Difference between revisions of "Jetson/L4T BSP development tips"

From eLinux.org
Jump to: navigation, search
(Ramdisk customization)
(Kernel update)
 
(34 intermediate revisions by the same user not shown)
Line 146: Line 146:
 
The original ram-disk image can be copied from device (/boot/initrd) or host SDK directory (Linux_for_Tegra/bootloader/l4t_initrd.img).
 
The original ram-disk image can be copied from device (/boot/initrd) or host SDK directory (Linux_for_Tegra/bootloader/l4t_initrd.img).
  
     1. Extract the initrd by following command: <pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">zcat xxx/initrd | cpio -idmv</pre>
+
     1. Extract the <span style="color:red">initrd</span> by following command:  
     2. Change all files in this directory owner as root: <pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">sudo chown root.root * -R</pre>
+
<pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">zcat xxx/initrd | cpio -idmv</pre>
 +
     2. Change all files in this directory owner as root:  
 +
<pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">sudo chown root.root * -R</pre>
 
     3. Make some private changes. (For example, to add some special echo message in init script.)
 
     3. Make some private changes. (For example, to add some special echo message in init script.)
     4. Repack the initrd: <pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">find . | cpio -o -H newc | gzip > ../initrd.debug</pre>
+
     4. Repack the <span style="color:red">initrd</span>:  
     5. Replace the file /boot/initrd with generated file initrd.debug in above step.
+
<pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">find . | cpio -o -H newc | gzip > ../initrd.debug</pre>
     6. Reboot the device, and check the kernel log to confirm the new initrd works.
+
     5. Replace the file <span style="color:red">/boot/initrd</span> with generated file <span style="color:red">initrd.debug</span> in above step.
 +
     6. Reboot the device, and check the kernel log to confirm the new <span style="color:red">initrd</span> works.
 +
 
  
 
In addition, the initrd can also be built-in to kernel image.
 
In addition, the initrd can also be built-in to kernel image.
  
     Extract the initrd by following command: zcat xxx/initrd | cpio -idmv
+
     1. Extract the initrd by following command:  
     Change all files in this directory owner as root: sudo chown root.root * -R
+
<pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">zcat xxx/initrd | cpio -idmv</pre>
     Make some private changes. (For example, to add some special echo message in init script.)
+
     2. Change all files in this directory owner as root:  
     Repack the initrd: find . | cpio -o -H newc ../initrd.debug.cpio
+
<pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">sudo chown root.root * -R</pre>
     Edit the kernel config, and add the following line:
+
     3. Make some private changes. (For example, to add some special echo message in init script.)
 
+
     4. Repack the initrd:  
 +
<pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">find . | cpio -o -H newc > ../initrd.debug.cpio</pre>
 +
     5. Edit the kernel config, and add the following line:
 +
<pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">
 
     CONFIG_BLK_DEV_INITRD=y
 
     CONFIG_BLK_DEV_INITRD=y
 
 
     CONFIG_INITRAMFS_SOURCE="/home/temp/initr.debug.cpio"
 
     CONFIG_INITRAMFS_SOURCE="/home/temp/initr.debug.cpio"
     Re-build the kernel, replace the kernel image in device.
+
</pre>
     Comment out the default INITRD in /boot/extlinux/extlinux.conf
+
     6. Re-build the kernel, replace the kernel image in device.
 +
     7. Comment out the default INITRD in /boot/extlinux/extlinux.conf
 +
<pre style="margin-left:30px; color:#B0B0B0; background-color:#111111; white-space:pre-wrap;">
 
     ...
 
     ...
 
     LINUX /boot/Image
 
     LINUX /boot/Image
 
     #INITRD /boot/initrd
 
     #INITRD /boot/initrd
 
     ...
 
     ...
     Reboot the device, and the kernel should use the built-in initrd.
+
</pre>
 +
     8. Reboot the device, and the kernel should use the built-in initrd.
 +
 
 +
== OVERLAY FS support in Jetson L4T  ==
 +
L4T BSP is using EXT4 file-system as root-FS in eMMC or SDCARD. For EXT4 file-system, there's high risk of file-system corruption if the device does not shut down correctly, like sudden power loss.
 +
'overlayfs' may help to avoid such file-system corruption, especially in embedded products.
 +
=== Introduction of overlayfs ===
 +
(Abstracted from https://en.wikipedia.org/wiki/OverlayFS)
 +
<div style="border-bottom:1px solid #abd5f5; background:#d0e5f5; padding:0.2em 0.5em;">In computing, OverlayFS is a union mount filesystem implementation for Linux. It combines multiple different underlying mount points into one, resulting in single directory structure that contains underlying files and sub-directories from all sources. Common applications overlay a read/write partition over a read-only partition, such as with LiveCDs and IoT devices with limited flash memory write cycles.</div>
 +
 
 +
(Abstracted from kernel/kernel-4.9/Documentation/filesystems/overlayfs.txt)
 +
<div style="border-bottom:1px solid #abd5f5; background:#d0e5f5; padding:0.2em 0.5em;">An overlay filesystem combines two filesystems - an 'upper' filesystem
 +
and a 'lower' filesystem.  When a name exists in both filesystems, the
 +
 
 +
object in the 'upper' filesystem is visible while the object in the
 +
 
 +
'lower' filesystem is either hidden or, in the case of directories,
 +
 
 +
merged with the 'upper' object.
 +
 
 +
...
 +
 
 +
The lower filesystem can be any filesystem supported by Linux and does
 +
 
 +
not need to be writable.  The lower filesystem can even be another
 +
 
 +
overlayfs.  The upper filesystem will normally be writable and if it
 +
 
 +
is it must support the creation of trusted.* extended attributes, and
 +
 
 +
must provide valid d_type in readdir responses, so NFS is not suitable.
 +
 
 +
General usage:
 +
 
 +
  mount -t overlay overlay -olowerdir=/lower,upperdir=/upper,\
 +
 
 +
workdir=/work /merged</div>
 +
 
 +
(Abstracted from https://unix.stackexchange.com/questions/324515/linux-filesystem-overlay-what-is-workdir-used-for-overlayfs)
 +
<div style="border-bottom:1px solid #abd5f5; background:#d0e5f5; padding:0.2em 0.5em;">The workdir option is required, and used to prepare files before they are switched to the overlay destination in an atomic action (the workdir needs to be on the same filesystem as the upperdir).</div>
 +
=== Add overlayfs support in L4T (verified in Jetson AGX devkit, SDK 32.5) ===
 +
==== Kernel update ====
 +
By default, overlayfs is built by module in kernel configuration. And it should be changed as built-in in kernel.
 +
<syntaxhighlight lang="console">--- ./kernel/kernel-4.9/arch/arm64/configs/tegra_defconfig.orig    2021-07-22 14:55:08.229066055 +0800
 +
+++ ./kernel/kernel-4.9/arch/arm64/configs/tegra_defconfig    2021-07-21 15:24:10.965243034 +0800
 +
@@ -1132,7 +1132,7 @@
 +
CONFIG_AUTOFS4_FS=y
 +
CONFIG_FUSE_FS=m
 +
CONFIG_CUSE=m
 +
-CONFIG_OVERLAY_FS=m
 +
+CONFIG_OVERLAY_FS=y
 +
CONFIG_VFAT_FS=y
 +
CONFIG_NTFS_FS=y
 +
CONFIG_NTFS_RW=y
 +
</syntaxhighlight>
 +
Rebuild the kernel, and boot the device with new kernel.
 +
 
 +
==== initrd update ====
 +
(A good reference: http://wiki.psuter.ch/doku.php?id=solve_raspbian_sd_card_corruption_issues_with_read-only_mounted_root_partition)
 +
Edit the script 'init' in initrd. This is the patch for ram-FS overlayfs:
 +
<syntaxhighlight lang="console">--- init.orig    2021-07-21 14:35:50.109250634 +0800
 +
+++ init    2021-07-22 16:36:52.948849123 +0800
 +
@@ -142,7 +142,8 @@
 +
        echo "ERROR: ${rootdev} not found" > /dev/kmsg;
 +
        exec /bin/bash;
 +
    fi
 +
-    mount /dev/${rootdev} /mnt/;
 +
+    #do not mount rootdev now
 +
+    #mount /dev/${rootdev} /mnt/;
 +
    if [ $? -ne 0 ]; then
 +
        echo "ERROR: ${rootdev} mount fail..." > /dev/kmsg;
 +
        exec /bin/bash;
 +
@@ -214,13 +215,27 @@
 +
# Disable luks-srv TA
 +
nvluks-srv-app -n > /dev/null 2>&1;
 +
 +
+#create /mnt as mount point
 +
+mount -t tmpfs inittemp /mnt;
 +
+mkdir /mnt/lower;
 +
+mkdir /mnt/rw;
 +
+mount -t tmpfs root-rw /mnt/rw;
 +
+mkdir /mnt/rw/upper;
 +
+mkdir /mnt/rw/work;
 +
+mkdir /mnt/newroot;
 +
+mount  -o ro /dev/mmcblk0p1 /mnt/lower;
 +
+mount -t overlay -o lowerdir=/mnt/lower,upperdir=/mnt/rw/upper,workdir=/mnt/rw/work overlayfs-root /mnt/newroot;
 +
+mkdir /mnt/proc;
 +
+mkdir /mnt/sys;
 +
+mkdir /mnt/dev;
 +
+
 +
echo "Rootfs mounted over ${rootdev}" > /dev/kmsg;
 +
mount -o bind /proc /mnt/proc;
 +
mount -o bind /sys /mnt/sys;
 +
mount -o bind /dev/ /mnt/dev;
 +
-cd /mnt;
 +
-cp /etc/resolv.conf etc/resolv.conf
 +
 +
-echo "Switching from initrd to actual rootfs" > /dev/kmsg;
 +
-mount --move . /
 +
+cd /mnt/newroot;
 +
+cp /etc/resolv.conf etc/resolv.conf
 +
+echo "Switching from initrd to actual rootfs (ro-root-fs debug)" > /dev/kmsg;
 +
exec chroot . /sbin/init 2;
 +
</syntaxhighlight>
 +
 
 +
Update the device to boot with new initrd. After the kernel's up, the mount information will look like:
 +
<syntaxhighlight lang="console">
 +
overlayfs-root on / type overlay (rw,relatime,lowerdir=/mnt/lower,upperdir=/mnt/rw/upper,workdir=/mnt/rw/work)
 +
sysfs on /sys type sysfs (rw,nosuid,nodev,noexec,relatime)
 +
proc on /proc type proc (rw,nosuid,nodev,noexec,relatime)
 +
...
 +
</syntaxhighlight>
 +
 
 +
(Refer to https://elinux.org/Jetson/L4T_BSP_development_tips#Ramdisk_customization for how to customize initrd in L4T SDK.)
 +
Now, every change in root-FS will be in ram-FS, and after the device reboots, all changes will be lost.
 +
In addition, a lot of memory will also be consumed if big files added/changed.
 +
 
 +
Also, non-volatile media, like USB-Disk can also be used as overlayfs, which can reserve the changes in root-FS.
 +
Here's the init script patch, which use sda1 (USB-Disk) as overlayfs:
 +
<syntaxhighlight lang="console">
 +
--- init.orig    2021-07-21 14:35:50.109250634 +0800
 +
+++ init.sda1.ok    2021-07-22 18:02:48.424700032 +0800
 +
@@ -142,7 +142,8 @@
 +
        echo "ERROR: ${rootdev} not found" > /dev/kmsg;
 +
        exec /bin/bash;
 +
    fi
 +
-    mount /dev/${rootdev} /mnt/;
 +
+    #do not mount rootdev now
 +
+    #mount /dev/${rootdev} /mnt/;
 +
    if [ $? -ne 0 ]; then
 +
        echo "ERROR: ${rootdev} mount fail..." > /dev/kmsg;
 +
        exec /bin/bash;
 +
@@ -214,13 +215,46 @@
 +
# Disable luks-srv TA
 +
nvluks-srv-app -n > /dev/null 2>&1;
 +
 +
+# Wait till sda1 ready
 +
+count=0;
 +
+if [ ! -e "/dev/sda1" ]; then
 +
+    while [ ${count} -lt 50 ]
 +
+    do
 +
+        sleep 0.2;
 +
+        count=`expr $count + 1`;
 +
+        if [ -e "/dev/sda1" ]; then
 +
+            break;
 +
+        fi
 +
+    done
 +
+fi
 +
+if [ -e "/dev/sda1" ]; then
 +
+    echo "sda1 found" > /dev/kmsg;
 +
+else
 +
+    echo "ERROR: sda1 not found" > /dev/kmsg;
 +
+    exec /bin/bash;
 +
+fi
 +
+
 +
+#create /mnt as mount point
 +
+mount -t tmpfs inittemp /mnt;
 +
+mkdir /mnt/lower;
 +
+mkdir /mnt/rw;
 +
+mount /dev/sda1 /mnt/rw;
 +
+mkdir /mnt/rw/upper;
 +
+mkdir /mnt/rw/work;
 +
+mkdir /mnt/newroot;
 +
+mount  -o ro /dev/mmcblk0p1 /mnt/lower;
 +
+mount -t overlay -o lowerdir=/mnt/lower,upperdir=/mnt/rw/upper,workdir=/mnt/rw/work overlayfs-root /mnt/newroot;
 +
+mkdir /mnt/proc;
 +
+mkdir /mnt/sys;
 +
+mkdir /mnt/dev;
 +
+
 +
echo "Rootfs mounted over ${rootdev}" > /dev/kmsg;
 +
mount -o bind /proc /mnt/proc;
 +
mount -o bind /sys /mnt/sys;
 +
mount -o bind /dev/ /mnt/dev;
 +
-cd /mnt;
 +
+
 +
+cd /mnt/newroot;
 +
cp /etc/resolv.conf etc/resolv.conf
 +
 +
-echo "Switching from initrd to actual rootfs" > /dev/kmsg;
 +
-mount --move . /
 +
+echo "Switching from initrd to actual rootfs (ro-root-fs debug)" > /dev/kmsg;
 +
exec chroot . /sbin/init 2;
 +
</syntaxhighlight>
 +
 
 +
With this script, after the kernel's up, the mount information is similar. But the root-FS changes will lie in mounted USB-Disk. So it will not be lost after reboot.
 +
After sda1 is mounted:
 +
<syntaxhighlight lang="console">
 +
sudo mount /dev/sda1 /media/sda
 +
</syntaxhighlight>
 +
 
 +
The content in /media/sda/upper will look like:
 +
<syntaxhighlight lang="console">
 +
├── upper
 +
│  ├── boot
 +
│  │  └── grub
 +
│  │      └── grubenv
 +
│  ├── dev
 +
│  ├── etc
 +
│  │  ├── asound.conf -> /etc/asound.conf.tegrahdat194ref
 +
│  │  ├── fstab
 +
│  │  ├── machine-id
 +
│  │  ├── rc.local
 +
│  │  └── X11
 +
│  │      └── xorg.conf -> /etc/X11/xorg.conf.t194_ref
 +
│  ├── home
 +
│  │  └── nvidia
 +
│  │      ├── Desktop
 +
│  │      │  └── nv_l4t_readme.desktop
 +
...
 +
</syntaxhighlight>
 +
 
 +
All changes in root-FS will still be there.
 +
Note the USB-Disk read/write throughput may have impact on system performance.
 +
In addition, if the device suddenly shutdown/power-off, the file-system in USB-Disk may, with some probabilities, be corrupted.
 +
Anyway, the content in EMMC root-FS should be good, and that makes it possible to scan/fix the file-system error or even format the USB-disk.

Latest revision as of 02:03, 12 August 2021

Preparation

First, please download BSP package from internet. It's preferred to download BSP package, instead of Jetpack.

Go to https://developer.nvidia.com/embedded/linux-tegra.

The following packages are necessary:

L4T Driver Package (BSP)

Sample Root Filesystem


Source code for kernel and other components:

L4T Driver Package (BSP) Sources


Toolchain for kernel building:

GCC 7.3.1 for 64 bit BSP and Kernel


Secure package if secure-boot is necessary:

Jetson Platform Fuse Burning and Secure Boot Documentation and Tools


Refer to https://docs.nvidia.com/jetson/l4t/index.html#page/Tegra%2520Linux%2520Driver%2520Package%2520Development%2520Guide%2Fquick_start.html

Then the device can be flashed by command line. With this method, user can have more controls for the BSP configuration, like pinmux, kernel/kernel DTB customization, etc.

BSP customization

PINMUX

How to find out the actual pinmux configuration file

Refer to https://docs.nvidia.com/jetson/l4t/index.html#page/Tegra%2520Linux%2520Driver%2520Package%2520Development%2520Guide%2Fmb1_platform_config_xavier.html%23wwpID0E0240HA

There are several types of Jetson reference boards which are supported in SDK.

Check the configuration file

For example, when the flash command is run:

 sudo ./flash.sh jetson-xavier mmcblk0p1

Check the configuration file: jetson-xavier.conf → p2822-0000+p2888-0004.conf → PINMUX_CONFIG="tegra19x-mb1-pinmux-p2888-0000-a04-p2822-0000-b01.cfg";

(Note: the value of PINMUX_CONFIG may be overwritten. So the later one should take effect.)

Or check the flash log

Run following command:

sudo ./flash.sh -r --no-flash jetson-xavier mmcblk0p1

And check the log:

...

copying pinmux_config(/home/Work/jetson_customer/32.4.3/Linux_for_Tegra/bootloader/t186ref/BCT/tegra19x-mb1-pinmux-p2888-0000-a04-p2822-0000-b01.cfg)... done.

...

How to update pinmux

There are several ways to customize the device PINMUX. The simple way is to generate PINMUX configuration through pre-defined excel. Another way is to edit the PINMUX configuration file directly, but that may need more knowledge about PINMUX setting for the chip.

Edit the excel and generate the Configuration

Refer to https://docs.nvidia.com/jetson/l4t/index.html#page/Tegra%2520Linux%2520Driver%2520Package%2520Development%2520Guide%2Fadaptation_and_bringup_xavier_nx.html%23wwpID0E0WL0HA

Search 'Pinmux Changes'

Also, another good reference: https://elinux.org/Jetson/AGX_Xavier_Update_Pinmux

Still, please make sure the PINMUX configuration file name is correct.

Edit the pinmux configuration file

That's the direct and simple way, assumed the developer is familiar with PINMUX setting.

Download technical reference manual from https://developer.nvidia.com/embedded/downloads#?search=TRM for different platforms. And those documents contain detailed information.

Edit the prod configuration

L4T document shows another way to override the PINMUX:

https://docs.nvidia.com/jetson/l4t/index.html#page/Tegra%2520Linux%2520Driver%2520Package%2520Development%2520Guide%2Fmb1_platform_config_xavier.html%23wwpID0E0A40HA

Same command as 2.1.1.2, and check the prod configuration in following log:

...

copying prod_config(/home/Work/jetson_customer/32.4.3/Linux_for_Tegra/bootloader/t186ref/BCT/tegra19x-mb1-prod-p2888-0000-p2822-0000.cfg)... done.

...

Follow the guide in above link.

How to update device PINMUX

After the pinmux configuration files are changed, re-flash the device:

 sudo ./flash.sh jetson-xavier mmcblk0p1        #flash the whole device

or

 sudo ./flash.sh -k MB1_BCT jetson-xavier mmcblk0p1      #only flash the MB1_BCT
How to verify the new PINUX configuration works

User can read the PINMUX registers to confirm the new configuration works.

For example,

In PINMUX configuration file, there's an entry:

pinmux.0x0c302030 = 0x00000540; # gen2_i2c_scl_pcc7: i2c2, tristate-disable, input-enable, io_high_voltage-disable, lpdr-enable

And in prod configuration file, another entry:

prod.0x0c302030.0x0000100 = 0x00000000; #gen2_i2c_scl_pcc7: LPDR disable

Run a physical memory access tool, like devmem2 in Jetson device:

root@nvidia-desktop:/home/nvidia# devmem2 0x0c302030
/dev/mem opened.
Memory mapped at address 0x7f84359000.
Value at address 0xC302030 (0x7f84359030): 0x440


Linux kernel

Kernel building

There are several resources in internet introducing how to build Jetson Linux kernel from source. But it seems that the NV-provided nvbuild.sh never works in my side. Here's the script I'm using for kernel building:

#!/bin/bash
set -e
export KERNEL_SRC_DIR=${HOME}/Work/jetson_sdk/32.4.3/source/Linux_for_Tegra/source/public/kernel-source/kernel/kernel-4.9
export CROSS_COMPILE=${HOME}/Tools/kernel-toolchain/gcc-linaro-7.3.1-2018.05-x86_64_aarch64-linux-gnu/bin/aarch64-linux-gnu-
export TEGRA_KERNEL_OUT=${HOME}/Work/jetson_sdk/32.4.3/source/Linux_for_Tegra/source/public/kernel-build
export ARCH=arm64
make -C $KERNEL_SRC_DIR ARCH=arm64 LOCALVERSION="-tegra" O=$TEGRA_KERNEL_OUT tegra_defconfig
make -C $KERNEL_SRC_DIR ARCH=arm64 LOCALVERSION="-tegra" O=$TEGRA_KERNEL_OUT -j8
make -C $KERNEL_SRC_DIR ARCH=arm64 INSTALL_MOD_PATH=$TEGRA_KERNEL_OUT/modules_install INSTALL_MOD_STRIP=1 O=$TEGRA_KERNEL_OUT modules_install -j8

Note the parameter 'INSTALL_MOD_STRIP=1' should be added, otherwise the built module will be quite large. For example, nvgpu.ko size will increase from 2.6MB to 89MB.

Ramdisk customization

The original ram-disk image can be copied from device (/boot/initrd) or host SDK directory (Linux_for_Tegra/bootloader/l4t_initrd.img).

   1. Extract the initrd by following command: 
zcat xxx/initrd | cpio -idmv
   2. Change all files in this directory owner as root: 
sudo chown root.root * -R
   3. Make some private changes. (For example, to add some special echo message in init script.)
   4. Repack the initrd: 
find . | cpio -o -H newc | gzip > ../initrd.debug
   5. Replace the file /boot/initrd with generated file initrd.debug in above step.
   6. Reboot the device, and check the kernel log to confirm the new initrd works.


In addition, the initrd can also be built-in to kernel image.

   1. Extract the initrd by following command: 
zcat xxx/initrd | cpio -idmv
   2. Change all files in this directory owner as root: 
sudo chown root.root * -R
   3. Make some private changes. (For example, to add some special echo message in init script.)
   4. Repack the initrd: 
find . | cpio -o -H newc > ../initrd.debug.cpio
   5. Edit the kernel config, and add the following line:
    CONFIG_BLK_DEV_INITRD=y
    CONFIG_INITRAMFS_SOURCE="/home/temp/initr.debug.cpio"
   6. Re-build the kernel, replace the kernel image in device.
   7. Comment out the default INITRD in /boot/extlinux/extlinux.conf
    ...
    LINUX /boot/Image
    #INITRD /boot/initrd
    ...
   8. Reboot the device, and the kernel should use the built-in initrd.

OVERLAY FS support in Jetson L4T

L4T BSP is using EXT4 file-system as root-FS in eMMC or SDCARD. For EXT4 file-system, there's high risk of file-system corruption if the device does not shut down correctly, like sudden power loss. 'overlayfs' may help to avoid such file-system corruption, especially in embedded products.

Introduction of overlayfs

(Abstracted from https://en.wikipedia.org/wiki/OverlayFS)

In computing, OverlayFS is a union mount filesystem implementation for Linux. It combines multiple different underlying mount points into one, resulting in single directory structure that contains underlying files and sub-directories from all sources. Common applications overlay a read/write partition over a read-only partition, such as with LiveCDs and IoT devices with limited flash memory write cycles.

(Abstracted from kernel/kernel-4.9/Documentation/filesystems/overlayfs.txt)

An overlay filesystem combines two filesystems - an 'upper' filesystem

and a 'lower' filesystem. When a name exists in both filesystems, the

object in the 'upper' filesystem is visible while the object in the

'lower' filesystem is either hidden or, in the case of directories,

merged with the 'upper' object.

...

The lower filesystem can be any filesystem supported by Linux and does

not need to be writable. The lower filesystem can even be another

overlayfs. The upper filesystem will normally be writable and if it

is it must support the creation of trusted.* extended attributes, and

must provide valid d_type in readdir responses, so NFS is not suitable.

General usage:

 mount -t overlay overlay -olowerdir=/lower,upperdir=/upper,\
workdir=/work /merged

(Abstracted from https://unix.stackexchange.com/questions/324515/linux-filesystem-overlay-what-is-workdir-used-for-overlayfs)

The workdir option is required, and used to prepare files before they are switched to the overlay destination in an atomic action (the workdir needs to be on the same filesystem as the upperdir).

Add overlayfs support in L4T (verified in Jetson AGX devkit, SDK 32.5)

Kernel update

By default, overlayfs is built by module in kernel configuration. And it should be changed as built-in in kernel.

--- ./kernel/kernel-4.9/arch/arm64/configs/tegra_defconfig.orig    2021-07-22 14:55:08.229066055 +0800
+++ ./kernel/kernel-4.9/arch/arm64/configs/tegra_defconfig    2021-07-21 15:24:10.965243034 +0800
@@ -1132,7 +1132,7 @@
 CONFIG_AUTOFS4_FS=y
 CONFIG_FUSE_FS=m
 CONFIG_CUSE=m
-CONFIG_OVERLAY_FS=m
+CONFIG_OVERLAY_FS=y
 CONFIG_VFAT_FS=y
 CONFIG_NTFS_FS=y
 CONFIG_NTFS_RW=y

Rebuild the kernel, and boot the device with new kernel.

initrd update

(A good reference: http://wiki.psuter.ch/doku.php?id=solve_raspbian_sd_card_corruption_issues_with_read-only_mounted_root_partition) Edit the script 'init' in initrd. This is the patch for ram-FS overlayfs:

--- init.orig    2021-07-21 14:35:50.109250634 +0800
+++ init    2021-07-22 16:36:52.948849123 +0800
@@ -142,7 +142,8 @@
         echo "ERROR: ${rootdev} not found" > /dev/kmsg;
         exec /bin/bash;
     fi
-    mount /dev/${rootdev} /mnt/;
+    #do not mount rootdev now
+    #mount /dev/${rootdev} /mnt/;
     if [ $? -ne 0 ]; then
         echo "ERROR: ${rootdev} mount fail..." > /dev/kmsg;
         exec /bin/bash;
@@ -214,13 +215,27 @@
 # Disable luks-srv TA
 nvluks-srv-app -n > /dev/null 2>&1;
 
+#create /mnt as mount point
+mount -t tmpfs inittemp /mnt;
+mkdir /mnt/lower;
+mkdir /mnt/rw;
+mount -t tmpfs root-rw /mnt/rw;
+mkdir /mnt/rw/upper;
+mkdir /mnt/rw/work;
+mkdir /mnt/newroot;
+mount  -o ro /dev/mmcblk0p1 /mnt/lower;
+mount -t overlay -o lowerdir=/mnt/lower,upperdir=/mnt/rw/upper,workdir=/mnt/rw/work overlayfs-root /mnt/newroot;
+mkdir /mnt/proc;
+mkdir /mnt/sys;
+mkdir /mnt/dev;
+
 echo "Rootfs mounted over ${rootdev}" > /dev/kmsg;
 mount -o bind /proc /mnt/proc;
 mount -o bind /sys /mnt/sys;
 mount -o bind /dev/ /mnt/dev;
-cd /mnt;
-cp /etc/resolv.conf etc/resolv.conf
 
-echo "Switching from initrd to actual rootfs" > /dev/kmsg;
-mount --move . /
+cd /mnt/newroot;
+cp /etc/resolv.conf etc/resolv.conf
+echo "Switching from initrd to actual rootfs (ro-root-fs debug)" > /dev/kmsg;
 exec chroot . /sbin/init 2;

Update the device to boot with new initrd. After the kernel's up, the mount information will look like:

overlayfs-root on / type overlay (rw,relatime,lowerdir=/mnt/lower,upperdir=/mnt/rw/upper,workdir=/mnt/rw/work)
sysfs on /sys type sysfs (rw,nosuid,nodev,noexec,relatime)
proc on /proc type proc (rw,nosuid,nodev,noexec,relatime)
...

(Refer to https://elinux.org/Jetson/L4T_BSP_development_tips#Ramdisk_customization for how to customize initrd in L4T SDK.) Now, every change in root-FS will be in ram-FS, and after the device reboots, all changes will be lost. In addition, a lot of memory will also be consumed if big files added/changed.

Also, non-volatile media, like USB-Disk can also be used as overlayfs, which can reserve the changes in root-FS. Here's the init script patch, which use sda1 (USB-Disk) as overlayfs:

--- init.orig    2021-07-21 14:35:50.109250634 +0800
+++ init.sda1.ok    2021-07-22 18:02:48.424700032 +0800
@@ -142,7 +142,8 @@
         echo "ERROR: ${rootdev} not found" > /dev/kmsg;
         exec /bin/bash;
     fi
-    mount /dev/${rootdev} /mnt/;
+    #do not mount rootdev now
+    #mount /dev/${rootdev} /mnt/;
     if [ $? -ne 0 ]; then
         echo "ERROR: ${rootdev} mount fail..." > /dev/kmsg;
         exec /bin/bash;
@@ -214,13 +215,46 @@
 # Disable luks-srv TA
 nvluks-srv-app -n > /dev/null 2>&1;
 
+# Wait till sda1 ready
+count=0;
+if [ ! -e "/dev/sda1" ]; then
+    while [ ${count} -lt 50 ]
+    do
+        sleep 0.2;
+        count=`expr $count + 1`;
+        if [ -e "/dev/sda1" ]; then
+            break;
+        fi
+    done
+fi
+if [ -e "/dev/sda1" ]; then
+    echo "sda1 found" > /dev/kmsg;
+else
+    echo "ERROR: sda1 not found" > /dev/kmsg;
+    exec /bin/bash;
+fi
+
+#create /mnt as mount point
+mount -t tmpfs inittemp /mnt;
+mkdir /mnt/lower;
+mkdir /mnt/rw;
+mount /dev/sda1 /mnt/rw;
+mkdir /mnt/rw/upper;
+mkdir /mnt/rw/work;
+mkdir /mnt/newroot;
+mount  -o ro /dev/mmcblk0p1 /mnt/lower;
+mount -t overlay -o lowerdir=/mnt/lower,upperdir=/mnt/rw/upper,workdir=/mnt/rw/work overlayfs-root /mnt/newroot;
+mkdir /mnt/proc;
+mkdir /mnt/sys;
+mkdir /mnt/dev;
+
 echo "Rootfs mounted over ${rootdev}" > /dev/kmsg;
 mount -o bind /proc /mnt/proc;
 mount -o bind /sys /mnt/sys;
 mount -o bind /dev/ /mnt/dev;
-cd /mnt;
+
+cd /mnt/newroot;
 cp /etc/resolv.conf etc/resolv.conf
 
-echo "Switching from initrd to actual rootfs" > /dev/kmsg;
-mount --move . /
+echo "Switching from initrd to actual rootfs (ro-root-fs debug)" > /dev/kmsg;
 exec chroot . /sbin/init 2;

With this script, after the kernel's up, the mount information is similar. But the root-FS changes will lie in mounted USB-Disk. So it will not be lost after reboot. After sda1 is mounted:

sudo mount /dev/sda1 /media/sda

The content in /media/sda/upper will look like:

├── upper
│   ├── boot
│   │   └── grub
│   │       └── grubenv
│   ├── dev
│   ├── etc
│   │   ├── asound.conf -> /etc/asound.conf.tegrahdat194ref
│   │   ├── fstab
│   │   ├── machine-id
│   │   ├── rc.local
│   │   └── X11
│   │       └── xorg.conf -> /etc/X11/xorg.conf.t194_ref
│   ├── home
│   │   └── nvidia
│   │       ├── Desktop
│   │       │   └── nv_l4t_readme.desktop
...

All changes in root-FS will still be there. Note the USB-Disk read/write throughput may have impact on system performance. In addition, if the device suddenly shutdown/power-off, the file-system in USB-Disk may, with some probabilities, be corrupted. Anyway, the content in EMMC root-FS should be good, and that makes it possible to scan/fix the file-system error or even format the USB-disk.