[PATCHES] Kexec syscall support, boots kernels from SD or USB (11/6/11, GB support)
Added statically-linked kexec to kexec_patches.tar.gz and example update.zips. Now works in stock recovery and CM7 CWM (with a kexec-patched kernel).
Added patches for the recently released GB sources to kexec_patches.tar.gz.
Attached is a set of patches (kexec_patches.tar.gz) against EC05, and the recently released GB sources, to implement kexec syscall support in the Epic's kernel. kexec enables the booting of kernels "directly" from the SD card or over USB without having to flash them to the device first. This allows us to easily use, test, and switch between many
kernels, not just the one (two with recovery) there's room for on flash.
When used in conjuction with modified init.rc scripts, this allows entire ROMs (with their own kernels) to run from SD card. In short, this allows us to run custom-kernel ROMs (e.g,. CyanogenMod) alongside each other or a stock kernel without having to flash back and forth.
Also attached is a modified version of the kexec userspace tool (also in kexec_patches.tar.gz, along with source and patches) that facilitates the proces of loading and kexecing a kernel image. Finally, attached is a demo EC05 kernel with kexec enabled (demo_kernel.tar.gz; mostly stock: RFS support only, testkeys recovery w/adbd, but does inlcude the keyboard patches), example update.zips that kexec an SD-card kernel from recovery--either as a normal boot (boot_zImage.zip) or recovery boot (boot_zImage_recovery.zip), and a script (patch_decomp_cachebufram.sh) to binary patch unmodified kernels to kexec boot faster.
Note, this thread is primarilly intended for kernel developers. Kexec probably won't be of great utility to end users until commonly-used kernels are patched. Also, although stock kernels can
be kexec'd, they need some init.rc modifications boot an entire ROM from SD card. Hopefully the fine folks here will come up with a user-friendly implementation of this work that's easy for everyone to use.
Flash a kexec-enabled kernel (e.g., the attached demo kernel) to either /dev/block/bml7 or /dev/block/bml8. For testing purposes, this kernel needs either "ro.secure=0" or "ro.debuggable=1" set in default.prop, and also needs recovery.rc/fota.rc modified to spawn the adbd service, so that an adb root shell is available while in recovery. Also copy the attached kexec tool to a convenient location on the device (e.g., /data/local/tmp/kexec).
Reboot into recovery. If the kexec kernel is installed to bml7, run "adb reboot recovery" while the phone is running. If installed to bml8, power down and boot into recovery by holding the volume-down, camera, and power buttons.
Make sure adb is running as root. If it's not, try running "adb root".
Find the kernel (zImage) you wish to boot. These can be extracted from a kernel update.zip or Odin .tar file, or use the demo kernel again.
Push the zImage into RAM (tmpfs) with:
adb push zImage /tmp
Now, open an adb root shell with "adb shell" and run the commands:
mount -ro remount /dev/block/stl6 /mnt/.lfs
mount -ro remount /dev/block/stl9 /system
mount -ro remount /dev/block/stl10 /data
mount -ro remount /dev/block/stl11 /cache
/data/local/tmp/kexec --load-hardboot --mem-min=0x50000000 --append="console=ttySAC2,115200 loglevel=4" /tmp/zImage
after which the phone will reboot, show the SAMSUNG logo, and eventually boot the kexec'd kernel. Do note that when booting unmodified kernels (see below), the SAMSUNG logo will persist for ~30 seconds longer than usual.
Also note that kexec performs an "abrupt" reboot, i.e., it doesn't shutdown the system normally. Hence it's important to kexec from recovery where few services are running. It's also prudent to remount file systems read-only and sync them to avoid any potential (although unlikely) of corruption.
In the future, kexec could be better integrated into the Android framework to allow for a clean shutdown. Otherwise, probably the best way to deploy kexec is through an update.zip file that boots a kernel from the SD card. See the attached example update.zips.
kexec is feature of Linux that allows it to directly execute (boot) a new kernel in place of itself, allowing Linux to effectively serve as its own bootloader.
The kexec procedure is two step. The first "kexec" command loads a zImage from disk, constructs parameters (e.g., the kernel command line), and stages it in memory, after which Linux continues to run as normal. The second "kexec" command tells Linux to execute (boot) the staged kernel.
In the standard implementation, Linux "soft boots" kexec'd kernels. That is, on "kexec -e" the running instance of Linux shuts-down all devices, drivers, and goes through the process of unloading itself as it does during a normal reboot. However, instead of invoking a hardware reboot, Linux, at the final stage of unloading itself, jumps to start executing the new kernel.
This soft boot process requires that Linux hardware drivers are fully capable
of unloading, reloading, and reinitializing the associated hardware without hardware-reboot or bootloader assistance. Since, for many built-in drivers, this capability is only used by kexec, hardware is often left in an unexpected or unknown state on unload, and thus the kexec'd kernel hangs on boot. Unfortunately this is the case with the Epic kernel, and soft booting doesn't work.
To work around this, the attached patches implement a "hard boot" method for kexecing kernels. Here, we use kexec to stage a kernel in memory as usual. On "kexec -e", Linux shuts-down as before, and at the very end of the unloading process it does two things: (i) scribble some information on how to boot the kexec'd kernel to a "special place" in memory, and (ii) performs a hardware
reboot, invoking the Epic bootloader as a normal reboot does.
On reboot, the bootloader loads the (previously-running) bml7 or bml8 kernel and starts executing it. Here, the hardboot patch modifies the Linux the zImage decompressor code to check the "special place" in memory to see if we're actually kexecing a different kernel. If so, it switches over to the other kernel, already staged elsewhere in memory.
Kernel Command Line:
Kexec (via hardboot) can
boot stock or non-kexec-modified custom kernels. However, unless the copy_atags patch is applied, they can only use the kernel command line provided by the bootloader, as opposed to the custom command line provided by kexec. Although this isn't a problem when kexecing from a bml7 boot kernel, kexecing from a bml8 kernel runs recovery (fota.rc) instead of a normal boot (init.rc).
With the copy_atags patch, the command line for the kexec'd kernel must be provided by with kexec's --append option. These are the command lines provided by the bootloader in normal boot and recovery scenarios, any of which may be used:
Normal boot (init.rc):
"adb reboot recovery" (recovery.rc):
bootmode=2 console=ttySAC2,115200 loglevel=4
Three-finger recovery boot (fota.rc):
bootmode=3 console=ttySAC2,115200 loglevel=4
Kexec booting of a stock or non-kexec-modified custom kernel is known take significantly longer than a regular boot, sitting at the SAMSUNG logo for 35 seconds instead of 8. The decomp_cachebufram patch resolves this issue for modified kernels. In addition, the attached patch_decomp_cachebufram.sh script will binary patch the decompressor code for any (to my knowledge) Epic kernel.
Many more details on the patches themselves are in the accompanying READMEs.
Kernel & kexec-tools patches: kexec_patches.tar.gz
Kexec EC05 demonstration kernel: demo_kernel.tar.gz
Recovery script to boot /sdcard/zImage (normal boot): boot_zImage.zip
Recovery script to boot /sdcard/zImage (recovery boot): boot_zImage_recovery.zip
Script to binary patch decompressor code: patch_decomp_cachebufram.sh