A zygote process is one that listens for spawn requests from a master process and forks itself in response. Generally they are used because forking a process after some expensive setup has been performed can save time and share extra memory pages.

More specifically, on Linux, it allows to:

  • Amortize the runtime and memory cost of the dynamic loader's relocations, which is respectively ~6 MB and 60 ms/GHz per process. See Appendix A and Appendix B.
  • Amortize the runtime and memory cost for initializing common libraries, such as ICU, NSS, the V8 snapshot and anything else in ContentMainRunnerImpl::Initialize(). With the above, this saves up to ~8 MB per process. See Appendix C.

Security-wise, the Zygote is responsible for setting up and bookkeeping the namespace sandbox.

Furthermore it is the only reasonable way to keep a reference to a binary and a set of shared libraries that can be exec‘ed. In the model used on Windows and Mac, renderers are exec’ed as needed from the chrome binary. However, if the chrome binary, or any of its shared libraries are updated while Chrome is running, we‘ll end up exec’ing the wrong version. A version x browser might be talking to a version y renderer. Our IPC system does not support this (and does not want to!).

So we would like to keep a reference to a binary and its shared libraries and exec from these. However, unless we are going to write our own ld.so, there's no way to do this.

Instead, we exec the prototypical renderer at the beginning of the browser execution. When we need more renderers, we signal this prototypical process (the zygote) to fork itself. The zygote is always the correct version and, by exec'ing one, we make sure the renderers have a different address space randomisation than the browser.

The zygote process is triggered by the --type=zygote command line flag, which causes ZygoteMain (in chrome/browser/zygote_main_linux.cc) to be run. The zygote is launched from content/browser/zygote_host/zygote_host_impl_linux.cc.

Signaling the zygote for a new renderer happens in chrome/browser/child_process_launcher.cc.

You can use the --zygote-cmd-prefix flag to debug the zygote process. If you use --renderer-cmd-prefix then the zygote will be bypassed and renderers will be exec'ed afresh every time.

Appendix A: Runtime impact of relocations

Measured on a Z620:

$ LD_DEBUG=statistics /opt/google/chrome-beta/chrome --help
runtime linker statistics:
  total startup time in dynamic loader: 73899158 clock cycles
    time needed for relocation: 56836478 clock cycles (76.9%)
       number of relocations: 4271
       number of relocations from cache: 11347
       number of relative relocations: 502740
    time needed to load objects: 15789844 clock cycles (21.3%)

56836478 clock cycles -> ~56 ms/GHz

Appendix B: Memory impact of relocations

$ readelf -WS /opt/google/chrome-beta/chrome
[Nr] Name              Type            Address          Off    Size   ES Flg Lk Inf Al
[25] .data.rel.ro      PROGBITS        0000000006a8b590 6a8a590 5b5500 00  WA  0   0 16
Note: 0x5b5500  -> 5.98 MB

Actual impact in terms of memory pages that get shared due to CoW:

$ cat /proc/.../smaps
7fbdd1c81000-7fbdd2233000 r--p 06a5d000 fc:00 665771     /opt/google/chrome-unstable/chrome
Shared_Dirty:       5796 kB

Appendix C: Overall memory impact

$ cat /proc/$PID_OF_ZYGOTE/smaps | grep Shared_Dirty | awk '{TOTAL += $2} END {print TOTAL}'
8092  # KB for dirty pages shared with other processes (mostly forked child processes).