Very useful speed optimization: make the system start faster

In embedded products, system startup speed is a very critical indicator. The optimization of system startup speed is usually called "Quick Start".

To optimize the system startup speed, you must first knowhow to count the system startup time.

The following introduces several methods of counting the time taken by kernel startup, as well as several methods of optimizing kernel startup speed.

1. Startup time-consuming statistics

printk time

Open kernel configuration:

kernel hacking --->
[*] Show timing information on printks

After opening, a timestamp will be displayed in front of each printk

Mainly used to measure various stages of the kernel startup process The time-consuming

initcall_debug

As we all know, kernel will execute different levels of when it starts initcall, and the time consumption of each initcall can also be counted.

在kernel的cmdline中加入参数initcall_debug=1：

initcall_debug=1
setargs_nand=setenv bootargs console=${console} earlyprintk=${earlyprintk} root=${nand_root} initcall_debug=${initcall_debug} init=${init}

开启后，就能打印每个initcall函数调用及耗时。

bootgraph

内核自带了一个工具用于统计启动时间：scripts/bootgraph.pl

使用该工具需要打开内核配置CONFIG_PRINTK_TIME=y，并且在cmdline中加上"initcall_debug=1"

系统启动之后，执行命令：

dmesg|perl $(kernel_dir)/script/bootgraph.pl > out.svg

用浏览器查看out.svg文件，可以看到内核启动过程中各个阶段的耗时。

这个工具有点类似于perf的火焰图，可以统计启动各阶段的耗时。

bootchart

除了内核自带的工具，也有开源的工具可用：bootchart。

bootchart is an open source software tool used for performance analysis of the Linux startup process. It automatically collects CPU usage, process and other information during the system startup process, and displays the analysis results graphically, which can be used to guide and optimize the system startup process.

• Modify kernel cmdline. Change init to "init=/sbin/bootchartd".
• collect information. bootchartd will collect information from /proc/stat, /proc/diskstat, /proc/[pid]/stat. After processing, save it as a bootchart.tgz file
• Convert the image. Use the pybootchartgui.py tool on pc to convert bootchart.tgz to bootchart.png to facilitate analysis

Finally, pictures will be produced for analysis, for example:

bootchar is mainly used to measure the time taken from mounting the file system to starting the main application

gpio Oscilloscope

You can find a GPIO that is idle during system startup and set GPIO## in the appropriate location #level.

measure the time-consuming of the entire startup, or the time-consuming of each stage. This method is also commonly used.

2. Kernel optimization method

kernel compression method

kernel has different compression formats, common ones are gz, xz, lzma, etc.

Different compression formats have different decompression speeds. By comparing the startup time and flash occupancy of different compression methods, choose one that meets the actual situation and optimize it.

Loading location

The kernel image can be self-decompressed by kernel or decompressed by uboot.

For kernel self-decompression, if the compressed kernel conflicts with the decompressed kernel address, it will be copied first Go to a safe place before unzipping to prevent self-covering. This requires time-consuming copying.

That is, setting the loading address and running address to different addresses can reduce time consumption.

Kernel clipping

Cutting the kernel is necessary. If the kernel image is too large, it will take a long time to decompress the kernel, so the kernel must Cut as much as possible.

Cutting the kernel can reduce the decompression time. Less initialization content will also reduce time consumption.

Therefore, when trimming the kernel, consider removing all unnecessary functions.

Preset lpj value

LPJ is loops_per_jiffy, which will be calculated every time it is started Once, but if no modifications are made, the value will be calculated the same every time it is started. You can directly provide the value to skip the calculation.

如下log所示，有skipped，lpj由timer计算得来，不需要再校准calibrate了。

[ 0.019918] Calibrating delay loop (skipped), value calculated using timer frequency.. 48.00 BogoMIPS (lpj=240000)

如果没有skipped，则可以在cmdline中添加lpj=xxx进行预设

initcall优化

如前面提到,initcall耗时是可以打印出来的，在cmdline中设置initcall_debug=1，即可打印跟踪所有内核初始化过程中调用的顺序以及耗时。

[ 0.021772] initcall sunxi_pinctrl_init+0x0/0x44 returned 0 after 9765 usecs
[ 0.067694] initcall param_sysfs_init+0x0/0x198 returned 0 after 29296 usecs
[ 0.070240] initcall genhd_device_init+0x0/0x88 returned 0 after 9765 usecs
[ 0.080405] initcall init_scsi+0x0/0x90 returned 0 after 9765 usecs
[ 0.090384] initcall mmc_init+0x0/0x84 returned 0 after 9765 usecs

根据打印信息，可以对耗时较多的initcall进行优化。

内核initcall_module并行

initcall有很多等级，但比较耗时的是module。

如果是多核，可以考虑将module_initcall并行执行来节省时间。

目前内核do_initcalls是一个一个按照顺序来执行，可以修改成新建内核线程来执行

减少pty/tty个数

加入initcall打印之后，发现pty/tty init耗时很多，可减少个数来缩短init时间。

initcall pty_init+0x0/0x3c4 returned 0 after 239627 usecs
initcall chr_dev_init+0x0/0xdc returned 0 after 36581 usecs

内核module

只把必须要加进内核的才编译进内核，其他的编译成模块。

例如将必要的clock、tty、pinctrl等编译进内核

3. Other optimizations

uboot

RISC-V architecture, you can consider removing uboot.

XIP

xip: eXecute In Place. That is, on-chip execution means that the CPU directly reads the program code from the memory for execution without reading it into the memory.

flash. When the system starts, the program is copied from flash to ddr for execution. xip technology does not need to copy the program to ddr, so the speed will be very fast.

This technology must be supported by the chip. You can check whether the description of SPI in the chip manual supports the XIP function.

4. Summary

The above optimization of system startup speed is ultimately provided by Some ideas, some methods.

To optimize the startup speed, generally speaking requires a deeper understanding of the startup of the entire system.

Optimization is endless, it needs to be optimized according to the goal, taking into account the startup speed and effect.