The device string has an existing size in rte_dev.h
use that instead of defining our own.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Use rte_pktmbuf_free_bulk instead of loop when freeing
packets.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Morten Brørup <mb@smartsharesystems.com>
Update the contributing guidelines to describe GitHub Actions first and
add a warning about Travis usage.
Fixes: 87009585e2 ("ci: hook to GitHub Actions")
Signed-off-by: David Marchand <david.marchand@redhat.com>
Acked-by: Aaron Conole <aconole@redhat.com>
In l3fwd no of transmit queues is calculated based on no of
lcores with which it is launched. Hence maximum no of tx
queues possible per port should depend on RTE_MAX_LCORE value.
Fixes: af75078fec ("first public release")
Cc: stable@dpdk.org
Signed-off-by: Harman Kalra <hkalra@marvell.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
When we're attaching fbarrays in secondary processes, we check for
whether the intended memory address for the fbarray is already in use by
some other, local fbarray. However, the check for end-overlap (i.e. to
see if our memory area's end overlaps with some other fbarray) is
incorrectly counting end offset as part of the overlap. Fix the check.
Fixes: 5b61c62cfd ("fbarray: add internal tailq for mapped areas")
Cc: stable@dpdk.org
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Tested-by: Zhihong Peng <zhihongx.peng@intel.com>
When no hugepages are found, we log a message about it, but we never
specify on which node. We also implicitly declare the page size based
on the directory name, but that's not very user friendly.
Fix both by changing the text of the message to note the NUMA node (if
applicable) and explicitly mention page size in kilobytes.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Reviewed-by: David Marchand <david.marchand@redhat.com>
The ticketlock test is fast and should be run all the time.
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Acked-by: Aaron Conole <aconole@redhat.com>
The Tx buffer may overflow when there is more than one port.
Fixes: 002ade70e9 ("app/test: measure cycles per packet in Rx/Tx")
Cc: stable@dpdk.org
Signed-off-by: Alvin Zhang <alvinx.zhang@intel.com>
Acked-by: Jeff Guo <jia.guo@intel.com>
Tested-by: Wei Ling <weix.ling@intel.com>
Implement support for the power management API by implementing a
`get_monitor_addr` function that will return an address of an RX ring's
status bit.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Implement support for the power management API by implementing a
`get_monitor_addr` function that will return an address of an RX ring's
status bit.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Acked-by: Jeff Guo <jia.guo@intel.com>
Implement support for the power management API by implementing a
`get_monitor_addr` function that will return an address of an RX ring's
status bit.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Add a simple API to allow getting the monitor conditions for
power-optimized monitoring of the Rx queues from the PMD, as well as
release notes information.
Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
Now that we have everything in a C file, we can store the information
about our sleep, and have a native mechanism to wake up the sleeping
core. This mechanism would however only wake up a core that's sleeping
while monitoring - waking up from `rte_power_pause` won't work.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Currently, the "sync" version of power monitor intrinsic is supposed to
be used for purposes of waking up a sleeping core. However, there are
better ways to achieve the same result, so remove the unneeded function.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Instead of passing around pointers and integers, collect everything
into struct. This makes API design around these intrinsics much easier.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Currently, the API documentation mandates that if the user wants to use
the power management intrinsics, they need to call the
`rte_cpu_get_intrinsics_support` API and check support for specific
intrinsics.
However, if the user does not do that, it is possible to get illegal
instruction error because we're using raw instruction opcodes, which may
or may not be supported at runtime.
Now that we have everything in a C file, we can check for support at
startup and prevent the user from possibly encountering illegal
instruction errors.
We also add return values to the API's as well, because why not.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Currently, power intrinsics are inline functions. Make them part of the
ABI so that we can have various internal data associated with them
without exposing said data to the outside world.
Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
Some Arm SoCs are not NUMA systems. Add the capability to disable NUMA
for cross build and disable NUMA in Arm cross files.
Signed-off-by: Juraj Linkeš <juraj.linkes@pantheon.tech>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Tested-by: Vimal Chungath <vcchunga@amazon.com>
Tested-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Add support for setting core count and numa nodes in cross files. The
values specified in cross files will override the default values.
Also add missing default values to Arm config.
Signed-off-by: Juraj Linkeš <juraj.linkes@pantheon.tech>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Tested-by: Vimal Chungath <vcchunga@amazon.com>
Tested-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Letting the compiler decide is going to yield the best results for
native builds, so use native machine args usable for both GCC and Clang.
Signed-off-by: Juraj Linkeš <juraj.linkes@pantheon.tech>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Tested-by: Vimal Chungath <vcchunga@amazon.com>
Tested-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Use generic configuration for the only build where it makes sense - the
generic build. For other builds, if we don't know either of implementer
ID or part number, the build is not supported.
Add part numbers to cross files where fallback to generic configuration
is assumed.
Signed-off-by: Juraj Linkeš <juraj.linkes@pantheon.tech>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Tested-by: Vimal Chungath <vcchunga@amazon.com>
Tested-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Use dictionary lookup instead of checking for existing variables,
iterating over all elements in the list or checking lists for optional
configuration. Move variable contents into the dictionary for variables
that would be referenced only once.
Fallback to generic part number if the discovered part number is
unknown.
Signed-off-by: Juraj Linkeš <juraj.linkes@pantheon.tech>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Tested-by: Vimal Chungath <vcchunga@amazon.com>
Tested-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Set flags in one loop. Append flags to a list and use the list in the
loop.
Signed-off-by: Juraj Linkeš <juraj.linkes@pantheon.tech>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Tested-by: Vimal Chungath <vcchunga@amazon.com>
Tested-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Change formatting so that it's more consistent and readable, add/modify
comments/stdout messages, move configuration options to more appropriate
places and make the order consistent according to these rules:
1. First list generic configuration options, then list options that may
be overwritten. List SoC-specific options last.
2. For SoC-specific options, list number of cores before the number of
NUMA nodes, to make it consistent with config/meson.build.
Signed-off-by: Juraj Linkeš <juraj.linkes@pantheon.tech>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Tested-by: Vimal Chungath <vcchunga@amazon.com>
Tested-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Remove variables that were either not used, referenced just once or not
needed.
Signed-off-by: Juraj Linkeš <juraj.linkes@pantheon.tech>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Tested-by: Vimal Chungath <vcchunga@amazon.com>
Tested-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
Rename Arm build variables and values so that they better conform to Arm
specifications. Also rename generically sounding variable to names that
better capture what the variables hold.
Rename machine_args_generic to part_number_config_arm since the
variable contains more than just the generic machine args and is used
mainly as the fallback arm configuration.
Rename the default machine args to generic machine args to reflect that.
The rest of the variables are self-explanatory.
Signed-off-by: Juraj Linkeš <juraj.linkes@pantheon.tech>
Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Jerin Jacob <jerinj@marvell.com>
Tested-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Tested-by: Vimal Chungath <vcchunga@amazon.com>
Tested-by: Dharmik Thakkar <dharmik.thakkar@arm.com>
The librte_cfgfile lib is functional on Windows.
Enable compilation of this lib for Windows.
Signed-off-by: Narcisa Vasile <navasile@linux.microsoft.com>
Add support for secondary processes in ioat devices. The update
allocates a memzone for a primary process or returns it in a
secondary process.
Signed-off-by: Kumar Amber <kumar.amber@intel.com>
Acked-by: Bruce Richardson <bruce.richardson@intel.com>
Currently bitmap line not empty check API assumes cache line
of 64B and only checks 8 slabs. Since in 128B cacheline, we
have 16 slabs per cacheline, rte_bitmap_clear() will mark
complete line as empty as soon as 8 slabs are empty thereby
breaking bitmap scan functionality. Fix it by defining new
__rte_bitmap_line_not_empty() for 128B cacheline platform.
Signed-off-by: Nithin Dabilpuram <ndabilpuram@marvell.com>
Acked-by: Cristian Dumitrescu <cristian.dumitrescu@intel.com>
Add a test case to test scan operation post clear of half
cacheline of slabs.
Also fix meson.build to include test_bitmap.c in the compilation.
Signed-off-by: Nithin Dabilpuram <ndabilpuram@marvell.com>
When a VF device is present, netvsc can send or receive packets over the
VF device. The VF device driver communicates directly with the PCI device
via the PF from the host hypervisor. This is faster than exchanging data
with netvsp via vmbus, i.e. syntheic path.
In Azure and Hyper-v environments, VF device can be hot added or hot
removed at anytime while guest VM is running. This patch improves netvsc
to support VF device hot add/remove.
1. netvsc monitors all system hot add activities over the PCI bus. When it
detects a VF device is added to the system and is managed under this
netvsc device, it asks EAL to probe and start this VF device, then it
attaches and switches data path to the VF device.
2. After a VF device is attached to netvsc, netvsc monitors this device on
hot remove. When this VF device is hot removed, netvsc switches data path
to synthetic, stops this VF device and removes it from EAL.
3. If any failure happens during a VF device hot remove or add, the netvsc
falls back to synthetic path for all data traffic.
Signed-off-by: Long Li <longli@microsoft.com>
In some cases, a device or infrastructure may want to enable hotplug
but application may also try and start hotplug as well. Therefore
change the monitor_started from a boolean into a reference count.
Signed-off-by: Long Li <longli@microsoft.com>
When running one test (via DPDK_TEST) the test program
would leave the terminal in raw mode. This was because
it was setting up cmdline to do interactive input.
The fix is to use cmdline_new() for the interactive case.
This also fixes a memory leak because the test
runner was never calling cmdline_free().
Fixes: 9b848774a5 ("test: use env variable to run tests")
Cc: stable@dpdk.org
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Tested-by: Harry van Haaren <harry.van.haaren@intel.com>
Trivial fix to for spelling errors and incorrect spacing.
No change to any built code.
Fixes: 7a61fc5d1b ("test/rwlock: add new test-cases")
Fixes: af75078fec ("first public release")
Cc: stable@dpdk.org
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
Explicitly cast void * to type * so that EAL headers may be compiled
as C or C++.
Fixes: e8428a9d89 ("eal/windows: add some basic functions and macros")
Cc: stable@dpdk.org
Signed-off-by: Tyler Retzlaff <roretzla@linux.microsoft.com>
Because Wei Hu has changed to a new job and the
email address (xavier.huwei@huawei.com) has expired,
we remove him from the hns3 maintainer list.
All patches signed-off-by Wei Hu will be copied to Lijun Ou.
Signed-off-by: Lijun Ou <oulijun@huawei.com>
When testing compilation and checking ABI compatibility,
there is no real need of static binaries eating disks.
The static linkage of applications was already well tested,
though the static examples tested with meson were limited to "l3fwd" only.
The static build test with make is limited to "helloworld" example.
The ABI compatibility is checked on shared libraries,
and there is no need to test again on similar builds.
A new parameter is added to the function "build",
so the ABI check is enabled only for native gcc and clang shared builds,
32-bit, generic armv8 and ppc cross compilations.
In other words, it is disabled for some static builds and some Arm ones.
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Acked-by: David Marchand <david.marchand@redhat.com>
Each core already comes with its local storage for mcslock (in its
stack), therefore there is no need to define an additional per-lcore
mcslock.
Fixes: 32dcb9fd2a ("test/mcslock: add MCS queued lock unit test")
Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Currently, when rte_service_init() fails at initialization, the
application always gets a ENOEXEC error code. For example, with testpmd,
this is displayed as:
Cannot init EAL: Exec format error
This error code does not describe the real issue. Instead, use the error
code returned by the function.
Fixes: e398245008 ("service: initialize with EAL")
Cc: stable@dpdk.org
Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
NetUIO device class and interface GUIDs are defined in system
headers starting from platform SDK v10.0.20253. Inspect SDK
version to avoid redefinition.
Pre-release SDKs do not promise compatibility and a narrow
subset of SDKs may still be subject to redefinition.
Fixes: c76ec01b45 (bus/pci: support netuio on Windows)
Cc: stable@dpdk.org
Signed-off-by: Tyler Retzlaff <roretzla@linux.microsoft.com>
Acked-by: Dmitry Kozlyuk <dmitry.kozliuk@gmail.com>
Acked-by: Ranjit Menon <ranjit.menon@intel.com>
Some figures had a title inside the picture but not in RST file.
As a consequence, some versions of Sphinx are emitting a warning.
Warning, treated as error:
doc/guides/prog_guide/graph_lib.rst:64:
no number is assigned for figure: figure-anatomy-of-a-node
The titles are moved from SVG to RST,
except for graph_mem_layout.svg where in-picture title must be kept.
Fixes: 4dc6d8e63c ("doc: add graph library guide")
Cc: stable@dpdk.org
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
Reviewed-by: Jerin Jacob <jerinj@marvell.com>
Device name parsing is done on all buses during device iterations at
either EAL or ethdev levels.
When a bus implements device name parsing slowly, all iterations are
impacted. Efficient implementation is important.
The DPAA bus device name parsing has two issues: it allocates dynamic
memory and uses snprintf without a real need for it. Both can be
avoided, which improves the parsing performance.
The function is also simpler and shorter.
Signed-off-by: Gaetan Rivet <grive@u256.net>
In some situations, we would get several ip fragments, which total
data length is less than min_ip_len(64) and padding with zeros.
We simulated intermediate fragments by modifying the MTU.
To illustrate the problem, we simplify the packet format and
ignore the impact of the packet header.In namespace2,
a packet whose data length is 1520 is sent.
When the packet passes tap2, the packet is divided into two
fragments: fragment A and B, similar to (1520 = 1510 + 10).
When the packet passes tap3, the larger fragment packet A is
divided into two fragments A1 and A2, similar to (1510 = 1500 + 10).
Finally, the bond interface receives three fragments:
A1, A2, and B (1520 = 1500 + 10 + 10).
One fragmented packet A2 is smaller than the minimum Ethernet
frame length, so it needs to be padded.
|---------------------------------------------------|
| HOST |
| |--------------| |----------------------------| |
| | ns2 | | |--------------| | |
| | |--------| | | |--------| |--------| | |
| | | tap1 | | | | tap2 | ns1| tap3 | | |
| | |mtu=1510| | | |mtu=1510| |mtu=1500| | |
| |--|1.1.1.1 |--| |--|1.1.1.2 |----|2.1.1.1 |--| |
| |--------| |--------| |--------| |
| | | | |
| |-----------------| | |
| | |
| |--------| |
| | bond | |
|--------------------------------------|mtu=1500|---|
|--------|
When processing the preceding packets above,
DPDK would aggregate fragmented packets A2 and B.
And error packets are generated, which padding(zero)
is displayed in the middle of the packet.
A2 + B:
0000 fa 16 3e 9f fb 82 fa 47 b2 57 dc 20 08 00 45 00
0010 00 33 b4 66 00 ba 3f 01 c1 a5 01 01 01 01 02 01
0020 01 02 c0 c1 c2 c3 c4 c5 c6 c7 00 00 00 00 00 00
0030 00 00 00 00 00 00 00 00 00 00 00 00 c8 c9 ca cb
0040 cc cd ce cf d0 d1 d2 d3 d4 d5 d6 d7 d8 d9 da db
0050 dc dd de df e0 e1 e2 e3 e4 e5 e6
So, we would calculate the length of padding, and remove
the padding in pkt_len and data_len before aggregation.
And also we have the fix for both ipv4 and ipv6.
Fixes: 7f0983ee33 ("ip_frag: check fragment length of incoming packet")
Cc: stable@dpdk.org
Signed-off-by: Yicai Lu <luyicai@huawei.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
As most NICs do not support segmentation for VXLAN-encapsulated
UDP/IPv4 packets, this patch adds VXLAN UDP/IPv4 GSO support.
Signed-off-by: Yi Yang <yangyi01@inspur.com>
Acked-by: Jiayu Hu <jiayu.hu@intel.com>
Allows i40e and mlx5 PMDs to compile on Windows and disable other drivers.
Disable few i40e warnings with Clang such as comparison of integers of
different signs and macro redefinitions.
Signed-off-by: Pallavi Kadam <pallavi.kadam@intel.com>
Reviewed-by: Ranjit Menon <ranjit.menon@intel.com>
Acked-by: Tal Shnaiderman <talshn@nvidia.com>
Signed-off-by: Thomas Monjalon <thomas@monjalon.net>
The file rte_random.c is required to build i40e PMD on Windows.
Add rte_rand variable to export file.
Redefine _m_prefetchw for Clang toolchain due to following error
with respect to conflicting types:
FAILED: lib/76b5a35@@rte_eal@sta/librte_eal_common_rte_random.c.obj
clang @lib/76b5a35@@rte_eal@sta/librte_eal_common_rte_random.c.obj.rsp
In file included from ../lib/librte_eal/common/rte_random.c:13:
In file included from ..\lib/librte_eal/include\rte_eal.h:20:
In file included from ..\lib/librte_eal/include\rte_per_lcore.h:25:
In file included from ..\lib/librte_eal/windows/include\pthread.h:21:
In file included from ..\lib/librte_eal/windows/include\rte_windows.h:27:
In file included from C:\Program Files (x86)\Windows Kits\10\include\
10.0.18362.0\um\windows.h:171:
In file included from C:\Program Files (x86)\Windows Kits\10\include\
10.0.18362.0\shared\windef.h:24:
In file included from C:\Program Files (x86)\Windows Kits\10\include\
10.0.18362.0\shared\minwindef.h:182:
C:\Program Files (x86)\Windows Kits\10\include\10.0.18362.0\um\
winnt.h:3324:1: error: conflicting types for '_m_prefetchw'
_m_prefetchw (
^
C:\Program Files\LLVM\lib\clang\10.0.0\include\prfchwintrin.h:50:1:
note: previous definition is here
_m_prefetchw(void *__P)
^
1 error generated.
Signed-off-by: Pallavi Kadam <pallavi.kadam@intel.com>
Reviewed-by: Ranjit Menon <ranjit.menon@intel.com>
Enable mlx5 common driver on Windows with clang compilation.
Signed-off-by: Ophir Munk <ophirmu@nvidia.com>
Signed-off-by: Tal Shnaiderman <talshn@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
While compiling with clang 11 the callers of the
__mlx5_bit_off macro warns on the cast of pointers to
unsigned long which is a smaller int type in Windows.
warning: cast to smaller integer type 'unsigned long'
from 'u8 (*)[16]' [-Wpointer-to-int-cast]
To resolve it the type is changed to uintptr_t to be
compatible for both Linux and Windows.
Fixes: 865a0c1567 ("net/mlx5: add Direct Verbs prepare function")
Cc: stable@dpdk.org
Signed-off-by: Tal Shnaiderman <talshn@nvidia.com>
Acked-by: Matan Azrad <matan@nvidia.com>
