iOS Crashpad Overview Design

iOS Limitations

Crashpad on other platforms captures exceptions out-of-process. The iOS sandbox, however, restricts applications from delegating work to separate processes. This limitation means Crashpad on iOS must combine the work of the handler and the client into the same process as the main application.

The Crashpad In-Process Handler

In-process handling comes with a number of limitations and difficulties. It is not possible to catch the specific Mach exception EXC_CRASH, so certain groups of crashes cannot be captured. This includes some major ones, like out-of-memory crashes. This also introduces difficulties in capturing all the relevant crash data and writing the minidump, as the process itself is in an unsafe state.

While handling an exception, the handler may not, for example:

  • Allocate memory.
  • Use libc, or most any library call.

While handling an exception, the handler may only:

  • Use audited syscalls.
  • access memory via vm_read.

In conjunction with Crashpad’s existing minidump writer and structural limitations of the minidump format, it is not possible to write a minidump immediately from the crash handler. Instead, an intermediate dump is written when a handler would normally write a minidump (such as during an exception or a forced dump without crashing). The intermediate dump file will be converted to a minidump on the next run (or when the application decides it's safe to do so).

During Crashpad initialization, the handler gathers basic system information and opens a pending intermediate dump adjacent to the Crashpad database.

The Crashpad IntermediateDump Format

Due to the limitations of in-process handling, an intermediate dump file is written during exceptions. The data is streamed to a file, which will be used to generate a final minidump when appropriate.

The file format is similar to binary JSON, supporting keyed properties, maps and arrays.

  • Property [key:int, length:int, value:intarray]
  • StartMap [key:int], followed by repeating Properties until EndMap
  • StartArray [key:int], followed by repeating Maps until EndArray
  • EndMap, EndArray, EndDocument

Similar to JSON, maps can contain other maps, arrays and properties.

The Crashpad In-Process Client

Other Crashpad platforms handle exceptions and upload minidumps out-of-process. On iOS, everything must happen in-process. Once started, the client will automatically handle exceptions and capture the crashed process state in an intermediate dump file. Converting that intermediate dump file into a minidump is likely not safe to do from within a crashed process, and uploading a minidump is definitely unsafe to do at crash time. Applications are expected to process intermediate dumps into pending minidumps and begin processing pending minidumps, possibly for upload, at suitable times following the next application restart.

ProcessIntermediateDumps

For performance and stability reasons applications may choose the correct time to convert intermediate dumps, as well as append metadata to the pending intermediate dumps. This is expected to happen during application startup, when suitable. After converting, a minidump will be written to the Crashpad database, similar to how other platforms write a minidump on exception handling. If uploading is enabled, this minidump will also be immediately uploaded. New intermediate dumps generated by exceptions or by CRASHPAD_SIMULATE_CRASH_AND_DEFER_PROCESSING will not be processed until the next call to ProcessIntermediateDumps. Conversely, CRASHPAD_SIMULATE_CRASH can be called when the client has no performance or stability concerns. In this case, intermediate dumps are automatically converted to minidumps and immediately eligible for uploading.

StartProcessingPendingReports

For similar reasons, applications may choose the correct time to begin uploading pending reports, such as when ideal network conditions exist. By default, clients start with uploading disabled. Applications should call this API when it is determined that it is appropriate to do so (such as on a few seconds after startup, or when network connectivity is appropriate).