dpdk.git
8 years agoapp/test: fix check for virtual ring device
Olivier Matz [Tue, 15 Mar 2016 12:25:29 +0000 (13:25 +0100)]
app/test: fix check for virtual ring device

This test expects that a vdev is instanciated on the command
line. If it's not the case, just skip this part.

Fixes: 4ea3801b3222 ("app/test: fix ring unit test")

Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
8 years agoapp/test: display command line when starting
Olivier Matz [Tue, 15 Mar 2016 12:25:22 +0000 (13:25 +0100)]
app/test: display command line when starting

It's useful for debug purposes to see the command line used by
autotest.py.

Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
8 years agoexamples/l2fwd-crypto: add NULL algorithms
Pablo de Lara [Tue, 22 Mar 2016 21:24:01 +0000 (21:24 +0000)]
examples/l2fwd-crypto: add NULL algorithms

NULL crypto operation is now supported, but l2fwd-crypto
was missing an update on the list of supported algorithms
that can be passed from command line.

Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
8 years agoexamples/l2fwd-crypto: add AES-GCM algorithm
Pablo de Lara [Tue, 22 Mar 2016 21:24:00 +0000 (21:24 +0000)]
examples/l2fwd-crypto: add AES-GCM algorithm

AES GCM is an algorithm for both ciphering and authentication,
but the authentication algorithm was missing in the
list of supported algorithms that can be passed from command line.

Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
8 years agoexamples/l2fwd-crypto: add SNOW 3G algorithms
Pablo de Lara [Tue, 22 Mar 2016 21:23:59 +0000 (21:23 +0000)]
examples/l2fwd-crypto: add SNOW 3G algorithms

Since SNOW3G UEA2/UIA2 are supported now by both HW and SW,
l2fwd-crypto may use them, extending the list of algorithms
parsed from command line.

Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
8 years agoexamples/l2fwd-crypto: fix command line help
Pablo de Lara [Thu, 24 Mar 2016 13:39:45 +0000 (13:39 +0000)]
examples/l2fwd-crypto: fix command line help

Previous cdev parameter was changed to cdev_type,
to select a crypto device type preference (HW/SW/ANY),
instead of the device itself (QAT/AESNI...).
Also deleted cdev duplicated parameter from the help.

Fixes: 27cf2d1b18e1 ("examples/l2fwd-crypto: discover capabilities")

Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
8 years agoeal/arm: disable NEON for 32-bit memcpy
Jan Viktorin [Sat, 19 Mar 2016 19:58:04 +0000 (20:58 +0100)]
eal/arm: disable NEON for 32-bit memcpy

The new flag CONFIG_RTE_ARCH_ARM_NEON_MEMCPY is used to enable memcpy
optimizations in EAL.
As it is not always the performance benefit, the feature is disabled.

Signed-off-by: Jan Viktorin <viktorin@rehivetech.com>
8 years agomk: improve ARM NEON detection
Jan Viktorin [Sat, 19 Mar 2016 19:58:03 +0000 (20:58 +0100)]
mk: improve ARM NEON detection

The __ARM_NEON declares that the arm_neon.h is available
which is not always true for the __ARM_NEON_FP.

 $ arm-linux-gnueabi-gcc -dM -E - < /dev/null  | grep "_FP\|_NEON"
 #define __ARM_FP 12
 #define __ARM_NEON_FP 4
 #define __VFP_FP__ 1

 $ arm-linux-gnueabi-gcc -mfpu=neon -dM -E - < /dev/null  | grep "_FP\|_NEON"
 #define __ARM_FP 12
 #define __ARM_NEON_FP 4
 #define __ARM_NEON__ 1
 #define __VFP_FP__ 1
 #define __ARM_NEON 1

 $ aarch64-linux-gnu-gcc -dM -E - < /dev/null | grep "NEON\|FP"
 #define __FP_FAST_FMAF 1
 #define __ARM_NEON 1
 #define __FP_FAST_FMA 1

 $ aarch64-thunderx-linux-gnu-gcc -dM -E - < /dev/null |grep "NEON\|FP"
 #define __ARM_FP 12
 #define __ARM_NEON_FP 12
 #define __FP_FAST_FMAF 1
 #define __ARM_NEON 1
 #define __FP_FAST_FMA 1

Signed-off-by: Jan Viktorin <viktorin@rehivetech.com>
Acked-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
8 years agoconfig: remove ARM NEON option
Jan Viktorin [Sat, 19 Mar 2016 19:58:02 +0000 (20:58 +0100)]
config: remove ARM NEON option

ARMv7 machines have usually the NEON available.
Customization of the -mfpu=neon must be done by hand or by defining
another machine rte.vars.mk.
So, the CONFIG_RTE_ARCH_ARM_NEON is useless (and confusing).

Signed-off-by: Jan Viktorin <viktorin@rehivetech.com>
8 years agoethdev: fix xstats size query with NULL
Stephen Hemminger [Tue, 22 Mar 2016 22:09:17 +0000 (15:09 -0700)]
ethdev: fix xstats size query with NULL

Normal usage of rte_eth_dev_xstats_get is to call twice. The
first time the function is called with portid, xstats = NULL
and n = 0; this returns the number of entries in the statistics
table that need to be allocated.

The problem is that the routine adds a count value to NULL (0)
and assumes that this is a valid pointer (it isn't). Device drivers
all have a check for NULL, and this no longer matches.

Fixes: d4fef8b0d5e5 ("ethdev: expose generic and driver specific stats in xstats")

Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
8 years agomk: fix gcc-5 version on Suse
Markos Chandras [Tue, 22 Mar 2016 17:13:36 +0000 (17:13 +0000)]
mk: fix gcc-5 version on Suse

In openSUSE Tumbleweed (and in any other SUSE distribution which
uses (or will use) gcc >= 5), gcc -dumpversion returns '5'. This is on
purpose as discussed in https://bugzilla.opensuse.org/show_bug.cgi?id=941428
As a result of which, the gcc-4.x comparison (40 against 5) does not
work leading to tons of warnings and failures during build.

This patch aims to change the way the gcc version is obtained by using
the gcc macros directly.
This is similar to what's being used in the Linux kernel. Querying the
GCC macros directly gives more accurate results compared to -dumpversion
which could vary across distributions.

Signed-off-by: Markos Chandras <mchandras@suse.de>
Acked-by: Thomas Monjalon <thomas.monjalon@6wind.com>
8 years agomk: fix missing librt dependencies
Daniel Mrzyglod [Tue, 22 Mar 2016 12:51:02 +0000 (13:51 +0100)]
mk: fix missing librt dependencies

For GLIBC < 2.17 it is necessery to add -lrt for linker
from glibc > 2.17 The `clock_*' suite of functions (declared in <time.h>) is now
available directly in the main C library. This affect Ubuntu 12.04 in i686
and other older Linux Distros).

Fixes: 4758404a3084 ("mk: fix eal shared library dependencies")

Signed-off-by: Daniel Mrzyglod <danielx.t.mrzyglod@intel.com>
8 years agomk: fix missing libm dependencies
Panu Matilainen [Mon, 21 Mar 2016 11:33:58 +0000 (13:33 +0200)]
mk: fix missing libm dependencies

Commit e86a699cf6b1 missed two further libm dependencies: ceil() used
by librte_meter is typically inlined so the missing dependency does not
actually cause failures, and librte_pmd_nfp is not built by default
so its easy to miss.

This causes duplicates in LDLIBS in many configurations so its vital
they are removed before passing to linker.

Fixes: e86a699cf6b1 ("mk: fix shared library dependencies on libm and librt")

Reported-by: Ferruh Yigit <ferruh.yigit@intel.com>
Signed-off-by: Panu Matilainen <pmatilai@redhat.com>
Tested-by: Ferruh Yigit <ferruh.yigit@intel.com>
8 years agomk: eliminate duplicates from libraries list
Panu Matilainen [Mon, 21 Mar 2016 11:33:57 +0000 (13:33 +0200)]
mk: eliminate duplicates from libraries list

Duplicates in LDLIBS can cause link failures from multiply defined
symbols, ensure all libraries are only mentioned once. Can't use
sorting for duplicate elimination as order is critical so awk one-liner
is used.

Signed-off-by: Panu Matilainen <pmatilai@redhat.com>
8 years agomk: restrict CPU flags list
Thomas Monjalon [Mon, 21 Mar 2016 20:12:55 +0000 (21:12 +0100)]
mk: restrict CPU flags list

When compiling each file, the CPU flags are given as RTE_MACHINE_CPUFLAG_*
and in the list RTE_COMPILE_TIME_CPUFLAGS.

RTE_MACHINE_CPUFLAG_* are used to check the CPU features when compiling.

The list RTE_COMPILE_TIME_CPUFLAGS is used only to check the CPU at
runtime in the function rte_cpu_check_supported(). So it is not needed to
define this list for every files.
That's why RTE_COMPILE_TIME_CPUFLAGS is removed from the common variable
MACHINE_CFLAGS and is added only to the CFLAGS of eal_common_cpuflags.c.

Signed-off-by: Thomas Monjalon <thomas.monjalon@6wind.com>
8 years agoexamples/l3fwd: remove old SSE4 comments
Thomas Monjalon [Mon, 21 Mar 2016 14:34:33 +0000 (15:34 +0100)]
examples/l3fwd: remove old SSE4 comments

The flag ENABLE_MULTI_BUFFER_OPTIMIZE has been removed so the
related comments are now useless.

Fixes: 268888b5b020 ("examples/l3fwd: modularize")

Signed-off-by: Thomas Monjalon <thomas.monjalon@6wind.com>
8 years agoexamples/performance-thread: fix build without SSE4.1
Thomas Monjalon [Wed, 3 Feb 2016 18:56:38 +0000 (19:56 +0100)]
examples/performance-thread: fix build without SSE4.1

clang reports these errors:

error: too many arguments to function call, expected 2, have 3
examples/performance-thread/l3fwd-thread/main.c:1220:1: note:
'l3fwd_simple_forward' declared here

examples/l3fwd/main.c:550:1: error: unused function 'send_packetsx4'

The function is used only when ENABLE_MULTI_BUFFER_OPTIMIZE is 1.

Fixes: d48415e1fee3 ("examples/performance-thread: add l3fwd-thread app")

Signed-off-by: Thomas Monjalon <thomas.monjalon@6wind.com>
8 years agoexamples/ip_pipeline: fix configuration parser
Fan Zhang [Mon, 14 Mar 2016 12:22:02 +0000 (12:22 +0000)]
examples/ip_pipeline: fix configuration parser

Fixes: 377cd98e0ab6 ("examples/ip_pipeline: add link identification")

Signed-off-by: Fan Zhang <roy.fan.zhang@intel.com>
Acked-by: Cristian Dumitrescu <cristian.dumitrescu@intel.com>
8 years agoapp/testpmd: fix build with icc
Tomasz Kulasek [Tue, 22 Mar 2016 14:52:09 +0000 (15:52 +0100)]
app/testpmd: fix build with icc

This patch fixes error #188: enumerated type mixed with another type,
when uint32_t is casted on enum type in icc.

Fixes: 05f1b9c82ec2 ("app/testpmd: add commands for L2 tunnel config")

Signed-off-by: Tomasz Kulasek <tomaszx.kulasek@intel.com>
8 years agoring: fix deadlock in zero object multi enqueue or dequeue
Lazaros Koromilas [Thu, 17 Mar 2016 15:49:43 +0000 (17:49 +0200)]
ring: fix deadlock in zero object multi enqueue or dequeue

Issuing a zero objects dequeue with a single consumer has no effect.
Doing so with multiple consumers, can get more than one thread to succeed
the compare-and-set operation and observe starvation or even deadlock in
the while loop that checks for preceding dequeues.  The problematic piece
of code when n = 0:

    cons_next = cons_head + n;
    success = rte_atomic32_cmpset(&r->cons.head, cons_head, cons_next);

The same is possible on the enqueue path.

Fixes: af75078fece3 ("first public release")

Signed-off-by: Lazaros Koromilas <l@nofutznetworks.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
8 years agolpm6: fix use after free
Christian Ehrhardt [Mon, 21 Mar 2016 14:06:11 +0000 (15:06 +0100)]
lpm6: fix use after free

In certain autotests lpm->max_rules turned out to be non initialized.
That was caused by a failing allocation for lpm->rules_tbl in rte_lpm6_create.
It then left the function via goto exit with lpm freed, but still a pointer
value being set.

In case of an allocation failure it resets lpm to NULL now, to avoid the
upper layers operate on that already freed memory.
Along that is also makes the RTE_LOG message of the failed allocation unique.

Fixes: 5c510e13a9cb ("lpm: add IPv6 support")

Signed-off-by: Christian Ehrhardt <christian.ehrhardt@canonical.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
8 years agolpm6: fix missing free
Christian Ehrhardt [Mon, 21 Mar 2016 14:06:12 +0000 (15:06 +0100)]
lpm6: fix missing free

lpm6 autotests failed with the default alloc of 512M Memory.
While >=2500M was a workaround it became clear while debugging that it
had a leak.
One could see a lot of output like:
  LPM Test tests6[i]: FAIL
  LPM: LPM memory allocation failed

It turned out that in rte_lpm6_free
- lpm might not be freed if it didn't find a te (early return)
- lpm->rules_tbl was not freed ever

Fixes: 899d8bc9b3b5 ("lpm: make tailq fully local")

Signed-off-by: Christian Ehrhardt <christian.ehrhardt@canonical.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
8 years agolpm: fix use after free
Christian Ehrhardt [Mon, 21 Mar 2016 14:06:14 +0000 (15:06 +0100)]
lpm: fix use after free

There were further chances for a use after free by returning an already
freed pointer in rte_lpm_create for v20 and v1604.
Along that is also makes the RTE_LOG messages of the failed allocations
unique.

Fixes: f1f7261838b3 ("lpm: add a new config structure for IPv4")

Signed-off-by: Christian Ehrhardt <christian.ehrhardt@canonical.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
8 years agolpm: fix missing free
Christian Ehrhardt [Mon, 21 Mar 2016 14:06:13 +0000 (15:06 +0100)]
lpm: fix missing free

In rte_lpm_free lpm might not be freed if it didn't find a te (early return)

The two lpm interfaces rte_lpm_free_v20 and rte_lpm_free_v1604 had a leak.
rte_lpm_free_v20 might have missed to free rules_tbl
rte_lpm_free_v1604 due to an early exit might have missed to free
rules_tbl and lpm itself.

Fixes: 899d8bc9b3b5 ("lpm: make tailq fully local")

Signed-off-by: Christian Ehrhardt <christian.ehrhardt@canonical.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
8 years agodoc: notify Rx errors not counting missed packets anymore
Thomas Monjalon [Tue, 22 Mar 2016 15:51:30 +0000 (16:51 +0100)]
doc: notify Rx errors not counting missed packets anymore

Fixes: 86057c9992f5 ("ethdev: remove missed packets from error counter")

Suggested-by: Maryam Tahhan <maryam.tahhan@intel.com>
Signed-off-by: Thomas Monjalon <thomas.monjalon@6wind.com>
8 years agoethdev: remove missed packets from error counter
Igor Ryzhov [Thu, 10 Mar 2016 13:03:30 +0000 (16:03 +0300)]
ethdev: remove missed packets from error counter

Comment for "ierrors" counter says that it counts erroneous received
packets. But for some reason "imissed" counter is added to "ierrors"
counter in most drivers.
It is a mistake, because missed packets are obviously not received.
This patch fixes it.

Fixes: 70bdb18657da ("ethdev: add Rx error counters for missed, badcrc and badlen packets")
Fixes: 6bfe648406b5 ("i40e: add Rx error statistics")
Fixes: 856505d303f4 ("cxgbe: add port statistics")

Signed-off-by: Igor Ryzhov <iryzhov@nfware.com>
Acked-by: Wenzhuo Lu <wenzhuo.lu@intel.com>
Acked-by: Rahul Lakkireddy <rahul.lakkireddy@chelsio.com>
Acked-by: Thomas Monjalon <thomas.monjalon@6wind.com>
8 years agomaintainers: sort examples
Thomas Monjalon [Tue, 22 Mar 2016 09:59:52 +0000 (10:59 +0100)]
maintainers: sort examples

Keep sorting examples and fix l2fwd-cat path.

Fixes: ab129e9065a5 ("examples/ptpclient: add minimal PTP client")
Fixes: f6baccbc2b3b ("examples/l2fwd-cat: add sample application for PQoS CAT and CDP")

Signed-off-by: Thomas Monjalon <thomas.monjalon@6wind.com>
8 years agoexamples/l2fwd-cat: add sample application for PQoS CAT and CDP
Wojciech Andralojc [Mon, 14 Mar 2016 13:46:59 +0000 (13:46 +0000)]
examples/l2fwd-cat: add sample application for PQoS CAT and CDP

This patch implements PQoS as a sample application.
PQoS allows management of the CPUs last level cache,
which can be useful for DPDK to ensure quality of service.
The sample app links against the existing 01.org PQoS library
(https://github.com/01org/intel-cmt-cat).

White paper demonstrating example use case "Increasing Platform Determinism
with Platform Quality of Service for the Data Plane Development Kit"
(http://www.intel.com/content/www/us/en/communications/increasing-platform-determinism-pqos-dpdk-white-paper.html)

Signed-off-by: Wojciech Andralojc <wojciechx.andralojc@intel.com>
Signed-off-by: Tomasz Kantecki <tomasz.kantecki@intel.com>
Signed-off-by: Marcel D Cornu <marcel.d.cornu@intel.com>
8 years agoexamples/l3fwd: fix exact match performance
Tomasz Kulasek [Fri, 18 Mar 2016 13:31:46 +0000 (14:31 +0100)]
examples/l3fwd: fix exact match performance

It seems that for the most use cases, previous hash_multi_lookup provides
better performance, and more, sequential lookup can cause significant
performance drop.

This patch sets previously optional hash_multi_lookup method as default.
It also provides some minor optimizations such as queue drain only on used
tx ports.

Fixes: 94c54b4158d5 ("examples/l3fwd: rework exact-match")
Fixes: dc81ebbacaeb ("lpm: extend IPv4 next hop field")
Fixes: 64d3955de1de ("examples/l3fwd: fix ARM build")

Reported-by: Qian Xu <qian.q.xu@intel.com>
Signed-off-by: Tomasz Kulasek <tomaszx.kulasek@intel.com>
8 years agoexamples/l3fwd: fix build with hash multi-lookup
Maciej Czekaj [Wed, 16 Mar 2016 06:23:24 +0000 (07:23 +0100)]
examples/l3fwd: fix build with hash multi-lookup

l3fwd does not compile with HASH_MULTI_LOOKUP.
2 issues:
 * in 64d395 mask0 changed type from xmm_t to rte_xmm_t
      -> use x field from rte_xmm_t
 * in dc81eb dst_port parameter changed to uint32_t
      -> change uint16_t dst_port to uin32_t dsp_port

Fixes: dc81ebbacaeb ("lpm: extend IPv4 next hop field")
Fixes: 64d3955de1de ("examples/l3fwd: fix ARM build")

Signed-off-by: Maciej Czekaj <maciej.czekaj@caviumnetworks.com>
8 years agolpm: fix pipeline apps
Michal Kobylinski [Thu, 17 Mar 2016 08:47:25 +0000 (09:47 +0100)]
lpm: fix pipeline apps

Updated ip_pipeline app is using new changes from LPM library
(Increased number of next hops and added new config structure
for LPM IPv4).

Fixes: f1f7261838b3 ("lpm: add a new config structure for IPv4")

Signed-off-by: Michal Kobylinski <michalx.kobylinski@intel.com>
Acked-by: Cristian Dumitrescu <cristian.dumitrescu@intel.com>
8 years agovhost: remove unnecessary memset when enqueueing
Yuanhan Liu [Wed, 16 Mar 2016 06:44:56 +0000 (14:44 +0800)]
vhost: remove unnecessary memset when enqueueing

We have to reset the virtio net hdr at virtio_enqueue_offload()
before, due to all mbufs share a single virtio_hdr structure:

struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, }, 0};

foreach (mbuf) {
virtio_enqueue_offload(mbuf, &virtio_hdr.hdr);

copy net hdr and mbuf to desc buf
}

However, after the vhost rxtx refactor, the code looks like:

copy_mbuf_to_desc(mbuf)
{
struct virtio_net_hdr_mrg_rxbuf virtio_hdr = {{0, }, 0}

virtio_enqueue_offload(mbuf, &virtio_hdr.hdr);

copy net hdr and mbuf to desc buf
}

foreach (mbuf) {
copy_mbuf_to_desc(mbuf);
}

Therefore, the memset at virtio_enqueue_offload() is not necessary
any more; remove it.

Signed-off-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Huawei Xie <huawei.xie@intel.com>
8 years agomk: fix linker script when re-building
Sergio Gonzalez Monroy [Wed, 16 Mar 2016 23:22:23 +0000 (23:22 +0000)]
mk: fix linker script when re-building

The linker script is generated by simply finding all libraries in
RTE_OUTPUT/lib.

The issue shows up when re-building the DPDK, hence already having a
linker script in that directory, resulting in the linker script
including itself.

That does not play well with the linker.

Simply filtering the linker script from all the found libraries solves
the problem.

Fixes: 948fd64befc3 ("mk: replace the combined library with a linker script")

Signed-off-by: Sergio Gonzalez Monroy <sergio.gonzalez.monroy@intel.com>
Acked-by: Panu Matilainen <pmatilai@redhat.com>
8 years agoversion: 16.04-rc1
Thomas Monjalon [Wed, 16 Mar 2016 20:44:50 +0000 (21:44 +0100)]
version: 16.04-rc1

Signed-off-by: Thomas Monjalon <thomas.monjalon@6wind.com>
8 years agodoc: add packet framework release notes
Jasvinder Singh [Mon, 14 Mar 2016 13:44:36 +0000 (13:44 +0000)]
doc: add packet framework release notes

This patch updates the release notes with the features that
have been added to ip_pipeline application.

Signed-off-by: Jasvinder Singh <jasvinder.singh@intel.com>
Acked-by: Cristian Dumitrescu <cristian.dumitrescu@intel.com>
8 years agopci: fix ioport support for uio_pci_generic on x86
David Marchand [Tue, 15 Mar 2016 06:29:13 +0000 (07:29 +0100)]
pci: fix ioport support for uio_pci_generic on x86

uio_pci_generic does not offer the same sysfs helpers as igb_uio.
In this case, ioport number can only be retrieved by parsing /proc/ioports.

Fixes: 756ce64b1ecd ("eal: introduce PCI ioport API")

Reported-by: Mauricio Vasquez B <mauricio.vasquezbernal@studenti.polito.it>
Signed-off-by: David Marchand <david.marchand@6wind.com>
8 years agopci: separate ioport handlers per UIO driver
David Marchand [Tue, 15 Mar 2016 06:29:10 +0000 (07:29 +0100)]
pci: separate ioport handlers per UIO driver

Prepare for fixes on x86 by separating igb_uio and uio_pci_generic cases.

Signed-off-by: David Marchand <david.marchand@6wind.com>
8 years agopci: align ioport special case for x86 in read/write/unmap
David Marchand [Tue, 15 Mar 2016 06:29:12 +0000 (07:29 +0100)]
pci: align ioport special case for x86 in read/write/unmap

Commit b8eb345378bd ("pci: ignore devices already managed in Linux when
mapping x86 ioport") did not update other parts of the ioport api.

The application is not supposed to call these read/write/unmap ioport
functions if map call failed but I prefer aligning the code for the sake
of consistency.

Signed-off-by: David Marchand <david.marchand@6wind.com>
8 years agopci: align ioport unmap error handling to ioport map
David Marchand [Tue, 15 Mar 2016 06:29:11 +0000 (07:29 +0100)]
pci: align ioport unmap error handling to ioport map

Same idea as commit bd80d4730aca ("pci: rework ioport map error handling").

Signed-off-by: David Marchand <david.marchand@6wind.com>
8 years agobonding: fix crash when no slave device
Bernard Iremonger [Mon, 7 Mar 2016 11:40:40 +0000 (11:40 +0000)]
bonding: fix crash when no slave device

If a bonded device is created when there are no slave devices
there is a loop in bond_ethdev_promiscuous_enable() which results
in a segmentation fault.

The solution is to initialise the current_primary_port to an
invalid port value when the bonded port is created.

Fixes: 2efb58cbab6e ("bond: new link bonding library")

Signed-off-by: Bernard Iremonger <bernard.iremonger@intel.com>
Acked-by: Ferruh Yigit <ferruh.yigit@intel.com>
8 years agobonding: do not activate slave twice
Eric Kinzie [Tue, 1 Mar 2016 17:32:02 +0000 (09:32 -0800)]
bonding: do not activate slave twice

The current code for detecting link during slave addition can cause a
slave interface to be activated twice -- once during slave_configure()
and again at the end of __eth_bond_slave_add_lock_free().  This will
either cause the active slave count to be incorrect or will cause the
802.3ad activation function to panic.  Ensure that the interface is not
activated more than once.

Fixes: 46fb43683679 ("bond: add mode 4")

Signed-off-by: Eric Kinzie <ekinzie@brocade.com>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Declan Doherty <declan.doherty@intel.com>
8 years agobonding: fix active slaves with no primary
Eric Kinzie [Tue, 1 Mar 2016 17:32:01 +0000 (09:32 -0800)]
bonding: fix active slaves with no primary

If the link state of a slave is "up" when added, it is added to the list
of active slaves but, even if it is the only slave, is not selected as
the primary interface.  Generally, handling of link state interrupts
selects an interface to be primary, but only if the active count is zero.
This change avoids the situation where there are active slaves but
no primary.

Fixes: 2efb58cbab6e ("bond: new link bonding library")

Signed-off-by: Eric Kinzie <ekinzie@brocade.com>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Declan Doherty <declan.doherty@intel.com>
8 years agobonding: do not ignore multicast in mode 4
Eric Kinzie [Tue, 1 Mar 2016 17:32:00 +0000 (09:32 -0800)]
bonding: do not ignore multicast in mode 4

The bonding PMD in mode 4 puts all enslaved interfaces into promiscuous
mode in order to receive LACPDUs and must filter unwanted packets
after the traffic has been "collected".  Allow broadcast and multicast
through so that ARP and IPv6 neighbor discovery continue to work.

Fixes: 46fb43683679 ("bond: add mode 4")

Signed-off-by: Eric Kinzie <ekinzie@brocade.com>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Declan Doherty <declan.doherty@intel.com>
8 years agobonding: copy entire config structure in mode 4
Eric Kinzie [Tue, 1 Mar 2016 17:31:59 +0000 (09:31 -0800)]
bonding: copy entire config structure in mode 4

Copy all needed fields from the mode8023ad_private structure in
bond_mode_8023ad_conf_get().  This help ensure that a subsequent call
to rte_eth_bond_8023ad_setup() is not passed uninitialized data that
would result in either incorrect behavior or a failed sanity check.

Fixes: 46fb43683679 ("bond: add mode 4")

Signed-off-by: Eric Kinzie <ekinzie@brocade.com>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Declan Doherty <declan.doherty@intel.com>
8 years agobonding: fix detach of slave devices
Bernard Iremonger [Wed, 10 Feb 2016 10:13:45 +0000 (10:13 +0000)]
bonding: fix detach of slave devices

Ensure that a bonded slave device is not detached,
until it is removed from the bonded device.

Fixes: 2efb58cbab6e ("bond: new link bonding library")
Fixes: a45b288ef21a ("bond: support link status polling")
Fixes: 494adb7f63f2 ("ethdev: add device fields from PCI layer")
Fixes: b1fb53a39d88 ("ethdev: remove some PCI specific handling")

Signed-off-by: Bernard Iremonger <bernard.iremonger@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
8 years agobonding: fix detach of bonded device
Bernard Iremonger [Wed, 10 Feb 2016 10:13:44 +0000 (10:13 +0000)]
bonding: fix detach of bonded device

Check that the bonded device has no slaves before detaching it.

Fixes: 8d30fe7fa737 ("bonding: support port hotplug")

Signed-off-by: Bernard Iremonger <bernard.iremonger@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
8 years agonull: remove duplicate fields in internal struct
Ferruh Yigit [Fri, 26 Feb 2016 16:58:09 +0000 (16:58 +0000)]
null: remove duplicate fields in internal struct

1- remove duplicate nb_rx/tx_queues fields from internals
2- remove duplicate numa_node field from internals

Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
Tested-by: Tetsuya Mukawa <mukawa@igel.co.jp>
Acked-by: Tetsuya Mukawa <mukawa@igel.co.jp>
Acked-by: Nicolás Pernas Maradei <nicolas.pernas.maradei@emutex.com>
8 years agoring: free rings when detaching device
Mauricio Vasquez B [Mon, 7 Mar 2016 16:20:31 +0000 (17:20 +0100)]
ring: free rings when detaching device

When a device is created with "CREATE" as action, new rings are
allocated for it, then it is a good practice to free them when the
rte_ethdev_dettach method is invoked by the application.

Rings are not freeded when "ATTACH" is used or when the device is
created by means of the rte_eth_from_rings function.

Signed-off-by: Mauricio Vasquez B <mauricio.vasquezbernal@studenti.polito.it>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
8 years agoring: clean up driver
Ferruh Yigit [Fri, 26 Feb 2016 16:58:08 +0000 (16:58 +0000)]
ring: clean up driver

Rename nb_rx/tx_queues fields in internals struct to max_rx/tx_queues
Updated fields required to keep max queue numbers configured. For current
queue number requirements data->nb_rx/tx_queues fields used.

Some checkpatch corrections and code clenaup.

Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
8 years agopcap: reduce duplication
Ferruh Yigit [Fri, 26 Feb 2016 16:58:07 +0000 (16:58 +0000)]
pcap: reduce duplication

1- Remove duplicate nb_rx/tx_queues fields from internals
2- Move duplicate code into a common function

Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
Acked-by: Nicolás Pernas Maradei <nicolas.pernas.maradei@emutex.com>
8 years agopcap: fix captured frame length
Dror Birkman [Thu, 28 Jan 2016 11:09:50 +0000 (13:09 +0200)]
pcap: fix captured frame length

The actual captured length is header.caplen, whereas header.len is
the original length on the wire.

Fixes: 4c173302c307 ("pcap: add new driver")

Signed-off-by: Dror Birkman <dror.birkman@lightcyber.com>
Acked-by: Nicolás Pernas Maradei <nicolas.pernas.maradei@emutex.com>
8 years agoaf_packet: make the device detachable
Wojciech Zmuda [Wed, 2 Mar 2016 11:55:41 +0000 (12:55 +0100)]
af_packet: make the device detachable

Allow dynamic deallocation of af_packet device through proper
API functions. To achieve this:
* set device flag to RTE_ETH_DEV_DETACHABLE
* implement rte_pmd_af_packet_devuninit() and expose it
  through rte_driver.uninit()
* copy device name to ethdev->data to make discoverable with
  rte_eth_dev_allocated()
Moreover, make af_packet init function static, as there is no
reason to keep it public.

Signed-off-by: Wojciech Zmuda <woz@semihalf.com>
Acked-by: Bernard Iremonger <bernard.iremonger@intel.com>
8 years agovmxnet3: support setting MAC address
Stephen Hemminger [Fri, 4 Mar 2016 18:08:02 +0000 (10:08 -0800)]
vmxnet3: support setting MAC address

Allow overriding the base mac address of the device.

Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Remy Horton <remy.horton@intel.com>
8 years agovmxnet3: fix VLAN filtering
Charles (Chas) Williams [Fri, 4 Mar 2016 18:08:01 +0000 (10:08 -0800)]
vmxnet3: fix VLAN filtering

During an MTU change, the adapter is restarted.  If hardware VLAN offload
is in use, this existing filter table would also be cleared.  Instead,
setup the shadow table once during device initialization and just update
during restart.

vmxnet3_dev_vlan_offload_set(dev, mask) was incorrectly treating the
mask parameter as the bitmask for vlan_strip and vlan_filter, whereas
the mask indicates only what has changed - the values for
vlan_stripping and vlan_filter needs to be taken from dev_conf.rxmode.

Fixes: f003fc383487 ("vmxnet3: enable vlan filtering")

Signed-off-by: Charles (Chas) Williams <ciwillia@brocade.com>
Signed-off-by: Nachiketa Prachanda <nprachan@brocade.com>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Remy Horton <remy.horton@intel.com>
8 years agovmxnet3: support jumbo frames
Stephen Hemminger [Fri, 4 Mar 2016 18:08:00 +0000 (10:08 -0800)]
vmxnet3: support jumbo frames

Add support for linking multi-segment buffers together to
handle Jumbo packets. The vmxnet3 API supports having header
and body buffer types. What this patch does is fill the primary
ring completely with header buffers and the secondary ring
with body buffers. This allows for non-jumbo frames to only
use one mbuf (from primary ring); and jumbo frames will have
first mbuf from primary ring and following mbufs from other
ring.

This could be optimized in future if the DPDK had API
to supply different sized mbufs (two pools) into driver.

Signed-off-by: Stephen Hemminger <shemming@brocade.com>
Acked-by: Remy Horton <remy.horton@intel.com>
Release note addition:
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
8 years agovmxnet3: announce offload capabilities
Yong Wang [Wed, 13 Jan 2016 02:08:37 +0000 (18:08 -0800)]
vmxnet3: announce offload capabilities

Signed-off-by: Yong Wang <yongwang@vmware.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
8 years agovmxnet3: support TSO
Yong Wang [Wed, 13 Jan 2016 02:08:36 +0000 (18:08 -0800)]
vmxnet3: support TSO

This commit adds vmxnet3 TSO support.

Verified with test-pmd (set fwd csum) that both tso and
non-tso pkts can be successfully transmitted and all
segmentes for a tso pkt are correct on the receiver side.

Signed-off-by: Yong Wang <yongwang@vmware.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
8 years agovmxnet3: add Tx L4 checksum offload
Yong Wang [Wed, 13 Jan 2016 02:08:35 +0000 (18:08 -0800)]
vmxnet3: add Tx L4 checksum offload

Support TCP/UDP checksum offload.

Signed-off-by: Yong Wang <yongwang@vmware.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
8 years agovmxnet3: rework Tx
Yong Wang [Wed, 13 Jan 2016 02:08:34 +0000 (18:08 -0800)]
vmxnet3: rework Tx

Clean up txNumDeferred usage.

Signed-off-by: Yong Wang <yongwang@vmware.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
8 years agovmxnet3: restore Tx data ring support
Yong Wang [Wed, 13 Jan 2016 02:08:33 +0000 (18:08 -0800)]
vmxnet3: restore Tx data ring support

Tx data ring support was removed in a previous change that
added multi-seg transmit.  This change adds it back.

According to the original commit (2e849373), 64B pkt
rate with l2fwd improved by ~20% on an Ivy Bridge
server at which point we start to hit some bottleneck
on the rx side.

I also re-did the same test on a different setup (Haswell
processor, ~2.3GHz clock rate) on top of the master
and still observed ~17% performance gains.

Fixes: 7ba5de417e3c ("vmxnet3: support multi-segment transmit")

Signed-off-by: Yong Wang <yongwang@vmware.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
8 years agovmxnet3: clean up typos and unused code
Yong Wang [Wed, 13 Jan 2016 02:08:32 +0000 (18:08 -0800)]
vmxnet3: clean up typos and unused code

Signed-off-by: Yong Wang <yongwang@vmware.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
8 years agovmxnet3: remove redundant function names in log
Huawei Xie [Wed, 10 Feb 2016 16:12:39 +0000 (16:12 +0000)]
vmxnet3: remove redundant function names in log

Signed-off-by: Huawei Xie <huawei.xie@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
8 years agovirtio: remove redundant function names in log
Huawei Xie [Wed, 10 Feb 2016 16:08:54 +0000 (16:08 +0000)]
virtio: remove redundant function names in log

Signed-off-by: Huawei Xie <huawei.xie@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
8 years agovirtio: optimize Tx enqueue
Stephen Hemminger [Fri, 4 Mar 2016 18:19:21 +0000 (10:19 -0800)]
virtio: optimize Tx enqueue

All the error checks in virtqueue_enqueue_xmit are already done
by the caller. Therefore they can be removed to improve performance.

Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Huawei Xie <huawei.xie@intel.com>
8 years agovirtio: use any layout on Tx
Stephen Hemminger [Fri, 4 Mar 2016 18:19:20 +0000 (10:19 -0800)]
virtio: use any layout on Tx

Virtio supports a feature that allows sender to put transmit
header prepended to data.  It requires that the mbuf be writeable, correct
alignment, and the feature has been negotiatied.  If all this works out,
then it will be the optimum way to transmit a single segment packet.

Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Huawei Xie <huawei.xie@intel.com>
8 years agovirtio: use indirect ring elements
Stephen Hemminger [Fri, 4 Mar 2016 18:19:19 +0000 (10:19 -0800)]
virtio: use indirect ring elements

The virtio ring in QEMU/KVM is usually limited to 256 entries
and the normal way that virtio driver was queuing mbufs required
nsegs + 1 ring elements. By using the indirect ring element feature
if available, each packet will take only one ring slot even for
multi-segment packets.

Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Acked-by: Huawei Xie <huawei.xie@intel.com>
8 years agovirtio: remove broadcast packets from multicast statistics
Igor Ryzhov [Fri, 26 Feb 2016 15:01:23 +0000 (18:01 +0300)]
virtio: remove broadcast packets from multicast statistics

Signed-off-by: Igor Ryzhov <iryzhov@nfware.com>
Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Applied with coding standards fixes:
Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
8 years agovirtio: fix descriptors pointing to the same buffer
Huawei Xie [Thu, 10 Dec 2015 16:07:32 +0000 (00:07 +0800)]
virtio: fix descriptors pointing to the same buffer

The virtio_net_hdr desc all pointed to the same buffer. It doesn't cause
issue because in the simple TX mode we don't use the header. This patch
makes the header desc point to different buffer.

Fixes: b4ae9c505f2e ("virtio: optimize ring layout")

Signed-off-by: Huawei Xie <huawei.xie@intel.com>
Acked-by: Jianfeng Tan <jianfeng.tan@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
8 years agovirtio: fix crash in statistics functions
Bernard Iremonger [Wed, 23 Dec 2015 09:45:19 +0000 (09:45 +0000)]
virtio: fix crash in statistics functions

This initialisation of nb_rx_queues and nb_tx_queues has been removed
from eth_virtio_dev_init.

The nb_rx_queues and nb_tx_queues were being initialised in
eth_virtio_dev_init before the tx_queues and rx_queues arrays were
allocated.

The arrays are allocated when the ethdev port is configured and the
nb_tx_queues and nb_rx_queues are initialised.

If any of the following functions were called before the ethdev
port was configured there was a segmentation fault because
rx_queues and tx_queues were NULL:

rte_eth_stats_get
rte_eth_stats_reset
rte_eth_xstats_get
rte_eth_xstats_reset

Fixes: 823ad647950a ("virtio: support multiple queues")

Signed-off-by: Bernard Iremonger <bernard.iremonger@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
8 years agovirtio: fix restart
Jianfeng Tan [Mon, 11 Jan 2016 06:16:13 +0000 (14:16 +0800)]
virtio: fix restart

Fix the issue that virtio device cannot be started after stopped.

The field, hw->started, should be changed by virtio_dev_start/stop instead
of virtio_dev_close.

Fixes: a85786dc816f ("virtio: fix states handling during initialization")

Reported-by: Pavel Fedin <p.fedin@samsung.com>
Signed-off-by: Jianfeng Tan <jianfeng.tan@intel.com>
Acked-by: Yuanhan Liu <yuanhan.liu@linux.intel.com>
Tested-by: Pavel Fedin <p.fedin@samsung.com>
8 years agoszedata2: support promiscuous and allmulticast modes
Matej Vido [Thu, 28 Jan 2016 22:23:48 +0000 (23:23 +0100)]
szedata2: support promiscuous and allmulticast modes

add functions for enabling/disabling promiscuous, allmulticast modes

Signed-off-by: Matej Vido <vido@cesnet.cz>
8 years agoszedata2: support link state operations
Matej Vido [Thu, 28 Jan 2016 22:23:47 +0000 (23:23 +0100)]
szedata2: support link state operations

Mmap PCI resource file and add inline functions for reading from and
writing to PCI resource address space.
Add description of IBUF and OBUF address space.
Add configuration option for setting which firmware type will be used.
Right address space values for IBUFs and OBUFs offsets are used
according to configuration option CONFIG_RTE_LIBRTE_PMD_SZEDATA2_AS.
Setting link up/down and getting info about link status is done through
mmapped PCI resource address space.

Signed-off-by: Matej Vido <vido@cesnet.cz>
8 years agoszedata2: change to physical device type
Matej Vido [Thu, 28 Jan 2016 22:23:46 +0000 (23:23 +0100)]
szedata2: change to physical device type

PMD was of type PMD_VDEV which means that PCI device is not recognised
automatically during EAL initialization, but it has to be created by
EAL option --vdev.
Now, PMD is of type PMD_PDEV which means that PCI device is probed
and recognised during EAL initialization automatically.
Path to szedata2 device file is matched with device and the count
of available RX and TX DMA channels is found out during device
initialization.
Initialization, starting and stopping of queues is changed to better
correspond with Ethernet device API model. Function callbacks
(rx|tx)_queue_(start|stop) are added. Unnecessary items are removed
from ethernet device private data structure.

Signed-off-by: Matej Vido <vido@cesnet.cz>
8 years agonfp: fix Tx queue reset
Alejandro Lucero [Mon, 14 Mar 2016 10:39:04 +0000 (10:39 +0000)]
nfp: fix Tx queue reset

When using start-stop functionality the per queue fields need to
be properly reset.

Fixes: b812daadad0d ("nfp: add Rx and Tx")

Signed-off-by: Alejandro Lucero <alejandro.lucero@netronome.com>
8 years agonfp: fix how Tx checksum is advertised to firmware
Alejandro Lucero [Thu, 3 Mar 2016 11:26:07 +0000 (11:26 +0000)]
nfp: fix how Tx checksum is advertised to firmware

Even with tx checksum offload available, do not set the flag by default.

Fixes: b812daadad0d ("nfp: add Rx and Tx")

Signed-off-by: Alejandro Lucero <alejandro.lucero@netronome.com>
8 years agonfp: fix variable type in Tx checksum offload
Alejandro Lucero [Thu, 3 Mar 2016 11:05:52 +0000 (11:05 +0000)]
nfp: fix variable type in Tx checksum offload

The mbuf ol_flags field was changed to uin64_t with DPDK version 1.8

Fixes: b812daadad0d ("nfp: add Rx and Tx")

Signed-off-by: Alejandro Lucero <alejandro.lucero@netronome.com>
8 years agonfp: fix non-x86 build
Thomas Monjalon [Sat, 6 Feb 2016 21:51:16 +0000 (22:51 +0100)]
nfp: fix non-x86 build

The file sys/io.h was included but it can be unavailable in some
non-x86 toolchains.
As others system includes in the file nfp_net.c, it seems useless,
so the easy fix is to remove them.

Signed-off-by: Thomas Monjalon <thomas.monjalon@6wind.com>
Acked-by: Alejandro Lucero <alejandro.lucero@netronome.com>
8 years agomlx5: fix Rx checksum offload in non L3/L4 packets
Yaacov Hazan [Thu, 3 Mar 2016 14:27:39 +0000 (15:27 +0100)]
mlx5: fix Rx checksum offload in non L3/L4 packets

Change rxq_cq_to_ol_flags() to set checksum flags according to packet type,
so for non L3/L4 packets the mbuf chksum_bad flags will not be set.

Fixes: 67fa62bc672d ("mlx5: support checksum offload")

Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
8 years agomlx5: add VLAN filtering for broadcast and IPv6 multicast
Adrien Mazarguil [Thu, 3 Mar 2016 14:27:38 +0000 (15:27 +0100)]
mlx5: add VLAN filtering for broadcast and IPv6 multicast

Unlike promiscuous and allmulticast flows, those should remain
VLAN-specific.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx5: remove redundant debug message
Adrien Mazarguil [Thu, 3 Mar 2016 14:27:37 +0000 (15:27 +0100)]
mlx5: remove redundant debug message

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx5: manage all special flow types at once
Adrien Mazarguil [Thu, 3 Mar 2016 14:27:36 +0000 (15:27 +0100)]
mlx5: manage all special flow types at once

This commit adds helpers to remove redundant code.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx5: check port is configured as ethernet device
Or Ami [Thu, 3 Mar 2016 14:27:35 +0000 (15:27 +0100)]
mlx5: check port is configured as ethernet device

If the port link layer is not Ethernet, notify the user.

Signed-off-by: Or Ami <ora@mellanox.com>
8 years agomlx5: fix possible crash during initialization
Or Ami [Thu, 3 Mar 2016 14:27:34 +0000 (15:27 +0100)]
mlx5: fix possible crash during initialization

RSS configuration should not be freed when priv is NULL.

Fixes: 2f97422e7759 ("mlx5: support RSS hash update and get")

Signed-off-by: Or Ami <ora@mellanox.com>
8 years agomlx: use aligned memory to register regions
Adrien Mazarguil [Thu, 3 Mar 2016 14:27:17 +0000 (15:27 +0100)]
mlx: use aligned memory to register regions

The first and last memory pool elements are usually cache-aligned but not
page-aligned, particularly when using huge pages.

Hardware performance can be improved significantly by registering memory
regions starting and ending on page boundaries.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx5: free buffers immediately after completion
Nelio Laranjeiro [Thu, 3 Mar 2016 14:27:16 +0000 (15:27 +0100)]
mlx5: free buffers immediately after completion

This lowers the amount of cache misses.

Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
8 years agomlx5: avoid lkey retrieval for inlined packets
Nelio Laranjeiro [Thu, 3 Mar 2016 14:27:15 +0000 (15:27 +0100)]
mlx5: avoid lkey retrieval for inlined packets

Improves performance as the lkey is not needed by hardware in this case.

Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
8 years agomlx5: process offload flags only when requested
Nelio Laranjeiro [Thu, 3 Mar 2016 14:27:14 +0000 (15:27 +0100)]
mlx5: process offload flags only when requested

Improve performance by processing offloads only when requested by the
application.

Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
8 years agomlx5: remove one indirection level from Rx/Tx
Nelio Laranjeiro [Thu, 3 Mar 2016 14:27:13 +0000 (15:27 +0100)]
mlx5: remove one indirection level from Rx/Tx

Avoid dereferencing pointers twice to get to fast Verbs functions by
storing them directly in RX/TX queue structures.

Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
8 years agomlx5: reorder Rx/Tx queue structure
Nelio Laranjeiro [Thu, 3 Mar 2016 14:27:12 +0000 (15:27 +0100)]
mlx5: reorder Rx/Tx queue structure

Remove padding and move important fields to the beginning for better
performance.

Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
8 years agomlx5: prefetch next Tx mbuf header and data
Nelio Laranjeiro [Thu, 3 Mar 2016 14:27:11 +0000 (15:27 +0100)]
mlx5: prefetch next Tx mbuf header and data

This change improves performance noticeably.

Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
8 years agomlx5: support Rx VLAN stripping
Yaacov Hazan [Thu, 3 Mar 2016 14:26:44 +0000 (15:26 +0100)]
mlx5: support Rx VLAN stripping

Allows HW to strip the 802.1Q header from incoming frames and report it
through the mbuf structure.

This feature requires MLNX_OFED >= 3.2.

Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx5: support flow director
Yaacov Hazan [Thu, 3 Mar 2016 14:26:43 +0000 (15:26 +0100)]
mlx5: support flow director

Add support for flow director filters (RTE_FDIR_MODE_PERFECT and
RTE_FDIR_MODE_PERFECT_MAC_VLAN modes).

This feature requires MLNX_OFED >= 3.2.

Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Raslan Darawsheh <rdarawsheh@asaltech.com>
8 years agomlx5: make flow steering rule generator more generic
Yaacov Hazan [Thu, 3 Mar 2016 14:26:42 +0000 (15:26 +0100)]
mlx5: make flow steering rule generator more generic

Upcoming flow director support will reuse this function to generate filter
rules.

Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx5: add special flows for broadcast and IPv6 multicast
Yaacov Hazan [Thu, 3 Mar 2016 14:26:41 +0000 (15:26 +0100)]
mlx5: add special flows for broadcast and IPv6 multicast

Until now, broadcast frames were handled like unicast. Moving the related
flow to the special flows table frees up the related unicast MAC entry.

The same method is used to handle IPv6 multicast frames.

Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx5: refactor special flows handling
Yaacov Hazan [Thu, 3 Mar 2016 14:26:40 +0000 (15:26 +0100)]
mlx5: refactor special flows handling

Merge redundant code by adding a static initialization table to manage
promiscuous and allmulticast (special) flows.

New function priv_rehash_flows() implements the logic to enable/disable
relevant flows in one place from any context.

Signed-off-by: Yaacov Hazan <yaacovh@mellanox.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx5: fix header generation in parallel builds
Adrien Mazarguil [Fri, 29 Jan 2016 10:31:36 +0000 (11:31 +0100)]
mlx5: fix header generation in parallel builds

Fixes: 771fa900b73a ("mlx5: introduce new driver for Mellanox ConnectX-4 adapters")

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx5: support setting primary MAC address
David Marchand [Tue, 5 Jan 2016 18:00:09 +0000 (19:00 +0100)]
mlx5: support setting primary MAC address

Signed-off-by: David Marchand <david.marchand@6wind.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx4: support setting primary MAC address
David Marchand [Tue, 5 Jan 2016 18:00:08 +0000 (19:00 +0100)]
mlx4: support setting primary MAC address

Signed-off-by: David Marchand <david.marchand@6wind.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
8 years agomlx4: ensure number of Rx queues is a power of 2
Robin Jarry [Thu, 3 Mar 2016 14:27:40 +0000 (15:27 +0100)]
mlx4: ensure number of Rx queues is a power of 2

In the documentation it is specified that the hardware only supports a
number of RX queues if it is a power of 2.

Since ibv_exp_create_qp may not return an error when the number of
queues is unsupported by hardware, sanitize the value in dev_configure.

Signed-off-by: Robin Jarry <robin.jarry@6wind.com>
8 years agomlx4: fix unneeded function error with clang 3.6
Bruce Richardson [Mon, 8 Feb 2016 22:11:49 +0000 (22:11 +0000)]
mlx4: fix unneeded function error with clang 3.6

When compiling with clang 3.6, the mlx4 driver gives the following error
message about an unneeded function.

  CC mlx4.o
.../drivers/net/mlx4/mlx4.c:136:20: fatal error: function
      'wr_id_t_check' is not needed and will not be emitted
[-Wunneeded-internal-declaration]
static inline void wr_id_t_check(void)
                   ^
1 error generated.

The function is to compile-time check the size of wr_id_t, so use
the standard DPDK BUILD_BUG_ON macro to do so in the init function
instead.

Fixes: 7fae69eeff13 ("mlx4: new poll mode driver")

Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
Acked-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>