RSC will flush its descriptors every time the interrupt throttle timer
expires.  In addition there are known issues with RSC when the rx-usecs
value is set too low.  As such we are forced to clear the RSC_ENABLED bit
and reset the adapter when the rx-usecs value is set too low.

However we do not need to clear the NETIF_F_LRO flag because it is used to
indicate that the user wants to leave the LRO feature enabled, and in fact
with this change we will now re-enable RSC as soon as the rx-usecs value is
increased and the flag is still set.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change makes it so that we perform link setup with interrupts
disabled. If the SFP has not been detected previously we will schedule the
SFP detection task to run in order to detect link.  By doing this we avoid
the possibility of interrupts firing in the middle of our link setup during
ixgbe_up_complete.

In addition this change makes it so that the multi-speed fiber setup and SFP
setup are not mutually exclusive.  The addresses issues seen in which a
link would only come up at 1G on some multi-speed fiber modules.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change adds a set of state flags to the rings that allow them to
independently function allowing for features like RSC, packet split, and
TX hang detection to be done per ring instead of for the entire device.

This is accomplished by re-purposing the flow director reinit_state member
and making it a global state instead since a long for a single bit flag is
a bit wasteful.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This is the start of work to sort out what belongs in the rings and what
belongs in the q_vector. Items like the CPU variable for make much more
sense in the q_vector since the CPU is a per-interrupt thing rather than a
per ring thing.
I also added a back-pointer from the ring to the q_vector.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change moves an adapter pointer into the private portion of the
pci_dev instead of a pointer to the netdev. The reason for this change is
because in most cases we just want the adapter anyway. In addition as we
start moving toward multiple netdevs per port we may want to move the
adapter pointer out of the netdevs entirely.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Missed some code that was left floating around in the DCB configuration
for the TXDCTL register.  As a result the register was being messed with in
two different spots when we only needed to do the change once.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

The main reason for this change is to keep the suspend/resume logic matched
up. The clear_interrupt_scheme function will disable MSI-X which will
effect the PCIe configuration space. Therefore we will want to do it before
we save state to avoid having the interrupt state restored by
pci_restore_state, and then trying to re-enable MSI/MSI-X interrupts via
ixgbe_setup_interrupt_scheme.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change places a netdev pointer directly into the ring structure. This
way we can avoid having to determine which netdev we are supposed to be
using and can just access the one on the ring directly.
As a result of this change further collapse of the code is possible by
dropping the adapter from ixgbe_alloc_rx_buffers, and the netdev pointer
from ixgbe_xmit_frame_ring_adv and ixgbe_maybe_stop_tx.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change moved some of the RX and TX stats into separate structures and
them placed those structures in a union in order to help reduce the size of
the ring structure.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change is meant to simplify DMA map/unmap by providing a device
pointer. As a result the adapter pointer can be dropped from many of
the calls.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change drops ring->head since it is not used in any hot-path and can
easily be determined using IXGBE_[RT]DH(ring->reg_idx).

It also changes ring->tail into a true pointer so we can avoid unnecessary
pointer math to find the location of the tail.

In addition I also dropped the setting of head and tail in
ixgbe_clean_[rx|tx]_ring. The only location that should be setting the head
and tail values is ixgbe_configure_[rx|tx]_ring and that is only while the
queue is disabled.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change re-orders alloc_rx_buffers to make better use of the packet
split enabled flag.  The new setup should require less branching in the
code since now we are down to fewer if statements since we either are
handling packet split or aren't.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

This change simplifies the work being done by the TX interrupt handler and
pushes it into the tx_map call. This allows for fewer cache misses since
the TX cleanup now accesses almost none of the skb members.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

There is no need to reset the adapter when changing the Rx checksum
settings. Since the only change is a software flag we can disable it
without needing to reset the entire adapter.
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

The maximum credits per traffic class only needs to be greater
then the TSO size for 82598 devices. The 82599 devices do not
have this requirement so only do this test for 82598 devices.
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Currently the high and low water marks for PFC are being set
conservatively for jumbo frames. This means the RX buffers
are being underutilized in the default 1500 MTU. This patch
fixes this so that the water marks are set as described in
the data sheet considering the MTU size.

The equation used is,

RTT * 1.44 + MTU * 1.44 + MTU

Where RTT is the round trip time and MTU is the max frame size
in KB. To avoid floating point arithmetic FC_HIGH_WATER is
defined

((((RTT + MTU) * 144) + 99) / 100) + MTU

This changes how the hardware field fc.low_water and
fc.high_water are used. With this change they are no longer
storing the actual low water and high water markers but are
storing the required head room in the buffer. This simplifies
the logic and we do not need to account for the size of the
buffer when setting the thresholds.

Testing with iperf and 16 threads showed a slight uptick in
throughput over a single traffic class .1-.2Gbps and a reduction
in pause frames. Without the patch a 30 second run would show
~10-15 pause frames being transmitted with the patch ~2-5 are
seen. Test were run back to back with 82599.

Note RXPBSIZE is in KB and low and high water marks fields are
also in KB. However the FCRT* registers are 32B granularity and
right shifted 5 into the register,

(((rx_pbsize - water_mark) * 1024) / 32) << 5

is the most explicit conversion here we simplify

(rx_pbsize - water_mark) * 32 << 5 = (rx_pbsize - water_mark) << 10

This patch updates the PFC thresholds and legacy FC thresholds.
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Update version string and copyright notice.
Signed-off-by: Greg Rose <gregory.v.rose@intel.com>
Tested-by: Emil Tantilov <emil.s.tantilov@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

"cat /proc/net/dev" uses RCU protection only.

Its quite possible we call a driver get_stats() method while device is
dismantling and freeing its data structures.

So get_stats() methods must be very careful not accessing driver private
data without appropriate locking.

In ixgbe case, we access rx_ring pointers. These pointers are freed in
ixgbe_clear_interrupt_scheme() and set to NULL, this can trigger NULL
dereference in ixgbe_get_stats64()

A possible fix is to use RCU locking in ixgbe_get_stats64() and defer
rx_ring freeing after a grace period in ixgbe_clear_interrupt_scheme()
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Reported-by: Tantilov, Emil S <emil.s.tantilov@intel.com>
Tested-by: Ross Brattain <ross.b.brattain@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>

Right now, fields in struct sock are not optimally ordered, because each
path (RX softirq, TX completion, RX user,  TX user) has to touch fields
that are contained in many different cache lines.

The really critical thing is to shrink number of cache lines that are
used at RX softirq time : CPU handling softirqs for a device can receive
many frames per second for many sockets. If load is too big, we can drop
frames at NIC level. RPS or multiqueue cards can help, but better reduce
latency if possible.

This patch starts with UDP protocol, then additional patches will try to
reduce latencies of other ones as well.

At RX softirq time, fields of interest for UDP protocol are :
(not counting ones in inet struct for the lookup)

Read/Written:
sk_refcnt   (atomic increment/decrement)
sk_rmem_alloc & sk_backlog.len (to check if there is room in queues)
sk_receive_queue
sk_backlog (if socket locked by user program)
sk_rxhash
sk_forward_alloc
sk_drops

Read only:
sk_rcvbuf (sk_rcvqueues_full())
sk_filter
sk_wq
sk_policy[0]
sk_flags

Additional notes :

- sk_backlog has one hole on 64bit arches. We can fill it to save 8
bytes.
- sk_backlog is used only if RX sofirq handler finds the socket while
locked by user.
- sk_rxhash is written only once per flow.
- sk_drops is written only if queues are full

Final layout :

[1] One section grouping all read/write fields, but placing rxhash and
sk_backlog at the end of this section.

[2] One section grouping all read fields in RX handler
   (sk_filter, sk_rcv_buf, sk_wq)

[3] Section used by other paths

I'll post a patch on its own to put sk_refcnt at the end of struct
sock_common so that it shares same cache line than section [1]

New offsets on 64bit arch :

sizeof(struct sock)=0x268
offsetof(struct sock, sk_refcnt)  =0x10
offsetof(struct sock, sk_lock)    =0x48
offsetof(struct sock, sk_receive_queue)=0x68
offsetof(struct sock, sk_backlog)=0x80
offsetof(struct sock, sk_rmem_alloc)=0x80
offsetof(struct sock, sk_forward_alloc)=0x98
offsetof(struct sock, sk_rxhash)=0x9c
offsetof(struct sock, sk_rcvbuf)=0xa4
offsetof(struct sock, sk_drops) =0xa0
offsetof(struct sock, sk_filter)=0xa8
offsetof(struct sock, sk_wq)=0xb0
offsetof(struct sock, sk_policy)=0xd0
offsetof(struct sock, sk_flags) =0xe0

Instead of :

sizeof(struct sock)=0x270
offsetof(struct sock, sk_refcnt)  =0x10
offsetof(struct sock, sk_lock)    =0x50
offsetof(struct sock, sk_receive_queue)=0xc0
offsetof(struct sock, sk_backlog)=0x70
offsetof(struct sock, sk_rmem_alloc)=0xac
offsetof(struct sock, sk_forward_alloc)=0x10c
offsetof(struct sock, sk_rxhash)=0x128
offsetof(struct sock, sk_rcvbuf)=0x4c
offsetof(struct sock, sk_drops) =0x16c
offsetof(struct sock, sk_filter)=0x198
offsetof(struct sock, sk_wq)=0x88
offsetof(struct sock, sk_policy)=0x98
offsetof(struct sock, sk_flags) =0x130
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

UDP sockets refcount is usually 2, unless an incoming frame is going to
be queued in receive or backlog queue.

Using atomic_inc_not_zero_hint() permits to reduce latency, because
processor issues less memory transactions.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Now vlan are lockless, we dont need special ndo_select_queue() logic.
dev_pick_tx() will do the multiqueue stuff on the real device transmit.
Suggested-by: Jesse Gross <jesse@nicira.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

vlan is a stacked device, like tunnels. We should use the lockless
mechanism we are using in tunnels and loopback.

This patch completely removes locking in TX path.

tx stat counters are added into existing percpu stat structure, renamed
from vlan_rx_stats to vlan_pcpu_stats.

Note : this partially reverts commit 2e59af3d (vlan: multiqueue vlan
device)
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

macvlan is a stacked device, like tunnels. We should use the lockless
mechanism we are using in tunnels and loopback.

This patch completely removes locking in TX path.

tx stat counters are added into existing percpu stat structure, renamed
from rx_stats to pcpu_stats.

Note : this reverts commit 2c114553 (macvlan: add multiqueue
capability)

Note : rx_errors converted to a 32bit counter, like tx_dropped, since
they dont need 64bit range.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Cc: Patrick McHardy <kaber@trash.net>
Cc: Ben Greear <greearb@candelatech.com>
Cc: Ben Hutchings <bhutchings@solarflare.com>
Acked-by: Patrick McHardy <kaber@trash.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

packet: Enhance AF_PACKET implementation to not require high order contiguous memory allocation (v4)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Version 4 of this patch.

Change notes:
1) Removed extra memset.  Didn't think kcalloc added a GFP_ZERO the way kzalloc did :)

Summary:
It was shown to me recently that systems under high load were driven very deep
into swap when tcpdump was run.  The reason this happened was because the
AF_PACKET protocol has a SET_RINGBUFFER socket option that allows the user space
application to specify how many entries an AF_PACKET socket will have and how
large each entry will be.  It seems the default setting for tcpdump is to set
the ring buffer to 32 entries of 64 Kb each, which implies 32 order 5
allocation.  Thats difficult under good circumstances, and horrid under memory
pressure.

I thought it would be good to make that a bit more usable.  I was going to do a
simple conversion of the ring buffer from contigous pages to iovecs, but
unfortunately, the metadata which AF_PACKET places in these buffers can easily
span a page boundary, and given that these buffers get mapped into user space,
and the data layout doesn't easily allow for a change to padding between frames
to avoid that, a simple iovec change is just going to break user space ABI
consistency.

So I've done this, I've added a three tiered mechanism to the af_packet set_ring
socket option.  It attempts to allocate memory in the following order:

1) Using __get_free_pages with GFP_NORETRY set, so as to fail quickly without
digging into swap

2) Using vmalloc

3) Using __get_free_pages with GFP_NORETRY clear, causing us to try as hard as
needed to get the memory

The effect is that we don't disturb the system as much when we're under load,
while still being able to conduct tcpdumps effectively.

Tested successfully by me.
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Acked-by: Maciej Żenczykowski <zenczykowski@gmail.com>
Reported-by: Maciej Żenczykowski <zenczykowski@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Using %pV reduces the number of printk calls and
eliminates any possible message interleaving from
other printk calls.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The changed functions do not modify the NL messages and/or attributes
at all. They should use const (similar to strchr), so that callers
which have a const nlmsg/nlattr around can make use of them without
casting.

While at it, constify a data array.
Signed-off-by: Jan Engelhardt <jengelh@medozas.de>
Signed-off-by: David S. Miller <davem@davemloft.net>

Fix ref count bug introduced by

commit 2de79570
Author: Lorenzo Colitti <lorenzo@google.com>
Date:   Wed Oct 27 18:16:49 2010 +0000

ipv6: addrconf: don't remove address state on ifdown if the address
is being kept

Fix logic so that addrconf_ifdown() decrements the inet6_ifaddr
refcnt correctly with in6_ifa_put().
Reported-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: John Fastabend <john.r.fastabend@intel.com>
Acked-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

ERROR: "netif_get_vlan_features" [drivers/net/xen-netfront.ko] undefined!
Reported-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: David S. Miller <davem@davemloft.net>

Fix data type of argument passed to pci_alloc_consistent and pci_free_consistent routines.
Signed-off-by: Vasanthy Kolluri <vkolluri@cisco.com>
Signed-off-by: Roopa Prabhu <roprabhu@cisco.com>
Signed-off-by: David Wang <dwang2@cisco.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Fallout from the TIOCGICOUNT work
Signed-off-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

br_port_get() renamed to br_port_get_rtnl() to make clear RTNL is held.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The macro br_port_exists() is not enough protection when only
RCU is being used. There is a tiny race where other CPU has cleared port
handler hook, but is bridge port flag might still be set.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Suggested by Eric's bridge RCU changes.
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Add br_should_route_hook_t typedef, this is the only way we can
get a clean RCU implementation for function pointer.

Move route_hook to location where it is used.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Add modern __rcu annotatations to bridge multicast table.
Use newer hlist macros to avoid direct access to hlist internals.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

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

Signed-off-by: Joe Perches <joe@perches.com>
Acked-by: Sjur Braendeland <sjur.brandeland@stericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

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

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