Why I did it
After docker_inram is enabled, the docker folder's default max size is 1.5G.
It's not big enough for some tests which need to install additional docker images or install extra packages.
Work item tracking
Microsoft ADO 24199761:
How I did it
add docker_inram into cmdline_allowlist
How to verify it
sudo sh -c 'echo "docker_inram_size=3000M" >> kernel-cmdline-append'
sudo reboot and check the docker folder size
Some devices running SONiC have a small storage device (2G and 4G mainly)
The SONiC image growth over time has made it impossible to install
2 images on a single device.
Some mitigations have been implemented in the past for some devices but
there is a need to do more.
One such mitigation is `docker_inram` which creates a `tmpfs` and
extracts `dockerfs.tar.gz` in it.
This all happens in the SONiC initramfs and by ensuring the installation
process does not extract `dockerfs.tar.gz` on the flash but keep the file as is.
This mitigation does a tradeoff by using more RAM to reduce the disk footprint.
It however creates new issues for devices with 4G of system memory since
the extracted `dockerfs.tar.gz` nears the 1.6G.
Considering debian upgrades (with dual base images) and the continuous
stream of features this is only going to get bigger.
This change introduces an alternative to the `tmpfs` by allowing a system
to extract the `dockerfs.tar.gz` inside a `zram` device thus bringing
compression in play at the detriment of performance.
Introduce 2 new optional kernel parameters to be consumed by SONiC initramfs.
- `docker_inram_size` which represent the max physical size of the
`zram` or `tmpfs` volume (defaults to DOCKER_RAMFS_SIZE)
- `docker_inram_algo` which is the method to use to extract the
`dockerfs.tar.gz` (defaults to `tmpfs`)
other values are considered to be compression algorithm for `zram`
(e.g `zstd`, `zlo-rle`, `lz4`)
Refactored the logic to mount the docker fs in the SONiC initramfs under
the `union-mount` script.
Moved the code into a function to make it cleaner and separated the
inram volume creation and docker extraction.
On Arista platform with a flash smaller or equal to 4GB set
`docker_inram_algo` to `zstd` which produces the best compression ratio
at the detriment of a slower write performance and a similar read
performance to other `zram` compression algorithms.
Enable docker_inram for all systems with 4GB or less of flash.
This is mandatory to allow these systems to store 2 SONiC images.
This change also fixes the missing docker_inram attribute when
installing a new image from SONiC.
Because the SWI image can ship with additional kernel parameters within
such as `sonic_fips=` this lead to a conflict.
To prevent the conflict, the extra kernel parameters from the SWI are
now stored in the file `kernel-cmdline-append` which isn't used anywhere.
Why I did it
Fix similar issue seen on #13739 but only for DCS-7050CX3-32S
How I did it
Add a kernel parameter to tell libata to disable NCQ
How to verify it
The message ata2.00: FORCE: horkage modified (noncq) should appear on the dmesg.
Test results using: fio --direct=1 --rw=randrw --bs=64k --ioengine=libaio --iodepth=64 --runtime=120 --numjobs=4
with NCQ
READ: bw=26.1MiB/s (27.4MB/s), 26.1MiB/s-26.1MiB/s (27.4MB/s-27.4MB/s), io=3136MiB (3288MB), run=120053-120053msec
WRITE: bw=26.3MiB/s (27.6MB/s), 26.3MiB/s-26.3MiB/s (27.6MB/s-27.6MB/s), io=3161MiB (3315MB), run=120053-120053msec
without NCQ
READ: bw=22.0MiB/s (23.1MB/s), 22.0MiB/s-22.0MiB/s (23.1MB/s-23.1MB/s), io=2647MiB (2775MB), run=120069-120069msec
WRITE: bw=22.2MiB/s (23.3MB/s), 22.2MiB/s-22.2MiB/s (23.3MB/s-23.3MB/s), io=2665MiB (2795MB), run=120069-120069msec
Why I did it
Some products might experience an occasional IO failure in the communication between CPU and SSD.
Based on some research it could be attributable to some device not handling ATA NCQ (Native Command Queue).
This issue currently affect 4 products:
DCS-7170-32C*
DCS-7170-64C
DCS-7060DX4-32
DCS-7260CX3-64
How I did it
This change disable NCQ on the affected drive for a small set of products.
How to verify it
When the fix is applied, these 2 patterns can be found in the dmesg.
ata1.00: FORCE: horkage modified (noncq)
NCQ (not used)
Test results using: fio --direct=1 --rw=randrw --bs=64k --ioengine=libaio --iodepth=64 --runtime=120 --numjobs=4
with NCQ (ata1.00: 61865984 sectors, multi 1: LBA48 NCQ (depth 32), AA)
READ: bw=33.9MiB/s (35.6MB/s), 33.9MiB/s-33.9MiB/s (35.6MB/s-35.6MB/s), io=4073MiB (4270MB), run=120078-120078msec
WRITE: bw=34.1MiB/s (35.8MB/s), 34.1MiB/s-34.1MiB/s (35.8MB/s-35.8MB/s), io=4100MiB (4300MB), run=120078-120078msec
without NCQ (ata1.00: 61865984 sectors, multi 1: LBA48 NCQ (not used))
READ: bw=31.7MiB/s (33.3MB/s), 31.7MiB/s-31.7MiB/s (33.3MB/s-33.3MB/s), io=3808MiB (3993MB), run=120083-120083msec
WRITE: bw=31.9MiB/s (33.4MB/s), 31.9MiB/s-31.9MiB/s (33.4MB/s-33.4MB/s), io=3830MiB (4016MB), run=120083-120083msec
Which release branch to backport (provide reason below if selected)
Why I did it
To address error sometimes seen when running sonic-mgmt test_stress_routes.py::test_announce_withdraw_route on 720DT-48S
How I did it
Update boot0 logic to set platform specific varlog size for 720DT-48S
How to verify it
Verified that /var/log size increased and error is no longer observed when running test
Why I did it
Fix some unreliability seen on emmc device with some AMD CPUs
How I did it
Added a kernel parameter to add quirks to
It depends on a sonic-linux-kernel change to work properly but will be a no-op without it.
The quirk added is SDHCI_QUIRK2_BROKEN_HS200 used to downgrade the link speed for the eMMC.
Many of these switches have had flash upgraded beyond 2G however, in
boot0 both were assigned 2GB for legacy reasons.
Remove the hardcoding of the flash size and let boot0 autodetect the available space.
Signed-off-by: Graham Hayes <gr@ham.ie>
Signed-off-by: Graham Hayes <gr@ham.ie>
Added initial set of config files to allow for booting and partial traffic testing in SONiC on the 720DT-48S.
How to verify it
- Switch boots
- show interfaces status shows links up on interfaces Ethernet24-51
- Traffic flows with no errors on interfaces Ethernet24-51
Why I did it
Support to use symbol links in platform folder to reduce the image size.
The current solution is to copy each lazy installation targets (xxx.deb files) to each of the folders in the platform folder. The size will keep growing when more and more packages added in the platform folder. For cisco-8000 as an example, the size will be up to 2G, while most of them are duplicate packages in the platform folder.
How I did it
Create a new folder in platform/common, all the deb packages are copied to the folder, any other folders where use the packages are the symbol links to the common folder.
Why platform.tar?
We have implemented a patch for it, see #10775, but the problem is the the onie use really old unzip version, cannot support the symbol links.
The current solution is similar to the PR 10775, but make the platform folder into a tar package, which can be supported by onie. During the installation, the package.tar will be extracted to the original folder and removed.
Why I did it
This change adds the support for Arista 7060dx5_64s and 7060px5_64s
How I did it
How to verify it
We verified the platform driver is working and the ports are up on 7060dx5_64s and 7060px5_64s.
Add most configuration files for the DCS-7050PX4-32S and DCS-7050DX4-32S.
This review only contains platform configuration files, dataplane ones will follow in future change.
Co-authored-by: Zhi Yuan (Carl) Zhao <zyzhao@arista.com>
If it is run during image install, it's not guaranteed that the
installation environment will have tune2fs available. Therefore, run it
during initramfs instead.
Signed-off-by: Saikrishna Arcot <sarcot@microsoft.com>
# Why I did it
Reduce the disk space taken up during bootup and runtime.
# How I did it
1. Remove python package cache from the base image and from the containers.
2. During bootup, if logs are to be stored in memory, then don't create the `var-log.ext4` file just to delete it later during bootup.
3. For the partition containing `/host`, don't reserve any blocks for just the root user. This just makes sure all disk space is available for all users, if needed during upgrades (for example).
* Remove pip2 and pip3 caches from some containers
Only containers which appeared to have a significant pip cache size are
included here.
Signed-off-by: Saikrishna Arcot <sarcot@microsoft.com>
* Don't create var-log.ext4 if we're storing logs in memory
Signed-off-by: Saikrishna Arcot <sarcot@microsoft.com>
* Run tune2fs on the device containing /host to not reserve any blocks for just the root user
Signed-off-by: Saikrishna Arcot <sarcot@microsoft.com>
- Use SfpOptoeBase by default to leverage new `sonic_xcvr` refactor
- Add support for `Woodleaf` product
- Move `libsfp-eeprom.so` to a different `.deb` package
- Add new logrotate configuration for arista logs
- Improve logging mechanism for the drivers (IO loglevel, fix syslog duplicates)
- Initialize chassis cards in parallel
- Refactor of `get_change_event` to fix interrupts treated as presence change
The Lodoga platform also matched crow which was hardcoding the flash
size to 3700. This change enables autodetect on Clearlake which in turns
allows autodetect for Lodoga.
The threshold was bumped from 3700 to 4000 because size computation can
differ slightly and report slightly above 3700.
Lodoga actually has a 8GB storage device.
LodogaSsd variant has a 30GB SSD drive.
However, in boot0 both were mishandled and assigned 4GB for legacy reasons.
Remove the hardcoding of the flash size and let boot0 autodetect the available space.
Master/202012 image size grew quite a bit. 3.7G harddrive can no longer hold one image and safely upgrade to another image. Every bit of harddrive space is precious to save now.
Also sh syntax seemingly changed, [ condition ] && action was a legit syntax in 201911 branch but it is an error when condition not met with 202012 or later images. Change the syntax to if statement to avoid the issue.
Signed-off-by: Ying Xie ying.xie@microsoft.com
Why I did it
Recent systemd upgrade from #7228 requires an extra cmdline parameter for dockerd to start properly.
Updating boot0 was missed as part of the systemd upgrade change.
How I did it
Just added the missing cmdline parameter in files/Aboot/boot0.j2
This change fixes#7372
How to verify it
Boot the image and dockerd should start normally.
- Add support for `DCS-7050SX3-48YC8` and `DCS-7050SX3-48C8` platform
- Add support for more variants of `DCS-7280CR3-32[PD]4`
- Add Supervisor to Linecard consutil support
- Complete Watchdog platform API support
- Fix some PSU behavior on `DCS-7050QX-32` and `DCS-7060CX-32S`
- Fix SEU management on `DCS-7060CX-32S`
- Allow kernel modules to build up to linux 5.10
- Rename led color `orange` to `amber`
- Miscellaneous fixes
A few issues where discovered with crashkernel on Arista platforms.
1) platforms using `docker_inram=on` would end up OOM in kdump environment.
This happens because the same initramfs is used by SONiC and the crashkernel.
With `docker_inram=on` the `dockerfs.tar.gz` is extracted in a `tmpfs` created for the occasion.
Since `dockerfs.tar.gz` weights more than 1.5G, it doesn't fit into the kdump environment and ends up OOM.
This OOM event can in turn trigger a panic.
2) Arista platforms with `secureboot` enabled would fail to load the crashkernel because the kernel parameter would be discarded on boot.
This happens because the `boot0` in secureboot mode is strict about kernel parameter injection.
3) The secureboot path allowlist would remove kernel crash reports.
4) The kdump service would fail on Arista products since `/boot/` is empty in `secureboot`
**- How I did it**
1) To prevent an OOM event in the crashkernel the fix is to avoid the codepaths in `union-mount` that create tmpfs and populate them. Some more codepath specific to Arista devices are also skipped to make the kdump process faster.
This relies on detecting that the initramfs is starting in a kdump environment and skipping some initialization.
The `/usr/sbin/kdump-config` tool appends a few kernel cmdline arguments when loading the crashkernel.
The most unique one is `systemd.unit=kdump-tools.service` which is used in a few initramfs hooks to set `in_kdump`.
2) To allow `kdump` to work in `secureboot` environment the cmdline generation in boot0 was slightly modified.
The codepath to load kernel parameters changed by SONiC is now running for booting in secure mode.
It was altered to prevent an append only behavior which would grow the `kernel-cmdline` at every reboot.
This ever growing behavior would lead `kexec` to fail to load the kernel due to a too long cmdline.
3) To get the kernel crash under /var/crash this path has to be added to `allowlist_paths`
4) The `/host/image-XXX/boot` folder is now populated in `secureboot` mode but not used.
**- How to verify it**
Regular boot:
- enable kdump
- enable docker_inram=on via kernel-params
- reboot
- generate a crash `echo c > /proc/sysrq-trigger`
- before: witness OOM events on the console
- after: crash kernel works and crash available under /var/crash
Secure boot:
- enable kdump
- reboot
- generate a crash `echo c > /proc/sysrq-trigger`
- before: witness no kdump
- after: crash kernel works and crash available under /var/crash
Co-authored-by: Boyang Yu <byu@arista.com>
To limit IO and space usage on the flash device the boot0 script makes sure the SWI is in memory.
Because SONiC maps /tmp on the flash, some logic is required to make sure of it.
However it is possible for some provisioning mechanism to already download the swi in a memory file system.
This case was not handled properly by the boot0 script.
It now detect if the image is on a tmpfs or a ramfs and keep it there if that is the case.
The cleanup method has been updated accordingly and will only cleanup
the mount path if it's below /tmp/ as to not affect user mounted paths.
- How I did it
Check the filesystem on which the SWI pointed by swipath lies.
If this filesystem is a ramfs or a tmpfs the move_swi_to_tmpfs becomes a no-op.
Made sure the cleanup logic would not behave unexpectedly.
- How to verify it
In SONiC:
Download the swi under /tmp and makes sure it gets moved to /tmp/tmp-swi which gets mounted for that purpose.
Make sure /tmp/tmp-swi gets unmounted once the install process is done.
Create a new mountpoint under /ram using either ramfs or tmpfs and download the swi there.
Install the swi using sonic-installer and makes sure the image doesn't get moved by looking at the logs.
- Enhance eeprom parsing robustness on corrupted fields
- Add chassis provisioning service
- Disable CPU sleep state on some systems
- Complete refactor for FanSlots
- Fix module unload while still in use
Current support for the 7060PX4-32 and 7060DX4 was broken.
With this change, ports are now linking fine.
Co-authored-by: Zhi Yuan Carl Zhao <zyzhao@arista.com>
- Merge chassis codebase upstream
- Add support for Otterlake supervisor
- Add support for NorthFace and Camp chassis
- Add support for Eldridge, Dragonfly and Brooks fabrics
- Add support for Clearwater2 and Clearwater2Ms linecards
- Add new arista Cli to power on/off cards
- Add new arista show Cli to inspect supervisor, chassis, fabrics and linecards
* Add secureboot support in boot0
* Initramfs changes for secureboot on Aboot devices
* Do not compress squashfs and gz in fs.zip
It doesn't make much sense to do so since these files are already
compressed.
Also not compressing the squashfs has the advantage of making it
mountable via a loop device.
* Add loopoffset parameter to initramfs-tools
* Support rw files allowlist for Sonic Secure Boot
* Improve the performance
* fix bug
* Move the config description into a md file
* Change to use a simple way to remove the blank line
* Support chmod a-x in rw folder
* Change function name
* Change some unnecessary words
Fix the issue of incorrectly skipping the convertfs hook when fast-reboot from EOS, by adding an extra kernel cmdline param "prev_os" to differentiate fast-reboot from EOS and from SONiC.
This is because we still do disk conversion for fast reboot from eos to sonic, like format the disk.
* Update arista driver submodule
* Add support for 7260CX3-64E in boot0
* Refactor boot0 platform specific definition
Make it easier to manage new sku
* Add support for 7050CX3-32S in boot0
Just contains the required boot0 information
* Add basic plugin support for DCS-7050CX3-32S
* Add port config for Arista-7050CX3-32S-C32
Co-authored-by: yurypm <yurypm@arista.com>
Co-authored-by: byu343 <byu@arista.com>
when device disk is small, do not unzip dockerfs.tar.gz on disk.
keep the tar file on the disk, unzip to tmpfs in the initrd phase.
enabled this for 7050-qx32
Signed-off-by: Guohan Lu <gulv@microsoft.com>
* Add boot0 support for the 7280CR3
* Add platform and plugins for 7280CR3
* Add port config for 7280CR3
* Add platform_reboot for 7280CR3
* Add support for 7280CR3-32D4 based on the 7280CR3-32P4
* Update arista driver submodules
- Introduce new 7280CR3-32P4
- Improve to the led plugin for OSFP
* Fix showing systemd shutdown sequence when verbose is set
* Fix creation of kernel-cmdline file
Sometimes boot0 prints error
"mv: can't preserve ownership of '/mnt/flash/image-arsonic.xxxx/kernel-cmdline': Operation not permitted"
* Improve flash space usage during installation
Some older systems only have 2GB of flash available. Installing a second
image on these can prove to be challenging.
The new installation process moves the installer swi to memory in order
to avoid free up space from the flash before uncompressing it there.
It removes all the flash space usage spike and also improves the IO
since the installation is no more reading and writting to the flash at
the same time.
* Add support of 7060CX-32S-SSD
* 7260CX3: use inventory powerCycle procedures
* 7050QX-32S: use inventory powerCycle procedures
* 7050QX-32: use inventory powerCycle procedures
* platform: arista: add common platform_reboot
Replace platform_reboot by a link to new common for devices already
using a similar script.
* 7060CX-32S: use inventory powerCycle procedures
* Install python smbus in pmon
Some platform plugin need the python smbus library to perform some actions.
This installs the dependency.
