When many cpus compete for sending frames on a given qdisc, the qdisc
spinlock suffers from very high contention.

The cpu owning __QDISC_STATE_RUNNING bit has same priority to acquire
the lock, and cannot dequeue packets fast enough, since it must wait for
this lock for each dequeued packet.

One solution to this problem is to force all cpus spinning on a second
lock before trying to get the main lock, when/if they see
__QDISC_STATE_RUNNING already set.

The owning cpu then compete with at most one other cpu for the main
lock, allowing for higher dequeueing rate.

Based on a previous patch from Alexander Duyck. I added the heuristic to
avoid the atomic in fast path, and put the new lock far away from the
cache line used by the dequeue worker. Also try to release the busylock
lock as late as possible.

Tests with following script gave a boost from ~50.000 pps to ~600.000
pps on a dual quad core machine (E5450 @3.00GHz), tg3 driver.
(A single netperf flow can reach ~800.000 pps on this platform)

for j in `seq 0 3`; do
  for i in `seq 0 7`; do
    netperf -H 192.168.0.1 -t UDP_STREAM -l 60 -N -T $i -- -m 6 &
  done
done
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Acked-by: NAlexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

In the worst case, when the first loop breaks an the end of the slave list,
the slave list is iterated through twice. This patch reduces this
function only to one loop. Also makes it simpler.
Signed-off-by: NJiri Pirko <jpirko@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This is stored but never restored. So remove this as it is useless.
Signed-off-by: NJiri Pirko <jpirko@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Move the code that copies slave's mac address in case that's the first slave into
bond_enslave. Ifenslave app does this also but that's not a problem. This is
something that should be done in bond_enslave, and it shound not matter from
where is it called.
Signed-off-by: NJiri Pirko <jpirko@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

- don't free bus->irq (obsoleted by ca816d98)
- don't dispose irqs (should be done in of_mdiobus_register())
- use fec-pointer consistently in transfer()
- use resource_size()
- cosmetic fixes
Signed-off-by: NWolfram Sang <w.sang@pengutronix.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch makes bonding_store_slaves function nicer and easier to understand.
Signed-off-by: NJiri Pirko <jpirko@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

(it's actually the same as v1)

Remove checks that duplicates similar checks in bond_enslave.
Signed-off-by: NJiri Pirko <jpirko@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

V1->V2: corrected res/ret use

For some reason, MTU handling (storing, and restoring) is taking  place in
bond_sysfs. The correct place for this code is in bond_enslave, bond_release.
So move it there.
Signed-off-by: NJiri Pirko <jpirko@redhat.com>
Signed-off-by: NJay Vosburgh <fubar@us.ibm.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NJiri Pirko <jpirko@redhat.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

__QDISC_STATE_RUNNING is always changed while qdisc lock is held.

We can avoid two atomic operations in xmit path, if we move this bit in
a new __state container.

Location of this __state container is carefully chosen so that fast path
only dirties one qdisc cache line.

THROTTLED bit could later be moved into this __state location too, to
avoid dirtying first qdisc cache line.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Define three helpers to manipulate QDISC_STATE_RUNNIG flag, that a
second patch will move on another location.
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Added support for NIC functions that work in non privileged mode where these
functions are privileged to do IO only, the control operations are handled via
privileged functions.
Bumped up version number to 5.0.3.
Signed-off-by: NAnirban Chakraborty <anirban.chakraborty@qlogic.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Following changes have been added to enable the adapter to work in
NIC partitioning mode where multiple PCI functions of an adapter port can
be configured to work as NIC functions. The first function that is enumerated on
the PCI bus assumes the role of management function which, besides being able
to do all the NIC functionality, can configure other NIC partitions. Other NIC
functions can be configured as privileged or non privileged functions.
Privileged function can not configure other NIC functions but can do all the
NIC functionality including any firmware initialization, chip reset etc. Non
privileged functions can do only basic IO. For chip reset etc, it depends on the
privilege or management function.

1. Added code to determine PCI function number independent of kernel API.
2. Added Driver - FW version 2.0 support.
3. Changed producer and consumer register offset calculation.
4. Added management and privileged operation modes for npar functions. A module
 parameter has been added to control it.
5. Added support for configuring the eswitch in the adapter.
Signed-off-by: NAnirban Chakraborty <anirban.chakraborty@qlogic.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

A single shared memory region used to communicate with firmware is
mapped into both PCI PFs of the SFC9020 and SFL9021.  Drivers must be
able to identify which port they are addressing in order to use the
correct sub-region.  Currently we use the PCI function number, but the
PCI address may be virtualised.  Use the CS_PORT_NUM register field
defined for just this purpose.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

rx_errors is defined as 'bad packets received', but we are currently
including various overflow errors as well.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Insert a structure at the start of the shared page that
tracks the dma mapping refcnt. DMA into the next cache
line of the (shared) page (plus EFX_PAGE_IP_ALIGN).

When recycling a page, check the page refcnt. If the
page is otherwise unused, then resurrect the other
receive buffer that previously referenced the page.
Be careful not to overflow the receive ring, since we
can now resurrect n receive buffers in a row.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

The cut-through design of the receive path means that packets that
fail to match the appropriate MAC filter are not discarded at the MAC
but are flagged in the completion event as 'to be discarded'.  On
networks with heavy multicast traffic, this can account for a
significant proportion of received packets, so it is worthwhile to
recycle the buffer immediately in this case rather than freeing it
and then reallocating it shortly after.

The only complication here is dealing with a page shared
between two receive buffers. In that case, we need to be
careful to free the dma mapping when both buffers have
been free'd by the kernel. This means that we can only
recycle such a page if both receive buffers are discarded.
Unfortunately, in an environment with 1500mtu,
rx_alloc_method=PAGE, and a mixture of discarded and
not-discarded frames hitting the same receive queue,
buffer recycling won't always be possible.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

- Pull the loop handling into efx_init_rx_buffers_(skb|page)
- Remove rx_queue->buf_page, and associated clean up code
- Remove unmap_addr, since unmap_addr is trivially calculable

This will allow us to recycle discarded buffers directly
from efx_rx_packet(), since will never be in the middle of
splitting a page.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Ensure that efx_fast_push_rx_descriptors() must only run
from efx_process_channel() [NAPI], or when napi_disable()
has been executed.

Reimplement the slow fill by sending an event to the
channel, so that NAPI runs, and hanging the subsequent
fast fill off the event handler. Replace the sfc_refill
workqueue and delayed work items with a timer. We do
not need to stop this timer in efx_flush_all() because
it's safe to send the event always; receiving it will
be delayed until NAPI is restarted.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Formerly, efx_test_eventq_irq() assumed it was the only user of
driver generated events. Allow it to interoperate with other users.

We can create more than 16 channels, so align event codes with
a multiple of 256 not 16.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

It's been observed that some phys (such as the qt2025c) can
do down-up-down-up transitions, presumably as pcs block lock
settles down.

The loopback selftest will start sending data immediately
after the link comes up. Work around this by waiting for
the link state to stay up for two consecutive polls, rather
than one.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

All of the ethtool code paths keep them in sync, but we need
to ensure they are sync'd at start of day. Matches the sft9001
driver.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Under certain conditions a PHY may backpressure Falcon B0
in such a way that flushes timeout. In normal circumstances
the phy poller would fix the PHY, and the flush could complete.

But efx_nic_flush_queues() is always called after efx_stop_all(),
so the poller has been stopped. Even if this weren't the case,
how long would we have to wait for the poller to fix this? And
several callers of efx_nic_flush_queues() are about to reset
the device anyway - so we don't need to do anything.

Work around this bug by scheduling a reset. Ensure that the
MAC is never rewired back into the datapath before the reset
runs (we already ignore all rx events anyway).
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

efx_pm_freeze() sets efx->state = STATE_FINI, which means
efx_reset_work() will abort any scheduled resets.

efx_pm_thaw() should reschedule efx_reset_work() again,
since a freeze/thaw will not have reset the hardware.

This bug was spotted by inspection - there is no real world example of
this happening.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Most of its members are constant capabilities, not configuration.  The
new name is also consistent with the name of the pointer to it in
struct efx_nic and the names of structures used by other PHY drivers.
Signed-off-by: NBen Hutchings <bhutchings@solarflare.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Add AR8151 v2.0 Gigabit 1000 support
Change jumbo frame size to 6K
Update L0s/L1 rountine
        when link speed is 100M or 1G, set L1 link timer to 4 for l1d_2 and l2c_b2
        set L1 link timer to 7 for l2c_b, set L1 link timer to 0xF for others.
Update atl1c_suspend routine
	just refactory the function, add atl1c_phy_power_saving routine,
	when Wake On Lan enable, this func will be called to save power,
	it will reautoneg PHY to 10/100M speed depend on the link
	partners link capability.
Update atl1c_configure_des_ring
        do not use l2c_b default SRAM configuration.
Signed-off-by: NJie Yang <Jie.Yang@atheros.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

As noticed by Julia Lawall, ipip6_tunnel_add_prl() incorrectly calls 
kzallloc(..., GFP_KERNEL) while a spinlock is held. She provided
a patch to use GFP_ATOMIC instead.

One possibility would be to convert this spinlock to a mutex, or
preallocate the thing before taking the lock.

After RCU conversion, it appears we dont need this lock, since 
caller already holds RTNL
Signed-off-by: NEric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch adds support for the IFF_ALLMULTI flag. Previously only the
IFF_PROMISC flag was supported.
Signed-off-by: NRichard Cochran <richard.cochran@omicron.at>
Acked-By: NKrzysztof Halasa <khc@pm.waw.pl>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

As mbox polling is done only in process context, it is better to
use schedule_timeout() instead of udelay().
Signed-off-by: NSathya Perla <sathyap@serverengines.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch adds cleanup code (things like unregistering irq,
disabling napi etc) to be_open() when an error occurs inside the
routine.
Signed-off-by: NSathya Perla <sathyap@serverengines.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Based on a patch from http://simon.baatz.info/wol-support-for-an983b/

Tested to resume from suspend by magic packet.
Signed-off-by: NSteven Walter <stevenrwalter@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

During the first suspend the chip would refuse to enter D3.  Subsequent
suspends worked okay.  During resume the chip is commanded into D0.
Doing so during initialization fixes the initial suspend.
Signed-off-by: NSteven Walter <stevenrwalter@gmail.com>
Signed-off-by: NGrant Grundler <grundler@parisc-linux.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NFlorian Fainelli <florian@openwrt.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

This patch adds support for using phylib and adds the required mdiobus driver
stubs. This allows for less code to be present in the driver and removes
the PHY status specific timer which is now handled by phylib directly.
Signed-off-by: NFlorian Fainelli <florian@openwrt.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

 - shuffle around functions to get rid of forward declarations
 - fix some CodingStyle and indentation issues
 - last but not least, get rid of the following CONFIG_MODULE=n warning:

	drivers/net/arcnet/capmode.c:52: warning: ‘capmode_proto’ defined but not used
Signed-off-by: NDaniel Mack <daniel@caiaq.de>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Signed-off-by: NDavid S. Miller <davem@davemloft.net>