Crash Collectors

For each major class of crash reports, we define a dedicated collector. This is a simple way to encapsulate all related logic in a single module.

When we run crash_reporter, depending on its mode, it simply iterates through all registered collectors.

The code isn‘t a real collector, it’s the base class to hold common logic for all collectors. Similarly, isn‘t a real collector, it’s the base class to hold common logic for all user related collectors.

The core_collector program is just a utility tool and not a collector in the sense of all these. It probably should have used a different naming convention.

Basic Operations

Each collector is designed to generate and queue crash reports. They get uploaded periodically by crash_sender.

Boot Collectors

These are the collectors that run once at boot. They are triggered via the crash-boot-collect.conf init service. They do not, by design, block the boot of the system. They are run in the background as a non-critical service.


This collects Boot Error Record Table (BERT) failures.

The dump collected might be referred to as bertdump.

  • Unhandled firmware errors that occurred in the previous boot are stored in the boot error region.
  • The Kernel ACPI sysfs interface generates the BERT table at /sys/firmware/acpi/tables/BERT and BERT data at /sys/firmware/acpi/tables/data/BERT.
  • During boot, if a BERT report exists, read them and create a report.


This is a meta crash collector: it collects already collected ephemeral crashes into persistent storage. This is useful for handling crash reports in situations where we may not have access to persistent storage (eg. early boot).


This collects EC (Chrome OS Embedded Controller) failures.

The program name is embedded-controller and might be referred to as eccrash.

  • The kernel driver cros_ec_debugfs.c sets up a debugfs path at /sys/kernel/debug/cros_ec/.
  • The driver probes the EC to see if it has any panic logs.
  • If the logs exist, the /sys/kernel/debug/cros_ec/panicinfo is created.
  • During boot, if that file exists, we read it and create a report.


This collects kernel (and BIOS) crashes that caused the system to reboot.

It is built on top of pstore and doesn't support any other data source. We currently support the ramoops and efi backend drivers.

The program name is kernel and might be referred to as kcrash.

  • The BIOS/AP firmware maintain some dedicated space to hold a snippet of the kernel log. They make sure to not clear it during reboot in case there's valid data.
    • For ramoops, CrOS firmware (e.g. coreboot) dedicate a chunk of RAM.
    • For efi, the EFI firmware provides data in its own NVRAM space.
  • While the kernel is running normally, a circular buffer is used to hold the most recent portion of the kernel log buffer (i.e. what printk writes to and what dmesg reads from).
  • When the kernel reboots unexpectedly (e.g. due to a panic, oops, or BUG()), that error message is saved by pstore to the persistent location.
  • If the watchdog reset the system, we won't have an explicit panic message, but we will have the last snippet of the kernel log buffer.
  • During the next boot, the firmware makes sure that space is not reset.
  • While the kernel boots, the pstore driver will check that common space to see if there are any valid records. All valid records are made available via files in /sys/fs/pstore/.
  • During userspace boot, those paths are checked and reports are created.
  • For panics the kernel handled, we'll read the logs from dmesg-ramoops-* & dmesg-efi-*, and generate a report for each one.
  • Stack traces created by the kernel are analyzed to create a stack for the server, as well as generate a hash/fingerprint to correlate other reports.
  • For watchdog resets, we‘ll first query the eventlog (from mosys) to see if the reset was actually due to that. Normally we’d query the watchdog driver directly, but not all platforms are able to support that properly via the kernel driver. We'll create a simpler report using the last snippet of the kernel log from console-ramoops-* and hope the events just before the reset are enough to triage the problem.
  • As records are processed, they get removed from the pstore area.
  • On systems with coreboot BIOS, we also collect the BIOS log. This may be helpful to debug crashes when the kernel interacted with runtime firmware.
  • coreboot maintains a ring buffer (the “CBMEM console”) of log messages in a memory area that is considered reserved by the kernel. The buffer is never erased unless the memory loses power (i.e. if the system fully shuts down), and is usually large enough to hold messages from several prior boots.
  • The collector will search the BIOS log for “banner” strings printed by coreboot on boot to determine where a reboot occured. It will only collect log lines from the boot prior to the current one (i.e. the one that the crash occured it).
  • On arm64 systems, we also attempt to collect runtime firmware (BIOS) crashes. This is done by the kernel collector since runtime firmware mostly does things when requested by the kernel, and errors in runtime firmware are usually triggered by how the kernel calls it. The BIOS generally doesn't log very much after boot, we need both the kernel and BIOS logs to understand the situation of a runtime firmware crash. Since the kernel collector already has the logic to collect both of these, it makes sense for it to handle BIOS crash collection as well.
  • On arm64, the runtime firmware is a piece of code called BL31 from the Arm Trusted Firmware project. BL31 logs crashes by dumping all CPU registers and knows how to append to the coreboot CBMEM console. Since we do not have the infrastructure to generate a full stack trace in firmware, we file these crash reports with a poor man's crash signature that just encodes the address of the program counter where the crash occured.


Collects unclean shutdown events.

  • On every boot, crash_reporter is run. It creates a file (/var/lib/crash_reporter/pending_clean_shutdown) to indicate that the system hasn't gone through a clean shutdown.
  • Upon clean shutdown (chromeos_shutdown), crash_reporter is run with the --clean_shutdown flag. The stateful partition file is removed to indicate the system has gone through a clean shutdown.
  • If during boot, the file already exists before crash_reporter attempts to create it, this signifies that the system hadn't shut down cleanly. A signal is enqueued for metrics_daemon to emit user metrics about this unclean shutdown.
  • No crash reports are otherwise generated for unclean shutdowns since it‘s not clear how we’d triage this in the first place (i.e. what to report).

Runtime Collectors

Here are the collectors that are triggered on demand while the OS is running. They are invoked either by the kernel or by other program.


Collects ARC++ crashes from programs inside the ARC++ container. This handles Android NDK and Java crashes. For Java crashes, it also collects ARCVM's one. It does not handle crashes from ARC++ support daemons that run outside of the container as those are collected like any other userland crash via the main user_collector.

arc_collector shares a lot of code with user_collector so it can overlay ARC++-specific processing details.


Collects crashes of Linux kernel of Android in ARCVM.

When the ARCVM Linux kernel crashes, it dumps logs to /sys/fs/pstore/dmesg-ramoops-0 in ARCVM. It's a pstore file, so the backend exists on Chrome OS as /home/root/<hash>/crosvm/*.pstore. arcvm_kernel_collector receives the content of this file from ArcCrashCollector and ARC bridge via Mojo (or possibly, directly reads the ring buffer in pstore file) and processes it.


Collects crashes of native binaries in ARCVM.

When a native binary crashes, Linux kernel detects the crash and invokes arc-native-crash-dispatcher via /proc/sys/kernel/core_pattern. arc-native-crash-dispatcher calls arc-native-crash-collector32 or arc-native-crash-collector64, and they dump crash file in /data/vendor/arc_native_crash_reports in ARCVM. A Java daemon ArcCrashCollector in ARCVM monitors this directory, and if new files appeared, then sends them to ARC bridge of Chrome browser via Mojo. Dump files are passed as FDs. And finally ARC bridge invokes crash_reporter with the FDs.


Collects Chrome browser crashes. The browser will hand us the minidump directly, so we only attach system metadata and queue it.

crash_reporter will be called by the kernel for Chrome crashes like any other user_collector crash, but we actually ignore these crashes. Chrome is supposed to catch the crash in its parent process and handle it itself; it links in Google Breakpad directly to do so. This is because Chrome is better suited to know what memory regions to ignore (e.g. large heaps or file memory maps or graphics buffers), as well as what metadata to attach (e.g. the last URL visited, whether the process was a renderer, browser, plugin, or other kind of process, chrome://flags, etc...). Otherwise Chrome coredumps can easily consume 3GB+ of memory!

This does mean the system may miss crashes if Chrome's handling itself is buggy.

In much older versions of Chrome OS (sometime before R40), Chrome would not only handle creating its own crash reports, it would also handle uploading them. We changed that behavior because Chrome's uploading is not as robust: it starts uploading immediately, lacks delays/rate limiting, it tries only once, and if it fails at all, it throws away the report entirely. By queueing the report with crash-reporter, it avoids all those problems.


This collects crash/error events triggered by udev events. It is invoked via the udev rules and relies heavily on callbacks in the crash_reporter_logs.conf file.

The program name is udev.

These reports are largely device specific as they try to capture whatever state the device/firmware needs to triage.

TODO: Add devcoredump details if we ever enable them.


Collects all userland crashes where the kernel dumps core. Basically any program that segfaults, aborts, violates a seccomp policy, or is otherwise unceremoniously killed.

  • When a process crashes, the kernel invokes crash_reporter with various important runtime attributes (e.g. the pid, the uid, etc...). The kernel writes a full core dump of the process to stdin.
  • At this point, the failing process is frozen until crash_reporter exits. That means any parent that is monitoring the child won't be notified until we finish processing. This is often a critical path operation if a service needs to be restarted.
    • Chrome reports are ignored normally; see the chrome_collector section for more details as to why.
  • The core2md is run to convert the full coredump to a minidump (.dmp). This process involves reading the core file contents to determine number of threads, register sets of all threads, and threads' stacks' contents. This is fundamental to our out-of-process design.
  • When a crash occurs, we consider the effective user ID of the process which crashed to determine where to save it. If the crashed process was running as chronos, we enqueue its crash to /home/user/<user_hash>/crash/ which is on the user-specific cryptohome when a user is logged in since it might have user PII in it. If the crashed process was running as any other user, we enqueue the crash in /var/spool/crash.
  • The name of the crashing program is used to determine if we should gather additional diagnostic information. crash_reporter_logs.conf contains a list of executables and shell commands to run to gather more details. Any output from them will automatically be attached to the crash report as a .log file.


Used to process crash reports generated inside VMs. This is mostly a wrapper around writing the right collection of files to the right directory, as most useful crash information has to be gathered inside the VM, but it has responsibility for gathering any VM logs stored on the host.

This collector writes to the new /home/root/<user_hash>/crash spool directory, as the daemons that interact directly with VMs to get crash information intentionally don't have the permissions required to access either of the existing spool directories.

Anomaly Detectors

The anomaly_detector service is spawned early during boot via anomaly-detector.conf. It runs a lexer on the syslog stream, and tries to match a set of regexes. A match triggers a collection or a D-Bus signal, depending on the regex.

  • Collection:

    • The matching lines are saved to a temp file under /tmp.
    • crash_reporter is invoked for a specific collector.
    • The temp file is read via stdin.
  • Signal:

    • A D-Bus signal is emitted on a specific service. Other processes may register for delivery of this signal. The service offers no methods.

See sections below for more details on each collector and signal.

The anomaly detector runs one collector at a time, and waits for it to finish running fully before processing more syslog entries.

As a special case, only the first instance of each kernel warning is collected during a session (from boot to shutdown or crash). A count of each warning is reported separately via a sparse UMA histogram.


Collects WARN() messages from anywhere in the depths of the kernel. Could be drivers, subsystems, or core logic.

The program name is kernel-warning or kernel-xxx-warning (where xxx is a common subsystem/area) and might be referred to as kcrash.

  • Whenever the kernel uses WARN() or WARN_ON(...) or any similar helper, it generates a standard log message including stack traces.
  • By default, kernel-warning is used everywhere, but the location of drivers in the backtrace are used to further refine the name.
  • The stack signature uses the same algorithm as the kernel_collector.


Collects SELinux policy violations.

The program name is selinux-violation.

  • Lines from the audit subsystem are processed.
  • Fields from each line are extracted (such as name= and scontext=) and used to create the magic signature.


Collects warnings from the init (e.g. Upstart) for non-ARC services that failed to startup or exited unexpectedly at runtime. This catches syntax errors in the init scripts and daemons that simply exit non-zero but didn't otherwise trigger an abort or crash.

The program name is service-failure.

  • Lines from init: are processed.
  • The standard upstart syntax is: <daemon> <job phase> process (<pid>) terminated with status <status>.
  • All non-normal exits are recorded this way.
  • The signature is constructed from the exit status and service name.


Similar to the above “service failure collector” except that it collects ARC services failures. ARC services are services with names started with “arc-”. A separate ARC services specific collector is created because the ARC services system log messages are kept in a separate file /var/log/arc.log.

The program name is arc-service-failure.

Out-Of-Memory kill signal (OOM kill)

On detection of OOM-kill attempts in the kernel, anomaly_detector sends a D-Bus signal on /org/chromium/AnomalyEventService. This is currently used by memd to collect a number of memory-manager related stats and events.

anomaly_detector does not try to confirm that the kill is successful.