This is the project for handling all crash related operations on the device. The intention is to be as low-overhead as possible while still maximizing the usefulness of crash reports and minimizing collection of any & all PII data.
For more background details, see the original design doc. This document focuses on the current state of the project.
Most bugs/features can be found via component:Internals>CrashReporting OS=Chrome.
No crashes get collected without explicit consent from the device owner. This is normally part of the OOBE setup flow. Consent is handled by the device owner and covers all users of that device. If no consent has been granted, then generally collectors will exit rather than do any further processing.
However, even if consent has been granted, there is one notable scenario where crash reports are still not uploaded -- guest mode. While system crashes are still collected (as they shouldn‘t have any user PII in them), browser and user crashes are not uploaded. This leads to a bit of a chicken & egg where real crashes in guest mode are not caught and uploaded for tracking, but we’re willing to accept that for the user privacy guarantees.
If consent is later revoked, we do not upload any crashes that had been queued.
For headless systems, the consent setting can be managed via metrics_client's
Setting up the crash reporter early in the boot process is tricky: before the encrypted stateful mount is setup, crash reporter does not have access to persistent storage to store crashes in. We run [crash-reporter-early-init.conf] early in the boot process to configure the kernel to pass crashes to us with an “early” flag to use a tmpfs path instead of a persistent location for storing crash reports.
After the encrypted stateful mount setup is complete, we run crash-reporter.conf which sets the crash path back to the persistent storage crash directory and initializes any system state that depends on it.
Then during boot, but not in the critical path, we run crash-boot-collect.conf to gather up any previous crashes in the system. For example, if the system had rebooted due to a kernel panic, or some firmware (like the EC) had crashed. This also includes crashes that occurred during early boot before persistent storage was available: such crashes are stored into
/run/crash_reporter/crash and collected into
/var/spool/crash once boot completes.
We background the anomaly_detector to monitor for system “anomalies” for which a notification mechanism is not available. This daemon operates by monitoring syslog messages via inotify. Depending on the anomaly, it generates crashes, UMA stats, and D-Bus signals.
At this point, if no crashes occur, nothing ever happens! And of course, since we never have bugs, that's precisely what happens!
If a userland crash occurs that would trigger a coredump, the kernel will execute crash_reporter directly and pass in the coredump and some metadata. We process it directly and queue the report in the right crash directory. The exact processing steps depend on the specific collector involved, so see the Collectors section for more details. See Crash Report Storage for more details on queuing.
If a Chrome (browser) process crashes, crash_reporter will be called by the kernel, but we actually ignore things at that point. Instead, Chrome itself will take care of processing that crash report before calling crash_reporter with the final crash report for queuing.
Each collector has a
All unittests are named
xxx_test.cc and contain tests for the corresponding
All init related scripts live in the init/ directory.
The core_collector program is kept separate because it‘s otherwise completely standalone and doesn’t share any code.
That covers pretty much all the files in here.
Here is a brief summary of each init script we provide. More details on each can be found in the sections below.
This section is triggered via the crash-reporter.conf init script.
We want to initialize boot collection as early as possible so as to minimize the time frame for crashes that we'd miss otherwise. Currently we initialize during
boot-services and after
syslog. This is because the crash_reporter code will use syslog for output/errors of its own, even during early init.
This means any userland crashes that occur before this point will be lost. Some non-exhaustive but significant highlights:
boot-servicesthat run in parallel (e.g.
The init step itself should be fairly quick as it only initializes internal state paths (e.g. under
/var), and it configures the various
/sys paths that the kernel has setup. Most notably, this is responsible for writing
/proc/sys/kernel/core_pattern which tells the kernel to execute us whenever a crash is detected.
For more details, see the core(5) man page.
/run. This would be active until we do the full init step described here. The reports would then be collected by the crash-boot-collect.conf step.
We store reports in a couple of different places.
/var/spool/crash/: Non-user (i.e. system) queued reports.
/home/chronos/<user_hash>/crash/: User-specific queued reports. Used when invoked as the user (e.g. by Chrome as
/home/root/<user_hash>/crash-reporter/: User-specific queued reports. Used when invoked as root (e.g. by the kernel to handle crashes for processes that were running as
/home/root/...support is in progress. See https://chromium-review.googlesource.com/1155107 for more details.
These directories are only written to by the crash-reporter at collection time. The crash_sender program will also delete reports after it uploads them.
These directories are only read by the crash_sender program for uploading, and by feedback reports. Test frameworks (e.g. autotest) also offload crash reports, but we don't need to provide special consideration for those.
We enforce a limit of about 32 crashes per spool directory. This is to avoid filling up the underlying storage especially if a daemon “goes crazy” and generates a lot of crashes quickly.
Every collector produces a unique name of sorts which is then looked up in the crash_reporter_logs.conf file. For example, user_collector uses the program's base file name, while udev_collector constructs a more complicated format to make it easier to match unique events.
The value of that setting is an arbitrary shell script which is then executed. The output is captured and automatically attached to the corresponding report.
This is most useful to collect data from device or program specific paths that otherwise would not be collected. For example, for kernel wifi driver warnings, we'll dump some of the low level PCI register state for those specific pieces of hardware.
The output of this snippet should be kept small as there are size limits to report uploads, and to minimize PII data leaking in.
Metric collection is currently being evaluated for their usefulness. As it stands currently, metrics are still in the codebase, but ultimately disabled just before being collected.
See https://crbug.com/754850 for more details.
If we re-enable the metrics, then we‘ll fill out this section of the doc. If we delete the code entirely, then we’ll delete this section too.
The crash_sender program is responsible for uploading reports to the server. It tries to be unobtrusive in a number of ways:
While there are multiple local crash queues, these are global limits.
Before we upload any report, we check to see if consent is still granted. This way we stop uploading reports right away if things change.
We also check if the network is online before uploading a crash report and stop uploading reports until the next run of crash_sender if we are offline.
If an upload fails for any reason (flaky network, system going to sleep, etc...), then the report is left in the local queue and retried later.
We attempt to upload the oldest reports first to minimize them going stale.
Network proxies are respected; see the Proxies section for more details.
After a report has been successfully uploaded, we add details to the
/var/log/chrome/Crash Reports/uploads.log file. This is used by Chrome itself in its internal
chrome://crashes page to provide crash information to the user. The exact format of this file is controlled by Chrome. Here are just some of the files involved:
uploads.loglags behind other platforms. Once crash_sender is converted to C++, we can re-evaluate improving this. See https://crbug.com/275910 for more details.
At runtime, there are a few different ways one can trigger uploading. Since crash_sender has internal locking, you don't have to worry about any of these methods clobbering or racing with any other crash component.
upload_crashescommand to trigger immediately.
chrome://crashespage has a “Start Uploading” link.
Once these crashes are uploaded, where do they go you might wonder? By default, they're sent to the Google Crash Server. For obvious privacy reasons, this server is only accessible to Google employees, and even then not all employees have access, only ones that need it to debug.
This is where the [list_proxies] tool comes in. It's a simple front end to sending queries to Chrome via D-Bus and then printing the results to stdout. This info can then be exported or passed to tools like
For Chromeless devices, this presents a bit of a challenge.
We won't cover (in depth) files covered by these topics:
Otherwise, this should be an exhaustive list of paths on the filesystem that this project uses in any fashion.
These paths are guaranteed to persist across boots.
/var/lib/crash_reporter/: Non-volatile state we need across reboots. Currently we use to cache previous OS details so we can correctly attribute previous kernel crashes to the previous OS version that was running.
/var/lib/crash_reporter/pending_clean_shutdown: Used by the unclean_shutdown_collector.
/var/lib/crash_sender_paused: Used by autotests to pause crash_sender. This should be moved to
/var/lib/crash_sender/: Non-volatile state that crash_sender maintains. Used to keep track of how many reports have been uploaded (and when) so we can regulate our limits. Do not add any additional files to this directory or it will break timestamp calculations. Add additional state information to the ‘state’ subdirectory instead. Currently we store a ‘client_id’ in the ‘state’ subdirectory for maintaining a persistent device identifier for coalescing crash reports by device. This ID should never be used for any other purpose.
/var/log/messages: anomaly_detector monitors this (read-only).
These spool dirs are covered in detail in Crash Report Storage.
/var/spool/crash/: System crash reports.
/home/chronos/<user_hash>/crash/: User-specific queued reports.
/home/root/<user_hash>/crash-reporter/: User-specific queued reports.
These paths are guaranteed to be reset at every boot, so we only store active runtime details here.
/run/crash_reporter/: Used by all crash-reporter tools (i.e. both crash_reporter and crash_sender) for runtime state. This should be moved to
/run/crash-reporter/as the project name.
/run/crash_reporter/crash-test-in-progress: Used by autotests to tell tools are being exercised by autotest and to adjust behavior accordingly.
/run/crash_reporter/mock-crash-sending: Used by autotests to tell crash_sender to mock out its behavior for testing purposes.
/run/lock/crash_sender: Used by crash_sender to guarantee only one upload instance is active at a time.
These are used to communicate with metrics_daemon.
/run/kernel-crash-detected: Used by crash_reporter to signal the metrics_daemon that a kernel crash occurred.
/run/unclean-shutdown-detected: Used by crash_reporter to signal the metrics_daemon that an unclean shutdown occurred.
/runfiles was done historically because the crash-reporter.conf init and crash-boot-collect.conf were one script that always executed early and before metrics_daemon started. However, now we can make crash-boot-collect.conf wait for metrics_daemon, we should be able to switch to its more standard existing IPC methods.
/tmp/ usage by tools would normally be quite concerning due to the ease of predictable file names, but crash_reporter and crash_sender both execute in a minijail with a unique empty
/tmp/ mount (and automatically freed on exit). That mount is visible only to the active process (and its children), so other processes in the system won‘t be able to inject content. This doesn’t mean the
/tmp/ usage shouldn't be improved, but it at least mitigates collisions (accidental or otherwise).
/tmp/crash_reporter/<pid>/: crash_reporter writes intermediate data here for the
<pid>that just crashed.
/tmp/crash_sender.XXXXXX/: crash_sender holds intermediate upload report status here and is cleared after every run.
/tmp/: core_collector writes intermediate
/procfiles directly here.
Here's any other random paths we handle.
/root/.leave_core: Normally crash_reporter will delete the large coredumps after converting them to minidumps. When this file exists, coredumps are saved in the spool dir too. Created on dev images automatically. This should be moved to
/etc/crash-reporter/leave-core, and duplicated in
/run/crash-reporter/so people don't have to modify the rootfs.
/mnt/stateful_partition/etc/collect_chrome_crashes: Used by autotests to let crash_reporter collect browser crashes directly (normally it ignores them and lets Chrome handle things). This should be moved to
Each crash report has a
.meta file with a unique basename. The
.meta file is a simple key-value store read by crash_sender. Each collector will generate this, and any supplemental files will use that same basename with other extensions. e.g. The main crash image will be a minidump but with a
.dmp extension. Supplemental log files will have a
However, crash_sender doesn't care about that. The
.meta file will have keys that point to those files directly.
The uploaded crashes are only minidumps (and related metadata). Actual stack walking and symbolification happens on the server. The symbols themselves are generated by CrOS builders by running cros_generate_breakpad_symbols in chromite on all the compiled programs for the board. This produces
.sym files which are used with Google Breakpad. All those files are then uploaded using upload_symbols in chromite.
This means the
.debug files generated at build time are used to generate all the
.sym files, but nothing after that.
Since these tools often run as root and work with untrusted data, security is a big concern. Check out security.md for more details.
For tips for hacking on this project, check out hacking.md.