In kni_allocate_mbufs(), we attempt to add max_burst (32) count of mbuf
always into alloc_q, which is excessively leading too many rte_pktmbuf_
free() when alloc_q is contending at high packet rate (for eg 10Gig data).
In a situation when alloc_q fifo can only accommodate very few (or zero)
mbuf, create only what needed and add in fifo.
With this patch, we could stop random network stall in KNI at higher packet
rate (eg 1G or 10G data between vEth0 and PMD) sufficiently exhausting
alloc_q on above condition. I tested i40e PMD for this purpose in ppc64le.
Signed-off-by: Gowrishankar Muthukrishnan <gowrishankar.m@linux.vnet.ibm.com>
Acked-by: Ferruh Yigit <ferruh.yigit@intel.com>
rte_pktmbuf_headroom() and rte_pktmbuf_tailroom() should be usable
with any segment, not only with headered ones, so is_header should be 0
when we call for sanity check inside them.
Fixes: af75078fece3 ("first public release")
Cc: stable@dpdk.org
Signed-off-by: Vasily Philipov <vasilyf@mellanox.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
There is no need for initializing the complete
packet buffer with zero as the packet data area will be
overwritten by the NIC Rx HW anyway.
The testpmd configures the packet mempool
with around 180k buffers with
2176B size. In existing scheme, the init routine
needs to memset around ~370MB vs the proposed scheme
requires only around ~22MB on 128B cache aligned system.
Useful in running DPDK in HW simulators/emulators,
where millions of cycles have an impact on boot time.
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
It is safe to enable LIBRTE_VHOST_NUMA by default for all
configurations where libnuma is already a default dependency.
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Acked-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Tested-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Currently EAL allocates hugepages one by one not paying attention
from which NUMA node allocation was done.
Such behaviour leads to allocation failure if number of available
hugepages for application limited by cgroups or hugetlbfs and
memory requested not only from the first socket.
Example:
# 90 x 1GB hugepages availavle in a system
cgcreate -g hugetlb:/test
# Limit to 32GB of hugepages
cgset -r hugetlb.1GB.limit_in_bytes=34359738368 test
# Request 4GB from each of 2 sockets
cgexec -g hugetlb:test testpmd --socket-mem=4096,4096 ...
EAL: SIGBUS: Cannot mmap more hugepages of size 1024 MB
EAL: 32 not 90 hugepages of size 1024 MB allocated
EAL: Not enough memory available on socket 1!
Requested: 4096MB, available: 0MB
PANIC in rte_eal_init():
Cannot init memory
This happens beacause all allocated pages are
on socket 0.
Fix this issue by setting mempolicy MPOL_PREFERRED for each hugepage
to one of requested nodes using following schema:
1) Allocate essential hugepages:
1.1) Allocate as many hugepages from numa N to
only fit requested memory for this numa.
1.2) repeat 1.1 for all numa nodes.
2) Try to map all remaining free hugepages in a round-robin
fashion.
3) Sort pages and choose the most suitable.
In this case all essential memory will be allocated and all remaining
pages will be fairly distributed between all requested nodes.
New config option RTE_EAL_NUMA_AWARE_HUGEPAGES introduced and
enabled by default for linuxapp except armv7 and dpaa2.
Enabling of this option adds libnuma as a dependency for EAL.
Fixes: 77988fc08dc5 ("mem: fix allocating all free hugepages")
Signed-off-by: Ilya Maximets <i.maximets@samsung.com>
Acked-by: Sergio Gonzalez Monroy <sergio.gonzalez.monroy@intel.com>
Acked-by: Hemant Agrawal <hemant.agrawal@nxp.com>
Acked-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Tested-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
This patch remove the deprecated functions as well as notice for
scheduler mode set/get API changes.
Signed-off-by: Fan Zhang <roy.fan.zhang@intel.com>
Acked-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
In one of the SNOW3G unit tests, insufficient memory
was allocated, leading to a buffer overflow.
Fixes: 2b52e1e95aab ("app/test: add SNOW 3G UEA2 with offset")
Cc: stable@dpdk.org
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Flag dev_started should be cleared after dev_stop() function call
because the flag is checked inside the dev_stop() function.
Fixes: d11b0f30df88 ("cryptodev: introduce API and framework for crypto devices")
Cc: stable@dpdk.org
Signed-off-by: Kirill Rybalchenko <kirill.rybalchenko@intel.com>
Acked-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Chain parameter can be CIPHER_HASH, HASH_CIPHER,
CIPHER_ONLY or HASH_ONLY, but only the first two
were shown in the application help.
Fixes: 1a75e9f3fadb ("examples/l2fwd-crypto: add cipher/hash only cases")
Cc: stable@dpdk.org
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
Instead of using duplicated functions to get the algorithm
strings, use the functions from the cryptodev library.
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
Instead of passing the authentication and cipher offset
from the test vectors, just get them from the IV and AAD lengths.
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
For AES/DES tests, there is no need to leave out
the first block from the crypto operation.
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
AAD should not point at IV for AES algorithms.
For AES-GCM, AAD will point at additional data in the mbuf.
For the other algorithms (such as 3DES CBC), AAD is not used.
Fixes: ffbe3be0d4b5 ("app/test: add libcrypto")
Cc: stable@dpdk.org
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Fiona Trahe <fiona.trahe@intel.com>
Some assert checks in the driver were
incorrect, but they are not necessary anyway,
as application will panic in any case.
Fixes: 0f548b50a160 ("crypto/aesni_mb: process crypto op on dequeue")
Cc: stable@dpdk.org
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
Add PCI probe/remove/init/uninit functions in a separate
file rte_cryptodev_pci.h, which do not use cryptodev driver,
in order to be removed in next commits.
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
Call rte_cryptodev_pmd_release_device() if probing a
PCI crypto device, instead of accessing the variables
directly. This will be useful when rte_cryptodev_pci_probe()
gets moved to a separate file.
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
Move all functions handling virtual devices to a separate
header file "rte_cryptodev_vdev.h", in order to leave only
generic functions for any device in the rest of the files.
It also creates the file "rte_cryptodev_pmd.c", with the
implementations of these functions.
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
Do not set PCI information in the device information structure
for any crypto device, just for the ones that are PCI, so
this is set internally in the PCI crypto PMDs (only QAT now).
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
rte_cryptodev_devices_get() function returns an array of devices
sharing the same driver.
Instead of having two different paths depending on the device being
virtual or physical, retrieve the driver name from rte_device structure.
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
rte_cryptodev_devices_get() function was parsing a crypto
device name as an argument, but the function actually
returns device identifiers of devices that share the
same crypto driver, so the argument should be driver name, instead.
Fixes: 38227c0e3ad2 ("cryptodev: retrieve device info")
Cc: stable@dpdk.org
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
When retrieving device information for a crypto driver,
driver name was only set when it was a PCI driver.
Getting the driver name from rte_device structure
allows rte_cryptodev_get_info() function to return it
regardless they are virtual or physical devices.
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
Only non virtual devices were storing the pointer to
rte_device structure in rte_cryptodev, which will be needed
to retrieve the driver name for any device.
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Acked-by: Declan Doherty <declan.doherty@intel.com>
Statistics period time option is parsed with -T argument,
but -t was accepted by mistake, instead.
Fixes: 387259bd6c67 ("examples/l2fwd-crypto: add sample application")
Cc: stable@dpdk.org
Signed-off-by: Pablo de Lara <pablo.de.lara.guarch@intel.com>
Currently when a malloc_elem is split after resizing, any padding
present in the elem is ignored. This causes the resized elem to be too
small when padding is present, and user data can overwrite the beginning
of the following malloc_elem.
Solve this by including the size of the padding when computing where to
split the malloc_elem.
Fixes: af75078fece3 ("first public release")
Signed-off-by: Jamie Lavigne <lavignen@amazon.com>
Acked-by: Sergio Gonzalez Monroy <sergio.gonzalez.monroy@intel.com>
Currently, pool resources are allocated statically
and are not freed. Results of that test can not run more than once.
Fix removes static dependency from test application and
now allocating and freeing resources dynamically.
Test runs for more than once.
Fixes: af75078fece3 ("first public release")
Signed-off-by: Santosh Shukla <santosh.shukla@caviumnetworks.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
When debug level logging enabled (--log-level=8) each driver failed to
probe the device printed, like:
EAL: Driver (net_ark) doesn't match the device
EAL: Driver (net_avp) doesn't match the device
EAL: Driver (net_bnxt) doesn't match the device
EAL: Driver (net_cxgbe) doesn't match the device
EAL: Driver (net_e1000_igb) doesn't match the device
EAL: Driver (net_e1000_igb_vf) doesn't match the device
EAL: Driver (net_e1000_em) doesn't match the device
EAL: Driver (net_ena) doesn't match the device
EAL: Driver (net_enic) doesn't match the device
EAL: Driver (net_fm10k) doesn't match the device
EAL: Driver (net_i40e) doesn't match the device
EAL: Driver (net_i40e_vf) doesn't match the device
....
Overall hundreds of similar lines printed, because all drivers printed
for all devices. This is too much noise and there is already a log
message printed when device matched.
Removing the debug log completely.
Signed-off-by: Ferruh Yigit <ferruh.yigit@intel.com>
The --exclude parameter must be passed before the input directory to
tar, otherwise it's silently ignored and the .doctrees directory is
installed by make install-doc.
Signed-off-by: Luca Boccassi <luca.boccassi@gmail.com>
Acked-by: John McNamara <john.mcnamara@intel.com>
Depending on the environment, make might echo the command being ran.
In mk/rte.sdkdoc.mk make is used to print the DPDK version to be
piped to doxygen. This causes the following to be written:
<div id="projectname">DPDK
 <span id="projectnumber">/usr/bin/make-f/build/dpdk-jYjqnr/
dpdk-16.11.2/mk/rte.sdkconfig.mkshowversion</span>
</div>
Use -s (--silent) to prevent echoing.
Signed-off-by: Luca Boccassi <luca.boccassi@gmail.com>
Acked-by: John McNamara <john.mcnamara@intel.com>
The function rte_eth_tx_done_cleanup() was missing in the map file
so it cannot be used by applications linking to shared libraries.
pktgen uses it since version 3.2.0.
Fixes: 44a718c457b5 ("ethdev: add API to free consumed buffers in Tx ring")
Cc: stable@dpdk.org
Signed-off-by: Luca Boccassi <luca.boccassi@gmail.com>
The NUMA node information for PCI devices provided through
sysfs is invalid for AMD Opteron(TM) Processor 62xx and 63xx
on Red Hat Enterprise Linux 6, and VMs on some hypervisors.
It is good to see more checking for valid values.
Typical wrong numa node in some VMs:
$ cat /sys/devices/pci0000:00/0000:00:18.6/numa_node
-1
Signed-off-by: Tonghao Zhang <nic@opencloud.tech>
Acked-by: Sergio Gonzalez Monroy <sergio.gonzalez.monroy@intel.com>
The error return code for rte_ring_sc_dequeue_bulk() and
rte_ring_mc_dequeue_bulk() function should be -ENOENT rather
than -ENOBUFS as described in the function description.
Fixes: cfa7c9e6fc1f ("ring: make bulk and burst return values consistent")
Cc: stable@dpdk.org
Signed-off-by: Anand B Jyoti <anand.b.jyoti@intel.com>
Acked-by: Olivier Matz <olivier.matz@6wind.com>
Defining the value 0 as default value for dequeue timeout
will help the application reduce the configuration setup
if the application is interested only in default
timeout value.
removed "min_dequeue_limit" negative testcase as
min_dequeue_limit value could be zero(which is
default timeout now) if driver has
dev_info->min_dequeue_timeout_ns = 1.
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
octeontx specific priority test expects priority of each
event queue to be a unique value. Verify that condition
before it processes to test the priority.
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Typically RTE_EVENT_OP_NEW issued by the producer
lcore. To reflect the write changes issued by the
producer lcore on worker lcore, an SMP write barrier
is required on producer enqueue. Fixing the missing
rte_smp_wmb() on enqueue with RTE_EVENT_OP_NEW.
Fixes: f10d322eff76 ("event/octeontx: support worker enqueue")
Cc: stable@dpdk.org
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Acked-by: Gage Eads <gage.eads@intel.com>
switch tag wait is a costly operation as it may
translate to IOB read if core swtag cache is not updated.
Do tag switch wait only when there is a tag request on
the same hardware work slot.
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Acked-by: Gage Eads <gage.eads@intel.com>
Introducing the burst mode capability flag to express the event device
is capable of operating in burst mode for enqueue(forward, release) and
dequeue operation. If the device is not capable, then the application
still uses the rte_event_dequeue_burst() and rte_event_enqueue_burst()
but PMD accepts only one event at a time which is any way transparent
with the current rte_event_*_burst API semantics.
It solves two purposes:
1) Fix performance regression on the PMD which supports only nonburst
mode, and this issue is two-fold.
Typically the burst_worker main loop consists of following pseudo code:
while(1)
{
uint16_t nb_rx = rte_event_dequeue_burst(ev,..);
for (i=0; i < nb_rx; i++) {
process(ev[i]);
if (is_release_required(ev[i]))
release_the_event(ev);
}
uint16_t nb_tx = rte_event_enqueue_burst(dev_id, port_id,
events, nb_rx);
while (nb_tx < nb_rx)
nb_tx += rte_event_enqueue_burst(dev_id, port_id,
events + nb_tx, nb_rx - nb_tx);
}
Typically the non_burst_worker main loop consists of following pseudo code:
while(1)
{
uint16_t nb_rx = rte_event_dequeue_burst(&ev, , 1);
if (!nb_rx)
continue;
process(ev);
while (rte_event_enqueue_burst(dev, port, &ev, 1) != 1);
}
Following overhead has been seen on nonburst mode capable PMDs with
burst mode version
- Extra explicit release(PMD does release on implicitly on next
dequeue) and thus avoids the cost additional driver function overhead.
- Extra "for" loop for event processing which compiler cannot detect at
runtime
2) Simplify the application configuration by avoiding the application to
find the correct enqueue and dequeue depth across different PMD.
If burst mode is not supported then, PMD can ignore depth field.
This will enable to write portable applications and makes
RFC eventdev_pipeline application works on OCTEONTX PMD
http://dpdk.org/dev/patchwork/patch/23799/
If an application wishes to get the maximum performance on nonburst
capable PMD then the application can write the code in a way that by
keeping packet processing function as inline functions and launch the
workers based on the capability.
The generic burst based worker still work on those PMDs without
any code change but this scheme needed only when the application wants
to gets the maximum performance out of nonburst capable PMDs.
This patch is based the on the real world test cases
http://dpdk.org/dev/patchwork/patch/24832/, Where without this scheme
20.9% performance drop observed per core.
See worker_wrapper(), perf_queue_worker(), perf_queue_worker_burst()
functions to use this scheme in a portable way without losing performance
on both sets of PMDs and achieving the portability.
http://dpdk.org/dev/patchwork/patch/24832/
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
Made libeventdev library independent of VDEV bus by moving vdev pmd
specific function to rte_eventdev_pmd_vdev.h header file. Eventdev VDEV
PMD can include that for generic eventdev VDEV init and uninit function
enablement.
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Made libeventdev library independent of PCI bus by moving pci pmd
specific function to rte_eventdev_pmd_pci.h header file. Eventdev PCI
PMD can include that for generic eventdev PCI probe and remove function
enablement.
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Remove rte_event_dev_close() from rte_event_pmd_release() function so
that rte_event_pmd_release() can be used in stateless way. This will
enable rte_event_pmd_vdev_uninit() function to avoid using
eventdev_globals global variable and the need for exposing the a
global variable to PMD.
Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
