On Wed, May 13, 2020 at 12:48:54PM -0700, Linus Torvalds wrote: > Looking at the current users of "probe_kernel_read()", it looks like > it's almost mostly things that just want a single byte or word. > > It's not 100% that: we definitely do several things that want the > "copy" semantics vs the "get" semantics: on the x86 side we have > CALL_INSN_SIZE and MAX_INSN_SIZE, and the ldttss_desc. > > But the bulk of them do seem to be a single value. > > I don't know if performance really matters here, but to me the whole > "most users seem to want to read a single value" is what makes me > think that maybe that should be the primary model, rather than have > the copy model be the primary one and then we implement the single > value case (badly) with a copy. > > It probably doesn't matter that much. I certainly wouldn't hold this > series up over it - it can be a future thing. I can make the get_kernel_nofault implementation suck a little less :) Note that the arch helper (we could call it unsafe_get_kernel_nofault) we still need to have a pagefault_disable / pagefault_enable pair around the calls. So maybe keep the get_kernel_nofault interface as-is (without the goto label), and prepare the arch helpers for being used similar to unsafe_get_user once all architectures are converted. And I can throw in a few patches to convert callers from the copy semantics to the get semantics.
