Using clib_bitmap_vec_validate makes free bitmap vector
to be x64 times bigger (assuming x86_64) than necessary
when non-zero and possible oom due (u32)(0 - 1) math with
zero alloc.
Fix it with clib_bitmap_validate which takes bit size, not
index and ensure at least one bit is allocated.
Type: fix
Change-Id: I7e191f4e2fb3722a06bb800e1d075f7c7e2dcec9
Signed-off-by: Vladislav Grishenko <themiron@yandex-team.ru>
In a case where one pounds on a single kvp in a KVP_AT_BUCKET_LEVEL
table, the code would sporadically return a transitional value (junk)
from a half-deleted kvp. At most, 64-bits worth of the kvp will be
written atomically, so using memset(...) to smear 0xFF's across a kvp
to free it left a lot to be desired.
Performance impact: very mild positive, thanks to FC for doing a
multi-thread host stack perf/scale test.
Added an ASSERT to catch attempts to add a (key,value) pair which
contains the magic "free kvp" value.
Type: fix
Signed-off-by: Dave Barach <dave@barachs.net>
Change-Id: I6a1aa8a2c30bc70bec4b696ce7b17c2839927065
Exported indentifiers in Go start with capital letters. Only few fields
in hs-test, which are being unmarshaled from yaml are required to be
exported. Every other field name or method name should start with
lower-case letter, to be consistent with this naming convention.
Type: test
Signed-off-by: Maros Ondrejicka <mondreji@cisco.com>
Change-Id: I7eab0eef9fd08a7890c77b6ce1aeb3fa4b80f3cd
Location changed and binary renamed to test_infra
Also it is built by default.
Type: improvement
Change-Id: I27cd97f274501ceb7a01213e2bc9676cea00f39c
Signed-off-by: Damjan Marion <damarion@cisco.com>
Used on intel client CPUs which suppport VAES instruction set without
AVX512
Type: improvement
Change-Id: I5f816a1ea9f89a8d298d2c0f38d8d7c06f414ba0
Signed-off-by: Damjan Marion <damarion@cisco.com>
DMA batch status was set by hardware. Its value may be variable between
cpus twice accesses. Saving the value of status can fix it.
Type: fix
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Change-Id: Ibc9337239555744a571685b486c986991c3e9b18
Recognize and drive google virtual ethernet (gve) in google cloud.
Type: feature
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Change-Id: Ia559615ac059cabbca5d10bcd4049e87beaad638
Fix the typo in the intrinsic name, which caused incorrect intrinsic to be used.
Type: fix
Signed-off-by: Andrew Yourtchenko <ayourtch@gmail.com>
Change-Id: Ib7fde14d12897e4d1bfb5a01f6d65025473e4f8e
Rename unused SESSION_IO_EVT_BUILTIN_TX to SESSION_IO_EVT_TX_MAIN and
leverage it for non-connected udp tx.
Non-connected udp sessions are listeners and are therefore allocated on
main thread. Consequently, whenever session queue node is not polling
main, tx events generated by external applications might be missed or
processed with some delay. To solve this, request that apps use
SESSION_IO_EVT_TX_MAIN tx events as opposed to SESSION_IO_EVT_TX and
send that to first worker as opposed to main.
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I5df5ac3dc80c0f192b2eefb1d465e9deefe8786b
Specify the number of max_batches when applying for dma config.
Skip this round when no batch available from vlib_dma_batch_new.
Type: improvement
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Change-Id: Ic6e0acf81ba4fc3ed33aea6ac6990ef841021c59
Allocate and initialize dma batch structure when adding dma config.
The number of required dma batches is set by max_batches parameter.
Thus dma batches are not allocated dynamically in worker thread.
Application need to check the return value of vlib_dma_batch_new.
Type: improvement
Signed-off-by: Marvin Liu <yong.liu@intel.com>
Change-Id: I5d05a67b59634cf2862a377d5ab77cb1040343ce
Make sure endpoint freelist is drained before alloc of fixed local
source port is tried.
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I302deee5609a463af8135185af71722ac8c55a27
Those CPUs are announcing VAES capability but they don't support AVX512.
Type: fix
Fixes: 73a60b2
Change-Id: I7b4be95e91bb6f367cd71461f1126690f3ecd988
Signed-off-by: Damjan Marion <damarion@cisco.com>
udp_output_get_connection handles correctly if the connection
is a listener whereas udp_connection_get does not which may lead
to a crash.
Type: fix
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I40b57287a8686820d29872cae2cfd6ae27a57c26
This avoids printing ldp debug messages while debug is disabled and vcl
is initializing.
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I5dfd1d59032db937fea146b6b84b8e26307a0de0
Process 8 packets perf batch in vlib_buffer_free_inline() when
CLIB_HAVE_VEC512 is enabled.
Type: improvement
Signed-off-by: Leyi Rong <leyi.rong@intel.com>
Change-Id: I78b8a525bce25ee355c9bf0e0f651698a8c45bda
Type: fix
In vpp, file descriptor handler closes the fd upon error
if there is no error handling function is registered.
This patch fixes the issue for af_packet interface by
registering the error handling function.
Errors will also be gracefully logged.
Signed-off-by: Mohsin Kazmi <sykazmi@cisco.com>
Change-Id: I260d780ac54ffd0199dcd6ca5b95e5afe957e968
Also make sure that only sessions with fifos try to set deq notification
flag on fifo
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I878c2d2e18bb98109ee03b42a4f0f8c48aa23e9f
For non-connected udp, when retrieving the subscriber session to send
the notification, it uses the current worker thread index whereas the
subscriber session is actually on the main thread. Using the worker
thread may cause a crash since the corresponding session may not be
valid in the worker thread context and even if it is valid, it is the
wrong session. This scenario is seen when the application forks
and adds subscribers to the worker thread session.
Type: fix
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: I236ee9d9ff9f3b2f7f9f8e782d70d1080aa1b627
Type: test
Add a helper wrapper script for vppctl called vppcli to vpp docker image
with proper cli socket path.
Signed-off-by: Filip Tehlar <ftehlar@cisco.com>
Change-Id: I1a51aa54bc91c1c812698501a56401c525d498e8
If not, worker threads may continue own loops after deinit and/or
thread0 exit with related crashes due no rpc capability, unmapped
shared memory, etc. Main loop exit handlers that uses barrier sync
will be happy too as long as recursive barrier sync is supported.
Type: feature
Signed-off-by: Vladislav Grishenko <themiron@yandex-team.ru>
Change-Id: I255a796b06936d96715683e3f062128060233dc6
format_udp_connection takes 2 arguments from the caller.
Type: fix
Signed-off-by: Steven Luong <sluong@cisco.com>
Change-Id: Ie618a809936a01c094982f9a8c81309826e0b087
Although removal from epoll means listener no longer accepts new
sessions, the accept queue built by vpp cannot be drained by stopping
the listener. Morover, some applications, e.g., nginx, might constantly
remove and add listeners to their epfds. Removing listeners in such
situations causes a lot of churn in vpp as segments and segment managers
need to be recreated.
Type: improvement
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: Ia412b3f8d50fbb4881a99ff024f798353b521af7
Otherwise if vcl epoll lt events are ignored by the app, libc and vcl mq
events are never drained.
Type: fix
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I1e22f6da46d56236c52714181f6c20dcb80a33a5
- avoid setting LD_PRELOAD for container
- save nginx error log to shared volume
- reduce test run time to 10s
- add vcl and ldp debug env variables to docker file. Default to
disabled.
Type: test
Signed-off-by: Florin Coras <fcoras@cisco.com>
Change-Id: I401ac74e7c0ebe87befedb44150b04f773f244ea
vec_alloc() does not mark vector as accessible contrary to
vec_validate().
Also removes redundant memset(0) as vector allocation always zeroed
new memory.
Type: fix
Change-Id: I8309831b964a618454ed0bebbcdec7ec21149414
Signed-off-by: Benoît Ganne <bganne@cisco.com>
