Hey Ted, On Tue, Jun 6, 2017 at 5:00 AM, Theodore Ts'o <tytso@xxxxxxx> wrote: > Note that crypto_rng_reset() is called by big_key_init() in > security/keys/big_key.c as a late_initcall(). So if we are on a > system where the crng doesn't get initialized until during the system > boot scripts, and big_key is compiled directly into the kernel, the > boot could end up deadlocking. > > There may be other instances of where crypto_rng_reset() is called by > an initcall, so big_key_init() may not be an exhaustive enumeration of > potential problems. But this is an example of why the synchronous > API, although definitely much more convenient, can end up being a trap > for the unwary.... Thanks for pointing this out. I'll look more closely into it and see if I can figure out a good way of approaching this. Indeed you're right -- that we have to be really quite careful every time we use the synchronous API. For this reason, I separated things out into the wait_for_random_bytes and then the wrapper around wait_for_random_bytes+get_random_bytes of get_random_bytes_wait. The idea here would be that drivers could place a single wait_for_random_bytes at some userspace entry point -- a configuration ioctl, for example -- and then try to ensure that all calls to get_random_bytes are ordered _after_ that wait_for_random_bytes call. While this pattern doesn't fix all cases of unseeded get_random_bytes calls -- we'll need to do some module loading order cleverness for that, as we discussed in the other thread -- I think this pattern will fix an acceptable amount of call sites, as seen here in this patchset, that it makes it worthwhile. Having it, too, I think would encourage other new drivers to think about when their calls to get_random_bytes happens, and if it's possible for them to defer it until after a userspace-blocking call to wait_for_random_bytes. Anyway, I'll look into and fix up the problem you mentioned. Looking forward to your feedback on the other patches here. Regards, Jason
