[PLAT-7804] Improve crash report writing performance with buffered output

[nickdowell]

Goal

Improve the performance of writing crash reports, to reduce the likelihood of apps being terminated (by watchdog or user) before finishing.

Design

Use write buffers to reduce the number of calls to write().

⚠️ Buffers are allocated on the stack since they cannot be malloc()'d at run-time due to the constrained crash-time environment. Thoughts on the safety of this are welcome - perhaps static allocation should be used instead at the expense of thread safety (which is not something we support in the crash handler.)

Changeset

• Implements write buffering in BSG_KSFile.c.
• Modifies BSG_KSCrashReport.c to use buffered output.
• Modifies BSG_KSCrashState.m to use buffered output.

Testing

Tested by updating unit tests BSG_KSCrashReportTests and KSCrashState_Tests.

Also verified by manually inspecting crash report files and running full E2E suite on CI. Bugs in file writing would manifest as unreported crashes.

The performance unit test case shows improvements of 40-60% depending on the device.

Profiling shows that around 95% of bsg_kscrashreport_writeStandardReport()'s time is now spent in bsg_symbolicate().

[github-actions]

Infer: No issues found 🎉

OCLint: No issues found 🎉

Bugsnag.framework binary size increased by 1,008 bytes from 1,298,728 to 1,299,736

Generated by 🚫 Danger

[kstenerud]

I think we need some isolated tests of BSG_KSFile to check all edge cases...

[kstenerud]

Allocations don't have to all happen on the stack. In a crashed context there will only be one thread that can use it at a time, so a globally pre-allocated buffer will suffice. On a recrash, you just reinit the buffering mechanism since the previous crash handler has been stopped in its tracks.

[nickdowell]

I think we need some isolated tests of BSG_KSFile to check all edge cases...

Added a unit test that covers all the non-error code.