The Zircon kernel's internal tracing system can be active on boot (and in fact is currently the default). This means that one can trace at least the kernel side of booting without extra effort. The data is already there, one just needs to collect it.
As long as the kernel's internal trace buffer is not rewound the data is available to be included in the trace. This is achieved by passing category “kernel:retain” to the
traceutil program. Note that the moment a trace is made without passing
kernel:retain then the ktrace buffer is rewound and the data is lost.
# ... Fuchsia boots ... host$ fx traceutil record --categories=kernel,kernel:retain \ --buffer-size=64 --duration=1s --stream
There are a few important things to note here.
The first thing to note is the categories passed:
kernel category tells the kernel to trace everything. In this example the kernel has already been tracing everything: that is the default on boot. It is specified here as a simple way to tell
ktrace_provider, which is the interface between the Fuchsia tracing system and the kernel, that kernel data is being collected. The
kernel:retain category tells
ktrace_provider not to rewind the kernel trace buffer at the start of tracing.
The second is the buffer size. The kernel‘s default trace buffer size is 32MB whereas the Fuchsia trace default buffer size is 4MB. Using a larger Fuchsia trace buffer size means there is enough space to hold the contents of the kernel’s trace buffer. There are some implementation quirks at play here. The kernel currently has its own trace format called “ktrace”. When tracing stops the
ktrace_provider program reads the kernel trace buffer and converts it to Fuchsia‘s trace format. Depending on circumstances the ktrace buffer format is a little more compact. That is why the above example provides a 64MB buffer even though the kernel’s buffer size was 32MB.
The third important thing to note is that in this example we just want to grab the current contents of the trace buffer, and aren't interested in tracing anything more. That is why a duration of one second is used. Ideally we would pass
traceutil currently interprets that as requesting the default which is ten seconds.
--stream arg just says to send the results directly to the development host instead of first writing them to disk on the target first.
The size of the kernel's trace buffer can be changed at boot time with the
ktrace.bufsize=N command line option, where
N is the size of the buffer in megabytes.
The choice of data to collect is controlled with the `ktrace.grpmask=0xNNN' command line option. The 0xNNN value is a bit mask of KTRACE_GRP_* values from //zircon/system/ulib/zircon-internal/include/lib/zircon-internal/ktrace.h. The default is 0xfff which enables all trace categories (or groups in ktrace parlance).
For more information on Zircon command line options see kernel_cmdline.