1960418 - Investigate rel.ro sharing on Android

Status: NEW

(GeckoView :: General, enhancement, P2)

Description

[Jeff Muizelaar [:jrmuizel]]

See what Chrome does here: https://chromium.googlesource.com/chromium/src/+/5f632cc8aa4c8a7138e8e6dfed3fc98254f37b89/docs/android_native_libraries.md#relro-sharing

Comment 1

[Markus Stange [:mstange]]

For a Firefox Nightly aarch64 build, readelf --segments libxul.so prints

  Type           Offset   VirtAddr           PhysAddr           FileSiz  MemSiz   Flg Align
[...]
  GNU_RELRO      0x932e640 0x0000000009336640 0x0000000009336640 0x59a828 0x59a9c0 R   0x1

and this segment encompasses the following sections:

  [19] .data.rel.ro      PROGBITS        0000000009336640 932e640 577348 00  WA  0   0 16
  [20] .fini_array       FINI_ARRAY      00000000098ad988 98a5988 000010 00  WA  0   0  8
  [21] .init_array       INIT_ARRAY      00000000098ad998 98a5998 0002b0 00  WA  0   0  8
  [22] .dynamic          DYNAMIC         00000000098adc48 98a5c48 000250 10  WA  9   0  8
  [23] .got              PROGBITS        00000000098ade98 98a5e98 00cb10 00  WA  0   0  8
  [24] .got.plt          PROGBITS        00000000098ba9a8 98b29a8 0164c0 00  WA  0   0  8
  [25] .relro_padding    NOBITS          00000000098d0e68 98c8e68 000198 00  WA  0   0  1

0x59a9c0 is 5876160, so the relro segment is responsible for 5.9MB of extra memory usage per Firefox process.

Comment 2

[Emilio Cobos Álvarez (:emilio)]

I thought that we fork()-without-exec'd on Android for stuff like content processes, so doesn't that give us that sharing for free?

Comment 3

[Jeff Muizelaar [:jrmuizel]]

It would, but we don't have that on Android. That requires the use of Zygote infrastructure that we have in bug 1859857.

Comment 4

[Markus Stange [:mstange]]

Do we understand how this optimization interacts with the App Zygote?

Specifically, does the App Zygote get us relro sharing effectively for free, or is relro sharing beneficial even if we have the App Zygote?

The source code for Chrome's relro sharing is here:

https://source.chromium.org/chromium/chromium/src/+/main:base/android/java/src/org/chromium/base/library_loader/Linker.java;l=31
https://source.chromium.org/chromium/chromium/src/+/main:base/android/linker/linker_jni.h;l=244

I've also found this page:

https://chromium.googlesource.com/chromium/src/+/HEAD/android_webview/docs/how-does-loading-work.md#Loading-native-code-with-RELRO-sharing

This makes me think that they're using relro sharing in order to save memory if the WebView is used by multiple different apps.

So my current thinking is that relro sharing is not needed for the purposes of saving memory in renderer processes of the same app, if you have an app zygote.