xref: /external/perfetto/docs/case-studies/memory.md

# Debugging memory usage on Android

## Prerequisites

* A host running macOS or Linux.
* [ADB](https://developer.android.com/studio/command-line/adb) installed and
  in PATH.
* A device running Android 11+.

If you are profiling your own app and are not running a userdebug build of
Android, your app needs to be marked as profileable or
debuggable in its manifest. See the [heapprofd documentation](
/docs/data-sources/native-heap-profiler.md#heapprofd-targets) for more
details on which applications can be targeted.

## dumpsys meminfo

A good place to get started investigating memory usage of a process is
`dumpsys meminfo` which gives a high-level overview of how much of the various
types of memory are being used by a process.

```bash
$ adb shell dumpsys meminfo com.android.systemui

Applications Memory Usage (in Kilobytes):
Uptime: 2030149 Realtime: 2030149

** MEMINFO in pid 1974 [com.android.systemui] **
                   Pss  Private  Private  SwapPss      Rss     Heap     Heap     Heap
                 Total    Dirty    Clean    Dirty    Total     Size    Alloc     Free
                ------   ------   ------   ------   ------   ------   ------   ------
  Native Heap    16840    16804        0     6764    19428    34024    25037     5553
  Dalvik Heap     9110     9032        0      136    13164    36444     9111    27333

[more stuff...]
```

Looking at the "Private Dirty" column of Dalvik Heap (= Java Heap) and
Native Heap, we can see that SystemUI's memory usage on the Java heap
is 9M, on the native heap it's 17M.

## Linux memory management

But what does *clean*, *dirty*, *Rss*, *Pss*, *Swap* actually mean? To answer
this question, we need to delve into Linux memory management a bit.

From the kernel's point of view, memory is split into equally sized blocks
called *pages*. These are generally 4KiB.

Pages are organized in virtually contiguous ranges called VMA
(Virtual Memory Area).

VMAs are created when a process requests a new pool of memory pages through
the [mmap() system call](https://man7.org/linux/man-pages/man2/mmap.2.html).
Applications rarely call mmap() directly. Those calls are typically mediated by
the allocator, `malloc()/operator new()` for native processes or by the
Android RunTime for Java apps.

VMAs can be of two types: file-backed and anonymous.

**File-backed VMAs** are a view of a file in memory. They are obtained passing a
file descriptor to `mmap()`. The kernel will serve page faults on the VMA
through the passed file, so reading a pointer to the VMA becomes the equivalent
of a `read()` on the file.
File-backed VMAs are used, for instance, by the dynamic linker (`ld`) when
executing new processes or dynamically loading libraries, or by the Android
framework, when loading a new .dex library or accessing resources in the APK.

**Anonymous VMAs** are memory-only areas not backed by any file. This is the way
allocators request dynamic memory from the kernel. Anonymous VMAs are obtained
calling `mmap(... MAP_ANONYMOUS ...)`.

Physical memory is only allocated, in page granularity, once the application
tries to read/write from a VMA. If you allocate 32 MiB worth of pages but only
touch one byte, your process' memory usage will only go up by 4KiB. You will
have increased your process' *virtual memory* by 32 MiB, but its resident
*physical memory* by 4 KiB.

When optimizing memory use of programs, we are interested in reducing their
footprint in *physical memory*. High *virtual memory* use is generally not a
cause for concern on modern platforms (except if you run out of address space,
which is very hard on 64 bit systems).

We call the amount a process' memory that is resident in *physical memory* its
**RSS** (Resident Set Size). Not all resident memory is equal though.

From a memory-consumption viewpoint, individual pages within a VMA can have the
following states:

* **Resident**: the page is mapped to a physical memory page. Resident pages can
  be in two states:
    * **Clean** (only for file-backed pages): the contents of the page are the
      same of the contents on-disk. The kernel can evict clean pages more easily
      in case of memory pressure. This is because if they should be needed
      again, the kernel knows it can re-create its contents by reading them from
      the underlying file.
    * **Dirty**: the contents of the page diverge from the disk, or (in most
      cases), the page has no disk backing (i.e. it's _anonymous_). Dirty pages
      cannot be evicted because doing so would cause data loss. However they can
      be swapped out on disk or ZRAM, if present.
* **Swapped**: a dirty page can be written to the swap file on disk (on most Linux
  desktop distributions) or compressed (on Android and CrOS through
  [ZRAM](https://source.android.com/devices/tech/perf/low-ram#zram)). The page
  will stay swapped until a new page fault on its virtual address happens, at
  which point the kernel will bring it back in main memory.
* **Not present**: no page fault ever happened on the page or the page was
  clean and later was evicted.

It is generally more important to reduce the amount of _dirty_ memory as that
cannot be reclaimed like _clean_ memory and, on Android, even if swapped in
ZRAM, will still eat part of the system memory budget.
This is why we looked at *Private Dirty* in the `dumpsys meminfo` example.

*Shared* memory can be mapped into more than one process. This means VMAs in
different processes refer to the same physical memory. This typically happens
with file-backed memory of commonly used libraries (e.g., libc.so,
framework.dex) or, more rarely, when a process `fork()`s and a child process
inherits dirty memory from its parent.

This introduces the concept of **PSS** (Proportional Set Size). In **PSS**,
memory that is resident in multiple processes is proportionally attributed to
each of them. If we map one 4KiB page into four processes, each of their
**PSS** will increase by 1KiB.

#### Recap

* Dynamically allocated memory, whether allocated through C's `malloc()`, C++'s
  `operator new()` or Java's `new X()` starts always as _anonymous_ and _dirty_,
  unless it is never used.
* If this memory is not read/written for a while, or in case of memory pressure,
  it gets swapped out on ZRAM and becomes _swapped_.
* Anonymous memory, whether _resident_ (and hence _dirty_) or _swapped_ is
  always a resource hog and should be avoided if unnecessary.
* File-mapped memory comes from code (java or native), libraries and resource
  and is almost always _clean_. Clean memory also erodes the system memory
  budget but typically application developers have less control on it.

## Memory over time

`dumpsys meminfo` is good to get a snapshot of the current memory usage, but
even very short memory spikes can lead to low-memory situations, which will
lead to [LMKs](#lmk). We have two tools to investigate situations like this

* RSS High Watermark.
* Memory tracepoints.

### RSS High Watermark

We can get a lot of information from the `/proc/[pid]/status` file, including
memory information. `VmHWM` shows the maximum RSS usage the process has seen
since it was started. This value is kept updated by the kernel.

```bash
$ adb shell cat '/proc/$(pidof com.android.systemui)/status'
[...]
VmHWM:    256972 kB
VmRSS:    195272 kB
RssAnon:  30184 kB
RssFile:  164420 kB
RssShmem: 668 kB
VmSwap:   43960 kB
[...]
```

### Memory tracepoints

NOTE: For detailed instructions about the memory trace points see the
      [Data sources > Memory > Counters and events](
      /docs/data-sources/memory-counters.md) page.

We can use Perfetto to get information about memory management events from the
kernel.

```bash
$ adb shell perfetto \
  -c - --txt \
  -o /data/misc/perfetto-traces/trace \
<<EOF

buffers: {
    size_kb: 8960
    fill_policy: DISCARD
}
buffers: {
    size_kb: 1280
    fill_policy: DISCARD
}
data_sources: {
    config {
        name: "linux.process_stats"
        target_buffer: 1
        process_stats_config {
            scan_all_processes_on_start: true
        }
    }
}
data_sources: {
    config {
        name: "linux.ftrace"
        ftrace_config {
            ftrace_events: "mm_event/mm_event_record"
            ftrace_events: "kmem/rss_stat"
            ftrace_events: "kmem/ion_heap_grow"
            ftrace_events: "kmem/ion_heap_shrink"
        }
    }
}
duration_ms: 30000

EOF
```

While it is running, take a photo if you are following along.

Pull the file using `adb pull /data/misc/perfetto-traces/trace ~/mem-trace`
and upload to the [Perfetto UI](https://ui.perfetto.dev). This will show
overall stats about system [ION](#ion) usage, and per-process stats to
expand. Scroll down (or Ctrl-F for) to `com.google.android.GoogleCamera` and
expand. This will show a timeline for various memory stats for camera.

![Camera Memory Trace](/docs/images/trace-rss-camera.png)

We can see that around 2/3 into the trace, the memory spiked (in the
mem.rss.anon track). This is where I took a photo. This is a good way to see
how the memory usage of an application reacts to different triggers.

## Which tool to use

If you want to drill down into _anonymous_ memory allocated by Java code,
labeled by `dumpsys meminfo` as `Dalvik Heap`, see the
[Analyzing the java heap](#java-hprof) section.

If you want to drill down into _anonymous_ memory allocated by native code,
labeled by `dumpsys meminfo` as `Native Heap`, see the
[Analyzing the Native Heap](#heapprofd) section. Note that it's frequent to end
up with native memory even if your app doesn't have any C/C++ code. This is
because the implementation of some framework API (e.g. Regex) is internally
implemented through native code.

If you want to drill down into file-mapped memory the best option is to use
`adb shell showmap PID` (on Android) or inspect `/proc/PID/smaps`.


## {#lmk} Low-memory kills

When an Android device becomes low on memory, a daemon called `lmkd` will
start killing processes in order to free up memory. Devices' strategies differ,
but in general processes will be killed in order of descending `oom_score_adj`
score (i.e. background apps and processes first, foreground processes last).

Apps on Android are not killed when switching away from them. They instead
remain *cached* even after the user finishes using them. This is to make
subsequent starts of the app faster. Such apps will generally be killed
first (because they have a higher `oom_score_adj`).

We can collect information about LMKs and `oom_score_adj` using Perfetto.

```protobuf
$ adb shell perfetto \
  -c - --txt \
  -o /data/misc/perfetto-traces/trace \
<<EOF

buffers: {
    size_kb: 8960
    fill_policy: DISCARD
}
buffers: {
    size_kb: 1280
    fill_policy: DISCARD
}
data_sources: {
    config {
        name: "linux.process_stats"
        target_buffer: 1
        process_stats_config {
            scan_all_processes_on_start: true
        }
    }
}
data_sources: {
    config {
        name: "linux.ftrace"
        ftrace_config {
            ftrace_events: "lowmemorykiller/lowmemory_kill"
            ftrace_events: "oom/oom_score_adj_update"
            ftrace_events: "ftrace/print"
            atrace_apps: "lmkd"
        }
    }
}
duration_ms: 60000

EOF
```

Pull the file using `adb pull /data/misc/perfetto-traces/trace ~/oom-trace`
and upload to the [Perfetto UI](https://ui.perfetto.dev).

![OOM Score](/docs/images/oom-score.png)

We can see that the OOM score of Camera gets reduced (making it less likely
to be killed) when it is opened, and gets increased again once it is closed.

## {#heapprofd} Analyzing the Native Heap

**Native Heap Profiles require Android 10.**

NOTE: For detailed instructions about the native heap profiler and
      troubleshooting see the [Data sources > Native heap profiler](
      /docs/data-sources/native-heap-profiler.md) page.

Applications usually get memory through `malloc` or C++'s `new` rather than
directly getting it from the kernel. The allocator makes sure that your memory
is more efficiently handled (i.e. there are not many gaps) and that the
overhead from asking the kernel remains low.

We can log the native allocations and frees that a process does using
*heapprofd*. The resulting profile can be used to attribute memory usage
to particular function callstacks, supporting a mix of both native and Java
code. The profile *will only show allocations done while it was running*, any
allocations done before will not be shown.

### {#capture-profile-native} Capturing the profile

Use the `tools/heap_profile` script to profile a process. If you are having
trouble make sure you are using the [latest version](
https://raw.githubusercontent.com/google/perfetto/master/tools/heap_profile).
See all the arguments using `tools/heap_profile -h`, or use the defaults
and just profile a process (e.g. `system_server`):

```bash
$ tools/heap_profile -n system_server

Profiling active. Press Ctrl+C to terminate.
You may disconnect your device.

Wrote profiles to /tmp/profile-1283e247-2170-4f92-8181-683763e17445 (symlink /tmp/heap_profile-latest)
These can be viewed using pprof. Googlers: head to pprof/ and upload them.
```

When you see *Profiling active*, play around with the phone a bit. When you
are done, press Ctrl-C to end the profile. For this tutorial, I opened a
couple of apps.

### Viewing the data

Then upload the `raw-trace` file from the output directory to the
[Perfetto UI](https://ui.perfetto.dev) and click on diamond marker that
shows.

![Profile Diamond](/docs/images/profile-diamond.png)

The tabs that are available are

* **space**: how many bytes were allocated but not freed at this callstack the
  moment the dump was created.
* **alloc\_space**: how many bytes were allocated (including ones freed at the
  moment of the dump) at this callstack
* **objects**: how many allocations without matching frees were sampled at this
  callstack.
* **alloc\_objects**: how many allocations (including ones with matching frees)
  were sampled at this callstack.

The default view will show you all allocations that were done while the
profile was running but that weren't freed (the **space** tab).

![Native Flamegraph](/docs/images/syssrv-apk-assets-two.png)

We can see that a lot of memory gets allocated in paths through
`ResourceManager.loadApkAssets`. To get the total memory that was allocated
this way, we can enter "loadApkAssets" into the Focus textbox. This will only
show callstacks where some frame matches "loadApkAssets".

![Native Flamegraph with Focus](/docs/images/syssrv-apk-assets-focus.png)

From this we have a clear idea where in the code we have to look. From the
code we can see how that memory is being used and if we actually need all of
it. In this case the key is the `_CompressedAsset` that requires decompressing
into RAM rather than being able to (_cleanly_) memory-map. By not compressing
these data, we can save RAM.

## {#java-hprof} Analyzing the Java Heap

**Java Heap Dumps require Android 11.**

NOTE: For detailed instructions about the Java heap profiler and
      troubleshooting see the [Data sources > Java heap profiler](
      /docs/data-sources/java-heap-profiler.md) page.

### {#capture-profile-java} Capturing the profile
We can get a snapshot of the graph of all the Java objects that constitute the
Java heap. We use the `tools/java_heap_dump` script. If you are having trouble
make sure you are using the [latest version](
https://raw.githubusercontent.com/google/perfetto/master/tools/java_heap_dump).

```bash
$ tools/java_heap_dump -n com.android.systemui

Dumping Java Heap.
Wrote profile to /tmp/tmpup3QrQprofile
This can be viewed using https://ui.perfetto.dev.
```

### Viewing the Data

Upload the trace to the [Perfetto UI](https://ui.perfetto.dev) and click on
diamond marker that shows.

![Profile Diamond](/docs/images/profile-diamond.png)

This will present a flamegraph of the memory attributed to the shortest path
to a garbage-collection root. In general an object is reachable by many paths,
we only show the shortest as that reduces the complexity of the data displayed
and is generally the highest-signal. The rightmost `[merged]` stacks is the
sum of all objects that are too small to be displayed.

![Java Flamegraph](/docs/images/java-flamegraph.png)

The tabs that are available are

* **space**: how many bytes are retained via this path to the GC root.
* **objects**: how many objects are retained via this path to the GC root.

If we want to only see callstacks that have a frame that contains some string,
we can use the Focus feature. If we want to know all allocations that have to
do with notifications, we can put "notification" in the Focus box.

As with native heap profiles, if we want to focus on some specific aspect of the
graph, we can filter by the names of the classes. If we wanted to see everything
that could be caused by notifications, we can put "notification" in the Focus box.

![Java Flamegraph with Focus](/docs/images/java-flamegraph-focus.png)

We aggregate the paths per class name, so if there are multiple objects of the
same type retained by a `java.lang.Object[]`, we will show one element as its
child, as you can see in the leftmost stack above.