished traversing the mmaped send ring buffer and don't care
about pending frames.

While doing so, we still unconditionally call an expensive
schedule() in the packet_current_frame() "error" path, which
is unnecessary in this case since it's enough to just quit
the function.

Also, in case MSG_DONTWAIT is not set, we should rather test
for need_resched() first and do schedule() only if necessary
since meanwhile pending frames could already have finished
processing and called skb destructor.
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Most people acquire PF_PACKET sockets with a protocol argument in
the socket call, e.g. libpcap does so with htons(ETH_P_ALL) for
all its sockets. Most likely, at some point in time a subsequent
bind() call will follow, e.g. in libpcap with ...

  memset(&sll, 0, sizeof(sll));
  sll.sll_family          = AF_PACKET;
  sll.sll_ifindex         = ifindex;
  sll.sll_protocol        = htons(ETH_P_ALL);

... as arguments. What happens in the kernel is that already
in socket() syscall, we install a proto hook via register_prot_hook()
if our protocol argument is != 0. Yet, in bind() we're almost
doing the same work by doing a unregister_prot_hook() with an
expensive synchronize_net() call in case during socket() the proto
was != 0, plus follow-up register_prot_hook() with a bound device
to it this time, in order to limit traffic we get.

In the case when the protocol and user supplied device index (== 0)
does not change from socket() to bind(), we can spare us doing
the same work twice. Similarly for re-binding to the same device
and protocol. For these scenarios, we can decrease create/bind
latency from ~7447us (sock-bind-2 case) to ~89us (sock-bind-1 case)
with this patch.

Alternatively, for the first case, if people care, they should
simply create their sockets with proto == 0 argument and define
the protocol during bind() as this saves a call to synchronize_net()
as well (sock-bind-3 case).

In all other cases, we're tied to user space behaviour we must not
change, also since a bind() is not strictly required. Thus, we need
the synchronize_net() to make sure no asynchronous packet processing
paths still refer to the previous elements of po->prot_hook.

In case of mmap()ed sockets, the workflow that includes bind() is
socket() -> setsockopt(<ring>) -> bind(). In that case, a pair of
{__unregister, register}_prot_hook is being called from setsockopt()
in order to install the new protocol receive handler. Thus, when
we call bind and can skip a re-hook, we have already previously
installed the new handler. For fanout, this is handled different
entirely, so we should be good.

Timings on an i7-3520M machine:

  * sock-bind-1:   89 us
  * sock-bind-2: 7447 us
  * sock-bind-3:   75 us

sock-bind-1:
  socket(PF_PACKET, SOCK_RAW, htons(ETH_P_IP)) = 3
  bind(3, {sa_family=AF_PACKET, proto=htons(ETH_P_IP), if=all(0),
           pkttype=PACKET_HOST, addr(0)={0, }, 20) = 0

sock-bind-2:
  socket(PF_PACKET, SOCK_RAW, htons(ETH_P_IP)) = 3
  bind(3, {sa_family=AF_PACKET, proto=htons(ETH_P_IP), if=lo(1),
           pkttype=PACKET_HOST, addr(0)={0, }, 20) = 0

sock-bind-3:
  socket(PF_PACKET, SOCK_RAW, 0) = 3
  bind(3, {sa_family=AF_PACKET, proto=htons(ETH_P_IP), if=lo(1),
           pkttype=PACKET_HOST, addr(0)={0, }, 20) = 0
Signed-off-by: Daniel Borkmann <dborkman@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Fixes: 9d8bf547 ("i40e: associate VMDq queue with VM type")
Reported-by: Stephen Hemminger <stephen@networkplumber.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

Recent commit 438e38fa
("gre_offload: statically build GRE offloading support") added
new module_init/module_exit calls to the gre_offload.c file.

The file is obj-y and can't be anything other than built-in.
Currently it can never be built modular, so using module_init
as an alias for __initcall can be somewhat misleading.

Fix this up now, so that we can relocate module_init from
init.h into module.h in the future.  If we don't do this, we'd
have to add module.h to obviously non-modular code, and that
would be a worse thing.  We also make the inclusion explicit.

Note that direct use of __initcall is discouraged, vs. one
of the priority categorized subgroups.  As __initcall gets
mapped onto device_initcall, our use of device_initcall
directly in this change means that the runtime impact is
zero -- it will remain at level 6 in initcall ordering.

As for the module_exit, rather than replace it with __exitcall,
we simply remove it, since it appears only UML does anything
with those, and even for UML, there is no relevant cleanup
to be done here.

Cc: Eric Dumazet <edumazet@google.com>
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Building resource_tracker.o triggers a GCC warning:
    drivers/net/ethernet/mellanox/mlx4/resource_tracker.c: In function 'mlx4_HW2SW_SRQ_wrapper':
    drivers/net/ethernet/mellanox/mlx4/resource_tracker.c:3202:17: warning: 'srq' may be used uninitialized in this function [-Wmaybe-uninitialized]
      atomic_dec(&srq->mtt->ref_count);
                     ^

This is a false positive. But a cleanup of srq_res_start_move_to() can
help GCC here. The code currently uses a switch statement where a plain
if/else would do, since only two of the switch's four cases can ever
occur. Dropping that switch makes the warning go away.

While we're at it, add some missing braces, and convert state to the
correct type.
Signed-off-by: Paul Bolle <pebolle@tiscali.nl>
Signed-off-by: David S. Miller <davem@davemloft.net>

Building resource_tracker.o triggers a GCC warning:
    drivers/net/ethernet/mellanox/mlx4/resource_tracker.c: In function 'mlx4_HW2SW_CQ_wrapper':
    drivers/net/ethernet/mellanox/mlx4/resource_tracker.c:3019:16: warning: 'cq' may be used uninitialized in this function [-Wmaybe-uninitialized]
      atomic_dec(&cq->mtt->ref_count);
                    ^

This is a false positive. But a cleanup of cq_res_start_move_to() can
help GCC here. The code currently uses a switch statement where an
if/else construct would do too, since only two of the switch's four
cases can ever occur. Dropping that switch makes the warning go away.

While we're at it, add some missing braces.
Signed-off-by: Paul Bolle <pebolle@tiscali.nl>
Signed-off-by: David S. Miller <davem@davemloft.net>

Aaron Brown says:

====================
Intel Wired LAN Driver Updates

This series contains updates to ixgbe and ixgbevf.

John adds rtnl lock / unlock semantics for ixgbe_reinit_locked()
which was being called without the rtnl lock being held.

Jacob corrects an issue where ixgbevf_qv_disable function does not
set the disabled bit correctly.

From the community, Wei uses a type of struct for pci driver-specific
data in ixgbevf_suspend()

Don changes the way we store ring arrays in a manner that allows
support of multiple queues on multiple nodes and creates new ring
initialization functions for work previously done across multiple
functions - making the code closer to ixgbe and hopefully more readable.
He also fixes incorrect fiber eeprom write logic.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

In this code we wanted to set the bit in IXGBE_SFF_SOFT_RS_SELECT_MASK to
the value in rs.  So we really needed a logical or rather than an and, this
patch makes that change.
Signed-off-by: Don Skidmore <donald.c.skidmore@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

This patch creates new functions for ring initialization,
ixgbevf_configure_tx_ring() and ixgbevf_configure_rx_ring(). The work done
in these function previously was spread between several other functions and
this change should hopefully lead to greater readability and make the code
more like ixgbe.  This patch also moves the placement of some older functions
to avoid having to write prototypes.  It also promotes a couple of debug
messages to errors.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: Don Skidmore <donald.c.skidmore@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

This will change how we store rings arrays in the adapter sturct.
We use to have a pointer to an array now we will be using an array
of pointers.  This will allow us to support multiple queues on
muliple nodes at some point we would be able to reallocate the rings
so that each is on a local node if needed.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: Don Skidmore <donald.c.skidmore@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

We had set the pci driver-specific data in ixgbevf_probe() as a type of
struct net_device, so we should use it as netdev in ixgbevf_suspend().
Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The ixgbevf_qv_disable function used by CONFIG_NET_RX_BUSY_POLL is broken,
because it does not properly set the IXGBEVF_QV_STATE_DISABLED bit, indicating
that the q_vector should be disabled (and preventing future locks from
obtaining the vector). This patch corrects the issue by setting the disable
state.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

ixgbe_service_task() is calling ixgbe_reinit_locked() without
the rtnl_lock being held. This is because it is being called
from a worker thread and not a rtnl netlink or dcbnl path.

Add rtnl_{un}lock() semantics. I found this during code review.
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Tested-by: Phil Schmitt <phillip.j.schmitt@intel.com>
Signed-off-by: Aaron Brown <aaron.f.brown@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

eth_type_trans() can read uninitialized memory as drivers
do not necessarily pull more than 14 bytes in skb->head before
calling it.

As David suggested, we can use skb_header_pointer() to
fix this without breaking some drivers that might not expect
eth_type_trans() pulling 2 additional bytes.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Srinivas Kandagatla says:

====================
net: stmmac PM related fixes.

During PM_SUSPEND_FREEZE testing, I have noticed that PM support in STMMAC is
partly broken. I had to re-arrange the code to do PM correctly. There were lot
of things I did not like personally and some bits did not work in the first
place. I thought this is the nice opportunity to clean the mess up.

Here is what I did:
 any
1> Test PM suspend freeze via pm_test
It did not work for following reasons.
 - If the power to gmac is removed when it enters in low power state.
stmmac_resume could not cope up with such behaviour, it was expecting the ip
register contents to be still same as before entering low power, This
assumption is wrong. So I started to add some code to do Hardware
initialization, thats when I started to re-arrange the code. stmmac_open
contains both resource and memory allocations and hardware initialization. I
had to separate these two things in two different functions.

These two patches do that
  net: stmmac: move dma allocation to new function
  net: stmmac: move hardware setup for stmmac_open to new function

And rest of the other patches are fixing the loose ends, things like mdio
reset, which might be necessary in cases likes hibernation(I did not test).

In hibernation cases the driver was just unregistering with subsystems and
releasing resources which I did not like and its not necessary to do this as
part of PM. So using the same stmmac_suspend/resume made more sense for
hibernation cases than using stmmac_open/release.
Also fixed a NULL pointer dereference bug too.

2> Test WOL via PM_SUSPEND_FREEZE
Did get an wakeup interrupt, but could not wakeup a freeze system.
So I had to add pm_wakeup_event to the driver.
net: stmmac: notify the PM core of a wakeup event. patch.

Also few patches like
  net: stmmac: make stmmac_mdio_reset non-static
  net: stmmac: restore pinstate in pm resume.
helps the resume function to reset the phy and put back the pins in default
state.

Changes since RFC:
	- Rebased to net-next on Dave's suggestion.

All these patches are Acked by Peppe.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

In PM_SUSPEND_FREEZE and WOL(Wakeup On Lan) case, when the driver gets a
wakeup event, either the driver or platform specific PM code should notify
the pm core about it, so that the system can wakeup from low power.

In cases where there is no involvement of platform specific PM, it
becomes driver responsibility to notify the PM core to wakeup the
system.

Without this WOL with PM_SUSPEND_FREEZE does not work on STi based SOCs.
Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@st.com>
Acked-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

This patch adds code to restore default pinstate of the pins when it
comes back from low power state. Without this patch the state of the
pins would be unknown and the driver would not work.

This patch also adds code to put the pins in to sleep state when the
driver enters low power state.
Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@st.com>
Acked-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

In hibernation freeze case the driver just releases the resources like
dma buffers, irqs, unregisters the drivers and during restore it does
register, request the resources. This is not really necessary, as part
of power management all the data structures are intact, all the
previously allocated resources can be used after coming out of low
power.

This patch uses the suspend and resume callbacks for freeze and
restore which initializes the hardware correctly without unregistering
or releasing the resources, this should also help in reducing the time
to restore.

Also this patch fixes a bug in stmmac_pltfr_restore and
stmmac_pltfr_freeze where it tries to get hold of platform data via
dev_get_platdata call, which would return NULL in device tree cases and
the next if statement would crash as there is no NULL check.
Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@st.com>
Acked-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The driver PM resume assumes that the IP is still powered up and the
all the register contents are not disturbed when it comes out of low
power suspend case. This assumption is wrong, basically the driver
should not consider any state of registers after it comes out of low
power. However driver can keep the part of the IP powered up if its a
wake up source. But it can not assume the register state of the IP. Also
its possible that SOC glue layer can take the power off the IP if its
not wake-up source to reduce the power consumption.

This patch re initializes hardware by calling stmmac_hw_setup function in
resume case.
Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@st.com>
Acked-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

This patch promotes stmmac_mdio_reset function from static to
non-static, so that power management functions can decide to reset if
the IP comes out from lowe power state specially hibernation cases.
Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@st.com>
Acked-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

This patch moves hardware setup part of the code in stmmac_open to a new
function stmmac_hw_setup, the reason for doing this is to make hw
initialization independent function so that PM functions can re-use it to
re-initialize the IP after returning from low power state.
This will also avoid code duplication across stmmac_resume/restore and
stmmac_open.
Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@st.com>
Acked-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

This patch moves dma resource allocation to a new function
alloc_dma_desc_resources, the reason for moving this to a new function
is to keep the memory allocations in a separate function. One more reason
it to get suspend and hibernation cases working without releasing and
allocating these resources during suspend-resume and freeze-restore
cases.
Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@st.com>
Acked-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

This patch removes gpio_free for reset line of the phy, driver stores
the gpio number in its private data-structure to use in future. As the
driver uses this pin in future this pin should not be freed.
Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@st.com>
Acked-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

This patch adds support to "max-speed" property which is a standard
Ethernet device tree property. max-speed specifies maximum speed
(specified in megabits per second) supported the device.

Depending on the clocking schemes some of the boards can only support
few link speeds, so having a way to limit the link speed in the mac
driver would allow such setups to work reliably.

Without this patch there is no way to tell the driver to limit the
link speed.
Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@st.com>
Acked-by: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Willy Tarreau says:

====================
Assorted mvneta fixes and improvements

this series provides some fixes for a number of issues met with the
mvneta driver, then adds some improvements. Patches 1-5 are fixes
and would be needed in 3.13 and likely -stable. The next ones are
performance improvements and cleanups :

  - driver lockup when reading stats while sending traffic from multiple
    CPUs : this obviously only happens on SMP and is the result of missing
    locking on the driver. The problem was present since the introduction
    of the driver in 3.8. The first patch performs some changes that are
    needed for the second one which actually fixes the issue by using
    per-cpu counters. It could make sense to backport this to the relevant
    stable versions.

  - mvneta_tx_timeout calls various functions to reset the NIC, and these
    functions sleep, which is not allowed here, resulting in a panic.
    Better completely disable this Tx timeout handler for now since it is
    never called. The problem was encountered while developing some new
    features, it's uncertain whether it's possible to reproduce it with
    regular usage, so maybe a backport to stable is not needed.

  - replace the Tx timer with a real Tx IRQ. As first reported by Arnaud
    Ebalard and explained by Eric Dumazet, there is no way this driver
    can work correctly if it uses a driver to recycle the Tx descriptors.
    If too many packets are sent at once, the driver quickly ends up with
    no descriptors (which happens twice as easily in GSO) and has to wait
    10ms for recycling its descriptors and being able to send again. Eric
    has worked around this in the core GSO code. But still when routing
    traffic or sending UDP packets, the limitation is very visible. Using
    Tx IRQs allows Tx descriptors to be recycled when sent. The coalesce
    value is still configurable using ethtool. This fix turns the UDP
    send bitrate from 134 Mbps to 987 Mbps (ie: line rate). It's made of
    two patches, one to add the relevant bits from the original Marvell's
    driver, and another one to implement the change. I don't know if it
    should be backported to stable, as the bug only causes poor performance.

  - Patches 6..8 are essentially cleanups, code deduplication and minor
    optimizations for not re-fetching a value we already have (status).

  - patch 9 changes the prefetch of Rx descriptor from current one to
    next one. In benchmarks, it results in about 1% general performance
    increase on HTTP traffic, probably because prefetching the current
    descriptor does not leave enough time between the start of prefetch
    and its usage.

  - patch 10 implements support for build_skb() on Rx path. The driver
    now preallocates frags instead of skbs and builds an skb just before
    delivering it. This results in a 2% performance increase on HTTP
    traffic, and up to 5% on small packet Rx rate.

  - patch 11 implements rx_copybreak for small packets (256 bytes). It
    avoids a dma_map_single()/dma_unmap_single() and increases the Rx
    rate by 16.4%, from 486kpps to 573kpps. Further improvements up to
    711kpps are possible depending how the DMA is used.

  - patches 12 and 13 are extra cleanups made possible by some of the
    simplifications above.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

The function return parameter is not used in mvneta_tx_done_gbe(),
where the function is called. This patch makes the function return
void.
Reviewed-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

mvneta_tx_done_gbe() return value and third parameter are no more
used. This patch changes the function prototype and removes a useless
variable where the function is called.
Reviewed-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

calling dma_map_single()/dma_unmap_single() is quite expensive compared
to copying a small packet. So let's copy short frames and keep the buffers
mapped. We set the limit to 256 bytes which seems to give good results both
on the XP-GP board and on the AX3/4.

The Rx small packet rate increased by 16.4% doing this, from 486kpps to
573kpps. It is worth noting that even the call to the function
dma_sync_single_range_for_cpu() is expensive (300 ns) although less
than dma_unmap_single(). Without it, the packet rate raises to 711kpps
(+24% more). Thus on systems where coherency from device to CPU is
guaranteed by a snoop control unit, this patch should provide even more
gains, and probably rx_copybreak could be increased.

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

Make use of build_skb() to allocate frags on the RX path. When frag size
is lower than a page size, we can use netdev_alloc_frag(), and we fall back
to kmalloc() for larger sizes. The frag size is stored into the mvneta_port
struct. The alloc/free functions check the frag size to decide what alloc/
free method to use. MTU changes are safe because the MTU change function
stops the device and clears the queues before applying the change.

With this patch, I observed a reproducible 2% performance improvement on
HTTP-based benchmarks, and 5% on small packet RX rate.

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

Currently, the mvneta driver tries to prefetch the current Rx
descriptor during read. Tests have shown that prefetching the
next one instead increases general performance by about 1% on
HTTP traffic.

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

At several places, we already know the value of the rx status but
we call functions which dereference the pointer again to get it
and don't need the descriptor for anything else. Simplify this
task by replacing the rx desc pointer by the status word itself.

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

Make mvneta_rxq_fill() use mvneta_rx_refill() instead of using
duplicate code.

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

Currently, mvneta_txq_bufs_free() calls mvneta_tx_done_policy() with
a non-null cause to retrieve the pointer to the next queue to process.
There are useless tests on the return queue number and on the pointer,
all of which are well defined within a known limited set. This code
path is fast, although not critical. Removing 3 tests here that the
compiler could not optimize (verified) is always desirable.

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

Right now the mvneta driver doesn't handle Tx IRQ, and relies on two
mechanisms to flush Tx descriptors : a flush at the end of mvneta_tx()
and a timer. If a burst of packets is emitted faster than the device
can send them, then the queue is stopped until next wake-up of the
timer 10ms later. This causes jerky output traffic with bursts and
pauses, making it difficult to reach line rate with very few streams.

A test on UDP traffic shows that it's not possible to go beyond 134
Mbps / 12 kpps of outgoing traffic with 1500-bytes IP packets. Routed
traffic tends to observe pauses as well if the traffic is bursty,
making it even burstier after the wake-up.

It seems that this feature was inherited from the original driver but
nothing there mentions any reason for not using the interrupt instead,
which the chip supports.

Thus, this patch enables Tx interrupts and removes the timer. It does
the two at once because it's not really possible to make the two
mechanisms coexist, so a split patch doesn't make sense.

First tests performed on a Mirabox (Armada 370) show that less CPU
seems to be used when sending traffic. One reason might be that we now
call the mvneta_tx_done_gbe() with a mask indicating which queues have
been done instead of looping over all of them.

The same UDP test above now happily reaches 987 Mbps / 87.7 kpps.
Single-stream TCP traffic can now more easily reach line rate. HTTP
transfers of 1 MB objects over a single connection went from 730 to
840 Mbps. It is even possible to go significantly higher (>900 Mbps)
by tweaking tcp_tso_win_divisor.

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Cc: Arnaud Ebalard <arno@natisbad.org>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

Marvell has not published the chip's datasheet yet, so it's very hard
to find the relevant bits to manipulate to change the IRQ behaviour.
Fortunately, these bits are described in the proprietary LSP patch set
which is publicly available here :

    http://www.plugcomputer.org/downloads/mirabox/

So let's put them back in the driver in order to reduce the burden of
current and future maintenance.

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

If a queue timeout is reported, we can oops because of some
schedules while the caller is atomic, as shown below :

  mvneta d0070000.ethernet eth0: tx timeout
  BUG: scheduling while atomic: bash/1528/0x00000100
  Modules linked in: slhttp_ethdiv(C) [last unloaded: slhttp_ethdiv]
  CPU: 2 PID: 1528 Comm: bash Tainted: G        WC   3.13.0-rc4-mvebu-nf #180
  [<c0011bd9>] (unwind_backtrace+0x1/0x98) from [<c000f1ab>] (show_stack+0xb/0xc)
  [<c000f1ab>] (show_stack+0xb/0xc) from [<c02ad323>] (dump_stack+0x4f/0x64)
  [<c02ad323>] (dump_stack+0x4f/0x64) from [<c02abe67>] (__schedule_bug+0x37/0x4c)
  [<c02abe67>] (__schedule_bug+0x37/0x4c) from [<c02ae261>] (__schedule+0x325/0x3ec)
  [<c02ae261>] (__schedule+0x325/0x3ec) from [<c02adb97>] (schedule_timeout+0xb7/0x118)
  [<c02adb97>] (schedule_timeout+0xb7/0x118) from [<c0020a67>] (msleep+0xf/0x14)
  [<c0020a67>] (msleep+0xf/0x14) from [<c01dcbe5>] (mvneta_stop_dev+0x21/0x194)
  [<c01dcbe5>] (mvneta_stop_dev+0x21/0x194) from [<c01dcfe9>] (mvneta_tx_timeout+0x19/0x24)
  [<c01dcfe9>] (mvneta_tx_timeout+0x19/0x24) from [<c024afc7>] (dev_watchdog+0x18b/0x1c4)
  [<c024afc7>] (dev_watchdog+0x18b/0x1c4) from [<c0020b53>] (call_timer_fn.isra.27+0x17/0x5c)
  [<c0020b53>] (call_timer_fn.isra.27+0x17/0x5c) from [<c0020cad>] (run_timer_softirq+0x115/0x170)
  [<c0020cad>] (run_timer_softirq+0x115/0x170) from [<c001ccb9>] (__do_softirq+0xbd/0x1a8)
  [<c001ccb9>] (__do_softirq+0xbd/0x1a8) from [<c001cfad>] (irq_exit+0x61/0x98)
  [<c001cfad>] (irq_exit+0x61/0x98) from [<c000d4bf>] (handle_IRQ+0x27/0x60)
  [<c000d4bf>] (handle_IRQ+0x27/0x60) from [<c000843b>] (armada_370_xp_handle_irq+0x33/0xc8)
  [<c000843b>] (armada_370_xp_handle_irq+0x33/0xc8) from [<c000fba9>] (__irq_usr+0x49/0x60)

Ben Hutchings attempted to propose a better fix consisting in using a
scheduled work for this, but while it fixed this panic, it caused other
random freezes and panics proving that the reset sequence in the driver
is unreliable and that additional fixes should be investigated.

When sending multiple streams over a link limited to 100 Mbps, Tx timeouts
happen from time to time, and the driver correctly recovers only when the
function is disabled.

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Cc: Ben Hutchings <ben@decadent.org.uk>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

Stats writers are mvneta_rx() and mvneta_tx(). They don't lock anything
when they update the stats, and as a result, it randomly happens that
the stats freeze on SMP if two updates happen during stats retrieval.
This is very easily reproducible by starting two HTTP servers and binding
each of them to a different CPU, then consulting /proc/net/dev in loops
during transfers, the interface should immediately lock up. This issue
also randomly happens upon link state changes during transfers, because
the stats are collected in this situation, but it takes more attempts to
reproduce it.

The comments in netdevice.h suggest using per_cpu stats instead to get
rid of this issue.

This patch implements this. It merges both rx_stats and tx_stats into
a single "stats" member with a single syncp. Both mvneta_rx() and
mvneta_rx() now only update the a single CPU's counters.

In turn, mvneta_get_stats64() does the summing by iterating over all CPUs
to get their respective stats.

With this change, stats are still correct and no more lockup is encountered.

Note that this bug was present since the first import of the mvneta
driver.  It might make sense to backport it to some stable trees. If
so, it depends on "d33dc73 net: mvneta: increase the 64-bit rx/tx stats
out of the hot path".

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

Better count packets and bytes in the stack and on 32 bit then
accumulate them at the end for once. This saves two memory writes
and two memory barriers per packet. The incoming packet rate was
increased by 4.7% on the Openblocks AX3 thanks to this.

Cc: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
Cc: Gregory CLEMENT <gregory.clement@free-electrons.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Tested-by: Arnaud Ebalard <arno@natisbad.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
Signed-off-by: David S. Miller <davem@davemloft.net>

None of these files are actually using any __init type directives
and hence don't need to include <linux/init.h>.   Most are just a
left over from __devinit and __cpuinit removal, or simply due to
code getting copied from one driver to the next.

This covers everything under drivers/net except for wireless, which
has been submitted separately.
Signed-off-by: Paul Gortmaker <paul.gortmaker@windriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>