MozillaWiki

Mobile/Fennec/Android/Profiling: Difference between revisions

Page Discussion

Mobile/Fennec/Android/Profiling (view source)

Revision as of 20:41, 9 January 2014

13,313 bytes added ,  9 January 2014
→‎traceview
 
(21 intermediate revisions by 2 users not shown)
Line 1: Line 1:
= traceview =
Java code can be effectively profiled with traceview, included in the Android SDK. See http://developer.android.com/guide/developing/tools/traceview.html
* The Fennec build needs to be "[http://developer.android.com/guide/topics/manifest/application-element.html#debug debuggable]". Nightlies are not, but local developer builds are.
* For profiling pageload, you can generally use [http://developer.android.com/tools/debugging/debugging-tracing.html Monitor/DDMS]. Just press the "start method tracing" button, load a page, wait, press the "stop method tracing" button.
* For profiling startup time, you really need to add Debug.startMethodTracing(...) and Debug.stopMethodTracing(). [http://people.mozilla.org/~mfinkle/fennec/profiles/startup-tracing.patch Sample patch]
* Or use the "am start" -P option, e.g.
adb shell am start -n org.mozilla.fennec_$USER/.App -S -P /data/local/tmp/fennec.trace
# wait for Fennec launch
adb pull /data/local/tmp/fennec.trace
# in Monitor, File > Open File > fennec.trace
= Built-in Profiler =
See https://developer.mozilla.org/en/Performance/Profiling_with_the_Built-in_Profiler#Profiling_Firefox_mobile
= oprofile =
= oprofile =


Line 59: Line 76:
    
    
   static void armv7_pmnc_stop(void)
   static void armv7_pmnc_stop(void)
If you will be getting callgraphs, apply the patch from {{bug|674986}}


Set the toolchain and compile: (the original kernel is built with 4.4.3 but NDK has 4.4.0; not sure if this matters...)
Set the toolchain and compile: (the original kernel is built with 4.4.3 but NDK has 4.4.0; not sure if this matters...)
Line 139: Line 158:


  echo "#! /bin/bash
  echo "#! /bin/bash
  rm -f -R oprofile
  rm -rf oprofile
  mkdir oprofile
  mkdir oprofile
  adb pull /data/oprofile oprofile/" > oppull
  adb pull /data/oprofile oprofile/" > oppull
Line 146: Line 165:
  adb shell mount -o remount,ro /system
  adb shell mount -o remount,ro /system


The sample rate is (CPU freuqency in MHz) / (CPU_CYCLES count) / 64. For the HTC G2, it's 800MHz / 100000 / 64 = 125 samples per second.
The sample rate is (CPU freuqency in MHz) / (CPU_CYCLES count) / 64. For the HTC G2, it's 800MHz / 125000 / 64 = 100 samples per second.


== Using ''oprofile'' on CyanogenMod 7 ==
== Using ''oprofile'' on CyanogenMod 7 ==
Line 367: Line 386:
This needs to be investigated, but it looks like ''oprofile'' compiled into the kernel does not have proper support for Nexus One hardware, and has to be fixed.
This needs to be investigated, but it looks like ''oprofile'' compiled into the kernel does not have proper support for Nexus One hardware, and has to be fixed.


I encountered the same behavior when running opcontrol on the Galaxy Nexus and found that defining CONFIG_HW_PERF_EVENTS in the kernel .config file fixed the problem on the Galaxy Nexus.
== Using oprofile on Galaxy Nexus ==
See also bug730900.
There is no oprofile support in the factory / production Android build on the Galaxy Nexus. The engineering build (ie. full_maguro-eng configuration) includes user-mode support for oprofile (opcontrol, etc.), but no kernel support for oprofile (/dev/oprofilefs is not present).
The approach used here is to build a custom kernel with oprofile support configured, patch the Android source tree with the new kernel and then build and flash an engineering build of Android.
Full instructions for building Android begin at http://source.android.com/source/initializing.html.
Here is the short version reflecting what I did, step-by-step:
<nowiki>
# Get Android sources, initialize environment
mkdir <myandroid>
cd <myandroid>
curl https://dl-ssl.google.com/dl/googlesource/git-repo/repo > repo
chmod +x repo
./repo init -u https://android.googlesource.com/platform/manifest -b android-4.0.1_r1.2
./repo sync
source build/envsetup.sh
lunch full_maguro-eng
<get the ITL41F binaries from http://code.google.com/android/nexus/drivers.html#maguroitl41f and extract them>
# Get kernel source, customize for oprofile, build kernel
mkdir <mykernel>
cd <mykernel>
git clone https://android.googlesource.com/kernel/omap.git
cd omap
git branch -a
git checkout -t remotes/origin/android-omap-tuna-3.0-mr0
export ARCH=arm
export CROSS_COMPILE=<myandroid>/prebuilt/linux-x86/toolchain/arm-eabi-4.4.3/bin/arm-eabi-
make tuna_defconfig
<edit .config, setting:
CONFIG_PROFILING=y
CONFIG_OPROFILE=y
CONFIG_HAVE_OPROFILE=y
CONFIG_HW_PERF_EVENTS=y
CONFIG_PERF_EVENTS=y
CONFIG_PERF_COUNTERS=y
/edit>
make
</nowiki>
The opcontrol project is in <myandroid>/external/oprofile/opcontrol; it includes special support for ARM_V7_A, but there is a bug in the makefile, Android.mk: -DWITH_ARM_V7_A is added to LOCAL_CFLAGS, but then LOCAL_CFLAGS is clobbered further down -- fix this with something like:
<nowiki>
LOCAL_C_INCLUDES := $(common_target_c_includes)
LOCAL_CFLAGS := $(common_target_cflags)
# Force ARM_V7_A
LOCAL_CFLAGS += -DWITH_ARM_V7_A
</nowiki>
<nowiki>
# Patch Android with my kernel, build, install images
cp arch/arm/boot/zImage <myandroid>/device/samsung/tuna/kernel
cd <myandroid>
make
<put phone into bootloader mode: power off, then hold volume up + volume down + power>
fastboot -w flashall
</nowiki>
Once the phone reboots, verify the new kernel:
<nowiki>
mylinux$ adb shell dmesg | grep oprofile
<6>[    2.772705] oprofile: using arm/armv7-ca9
</nowiki>
Now everything appears to be working, but I cannot seem to collect samples!
<nowiki>
    root@android:/data/local # opcontrol --reset --kernel-range=0xc004d000,0xc0752000 --event=CPU_CYCLES:1 --setup --status --verbose-log=all --verbose
    list_events = 0
    setup = 1
    Configure /dev/oprofile/0/user (1)
    Configure /dev/oprofile/0/kernel (1)
    Configure /dev/oprofile/0/unit_mask (0)
    Configure /dev/oprofile/0/enabled (1)
    Configure /dev/oprofile/0/count (1)
    Configure /dev/oprofile/0/event (255)
    Configure /dev/oprofile/1/enabled (0)
    Configure /dev/oprofile/2/enabled (0)
    Configure /dev/oprofile/3/enabled (0)
    Configure /dev/oprofile/4/enabled (0)
    Starting oprofiled...
    command: oprofiled --session-dir=/data/oprofile --events=CPU_CYCLES:255:0:1:0:1:1 --no-vmlinux -r 0xc004d000,0xc0752000 --verbose=all
    Events: CPU_CYCLES:255:0:1:0:1:1
    Using 2.6+ OProfile kernel interface.
    Using log file /data/oprofile/samples/oprofiled.log
    Ready
    Driver directory: /dev/oprofile
    Session directory: /data/oprofile
    Counter 0:
        name: CPU_CYCLES
        count: 1
    Counter 1 disabled
    Counter 2 disabled
    Counter 3 disabled
    Counter 4 disabled
    oprofiled pid: 1528
    profiler is not running
      cpu1        0 samples received
      cpu1        0 samples lost overflow
      cpu1        0 samples invalid eip
      cpu1        0 backtrace aborted
      cpu0        0 samples received
      cpu0        0 samples lost overflow
      cpu0        0 samples invalid eip
      cpu0        0 backtrace aborted
   
    root@android:/data/local #
    root@android:/data/local # opcontrol --start
    root@android:/data/local # opcontrol --status --verbose
    list_events = 0
    setup = 0
    Driver directory: /dev/oprofile
    Session directory: /data/oprofile
    Counter 0:
        name: CPU_CYCLES
        count: 1
    Counter 1 disabled
    Counter 2 disabled
    Counter 3 disabled
    Counter 4 disabled
    oprofiled pid: 1528
    profiler is running
      cpu1        0 samples received
      cpu1        0 samples lost overflow
      cpu1        0 samples invalid eip
      cpu1        0 backtrace aborted
      cpu0        0 samples received
      cpu0        0 samples lost overflow
      cpu0        0 samples invalid eip
      cpu0        0 backtrace aborted
   
    root@android:/data/local # opcontrol --stop
    root@android:/data/local # ls -lR /data/oprofile
   
    /data/oprofile:
    -rw-rw-rw- root    root          724 2012-03-07 16:43 abi
    -rw-rw-rw- root    root            2 2012-03-07 16:46 complete_dump
    -rw-rw-rw- root    root            4 2012-03-07 16:43 lock
    prw-rw-rw- root    root              2012-03-07 16:43 opd_pipe
    drwxr-xr-x root    root              2012-03-07 16:43 samples
   
    /data/oprofile/samples:
    -rw-r--r-- root    root        57424 2012-03-07 16:46 oprofiled.log
    root@android:/data/local # tail /data/oprofile/samples/oprofiled.log
    ...
    CPU_SWITCH to 1
    CPU_SWITCH to 0
    CPU_SWITCH to 1
    CPU_SWITCH to 0
    CPU_SWITCH to 1
</nowiki>


== Using ''oprofile'' ==
== Using ''oprofile'' ==
Line 420: Line 597:


Patch PowerTOP using this patch:
Patch PowerTOP using this patch:
'''TODO''': Investigate other changes to PowerTOP source to make it run better on Android


  diff --git a/powertop-1.13/Makefile b/powertop-1.13/Makefile
  diff --git a/powertop-1.13/Makefile b/powertop-1.13/Makefile
Line 613: Line 792:
  adb shell
  adb shell
  TERMINFO=/system/etc/terminfo /data/powertop
  TERMINFO=/system/etc/terminfo /data/powertop
= systemtap =
== Compiling systemtap ==
Think of the children, get the [http://www.jnchen.com/_media/projects/mozilla/moz-systemtap-1.5.tar.bz2 prebuilt version]. Extract it to your host machine, and to /data on your device.
Basically running systemtap consists of the host portion and the target portion. For the host portion you use a standard systemtap installation to cross compile your script into a kernel module. For the target portion you transfer the kernel module to your device and run it on the device.
The host portion you can get through apt-get. The target portion you have to compile (use the [http://www.codesourcery.com/sgpp/lite/arm/portal/subscription?@template=lite CodeSourcery ARM GNU/Linux toolchain]):
LDFLAGS=-static \
  ac_cv_file__usr_include_avahi_common=no \
  ac_cv_file__usr_include_avahi_client=no \
  ac_cv_file__usr_include_nspr=no \
  ac_cv_file__usr_include_nspr4=no \
  ac_cv_file__usr_include_nss=no \
  ac_cv_file__usr_include_nss3=no \
  ./configure --prefix=/data/systemtap --host=arm-none-linux-gnueabi \
    --disable-translator --disable-nls
make install
Note that prefix has to be /data/systemtap (or another path that matches Android file system) so you need to make a /data directory on your host machine before you run make install.
You might need libelf. Get the latest elfutils release, then compile:
./configure --host=arm-none-linux-gnueabi --disable-nls \
  --prefix=/PATH/TO/TOOLCHAIN/arm-none-linux-gnueabi/libc/usr
cd libelf
make install
Where /PATH/TO/TOOLCHAIN is the path to your CodeSourcery toolchain.
== Using systemtap ==
Before you start, you need the source of your device's kernel. See the oprofile section for one example way to get the kernel source.
To compile your script on the host, make sure you have the NDK toolchain (r5 or above), then run
stap -gv -a arm -B CROSS_COMPILE=arm-linux-androideabi- \
  -r /PATH/TO/KERNEL -R /PATH/TO/SYSTEMTAP/RUNTIME \
  -m NAME -B CFLAGS_MODULE="-DMODULE -fno-pic" NAME.stp
where NAME is the name of your script, /PATH/TO/KERNEL is the path to your kernel source, and /PATH/TO/SYSTEMTAP/RUNTIME is the path to your target systemtap runtime (for example if you're using the prebuilt version above, /PATH/TO/SYSTEMTAP/RUNTIME = systemtap/share/systemtap/runtime)
adb push your kernel module to your device, then run
staprun -x PID NAME.ko
where PID is the pid of fennec process
== Sample systemtap scripts ==
=== glandium's I/O tracking script ===
global targetpid;
global file_path;
probe begin {
  targetpid = target();
}
probe kernel.function("__do_page_cache_readahead") {
  if (targetpid == pid())
    file_path[tid()] = d_path(&$filp->f_path);
}
probe kernel.function("do_mpage_readpage") {
  if (targetpid == pid() && (tid() in file_path)) {
    now = gettimeofday_us();
    printf("%d %s %d\n", now, file_path[tid()], $page->index*4096);
  }
}
probe kernel.function("__do_page_cache_readahead").return {
  if (targetpid == pid())
    delete file_path[tid()];
}
=== jchen's I/O tracking script ===
Note that this script accesses the hardware timer directly to overcome limitations in the systemtap gettimeofday() resolution (basically in systemtap, gettimeofday() is limited to jiffy resolution).
This also means to run it on a processor other than a MSM7x30, get_rawtime() will need to change.
%{
#include <arch/arm/include/asm/io.h>
#define MSM_CSR_BASE ((void*)0xF8001000)
#define MSM_TMR_BASE MSM_CSR_BASE
#define MSM_GPT_BASE (MSM_TMR_BASE + 0x04)
#define MSM_DGT_BASE (MSM_TMR_BASE + 0x24)
#define TIMER_COUNT_VAL 0x0004
%}
function get_rawtime:long () %{ /* pure */ /* unprivileged */
  cycles_t c = readl(MSM_DGT_BASE + TIMER_COUNT_VAL);
  THIS->__retvalue = ((int64_t) (int32_t) c) - global.s_origin_time;
%}
global parentid, childid;
global start_time, total_time;
global origin_time;
global creation_time;
probe begin {
  parentid = 0;
  childid = 0;
  origin_time = get_rawtime()
}
function is_parent:long () {
  if (pid() == parentid)
    return parentid;
  if (0 == parentid && isinstr(execname(), "fennec")) {
    parentid = pid();
    return parentid;
  }
  return 0;
}
function is_child:long () {
  if (pid() == childid)
    return childid;
  if (0 == childid && isinstr(execname(), "plugin")) {
    childid = pid();
    return childid;
  }
  return 0;
}
probe syscall.fork.return {
  if (is_parent() || is_child()) {
    if ($return != 0) {
      thdid = is_parent() ? $return : -$return;
      creation_time[thdid] = get_rawtime();
      total_time[thdid, "", creation_time[thdid] / 8192] = -1;
    }
  }
}
probe syscall.execve.return {
  if (is_parent() || is_child()) {
    thdid = is_parent() ? tid() : -tid();
    creation_time[thdid] = get_rawtime();
    total_time[thdid, "", creation_time[thdid] / 8192] = -1;
  }
}
probe syscall.exit {
  if (is_parent() || is_child()) {
    thdid = is_parent() ? tid() : -tid();
    total_time[thdid, "", creation_time[thdid] / 8192] =
        get_rawtime() - creation_time[thdid];
  }
}
probe vfs.read {
  if (bytes_to_read > 0 && (is_parent() || is_child())) {
    thdid = is_parent() ? tid() : -tid();
    start_time[thdid] = get_rawtime();
  }
}
probe vfs.read.return {
  if (bytes_to_read > 0 && (is_parent() || is_child())) {
    thdid = is_parent() ? tid() : -tid();
    fn = "rd:" . __file_filename(file);
    st = start_time[thdid];
    total_time[thdid, fn, st / 8192] += get_rawtime() - st;
  }
}
probe kernel.function("filemap_fault") {
  if (is_parent() || is_child()) {
    thdid = is_parent() ? tid() : -tid();
    start_time[thdid] = get_rawtime();
  }
}
probe kernel.function("filemap_fault").return {
  if (is_parent() || is_child()) {
    thdid = is_parent() ? tid() : -tid();
    fn = "mm:" . __file_filename($vma->vm_file);
    st = start_time[thdid];
    total_time[thdid, fn, st / 8192] += get_rawtime() - st;
  }
}
probe end {
  foreach ([t, fn, st] in total_time) {
    printf("%d, %s, %d, %d\n", t, fn, st, total_time[t, fn, st]);
  }
}
Nchen
Confirmed users
507

edits

Retrieved from "https://wiki.allizom.org/Special:MobileDiff/325292...875072"