This is only test but I think we should also destroy key if key generation fails in the middle.

it seems to me that this is the only point in that function for which there is a return instead of goto exit. However this seems ok to me since if psa_generate_key fails then there should be nothing to destroy. Other failures destroy the key. What am I missing?

Yes, that is all true. I'm not sure if we shouldn't also destroy key here. But looking at the description of psa_generate_key:

On success, an identifier for the newly created key.

It seems that current version is correct.

Hmm, I would expect psa_generate_key() to act in an atomic way: either it succeeds, or it leave the key slot unchanged, so that callers don't need to call destroy on failure. However I checked the documentation and it doesn't say explicitly. I checked other uses in the library, and they seem to only call destroy when generate succeeded but an error occurred later. So I think the code is fine as it is, but out of precaution I'll ask on the PSA Crypto API channel.

However I checked the documentation and it doesn't say explicitly.

Well, I apparently didn't get enough sleep last night, because the documentation does say: "key On success, an identifier for the newly created key. PSA_KEY_ID_NULL on failure." So, we don't need to call destroy() on failure here. (It would however be legal, as psa_destroy_key(PSA_KEY_ID_NULL) is guaranteed to do nothing an return success.)

It's a general principle that a PSA function doesn't affect the system state on failure. (It might leave its in-out arguments in a changed state on failure, however: multipart operation functions put the operation object in an error state when they fail.)

psa_generate_key and the other key creation functions are guaranteed to set the key_id output argument to 0 on failure. So it's guaranteed that calling psa_destroy_key(key_id) is always correct, but it's also guaranteed that it's a no-op if the creation function fails.

Well, I should refresh the page more often when replying to comments :) I hadn't seen your replies when I wrote mines.