- Why I did it
New implementation of Nvidia platform_wait due to:
1. sysfs deprecated by hw-mgmt
2. new dependencies to SDK
3. For CMIS host management mode
- How I did it
wait hw-management ready
wait SDK sysfs nodes ready
- How to verify it
manual test
unit test
sonic-mgmt regression
- Why I did it
Fix issue xcvrd crashes due to cannot import name 'initialize_sfp_thermal':
Nov 27 09:47:16.388639 sonic ERR pmon#xcvrd: Exception occured at CmisManagerTask thread due to ImportError("cannot import name 'initialize_sfp_thermal' from partially initialized module 'sonic_platform.thermal' (most likely due to a circular import) (/usr/local/lib/python3.9/dist-packages/sonic_platform/thermal.py)")
- How I did it
Add lock for creating SFP object
- How to verify it
Unit test
Manual Test
- Why I did it
When module is totally under software control, driver cannot get module temperature/temperature threshold from firmware. In this case, sonic needs to get temperature/temperature threshold from EEPROM. In this PR, a thread thermal updater is created to update module temperature/temperature threshold while software control is enabled.
- How I did it
Query ASIC temperature from SDK sysfs and update hw-management-tc periodically
Query Module temperature from EEPROM and update hw-management-tc periodically
- How to verify it
Manual test
New Unit tests
- Why I did it
When module is under software control (CMIS host management enabled), EEPROM should be controlled by software and there should be no limitation for any write operation.
- How I did it
Remove EEPROM write limitation if a module is under software control
- How to verify it
Manual test
UT
- Why I did it
Update SAI version to SAIBuild2311.26.0.28
Fixed issues
1. Traffic with unicast destination ip and multicast destination mac wasn't properly dropped
2. When working with SAI_DEFAULT_SWITCHING_MODE_STORE_FORWARD key/value enabled, trying to add a LAG member to a LAG which is created after warm boot initial configuration phase ended, will fail.
3. Optional feature of Port IP counters (SAI_PORT_STAT_IP*) , enabled by SAI XML per-port-ip-counter-enabled config node, wasn't initialized properly.
4. Creating BFD session for non default VRF fails (SAI_BFD_SESSION_ATTR_VIRTUAL_ROUTER != SAI_SWITCH_ATTR_DEFAULT_VIRTUAL_ROUTER_ID).
5. The default value for port FEC during switch init for Spectrum3 was initialized as 'auto' and not aligned to SAI header default 'none'. Note if setups has invalid configuration and relied previously on auto, now it might be necessary for the user to provide explicit valid value for SAI_PORT_ATTR_FEC_MODE
Update SDK/FW version to 4.6.2134/2012.2134
Fixed issues:
1. Updated SN3700C to enable limit to 100G speed.
2. Recovering from Low power mode might ends with port down.
- How I did it
Updating the versions in makefile
- How to verify it
Confirm issues fixed and run sonic-mgmt tests
- Why I did it
For cmis host management mode, the prevous sysfs cannot be used for low power mode setting. This PR reuses existing low power mode implementation in sonic_xcvr package when CMIS host management mode is enabled
- How I did it
Use sonic_xcvr low power mode implementation when CMIS host management mode is enabled.
- How to verify it
Manual test for CMIS host management mode
Regression test for old mode and backward compatible test
A W/A to overcome delay of about 20 sec on login due to MFT bash autocompletion bug.
Should be reverted once a formal solution will be available in future MFT release.
- Why I did it
To overcome SN2700 20 sec delay on login
- How I did it
Removed MFT bash autocompletion part
- How to verify it
1. Build a mellanox image
2. Verify no such links after system boot.
Signed-off-by: Nazarii Hnydyn <nazariig@nvidia.com>
- Why I did it
Enable CMIS host management for Mellanox devices which are expected to support the feature
- How I did it
new thread in a new file and changing logic in platform code in chassis.py which is calling this thread from get_change_event()
this thread in the new file handles the state machine per port.
first the static detection takes place once the thread is up (during switch bootup sequence), until final decision if it's FW control or SW control module.
After it ends, the dynamic detection takes place, listening to changes in the sysfs fds, per port,
so it will be able to detect plug in or out events of a cable.
- How to verify it
Enhanced unit tests
run sonic mgmt on Nvidia SN4700 with CMIS host management enabled
- Why I did it
Provide a dummy implementation for SFP error description when CMIS host management is enabled. A future feature shall be raised to implement SFP error description for such mode.
- How I did it
if SFP is under software control, provide "Not supported" as error description
if SFP is under initialization, provide "Initializing" as error description
- How to verify it
unit test
- Why I did it
The current low power mode setting implementation requests the user to set the port to admin down first before toggling LP mode, this is not backward compatible, now revert it to the old way so that the user can toggle the LP mode regardless of the port admin status.
- How I did it
Revert the recent changes related to LPM in PR #14130 and #16545
- How to verify it
Run all sfputil and SFP platform API related tests on all the Mellanox platforms.
Signed-off-by: Kebo Liu <kebol@nvidia.com>
Why I did it
Update SDK/SAI and FW for Mellanox Platform
How I did it
Update SDK/FW to v4.6.2104/v2012.2104
Fixed Issues:
Some of the Warmboot related files which were created by SDK during switch create are now generated during pre shutdown flow
New Features:
Debian 12 and kernel 6.1 support
Update SAI
New Features:
Auto Fec Support
FDB entries are now restored after warmboot to prevent temporary system flooding.
Minor Enhancement and Bug Fix in integrate-mlnx-sdk
How to verify it
Build Image and run tests
Signed-off-by: Vivek Reddy <vkarri@nvidia.com>
Why I did it
Add platform support for Debian 12 (Bookworm) on Mellanox Platform
How I did it
Update hw-management to v7.0030.2008
Deprecate the sfp_count == module_count approach in favour of asic init completion
Ref: Mellanox/hw-mgmt@bf4f593
Add xxd package to base image which is required by hw-management scripts
Add the non-upstream flag into linux kernel cache options
Update the thermalctl logic based on new sysfs attributes
Fix the integrate-mlnx-hw-mgmt script to not populate the arm64 Kconfig
How to verify it
Build kernel and run platform tests
Signed-off-by: Vivek Reddy <vkarri@nvidia.com>
Co-authored-by: Junchao-Mellanox <junchao@nvidia.com>
Co-authored-by: Junchao-Mellanox <57339448+Junchao-Mellanox@users.noreply.github.com>
Signed-off-by: Vivek Reddy <vkarri@nvidia.com>
[Nvidia] Enable iproute2 & fix mft build (#16)
* Enable iproute2 as the SDK is also built
Signed-off-by: Vivek Reddy <vkarri@nvidia.com>
* [Nvidia] Dont use mkbmdeb method of dkms to build the package
Signed-off-by: Vivek Reddy <vkarri@nvidia.com>
* Added linux image to the Depends section of mft
Signed-off-by: Vivek Reddy <vkarri@nvidia.com>
[Nvidia] [Bookworm] Separate KERNEL_MFT into a new target (#16782)
* [Nvidia] Seperate KERNEL_MFT into a new target because of kernel header dependency
Signed-off-by: Vivek Reddy <vkarri@nvidia.com>
* Update linux-kernel submodule
Signed-off-by: Vivek Reddy <vkarri@nvidia.com>
* Fix paralell build problem
Signed-off-by: Vivek Reddy <vkarri@nvidia.com>
---------
Signed-off-by: Vivek Reddy <vkarri@nvidia.com>
- Why I did it
Support running hw-management service on MSN4700 emulation platform.
- How I did it
Use physical EEPROM instead of the fake one
Do not skip PSUd, PCId, thermal control daemon
Adjust PCIe and thermal configuration files
Adjust platform.json for different chassis names and thermals
Remove a patch to hw-management in order to enable it
- How to verify it
Run Nvidia simulation on SN4700 (ASIC and Platform)
Signed-off-by: Stephen Sun <stephens@nvidia.com>
- Why I did it
In order to activate FW after it was upgraded need to perform reboot.
If reboot wasn't performed and user need to upgrade to another SONiC image then it will fail.
The reason for that is that during SONiC upgrade new FW should be installed but it will fail because previously installed FW wasn't activated.
In order to allow 2nd FW upgrade without reboot in-between need to reactivate FW image.
This change handles such flow.
Example of issue scenario:
User installed SONiC image on the switch
Then for some reason FW was upgraded by user or script but reboot was not performed to activate it.
After that upgrade to new SONiC image will fail because new image need to install FW but it fails due to previous one wasn't activated.
- How I did it
In "mlnx-fw-upgrade" script check if FW upgrade failed with the error that FW was already installed but reboot was not performed.
If so then perform FW image reactivation and try to upgrade FW again.
- How to verify it
Install SONiC image on the switch
Then upgrade FW but don't perform reboot.
After that upgrade to new SONiC image and check that upgrade was successfull.
Signed-off-by: Volodymyr Samotiy <volodymyrs@nvidia.com>
- Why I did it
Add an ability to add arm64 mellanox specific kconfig using the integration tool
Fix the existing duplicate kconfig problem by using the vanilla .config
Add an ability to patch kconfig-inclusions file. Renamed series.patch to external-changes.patch to reflect the behavior
NOTE: Min hw-mgmt version to use with these changes: V.7.0030.2000 not yet upstream but required prio to it.
This option will be enabled one the new hw mgmt will be upstream.
Depends on sonic-net/sonic-linux-kernel#336
Signed-off-by: Vivek Reddy Karri <vkarri@nvidia.com>
- Why I did it
hw-management renamed PSU temperature related sysfs:
psu1_temp -> psu1_temp1
psu2_temp -> psu2_temp1
psu1_temp_max -> psu1_temp1_max
psu2_temp_max -> psu2_temp1_max
This PR is to align the change in SONiC.
- How I did it
Use new sysfs node for PSU temperature and PSU temperature threshold
- How to verify it
Manual test
sonic-mgmt Regression test
Why I did it
SONiC service determine-reboot-cause might run before driver creating reset cause files. In that case, the reset cause will be "Unknown". This PR introduces a wait mechanism to wait for reset cause sysfs files ready.
How I did it
/run/hw-management/config/reset_attr_ready is the file to indicate all reset cause files are ready. In chassis.get_reboot_cause function, it waits /run/hw-management/config/reset_attr_ready for up to 45 seconds.
How to verify it
Manual test on master/202211/202205
Sfp api can now be called from the host which doesn't have the python_sdk_api installed. Also, sfp api has been migrated to use sysfs instead of sdk handle.
Signed-off-by: Vivek Reddy Karri <vkarri@nvidia.com>
- Why I did it
SAI bug Fixes
1. When creating an ACL rule with SAI_ACL_ENTRY_ATTR_FIELD_SRC_IP/SAI_ACL_ENTRY_ATTR_FIELD_DST_IP enabled, and then disabling the field by setting enable=false, a match on L3_type=IPv4 will remain programmed for the rule Issue resolved after the fix
2. Allow the max scale of virtual routers to be configure for SPC-1, SPC-2, SPC-3 which is 255 when fastboot enable and 511 when fastboot disable
3. Remove default hash key of SRC_MAC, DST_MAC and ETH_TYPE
SAI features
1. Port init profile
2. Dual ToR Active-Standby | Additional MAC support
SDK/FW bug fixes
1. When preforming fast boot from an old SDK version (currently installed) to a newer one (target version), and the system was initially loaded with a new SDK version (past version), and the system has not been wiped, under specific conditions, the fast boot would use the past version's data and may fail.
- How I did it
Update SAI version to SAIBuild2211.25.1.4
Update SDK/FW version to 4.6.1062/2012.1062
- Why I did it
1. Update Mellanox HW-MGMT package to newer version V.7.0030.1011
2. Replace the SONiC PMON Thermal control algorithm with the one inside the HW-MGMT package on all Nvidia platforms
3. Support Spectrum-4 systems
- How I did it
1. Update the HW-MGMT package version number and submodule pointer
2. Remove the thermal control algorithm implementation from Mellanox platform API
3. Revise the patch to HW-MGMT package which will disable HW-MGMT from running on SIMX
4. Update the downstream kernel patch list
Signed-off-by: Kebo Liu <kebol@nvidia.com>
### Why I did it
1. Enhance the diagnosis information collecting mechanism
- If the option `-v` is fed, it will pass additional diagnosis flags to mlxfwmanager
- Collect all the output from mlxfwmanager and print them to syslog if it fails
2. Abort syncd in case waiting for device or upgrading firmware fails
Signed-off-by: Stephen Sun <stephens@nvidia.com>
### How I did it
#### How to verify it
Regression and manual test
- Why I did it
Because the Spectrum4 devices don't support mlxtrace utility.
- How I did it
Edit sai.profile and remove mlxtrace_spectrum4_itrace_*.cfg.ext files
Signed-off-by: vadymhlushko-mlnx <vadymh@nvidia.com>
Why I did it
To avoid errors when the sfputil show error-status -hw is called from the host OS (not from the pmon docker).
How I did it
Remove the self.sdk_handle parameter from the _get_module_info() function.
How to verify it
Execute the sfputil show error-status -hw
Signed-off-by: vadymhlushko-mlnx <vadymh@nvidia.com>
- Why I did it
watchdogutil uses platform API watchdog instance to control/query watchdog status. In Nvidia watchdog status, it caches "armed" status in a object member "WatchdogImplBase.armed". This is not working for CLI infrastructure because each CLI will create a new watchdog instance, the status cached in previous instance will totally lose. Consider following commands:
admin@sonic:~$ sudo watchdogutil arm -s 100 =====> watchdog instance1, armed=True
Watchdog armed for 100 seconds
admin@sonic:~$ sudo watchdogutil status ======> watchdog instance2, armed=False
Status: Unarmed
admin@sonic:~$ sudo watchdogutil disarm =======> watchdog instance3, armed=False
Failed to disarm Watchdog
- How I did it
Use sysfs to query watchdog status
- How to verify it
Manual test
Unit test
- Why I did it
Change Mellanox platform API implementation to use ASIC driver sysfs for the module operational state and status error fields.
- How I did it
Modify the platform/mellanox/mlnx-platform-api/sonic_platform/sfp.py file by change the call of sx_mgmt_phy_module_info_get() SDK API to sysfs
- How to verify it
Simulate the unplug cable event
Check the CLI output
sfputil show presence
sfputil show error-status -hw
Simulate the plug cable event
Repeat 2 step
Signed-off-by: vadymhlushko-mlnx <vadymh@nvidia.com>
SONiC changes:
1. Support Spectrum4 ASIC FW binary building.
2. Support new SDK sx-obj-desc lib building since new SAI need it.
3. Remove SX_SCEW debian package from Mellanox SDK build since we are no longer using it (we use libxml2 instead).
4. Update SAI, SDK, FW to version 4.6.1020/2012.1020/SAIBuild2305.25.0.3
SDK/FW bug fixes
1. In SPC-1 platforms: Fastboot mode is not operational for Split port with Force mode in 50G speed
SFP modules are kept in disabled state after set LPM (low power mode) on/off for at least 3 minutes.
2. When preforming fast boot from an old SDK version (currently installed) to a newer one (target version), and the system was initially loaded with a new SDK version (past version), and the system has not been wiped, under specific conditions, the fast boot would use the past version's data and may fail.
SDK/FW Features
1. On SN2700 all ports can support y cable by credo
SAI bug Fixes
1. When creating an ACL rule with SAI_ACL_ENTRY_ATTR_FIELD_SRC_IP/SAI_ACL_ENTRY_ATTR_FIELD_DST_IP enabled, and then disabling the field by setting enable=false, a match on L3_type=IPv4 will remain programmed for the rule Issue resolved after the fix
2. Allow the max scale of virtual routers to be configure for SPC-1, SPC-2, SPC-3 when fastboot enable
3. Remove default hash key of SRC_MAC, DST_MAC and ETH_TYPE
SAI features
1. Port init profile
- How I did it
Update SDK/FW/SAI make files
- How to verify it
Run full sonic-mgmt regression on Mellanox platform
Signed-off-by: Kebo Liu <kebol@nvidia.com>
- Why I did it
Update Mellanox MFT tool to version 4.25.0-62
- How I did it
Update the MFT tool make file
- How to verify it
Run full sonic-mgmt regression.
Signed-off-by: Kebo Liu <kebol@nvidia.com>
- Why I did it
Increase UT coverage for Nvidia platform API code
Work item tracking
Microsoft ADO (number only):
- How I did it
Focus on low coverage file:
1. component.py
2. watchdog.py
3. pcie.py
- How to verify it
Run the unit test, the coverage has been changed from 70% to 90%
- Why I did it
Added the fwtrace config files in order to be able to call the mlxstrace utility during the show techsupport dump.
Work item tracking
Microsoft ADO (number only):
- How I did it
Added fwtrace config files. Added path to these files to sai.profile for each mlnx device.
- How to verify it
Execute the show techsupport command and check if mlxstrace output is in system dump.
Signed-off-by: vadymhlushko-mlnx <vadymh@nvidia.com>
= Why I did it
To optimize Mellanox platform SAI build
- How I did it
SAI debs are now downloaded as Spectrum-SDK-Drivers-SONiC-Bins release.
- How to verify it
Configure/build for Mellanox platform, check the image and ensure that correct SAI debs are included.
- Why I did it
The reset cause "reset_from_comex" has been removed by hw-management, hence removing it from platform API code
- How I did it
Remove reset_from_comex from reboot cause mapping
- How to verify it
Manual test
- Why I did it
BIOS on new generation switch can come with a file type of cap or cab. Needs to add support to these file type.
Also ONIE version on new devices can have a suffix of 'dev'.
- How I did it
Added cap & cab as possible component extensions for ComponentBIOS.
Update the ONIE version regex to include dev signed versions.
- How to verify it
Update BIOS.
Why I did it
Sharing the storage of syncd with other proprietary application extensions allows them to communicate with syncd in differnt ways.
If one container wants to pass some information to syncd then shared storage can be used. However, today the shared storage isn't cleaned on restarts making it possible for syncd to read out-of-date information generated in the past.
NOTE: No plans to use it for standard SONIC dockers and we are working on removing the SDK dependency from PMON docker
How I did it
Implemented new service to clean the shared storage.
How to verify it
Do reboot/fast-reboot/warm-reboot/config-reload/systemctl restart swss and verify /tmp/ is cleaned after each restart in syncd container.
Signed-off-by: Stepan Blyschak <stepanb@nvidia.com>
- Why I did it
SDK patches for iproute2 were added to SONiC tree as a temporary solution.
Now that SDK with the patches is available, I have removed the patches from SONiC tree and we consume them from SDK github during compilation.
- How I did it
During build we download SDK iproute2 patches from SDK github (or from the URL provided by user if compiling SDK from sources) and apply them before compilation.
- How to verify it
Compile and load on switch, verify interfaces network devices created successfully.
Verify LLDP shows connections to neighbors.
Verify ping between 2 hosts over 2 router ports is successful.
- Why I did it
Adjust the warning threshold implementation according to the latest algorithm update
- How I did it
Modify power warning and critical thresholds methods
- How to verify it
Unit test updated to cover the change
Signed-off-by: Stephen Sun <stephens@nvidia.com>
- Why I did it
Add the patchwork link to the commit description for non-upstream patches if present
- How I did it
Parse the patchwork/<patch_name>.txt file from hw-mgmt
- Why I did it
Bug fix:
- * I2C bus is stuck - Unable to probe I2C bus 2-0048, which causes /var/run/hw-management/config/sfp_counter, module_counter to be zero and pmon docker unable to start.
- How I did it
Update HW-MGMT package version in the make file
Update HW-MGMT submodule pointer
-How to verify it
Run full sonic-mgmt regression
Signed-off-by: Kebo Liu <kebol@nvidia.com>
#### Why I did it
Facilitate Automatic integration of sdk kernel patches into SONiC.
**Inputs to the Script:**
1) `MLNX_SDK_VERSION` Eg: `4.5.4206`
2) `MLNX_SDK_ISSU_VERSION` Eg: `101`
**Note: If nothing is provided the one already present in the sdk.mk file is used**
3) `MLNX_SDK_SOURCE_BASE_URL:`
**Note: If nothing is provided the upstream sdk drivers url is used**
4) `CREATE_BRANCH: (y|n)` Creates a branch instead of a commit (optional, default: n)
5) `BRANCH_SONIC`: Only relevant when CREATE_BRANCH is y. `Default: master`.
Note: These should be provided through `SONIC_OVERRIDE_BUILD_VARS ` parameter
**Output:**
1) Script creates a commit in sonic-linux-kernel with any updates to sdk-kernel patches in sonic in accordance with the version provided by `MLNX_SDK_VERSION`
**Note: Script Doesn't commit anything to linux-kernel when there aren't any changes required..**
#### How I did it
1) Added a new make target which can be invoked by calling `make integrate-mlnx-sdk`
```
user@server:/sonic-buildimage/src/sonic-linux-kernel$ git rev-parse --abbrev-ref HEAD
master_6f38dca_integrate_4.5.4206
user@server:/sonic-buildimage/src/sonic-linux-kernel$ git log --oneline -n 1
d64d1e7 (HEAD -> master_6f38dca_integrate_4.5.4206) Intgerate MLNX SDK 4.5.4206 Kernel Patches
```
Changes made will be summarized under `sonic-buildimage/integrate-mlnx-sdk_user.out` file. Debugging and troubleshooting output is written to `sonic-buildimage/integrate-mlnx-sdk.log` files
[log_files.zip](https://github.com/sonic-net/sonic-buildimage/files/11226441/log_files.zip)
#### Limitations:
1) Assumes that the sdk kernel patches are always upstreamed
#### How to verify it
Build the Kernel and test
Why I did it
Align with SAI headers v1.12.0
Work item tracking
Microsoft ADO (number only):
How I did it
Update Mellanox SAI submodule
How to verify it
Compile SONiC image
- Why I did it
The current implementation of SFP reset, LPM, present relies on SDK API. This PR moves the implementation to SDK sysfs. By this PR, it gains following benefit:
1. SDK sysfs provides better performance.
2. Host side and container side share the same code.
3. Code is much cleaner.
- How I did it
Use SDK sysfs to implement SFP reset, LPM, present.
- How to verify it
1. Manual test.
2. Unit test.
SDK/FW Fixed Issues:
• When a system has more than 256 ACL entries, on rare occasion, removing/adding entries may cause some ACL entries not to work.
• When using mirror session policer on spectrum-2, spectrum-3, the actual CIR was 1.28 times more than the configured CIR value
• After warm boot process, when enabling ECN marking and the port is in split mode, traffic sent to the port under congestion (for example, when connecting two ports with a total speed of 50GbE to a single 25GbE port) is not marked.
• Warm boot might fail if the key value SAI_KEY_ACCUMULATED_FLOW_COUNTER_UNITS_IN_KB is set
• If counters are bound to an next hop group, there is a probability the next API calls that modify the next-hop group members will fail.
• In Spectrum platforms Fastboot mode is not operational for Split port with Force mode in 50G speed
• When fine grain next hop group has a size of 2K or 4K members, and group is removed, FW will remove only (size % 2048) members, resulting in leakage of KVD resources
• When reading some port statistics, or bulk reading some Queue or PG statistics, and in parallel reading or writing other counters, FW may, in rare cases, get stuck
• SN2201 Module 1 is considered to be present/linked while no cable/module is plugged
• On Spectrum-3 when port configure to 400G FW might stuck after running mlxlink while 400G interface connected and swap between upper and lower 4 lanes
SAI New features:
• ACL: Added support for an ACL match on the AETH field (SAI_ACL_TABLE_ATTR_FIELD_AETH_SYNDROME, SAI_ACL_ENTRY_ATTR_FIELD_AETH_SYNDROME) to count RoCE NAK and CNP packets.
• PLL Status: Added a new logging entry that alerts the user upon a PLL lock loss event.
• Dual ToR - Additional MAC Address: Added support for setting a MAC address for the router interface which is not part of the 10 bit MAC address available for RIFs on Spectrum-1, as part of the Dual ToR scenario.
• Dual ToR: DSCP Remapping Added support for tunnel QoS maps as part of the Dual TOR scenario.
SAI Fixed issues:
• When setting a WRED profile attribute for a color that was not enabled during the profile create time, an error would be returned. After the fix, a default profile is create on such scenario and the set attribute is applied on top of it
• When calling the flush FDB by using the SAI_FDB_FLUSH_ATTR_BRIDGE_PORT_ID attribute, the bridge bv_id value was filled on the notification callback where it should have been left empty.
Signed-off-by: Kebo Liu <kebol@nvidia.com>
- Why I did it
As a LED indicator to help user to find switch location in the lab, UID LED is a useful LED in Mellanox switch.
- How I did it
I add a new member _led_uid in Mellanox/Chassis.py, and extend Mellanox/led.py to support blue color.
Relevant platform-common PR sonic-net/sonic-platform-common#369
- How to verify it
Add unit test cases in test.py, and do manual test including turn-on/off/show uid led.
Signed-off-by: David Xia <daxia@nvidia.com>
