blob: 4c30851ee8f6545f3be3d62b447e989afe32bd27 [file] [log] [blame] [view]
# Adding tracing to device drivers
This document describes how to add tracing to device drivers.
## Overview
Please read [Fuchsia tracing system](/docs/concepts/tracing/README.md)
for an overview of tracing.
## Trace Provider
Drivers don't have to specify a Trace Provider, the devhost process
via `libdriver.so` provides it. It is mentioned here in case the topic
comes up.
## Adding trace records
### Source additions
Trace records are easiest to add by invoking the `TRACE_*()` macros
from `ddk/trace/event.h`.
There are various kinds of trace records that can be emitted.
Please see `trace/internal/event_common.h` for a description
of the various macros.
Looking up macro documentation from internal implementation files
is a temporary situation. Ultimately such documentation will live
in a more appropriate place.
Example:
```c++
#include <ddk/trace/event.h>
void DoSomething(int a, std::string b) {
TRACE_DURATION("example:example1", "DoSomething", "a", a, "b", b);
// Do something
}
```
The first two arguments to most macros are the "category" and the
event name. Here they are "example:example1" and "DoSomething" respectively.
Trace categories are how the tracing system lets the user specify
what data to collect. If a category is not requested by the user
then the data is not collected.
Categories don't need to be unique across the driver.
One typically groups several events under the same category.
By convention categories have the format
"<provider-name>:<category-name>[:<subcategory1-name>...]".
"<provider-name>" for drivers should generally by the driver name.
This is done to avoid collisions in category names across the
entire system. A potential augmentation to this convention is to prefix
all driver categories with "driver:". E.g., "driver:ethernet:packets".
Avoiding naming collisions with other trace providers is important,
otherwise the user may ask for a particular category and get completely
unrelated data from a different trace provider.
The event name is included in the trace to describe what the event
is about. It is typically unique for each event.
### BUILD.gn additions
The following addition to your driver's `BUILD.gn` target is needed to
pick up tracing support:
```gn
driver_module("my_driver") {
deps = [
...
"//zircon/public/lib/trace"
"//zircon/public/lib/trace-driver",
]
}
```
## Building with tracing
The following needs to be passed to fx set in order to trace drivers
that are loaded during boot: `--with-base=//garnet/packages/prod:tracing`.
```sh
$ fx set ${PRODUCT}.${BOARD} --with-base=//garnet/packages/prod:tracing
$ fx build
```
The issue is that without this option then TraceManager won't be present
when the driver starts and thus the driver won't be able to participate
in tracing when TraceManager is started later.
See the documentation for (fx)[../../../docs/development/workflows/fx.md]
or even just the output of `fx help` and especially `fx help set` for further
documentation of running `fx` in general and `fx set` specifically.
## Booting with tracing
To be conservative, tracing uses a kernel command line flag to enable it:
`driver.tracing.enable=1`.
`driver.tracing.enable=1` is the default. To disable participation
of drivers in Fuchsia tracing, boot the kernel with `driver.tracing.enable=0`.
Then boot. See the documentation for your hardware or qemu for instructions
for booting the device. Tracing doesn't require anything special during boot.
## Using tracing
Once the system is booted you can collect traces on the target and
then manually copy them to your development host.
These examples use the category from the source additions described above.
Example:
```sh
fuchsia$ trace record --categories=example:example1,kernel:sched,kernel:meta
host$ fx cp --to-host /data/trace.json trace.json
```
However, it's easier to invoke the `traceutil` program on your development
host and it will copy the files directly to your host and prepare them for
viewing with the Chrome trace viewer.
```sh
host$ fx traceutil record \
--categories=example:example1,kernel:sched,kernel:meta
```
The categories `kernel:sched,kernel:meta` should always be present if you
want to visualize the results. The visualizer wants to associate trace data
with threads and processes, and thus it needs the data provided by the kernel
via these categories.
## Further Reading
See the [Tracing Documentation](/docs/development/tracing/README.md)
for further info.