-
Notifications
You must be signed in to change notification settings - Fork 213
OE4T Meeting Notes 2022 06 09
Dan Walkes edited this page Jun 11, 2022
·
2 revisions
8
- Concerned about losing search history on slack chats with the free account. Billing would be around $200 per month with current active users, up to $800/month if all members were active.
- Important to have an app that’s on all the time, can be installed in Linux
- Ideally find a place which is more of a hangout for embedded stuff, might be best to use IRC.
- Dan will lead investigation of gittr and possibly other tools, experiment with these as alternatives to slack.
- Follow up on discussion from last month.
- Doesn’t work on 32.7.1, ultimately backported to 32.6.1. Will revisit upstream approach later.
- NVIDIA's UUID path is too complicated. Current approach as discussed with others - taking out UUID completely.
- Still need to figure out a way to solve for arbitrary partitions.
- Really like the sysinstaller design used by Matt. Question is how to bring that into tegra demo distro and meta-tegra. One thing that could help is to switch from keystore to lukserv application. Lukserv application is the client application that talks to trusted application. Switch to one context to generate one passphrase for the entire system and not one passphrase per partition. This makes the cryptab able to be static instead of dynamic. Otherwise need to know at flash time what all the UUIDs are. Instead use the part labels which are already known ahead of time.
- Biggest concern about NVIDIA implementation is the amount of stuff happening during flash time. Required if you have anything in LUKS boot you need to make unencrypted.
- Simplest use cases are working, A/B and more complicated use cases are still a work in progress.
- Some things are really obvious like using custom flash layout, things are straightforward. The complicated piece is figuring out what to put into tegraflash class and imagetypes.tegra, and can we avoid putting things there.
- Tegraflash helper script changes from NVIDIA can be avoided, since you boot into an OS and install from there. Too hard-coded to the Ubuntu OS to be useful most likely.
- Need to refactor initrd script form NVIDIA which assumes UUID implementation. Probably not a big deal.
- Tried to do stock tegra-test-distro, however tegra-bootinfo segfaulted on first run.
- Most of the work is done. See https://github.com/OE4T/meta-tegra-community/pull/27
- Not using SONAME for libraries, breaks all the shared library linking.
- Ilies will add an NVIDIA developer forum post for the issues noticed there.
- Jetpack 5 docker images don’t fit in emmc flash
- May make sense to make a smaller demo image on tegra-demo-distro with minimal components for docker build.
- Ran into compile errors building the kernel with latest meta-openembedded and poky. Latest build from test distro worked fine. Matt has been holding consistent at kirkstone instead of master for the jetpack 5 branch.
- Most testing and fixes from openembedded are happening on kirkstone, possibly some stuff broken in master.
- If boot slots get unaligned once they stay unaligned. Not self healing.
- Plan to submit a PR for reverting slots after update failure
- Has been an issue historically on other yocto projects - see https://github.com/t-kuha/recipes-ml/tree/2021.2/pytorch and https://stackoverflow.com/questions/56956555/yocto-warrior-bitbake-recipe-for-pytorch-for-nvidia-jetson-nano
- Some interest in slack about this - see https://oe4t.slack.com/archives/C01BTRXHGQ5/p1654680351774349
- Might make sense to integrate with existing layers for machine learning.