Crash Events

Crash Events refers to a special subsystem of Gecko that aims to capture events of interest related to process crashing and hanging.

When an event worthy of recording occurs, a file containing that event’s information is written to a well-defined location on the filesystem. The Gecko process periodically scans for produced files and consolidates information into a more unified and efficient backend store.

Crash Event Files

When a crash-related event occurs, a file describing that event is written to a well-defined directory. That directory is likely in the directory of the currently-active profile. However, if a profile is not yet active in the Gecko process, that directory likely resides in the user’s app data directory (UAppData from the directory service).

The filename of the event file is not relevant. However, producers need to choose a filename intelligently to avoid name collisions and race conditions. Since file locking is potentially dangerous at crash time, the convention of generating a UUID and using it as a filename has been adopted.

File Format

All crash event files share the same high-level file format. The format consists of the following fields delimited by a UNIX newline (n) character:

  • String event name (valid UTF-8, but likely ASCII)

  • String representation of integer seconds since UNIX epoch

  • Payload

The payload is event specific and may contain UNIX newline characters. The recommended method for parsing is to split at most 3 times on UNIX newline and then dispatch to an event-specific parsed based on the event name.

If an unknown event type is encountered, the event can safely be ignored until later. This helps ensure that application downgrades (potentially due to elevated crash rate) don’t result in data loss.

The format and semantics of each event type are meant to be constant once that event type is committed to the main Firefox repository. If new metadata needs to be captured or the meaning of data captured in an event changes, that change should be expressed through the invention of a new event type. For this reason, event names are highly recommended to contain a version. e.g. instead of a Gecko process crashed event, we prefer a Gecko process crashed v1 event.

Event Types

Each subsection documents the different types of crash events that may be produced. Each section name corresponds to the first line of the crash event file.

Currently only main process crashes produce event files. Because crashes and hangs in child processes can be easily recorded by the main process, we do not foresee the need for writing event files for child processes, design considerations below notwithstanding.

crash.main.3

This event is produced when the main process crashes.

The payload of this event is delimited by UNIX newlines (n) and contains the following fields:

  • The crash ID string, very likely a UUID

  • One line holding the crash metadata serialized as a JSON string

crash.main.2

This event is produced when the main process crashes.

The payload of this event is delimited by UNIX newlines (n) and contains the following fields:

  • The crash ID string, very likely a UUID

  • 0 or more lines of metadata, each containing one key=value pair of text

This event is obsolete.

crash.main.1

This event is produced when the main process crashes.

The payload of this event is the string crash ID, very likely a UUID. There should be UUID.dmp and UUID.extra files on disk, saved by Breakpad.

This event is obsolete.

crash.submission.1

This event is produced when a crash is submitted.

The payload of this event is delimited by UNIX newlines (n) and contains the following fields:

  • The crash ID string

  • “true” if the submission succeeded or “false” otherwise

  • The remote crash ID string if the submission succeeded

Aggregated Event Log

Crash events are aggregated together into a unified event log. Currently, this log is really a JSON file. However, this is an implementation detail and it could change at any time. The interface to crash data provided by the JavaScript API is the only supported interface.

Design Considerations

There are many considerations influencing the design of this subsystem. We attempt to document them in this section.

Decoupling of Event Files from Final Data Structure

While it is certainly possible for the Gecko process to write directly to the final data structure on disk, there is an intentional decoupling between the production of events and their transition into final storage. Along the same vein, the choice to have events written to multiple files by producers is deliberate.

Some recorded events are written immediately after a process crash. This is a very uncertain time for the host system. There is a high liklihood the system is in an exceptional state, such as memory exhaustion. Therefore, any action taken after crashing needs to be very deliberate about what it does. Excessive memory allocation and certain system calls may cause the system to crash again or the machine’s condition to worsen. This means that the act of recording a crash event must be very light weight. Writing a new file from nothing is very light weight. This is one reason we write separate files.

Another reason we write separate files is because if the main Gecko process itself crashes (as opposed to say a plugin process), the crash reporter (not Gecko) is running and the crash reporter needs to handle the writing of the event info. If this writing is involved (say loading, parsing, updating, and reserializing back to disk), this logic would need to be implemented in both Gecko and the crash reporter or would need to be implemented in such a way that both could use. Neither of these is very practical from a software lifecycle management perspective. It’s much easier to have separate processes write a simple file and to let a single implementation do all the complex work.

Idempotent Event Processing

Processing of event files has been designed such that the result is idempotent regardless of what order those files are processed in. This is not only a good design decision, but it is arguably necessary. While event files are processed in order by file mtime, filesystem times may not have the resolution required for proper sorting. Therefore, processing order is merely an optimistic assumption.

Aggregated Storage Format

Crash events are aggregated into a unified data structure on disk. That data structure is currently LZ4-compressed JSON and is represented by a single file.

The choice of a single JSON file was initially driven by time and complexity concerns. Before changing the format or adding significant amounts of new data, some considerations must be taken into account.

First, in well-behaving installs, crash data should be minimal. Crashes and hangs will be rare and thus the size of the crash data should remain small over time.

The choice of a single JSON file has larger implications as the amount of crash data grows. As new data is accumulated, we need to read and write an entire file to make small updates. LZ4 compression helps reduce I/O. But, there is a potential for unbounded file growth. We establish a limit for the max age of records. Anything older than that limit is pruned. We also establish a daily limit on the number of crashes we will store. All crashes beyond the first N in a day have no payload and are only recorded by the presence of a count. This count ensures we can distinguish between N and 100 * N, which are very different values!