Contents
|
In the course of development, Google developers made some changes to the Linux kernel. The amount of changes is not extremely large, and is on the order of changes that are customarily made to the Linux kernel by embedded developers (approximately 250 patches, with about 3 meg. of differences in 25,000 lines). The changes include a variety of large and small additions, ranging from the wholesale addition of a flash filesystem (YAFFS2), to very small patches to augment Linux security (paranoid networking patches).
Various efforts have been made over the past few years to submit these to changes to mainline (mostly by Google engineers, but also by others), with not much success so far.
A very good overview of the changes is available in a talk by John Stultz at ELC 2011. (The talk has a somewhat misleading name.)
Some changes were temporarily adding the "staging" driver area in the stock kernel, but were removed due to lack of support. See Greg KH blog post on -staging for 2.6.33, where he announces to remove various Android drivers from -staging.
Several groups and individuals are working to get kernel changes from Android mainlined into the Linux kernel.
Please see Android Mainlining Project for more information.
Here is a list of changes/addons that the Android Project made to the linux kernel. As of September, 2011, these kernel changes are not part of the standard kernel and are only available in the Android kernel trees in the Android Open Source project.
This list does not include board- or platform-specific support or drivers (commonly called "board support").
Binder is an Android-specific interprocess communication mechanism, and remote method invocation system.
See Android Binder
According to the Kconfig help "The ashmem subsystem is a new shared memory allocator, similar to POSIX SHM but with different behavior and sporting a simpler file-based API."
Apparently it better-supports low memory devices, because it can discard shared memory units under memory pressure.
To use this, programs open /dev/ashmem, use mmap() on it, and can perform one or more of the following ioctls:
From a thread on android-platform source
You can create a shared memory segment using:
fd = ashmem_create_region("my_shm_region", size); if(fd < 0) return -1; data = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); if(data == MAP_FAILED) goto out;
In the second process, instead of opening the region using the same name, for security reasons the file descriptor is passed to the other process via binder IPC.
The libcutils interface for ashmem consists of the following calls: (found in system/core/include/cutils/ashmem.h)
The pmem driver is used to manage large (1-16+MB) physically contiguous regions of memory shared between userspace and kernel drivers (dsp, gpu, etc). It was written specifically to deal with hardware limitations of the MSM7201A, but could be used for other chipsets as well. For now, you're safe to turn it off on x86.
David Sparks wrote the following: source
2. ashmem and pmem are very similar. Both are used for sharing memory between processes. ashmem uses virtual memory, whereas pmem uses physically contiguous memory. One big difference is that with ashmem, you have a ref-counted object that can be shared equally between processes. For example, if two processes are sharing an ashmem memory buffer, the buffer reference goes away when both process have removed all their references by closing all their file descriptors. pmem doesn't work that way because it needs to maintain a physical to virtual mapping. This requires the process that allocates a pmem heap to hold the file descriptor until all the other references are closed. 3. You have the right idea for using shared memory. The choice between ashmem and pmem depends on whether you need physically contiguous buffers. In the case of the G1, we use the hardware 2D engine to do scaling, rotation, and color conversion, so we use pmem heaps. The emulator doesn't have a pmem driver and doesn't really need one, so we use ashmem in the emulator. If you use ashmem on the G1, you lose the hardware 2D engine capability, so SurfaceFlinger falls back to its software renderer which does not do color conversion, which is why you see the monochrome image.
Informally known as the Viking Killer, the OOM handler simply kills processes as available memory becomes low. The kernel module follows rules for this that are supplied from user space in two ways:
1. init writes information about memory levels and associated classes:
write /sys/module/lowmemorykiller/parameters/adj 0,1,2,4,7,15 write /sys/module/lowmemorykiller/parameters/minfree 2048,3072,4096,6144,7168,8192
2. User space sets the oom_adj of processes to put them in the correct class for their current operation. This redefines the meaning of oom_adj from that used by the standard OOM killer to something that is more aggressive and controlled.
These oom_adj levels end up being based on the process lifecycle defined here: http://developer.android.com/guide/topics/fundamentals.html#proclife
This is the kernel implementation to support Android's AlarmManager. It lets user space tell the kernel when it would like to wake up, allowing the kernel to schedule that appropriately and come back (holding a wake lock) when the time has expired regardless of the sleep state of the CPU.
Note that POSIX Alarm timers, which implement this functionality (but not identically), was accepted into mainline Linux in kernel version 3.0.
See Waking Systems from Suspend and http://lwn.net/Articles/439364/
Generic gpio is a mechanism to allow programs to access and manipulate gpio registers from user space.
Timed output/gpio is a system to allow chaning a gpio pin and restore it automatically after a specified timeout. Seedrives/misc/timed_output.canddrives/misc/timed_gpio.cThis expose a user space interface used by the vibrator code.
On ADP1, there is a driver at:
# cd /sys/bus/platform/drivers/timed-gpio # ls -l --w------- 1 0 0 4096 Nov 13 02:11 bind lrwxrwxrwx 1 0 0 0 Nov 13 02:11 timed-gpio -> ../../../../devices/platform/timed-gpio --w------- 1 0 0 4096 Nov 13 02:11 uevent --w------- 1 0 0 4096 Nov 13 02:11 unbind
Also, there is a device at:
# cd /sys/devices/platform/timed-gpio # ls -lR .: lrwxrwxrwx 1 0 0 0 Nov 13 01:34 driver -> ../../../bus/platform/drivers/timed-gpio -r--r--r-- 1 0 0 4096 Nov 13 01:34 modalias drwxr-xr-x 2 0 0 0 Nov 13 01:34 power lrwxrwxrwx 1 0 0 0 Nov 13 01:34 subsystem -> ../../../bus/platform -rw-r--r-- 1 0 0 4096 Nov 13 01:34 uevent ./power: -rw-r--r-- 1 0 0 4096 Nov 13 01:34 wakeup
This allows saving the kernel printk messages to a buffer in RAM, so that after a kernel panic they can be viewed in the next kernel invocation, by accessing /proc/last_kmsg.
[Would be good to get more details on how to set this up and use it here!] [I guess this is something like pramfs?]
Here is a miscellaneous list of other kernel changes in the mistral Android kernel:
The file Documentation/android.txt has a list of required configuration options for a kernel to support an Android system.