Google Git
Sign in
android / platform / docs / source.android.com / 179abc8d13154ca5242e77a80145c479a9548e9d / . / en / devices / tech / perf / boot-times.html
blob: 197b7ff1b203be65b8a7564bdb3f20bcc1730bbd [file] [log] [blame]
<html devsite>
<head>
<title>Optimizing Boot Times</title>
<meta name="project_path" value="/_project.yaml" />
<meta name="book_path" value="/_book.yaml" />
</head>
<body>
<!--
Copyright 2017 The Android Open Source Project
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
<p>
This document provides partner guidance for improving boot times for specific
Android devices. Boot time is an important component of system performance as
users must wait for boot to complete before they can use the device. For devices
such as cars where cold boot-up happens more frequently, having a quick boot
time is critical (no one likes waiting for dozens of seconds just to input a
navigation destination).
</p>
<p>
Android 8.0 allows for reduced boot times by supporting several improvements
across a range of components. The following table summarizes these performance
improvements (as measured on a Google Pixel and Pixel XL devices).
</p>
<table>
<tr>
<th>Component</th>
<th>Improvement</th>
</tr>
<tr>
<td>Bootloader
</td>
<td>
<ul>
<li>Saved 1.6s by removing UART log
<li>Saved 0.4s by changing to LZ4 from GZIP</li>
</ul>
</td>
</tr>
<tr>
<td>Device kernel
</td>
<td>
<ul>
<li>Saved 0.3s by removing unused kernel configs and reducing driver size
<li>Saved 0.3s with dm-verity prefetch optimization
<li>Saved 0.15s to remove unnecessary wait/test in driver
<li>Saved 0.12s to remove CONFIG_CC_OPTIMIZE_FOR_SIZE</li>
</ul>
</td>
</tr>
<tr>
<td>I/O tuning
</td>
<td>
<ul>
<li>Saved 2s on normal boot
<li>Saved 25s on first boot</li>
</ul>
</td>
</tr>
<tr>
<td>init.*.rc
</td>
<td>
<ul>
<li>Saved 1.5s by paralleling init commands
<li>Saved 0.25s by starting zygote early
<li>Saved 0.22s by cpuset tune</li>
</ul>
</td>
</tr>
<tr>
<td>Boot animation
</td>
<td>
<ul>
<li>Started 2s earlier on boot without fsck triggered, much bigger on boot with
fsck triggered boot
<li>Saved 5s on Pixel XL with immediate shutdown of boot animation</li>
</ul>
</td>
</tr>
<tr>
<td>SELinux policy
</td>
<td>Saved 0.2s on by genfscon
</td>
</tr>
</table>
<h2 id="optimizing-bootloader">Optimizing Bootloader</h2>
<p>
To optimize bootloader for improved boot times:
</p>
<ul>
<li>For logging:
<ul>
<li>Disable log writing to UART as it can take a long time with lots of
logging. (On the Google Pixel devices, we found it slows the bootloader 1.5s).</li>
<li>Log only error situations and consider storing other information to memory
with a separate mechanism to retrieve.</li>
</ul>
</li>
<li>For kernel decompression, considering using LZ4 for contemporary hardware
instead of GZIP (example <a class="external"
href="https://patchwork.kernel.org/patch/6810841/">patch</a>). Keep in mind that
different kernel compression options can have different loading and
decompression times, and some options may work better than others for your
specific hardware.</li>
<li>Check unnecessary wait times for debouncing/special mode entry and minimize
them.</li>
<li>Pass boot time spent in bootloader to kernel as cmdline.</li>
<li>Check CPU clock and consider parallelization (requires multi-core support)
for kernel loading and initializing I/O.</li>
</ul>
<h2 id="optimizing-kernel">Optimizing Kernel</h2>
<p>
Use the following tips to optimize the kernel for improved boot times.
</p>
<h3 id="minimizing-device-defconfig">Minimizing device defconfig</h3>
<p>
Minimizing kernel config can reduce kernel size for faster loading
decompression, initialization and smaller attack surfaces. To optimize the
device defconfig:
</p>
<ul>
<li><strong>Identify unused drivers</strong>. Review the <code>/dev</code> and
<code>/sys</code> directories and look for nodes with general SELinux labels
(which indicates those nodes are not configured to be accessible by user space).
Remove any such nodes if found.
<li><strong>Unset unused CONFIGs</strong>. Review the .config file generated by
kernel build to explicitly unset any unused CONFIG that was turned on by
default. For example, we removed the following unused CONFIGs from the Google
Pixel:
<pre
class="prettyprint">
CONFIG_ANDROID_LOGGER=y
CONFIG_IMX134=y
CONFIG_IMX132=y
CONFIG_OV9724=y
CONFIG_OV5648=y
CONFIG_GC0339=y
CONFIG_OV8825=y
CONFIG_OV8865=y
CONFIG_s5k4e1=y
CONFIG_OV12830=y
CONFIG_USB_EHCI_HCD=y
CONFIG_IOMMU_IO_PGTABLE_FAST_SELFTEST=y
CONFIG_IKCONFIG=y
CONFIG_RD_BZIP2=y
CONFIG_RD_LZMA=y
CONFIG_TI_DRV2667=y
CONFIG_CHR_DEV_SCH=y
CONFIG_MMC=y
CONFIG_MMC_PERF_PROFILING=y
CONFIG_MMC_CLKGATE=y
CONFIG_MMC_PARANOID_SD_INIT=y
CONFIG_MMC_BLOCK_MINORS=32
CONFIG_MMC_TEST=y
CONFIG_MMC_SDHCI=y
CONFIG_MMC_SDHCI_PLTFM=y
CONFIG_MMC_SDHCI_MSM=y
CONFIG_MMC_SDHCI_MSM_ICE=y
CONFIG_MMC_CQ_HCI=y
CONFIG_MSDOS_FS=y
# CONFIG_SYSFS_SYSCALL is not set
CONFIG_EEPROM_AT24=y
# CONFIG_INPUT_MOUSEDEV_PSAUX is not set
CONFIG_INPUT_HBTP_INPUT=y
# CONFIG_VGA_ARB is not set
CONFIG_USB_MON=y
CONFIG_USB_STORAGE_DATAFAB=y
CONFIG_USB_STORAGE_FREECOM=y
CONFIG_USB_STORAGE_ISD200=y
CONFIG_USB_STORAGE_USBAT=y
CONFIG_USB_STORAGE_SDDR09=y
CONFIG_USB_STORAGE_SDDR55=y
CONFIG_USB_STORAGE_JUMPSHOT=y
CONFIG_USB_STORAGE_ALAUDA=y
CONFIG_USB_STORAGE_KARMA=y
CONFIG_USB_STORAGE_CYPRESS_ATACB=y
CONFIG_SW_SYNC_USER=y
CONFIG_SEEMP_CORE=y
CONFIG_MSM_SMEM_LOGGING=y
CONFIG_IOMMU_DEBUG=y
CONFIG_IOMMU_DEBUG_TRACKING=y
CONFIG_IOMMU_TESTS=y
CONFIG_MOBICORE_DRIVER=y
# CONFIG_DEBUG_PREEMPT is not set
</pre>
</li>
</ul>
<ul>
<li><strong>Remove CONFIGs that lead to unnecessary test runs on every
boot</strong>. While useful in development, such configs (i.e.
CONFIG_IOMMU_IO_PGTABLE_FAST_SELFTEST) should be removed in a production
kernel.
</li>
</ul>
<h3 id="minimizing-driver-size">Minimizing driver size</h3>
<p>
Some drivers in the device kernel can be removed if the function is not used to
reduce kernel size further. For example, if WLAN is connected through PCIe, the
SDIO support is not used and should be removed during compile time. For details,
refer to the Google Pixel kernel: net: wireless: cnss: add option to disable
SDIO support.
</p>
<h3 id="removing-compiler-optimization-for-size">Removing compiler optimization
for size</h3>
<p>
Remove the kernel config for CONFIG_CC_OPTIMIZE_FOR_SIZE. This flag was
originally introduced when the assumption was that smaller code size would yield
hot cache hit (and thus be faster). However, this assumption is no longer valid
as modern mobile SoCs have become more powerful.
</p>
<p>
In addition, removing the flag can enable the compiler warning for uninitialized
variables, which is suppressed in Linux kernels when the
CONFIG_CC_OPTIMIZE_FOR_SIZE flag is present (making this change alone has helped
us uncover many meaningful bugs in some Android device drivers).
</p>
<h3 id="deferring-initialization">Deferring initialization</h3>
<p>
Many processes launch during boot, but only components in critical path
(bootloader > kernel > init > file system mount > zygote > system server)
directly affect boot time. Profile <strong>initcall</strong> during kernel
booting to identify peripheral/components that are slow and not critical
to start init process, then delay those peripherals/components until later
in the boot process by moving into loadable kernel modules. Moving to
asynchronous device/driver probe can also help to parallel slow components
in kernel > init critical path.
</p>
<pre
class="prettyprint">
BoardConfig-common.mk:
BOARD_KERNEL_CMDLINE += initcall_debug ignore_loglevel
driver:
.probe_type = PROBE_PREFER_ASYNCHRONOUS,
</pre>
<p class="note">
<strong>Note:</strong> Driver dependencies must be resolved carefully by adding
<code>EPROBEDEFER</code> support.
</p>
<h2 id="optimizing-i-o-efficiency">Optimizing I/O efficiency</h2>
<p>
Improving I/O efficiency is critical to making boot time faster, and reading
anything not necessary should be deferred until after boot (on a Google Pixel,
about 1.2GB of data is read on boot).
</p>
<h3 id="tuning-the-filesystem">Tuning the filesystem</h3>
<p>
The Linux kernel read ahead kicks in when a file is read from beginning or when
blocks are read sequentially, making it necessary to tune I/O scheduler
parameters specifically for booting (which has a different workload
characterization than normal applications).
</p>
<p>
Devices that support seamless (A/B) updates benefit greatly from filesystem
tuning on first time boot (e.g. 20s on Google Pixel). An an example, we tuned
the following parameters for the Google Pixel:
</p>
<pre
class="prettyprint">
on late-fs
# boot time fs tune
# boot time fs tune
write /sys/block/sda/queue/iostats 0
write /sys/block/sda/queue/scheduler cfq
write /sys/block/sda/queue/iosched/slice_idle 0
write /sys/block/sda/queue/read_ahead_kb 2048
write /sys/block/sda/queue/nr_requests 256
write /sys/block/dm-0/queue/read_ahead_kb 2048
write /sys/block/dm-1/queue/read_ahead_kb 2048
on property:sys.boot_completed=1
# end boot time fs tune
write /sys/block/sda/queue/read_ahead_kb 512
...
</pre>
<h3 id="miscellaneous">Miscellaneous</h3>
<ul>
<li>Turn on the dm-verity hash prefetch size using kernel config
DM_VERITY_HASH_PREFETCH_MIN_SIZE (default size is 128).
<li>For better file system stability and a dropped forced check that occurs on
every boot, use the new ext4 generation tool by setting TARGET_USES_MKE2FS in
BoardConfig.mk.</li>
</ul>
<h3 id="analyzing-i-o">Analyzing I/O</h3>
<p>
To understand I/O activities during boot, use kernel ftrace data (also used by
systrace):
</p>
<pre
class="prettyprint">trace_event=block,ext4 in BOARD_KERNEL_CMDLINE</pre>
<p>
To breakdown file access for each file, make the following changes to the kernel
(development kernel only; do not use in production kernels):
</p>
<pre
class="prettyprint">
diff --git a/fs/open.c b/fs/open.c
index 1651f35..a808093 100644
--- a/fs/open.c
+++ b/fs/open.c
@@ -981,6 +981,25 @@
}
EXPORT_SYMBOL(file_open_root);
+static void _trace_do_sys_open(struct file *filp, int flags, int mode, long fd)
+{
+ char *buf;
+ char *fname;
+
+ buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
+ if (!buf)
+ return;
+ fname = d_path(&filp-&lt;f_path, buf, PAGE_SIZE);
+
+ if (IS_ERR(fname))
+ goto out;
+
+ trace_printk("%s: open(\"%s\", %d, %d) fd = %ld, inode = %ld\n",
+ current-&lt;comm, fname, flags, mode, fd, filp-&lt;f_inode-&lt;i_ino);
+out:
+ kfree(buf);
+}
+
long do_sys_open(int dfd, const char __user *filename, int flags, umode_t mode)
{
struct open_flags op;
@@ -1003,6 +1022,7 @@
} else {
fsnotify_open(f);
fd_install(fd, f);
+ _trace_do_sys_open(f, flags, mode, fd);
</pre>
<p>
Use the following scripts to help with analyzing boot performance.
</p>
<ul>
<li><code>system/extras/boottime_tools/bootanalyze/bootanalyze.py</code>
Measures boot time with a breakdown of important steps in the boot process.
<li><code>system/extras/boottime_tools/io_analysis/check_file_read.py
boot_trace</code> Provides access information per each file.
<li><code>system/extras/boottime_tools/io_analysis/check_io_trace_all.py
boot_trace</code> Gives system-level breakdown.</li>
</ul>
<h2 id="optimizing-init-*-rc">Optimizing init.*.rc</h2>
<p>
Init is the bridge from the kernel till the framework is established, and
devices usually spend a few seconds in different init stages.
</p>
<h3 id="running-tasks-in-parallel">Running tasks in parallel</h3>
<p>
While the current Android init is more or less a single threaded process, you
can still perform some tasks in parallel.
</p>
<ul>
<li>Execute slow commands in a shell script service and join that later by
waiting for specific property. Android 8.0 supports this use case with a new
<code>wait_for_property</code> command.
<li>Identify slow operations in init. The system logs the init command
exec/wait_for_prop or any action taking a long time (in Android 8.0, any command
taking more than 50 ms). For example:
<pre
class="prettyprint">init: Command 'wait_for_coldboot_done' action=wait_for_coldboot_done returned 0 took 585.012ms</pre>
<p>
Reviewing this log may indicate opportunities for improvements.
</p>
<li>Start services and enable peripheral devices in critical path early. For
example, some SOCs require starting security-related services before starting
SurfaceFlinger. Review the system log when ServiceManager returns "wait for
service" — this is usually a sign that a dependent service must be started
first.
<li>Remove any unused services and commands in init.*.rc. Anything not used in
early stage init should be deferred to boot completed.</li></ul>
</li>
</ul>
<p class="note">
<strong>Note:</strong> Property service is part of init process, so calling
<code>setproperty</code> during boot can lead a long delay if init is busy in
builtin commands.
</p>
<h3 id="using-scheduler-tuning">Using scheduler tuning</h3>
<p>
Use scheduler tuning for early boot. Example from a Google Pixel:
</p>
<pre
class="prettyprint">on init
# update cpusets now that processors are up
write /dev/cpuset/top-app/cpus 0-3
write /dev/cpuset/foreground/cpus 0-3
write /dev/cpuset/foreground/boost/cpus 0-3
write /dev/cpuset/background/cpus 0-3
write /dev/cpuset/system-background/cpus 0-3
# set default schedTune value for foreground/top-app (only affects EAS)
write /dev/stune/foreground/schedtune.prefer_idle 1
write /dev/stune/top-app/schedtune.boost 10
write /dev/stune/top-app/schedtune.prefer_idle 1</pre>
<p>
Some services may need a priority boost during boot. Example:
</p>
<pre
class="prettyprint">
init.zygote64.rc:
service zygote /system/bin/app_process64 -Xzygote /system/bin --zygote --start-system-server
class main
priority -20
user root
...</pre>
<h3 id="starting-zygote-early">Starting zygote early</h3>
<p>
Devices with file-based encryption can start zygote earlier at the zygote-start
trigger (by default, zygote is launched at class main, which is much later than
zygote-start). When doing this, make sure to allow zygote to run in all CPUs (as
the wrong cpuset setting may force zygote to run in specific CPUs).
</p>
<h3 id="disable-power-saving">Disable power saving</h3>
<p>
During device booting, power saving setting for components like UFS and/or CPU
governor can be disabled.
</p>
<p class="aside"><strong>Caution:</strong> Power saving should be enabled in
charger mode for efficiency.</p>
<pre
class="prettyprint">
on init
# Disable UFS powersaving
write /sys/devices/soc/${ro.boot.bootdevice}/clkscale_enable 0
write /sys/devices/soc/${ro.boot.bootdevice}/clkgate_enable 0
write /sys/devices/soc/${ro.boot.bootdevice}/hibern8_on_idle_enable 0
write /sys/module/lpm_levels/parameters/sleep_disabled Y
on property:sys.boot_completed=1
# Enable UFS powersaving
write /sys/devices/soc/${ro.boot.bootdevice}/clkscale_enable 1
write /sys/devices/soc/${ro.boot.bootdevice}/clkgate_enable 1
write /sys/devices/soc/${ro.boot.bootdevice}/hibern8_on_idle_enable 1
write /sys/module/lpm_levels/parameters/sleep_disabled N
on charger
# Enable UFS powersaving
write /sys/devices/soc/${ro.boot.bootdevice}/clkscale_enable 1
write /sys/devices/soc/${ro.boot.bootdevice}/clkgate_enable 1
write /sys/devices/soc/${ro.boot.bootdevice}/hibern8_on_idle_enable 1
write /sys/class/typec/port0/port_type sink
write /sys/module/lpm_levels/parameters/sleep_disabled N</pre>
<h3 id="defer-non-critical-initialization">Defer non-critical initialization</h3>
<p>
Non-critical initialization such as ZRAM can be deferred to boot_complete.
</p>
<pre
class="prettyprint">
on property:sys.boot_completed=1
# Enable ZRAM on boot_complete
swapon_all /vendor/etc/fstab.${ro.hardware}</pre>
<h2 id="optimizing-boot-animation">Optimizing boot animation</h2>
<p>
Use the following tips to optimize the boot animation.
</p>
<h3 id="configuring-early-start">Configuring early start</h3>
<p>
Android 8.0 enables starting boot animation early, before mounting userdata
partition. However, even when using the new ext4 tool chain in Android 8.0, fsck
is still triggered periodically due to safety reasons, causing a delay in
starting the bootanimation service.
</p>
<p>
To make bootanimation start early, split the fstab mount into two phases:
</p>
<ul>
<li>In the early phase, mount only the partitions (such as
<code>system/</code> and <code>vendor/</code>) that don't require run
checks, then start boot animation services and its dependencies (such as
servicemanager and surfaceflinger).
<li>In the second phase, mount partitions (such as <code>data/</code>) that
do require run checks.</li>
</ul>
<p>
Boot animation will be started much faster (and in constant time) regardless of
fsck.
</p>
<h3 id="finishing-clean">Finishing clean</h3>
<p>
After receiving the exit signal, bootanimation plays the last part, the length
of which can slow boot time. A system that boots quickly has no need for lengthy
animations which could effectively hide any improvements made. We recommend
making both the repeating loop and finale short.
</p>
<h2 id="optimizing-selinux">Optimizing SELinux</h2>
<p>
Use the following tips to optimize SELinux for improved boot times.
</p>
<ul>
<li><strong>Use clean regular expressions (regex)</strong>. Poorly-formed regex
can lead to a lot of overhead when matching SELinux policy for
<code>sys/devices</code> in <code>file_contexts</code>. For example, the regex
<code>/sys/devices/.*abc.*(/.*)?</code> mistakenly forces a scan of all
<code>/sys/devices</code> subdirectories that contain "abc", enabling matches
for both <code>/sys/devices/abc</code> and <code>/sys/devices/xyz/abc</code>.
Improving this regex to <code>/sys/devices/[^/]*abc[^/]*(/.*)?</code> will
enable a match only for <code>/sys/devices/abc</code>.
<li><strong>Move labels to </strong><a
href="https://selinuxproject.org/page/FileStatements#genfscon">genfscon</a>.
This existing SELinux feature passes file-matching prefixes into the kernel in
the SELinux binary, where the kernel applies them to kernel-generated
filesystems. This also helps fix mislabeled kernel-created files, preventing
race conditions that can occur between userspace processes attempting to access
these files before relabeling occurs.</li>
</ul>
<h2 id="tool-and-methods">Tool and methods</h2>
<p>
Use the following tools to help you collect data for optimization targets.
</p>
<h3 id="bootchart">Bootchart</h3>
<p>
Bootchart provides CPU and I/O load breakdown of all processes for the whole
system. It doesn't require rebuilding system image and can be used as a quick
sanity check before diving into systrace.
</p>
<p>
To enable bootchart:
</p>
<pre>
<code class="devsite-terminal">adb shell 'touch /data/bootchart/enabled'</code>
<code class="devsite-terminal">adb reboot</code>
</pre>
<p>
After boot up, fetch boot chart:
</p>
<pre
class="prettyprint">$ANDROID_BUILD_TOP/system/core/init/grab-bootchart.sh</pre>
<p>
When finished, delete <code>/data/bootchart/enabled</code> to prevent collecting
the date every time.
</p>
<h3 id="systrace">Systrace</h3>
<p>
Systrace allows collecting both kernel and Android traces during boot up.
Visualization of systrace can help in analyzing specific problem during the
boot-up. (However, to check the average number or accumulated number during the
entire boot, it is easier to look into kernel trace directly).
</p>
<p>
To enable systrace during boot-up:
</p>
<ul>
<li> In <code>frameworks/native/atrace/atrace.rc</code>, change:
<pre
class="prettyprint">write /sys/kernel/debug/tracing/tracing_on 0</pre>
<p>
To:
</p>
<pre
class="prettyprint">#write /sys/kernel/debug/tracing/tracing_on 0</pre>
</li>
<p>
This enables tracing (which is disabled by default).
</p>
<li>In the <code>device.mk</code> file, add the following line:
<pre
class="prettyprint">PRODUCT_PROPERTY_OVERRIDES += debug.atrace.tags.enableflags=802922</pre>
</li>
<li>In the device <code>BoardConfig.mk</code> file, add the following:
<pre
class="prettyprint">BOARD_KERNEL_CMDLINE := ... trace_buf_size=64M trace_event=sched_wakeup,sched_switch,sched_blocked_reason,sched_cpu_hotplug</pre>
</li>
<p>
For detailed I/O analysis, also add block and ext4.
</p>
<li>In the device-specific <code>init.rc</code> file, make the following
changes:
<ul>
<li><code>on property:sys.boot_completed=1</code> (this stops tracing on boot
complete)</li>
<li><code>write /d/tracing/tracing_on 0</code></li>
<li><code>write /d/tracing/events/ext4/enable 0</code></li>
<li><code>write /d/tracing/events/block/enable 0</code></li>
</ul>
</li>
</ul>
<p>
After boot up, fetch trace:
</p>
<pre class="devsite-terminal">adb root && adb shell "cat /d/tracing/trace" &lt; boot_trace
./external/chromium-trace/catapult/tracing/bin/trace2html boot_trace --output boot_trace.html
</pre>
<p class="note">
<strong>Note:</strong> Chrome cannot handle files that are too large. Consider
cutting the <code>boot_trace</code> file using <code>tail</code>,
<code>head</code>, or <code>grep</code> for necessary portions. And
I/O analysis often requires analyzing the captured <code>boot_trace</code>
directly, as there are too many events.
</p>
</body>
</html>