Per the TWS session...
- Some simple treatment of the contents of pool.h
- Changing some \brief commands to @brief. (will do a more
complete pass at this later.)
Change-Id: I050ee69c59c4b572ac295b5f86940b7f4c934cd9
Signed-off-by: Chris Luke <chrisy@flirble.org>
mheap_alloc allocates memory_size bytes of memory and returns
a page-aligned memory space prefixed with a word-aligned
header of type mheap_t.
This header includes the actual usable space size, but
when the requested size is too small, the computed size
was incorrect (infinite).
mheap_alloc now returns 0 in such cases.
With help from Yoann Desmouceaux.
Change-Id: I00af63d573d6939aca53dbe7ff612b726bd8f0df
Signed-off-by: Pierre Pfister <ppfister@cisco.com>
This change-set enables plugins to add themselves to the ip4/ip6
feature subgraphs without having to modify core vpp engine code
at all. Add VNET_IP4/IP6_UNICAST/MULTICAST_FEATURE_INIT macros
which express the required ordering constraints, and off you go.
Along the way, added an implementation of Warshall's algorithm to
vppinfra; to compute the positive transitive closure of a relation. In
this case, the relation is "feature A runs before feature B."
With that in hand, ip_feature_init_cast(...) computes a partial order
across the set of configured feature subgraph nodes.
In unit-testing, we discovered VPP-145 - ip4/6 inacl wiped out
vnet_buffer(b)->ip>current_config_index, which exists in main. So, we
fixed that by moving b->trace_index, adding b->current_config_index,
and removing the ip opaque union current_config_index.
Change-Id: Iff132116f66413dc6b31ac3377198c7a32d51f48
Signed-off-by: Dave Barach <dave@barachs.net>
New parameter allows specifying which worker threads will process rx
queues. Parameter arguments is list of cores and number of worker specified
must be equal to the number of rx queues configured (num-rx-queues). If
num-rx-queues is not specified, it will be automatically set to
number of workers.
Sample config:
dpdk {
dev 0000:86:00.0 {
workers 2,3
num-rx-queues 2
}
}
Change-Id: I88bc381e0e542eb02def09a726c6f04de3e1ae17
Signed-off-by: Damjan Marion <damarion@cisco.com>
worker-handoff node is universal node which taakes packets
from the input node and hands them over to worker threads.
Currently it supports flow hashing based on ipv4, ipv6 and
mpls headers.
New cli:
set interface handoff <intrerface-name> workers <list>
e.g.
set interface handoff TenGigabitEthernet2/0/0 workers 3-6,9-10
Change-Id: Iaf0df83e69bb0e84969865e0e1cdb000b0864cf5
Signed-off-by: Damjan Marion <damarion@cisco.com>
* compiler -march= parameter is changed from native to corei7
so code is always genereted with instructions which are available
on the Nehalem microarchitecture (up to SSE4.2)
* compiler -mtune= parameter is added so code is optimized for
corei7-avx which equals to Sandy Bridge microarchitecture
* set of macros is added which allows run-time detection of available
cpu instructions (e.g. clib_cpu_supports_avx())
* set of macros is added which allows us to clone graph node funcitons
where cloned function is optmized for different microarchitecture
Those macros are using following attributes:
__attribute__((flatten))
__attribute__((target("arch=core-avx2)))
I.e. If applied to foo_node_fn() macro will generate cloned
functions foo_node_fn_avx2() and foo_node_fn_avx512() (future)
It will also generate function void * foo_node_fn_multiarch_select()
which detects available instruction set and returns pointer to the
best matching function clone.
Change-Id: I2dce0ac92a5ede95fcb56f47f3d1f3c4c040bac0
Signed-off-by: Damjan Marion <damarion@cisco.com>
- Configures Doxygen.
- Adds a source filter to do magic on our use of the preprocessor to do
constructor stuff to make Doxygen grok it better.
- Adds a convenience helper to the root Makefile.
- Adds a README.md to the root directory (and which Doxygem uses as its
"mainpage".
- Add several other documentative files.
- Currently using SVG for call graphs, though this may have a
load-time performance impact in browsers.
Change-Id: I25fc6fb5bf634319dcb36a7f0e32031921c125ac
Signed-off-by: Chris Luke <chrisy@flirble.org>
gcc version 4.9.2 (Raspbian 4.9.2-10)
Tested on Linux raspberrypi 4.4.6-v7+ #875 SMP Tue Apr 12 16:33:02 BST 2016 armv7l GNU/Linux
CPUs may be little or big endian, detect with gcc flags, not the processor architecture
Add a new flag $(PLATFORM)_uses_openssl which allows to disable the link with openssl lib.
vlib/vlib/threads.c:
startup.conf must:
- specify the heapsize as we don't have hugepages on raspbian
cpu {
main-core 3
}
heapsize 64M
Corrects in various files the assumption uword == u64 and replaces 'u64' cast with 'pointer_to_uword' and 'uword_to_pointer' where appropriate.
256 CPUs may create an OOM when testing with small memory footprint ( heapsize 64M ), allows the number of VLIB_MAX_CPUS to be set in platforms/*.mk
vppinfra/vppinfra/longjmp.S:
ARM - copy r1 (1st parameter of the setjmp call) to r0 (return value)
vppinfra/vppinfra/time.h:
On ARMv7 in AArch32 mode, we can access to a 64bit register to retreive the cycles count.
gcc on rpi only declare ARM_ARCH 6. Override this info, and check if it is possible to use 'mrrc'.
/!\ the time function will NOT work without allowing the user mode access to the PMU.
You may download the source of the kmod here:
https://github.com/christophefontaine/arm_rdtsc
Change-Id: I8142606436d9671a184133b935398427f08a8bd2
Signed-off-by: Christophe Fontaine <christophe.fontaine@qosmos.com>
"make check" in build-root/build-$tag-$arch/vppinfra should now work on a
machine with 3G of spare memory
Change-Id: Ie6a22e8496b56b93aa90fdc81e3e0c10dc7514fb
Signed-off-by: Florin Coras <fcoras@cisco.com>
This fixes issue observed on Ubuntu 16.04 where
dynamic loader is not finding correct instance of
specific structure.
Change-Id: I618d0933c7e171b8a9b40495b36894785af7790a
Signed-off-by: Damjan Marion <damarion@cisco.com>
Note that compiling -O3 doesn't improve performance as of this
writing, might as well clean up warnings in any event.
Change-Id: Ic2f4982d12fbbf36f5324075183982731759dc94
Signed-off-by: Dave Barach <dave@barachs.net>
It also includes check to ensure that number of
per-cpu mheaps is not lower than number of cpus.
Change-Id: Ibc68b34dda130f922243f9ea15b03e44bbcac269
Signed-off-by: Damjan Marion <damarion@cisco.com>
vec_sort macro was using gcc proprietary nested functions that
require a executable stack and they are considered as unsafe.
Also, nested functions are not supported by other compilers.
vec_sort_with_function() should be used instead.
Change-Id: I05959da63d222ec71c090ba63420b427ce10c79b
Signed-off-by: Damjan Marion <damarion@cisco.com>
Fixes "cpu_config: no such thread type 'corelist-workers'" issue
Change-Id: Ic0309ee62859ac73fd58c57b5f630aff5daf1775
Signed-off-by: Damjan Marion <damarion@cisco.com>
