I've been struggling with this problem as well, but I've managed to seemingly fix it.
1. Perform all your RequestIntoNativeArrays that you want (I do all mine at the end of the frame)
2. Call AsyncGPUReadback.WaitAllRequests();

Things that will cause the problem (with these it fails every 2nd frame for me):
1. Trying to access the NativeArray between the RequestIntoNativeArray and AsyncGPUReadback.WaitAllRequests
2. Calling AsyncGPUReadback.WaitAllRequests when you haven't submitted any new requests that frame. It seems it mistakenly tries to re-access a previous request or something

Errors are still thrown Unity 2021.3 Editor, but not in builds according to my tests. Do note that the readback is actually completed without issue, and only errors are thrown.

Still broken.

AsyncGPUReadback - NativeArray does not have read/write access

AsyncGPUReadback - NativeArray should not be undisposable

Still observed in 2021.3.31

Claimed to have been fixed in issue #1295472 , YMMV.

Please fix this bug, so I can use it for occlusion culling with depth texture.

Am I correct in my understanding that this bug means there is essentially no reason to use RequestIntoNativeArray rather than Request? Both methods seem to force you to allocate a new array every time you have a new request.

Thanks for the quote, Gravenstein! I didn't see this issue on the tracker (looked at a couple of similar ones). There is nothing really more to reproducing it other than those steps, indeed. The only mention to be had here is that we are not getting the error in the issue but "AsyncGPUReadback - NativeArray does not have read/write access" and "AsyncGPUReadback - NativeArray should not be undisposable" respectively (both thrown for the same call).

Please, fix it, this bug is like three years old.

A more detailed guide to reproduce it. It was written by Hypergeanu on the forums, but we are having the exact same issue in the same way.

- Create persistent NativeArray (only once);
- Call AsyncGPUReadback.RequestIntoNativeArray;
- Wait for completion;
- Create and schedule a job that reads from the array (after that gpu request is done);
- Run the job, call complete on the job handle (the job transfers the data from the nativearray into a different one, so the usage cycle of the nativearray used for gpu request ends here).

The sequence above executes with no errors for the first pass, but when it is called again (the buffer stays the same, its purpose is to perform these requests and job processing repeatedly using the pre-allocated memory), the specified errors are thrown when the array is touched by the gpu async request.