Regarding global error:

From the hid_error documentation:

This function is thread-safe, and error messages are thread-local.

The only backend that tries to implement this correctly is hidraw.
Nevertheless it has an issue too (a memory leak under some conditions).

While I understand your best inentions of preparing this PR, I'd rather want to accept an implementation that strictly follows all requirements as per documentation: uses a thread-local storage of some kind, and correctly releases all the error strings when not required anymore.

Or we need to change initial requirement, to make it easier to implement.

I don't see a perfect solution as of yet.

The device error improvement I'll gladly accept.

Hmm, yeah, that makes sense. I'll alter the PR to remove the global error handling.

I noticed when I tried adding the device error to the windows error that there was some heap corruption on freeing the memory via LocalFree() in free_hid_device(). I couldn't tell if it was introduced by me or the heap corruption was happening somewhere else and my changes exposed it somehow. I debugged in Visual Studios and traced the LocalFree()s on the error string storage and didn't see any double frees anywhere.

We've been having issues with a downstream library (hidapi-rs) where after a HID device reset, everything crashes. We have that library pinned to a certain version which seems to use hidapi 0.11.0 (0ec60c0).

Anyway, I'll remove the global error handling, do some more testing, and push the changes to this PR.

Alright, removed global error messaging. I also had to switch the device specific error message to calloc and free since the Win32 LocalAlloc and LocalFree were doing some heap corruption (even though it's supposed to be thread local).

Yes, these are good recommendations.

The usage of these wide chars makes the concatenation annoying (can't just strcpy to the FormatMessageW buffer). Also, FormatMessageW requires a real buffer to write to in order to return the error string length. Seems I'm going to have to tinker with this more to get it right.

The usage of these wide chars makes the concatenation annoying (can't just strcpy to the FormatMessageW buffer)

I believe we can easily and safely replace the type for op to be const wchar_t*.

requires a real buffer to write to in order to return the error string length

That was a surprise for me. Most WinAPI functions allow passing NULL for buffer to get the required buffer size.

I took liberty to update the implementation on my own