executor: always mmap input_data

[blackgnezdo]

mmap'ing over bss is asking for trouble and openbsd delivered them with mimmutable changes.

[blackgnezdo]

I only tested on OpenBSD which (just like linux) defines SYZ_EXECUTOR_USES_SHMEM. It'd be good for at least one of Akaros, Fuchsia, or Windows to test this.

[dvyukov]

I don't see any issues with this. I think this is OK.

We randomize address of output region so that the test program does not corrupt it by guessing its address. But the input region is mapped as PROT_READ only (at least in shmem mode), so the test program shouldn't corrupt it.

There is potential concern of taking the address of the data region and locating input adjacent to the data region (which can affect kernel behavior wrt the data region in some ways). But the data region is already mapped at this point (in os_init). And at least on Linux we surround the data region with protected pages to avoid the effect of having something unpredictable around (thus changing behavior of the kernel in some corner cases). So this should not be an issue as well.

[dvyukov]

The ci failures look related.

[codecov]

Codecov Report

Merging #3466 (ff5833e) into master (4564542) will not change coverage.
The diff coverage is n/a.

[blackgnezdo]

This should now be ready for review.

[mptre]

Is this fixing an actual issue on OpenBSD or just a refactoring/cleanup? __attribute__((section(".openbsd.mutable"))) circumvents mimmutable on OpenBSD and I cannot see anything wrong with the current mmap invocation.

[blackgnezdo]

This was triggered by MAP_FIXED being temporarily broken in OpenBSD. There's now less platform-specific special sauce when we stick to plain mmap instead of mapping over bss that needs to be magically marked to hide it from mimmutable.