On Sat, May 09, 2020 at 05:57:56PM -0700, Linus Torvalds wrote: > On Sat, May 9, 2020 at 5:51 PM Tetsuo Handa > <penguin-ker...@i-love.sakura.ne.jp> wrote: > > > > I think that this access_ok() check helps reducing partial writes (either > > "whole amount was processed" or "not processed at all" unless -ENOMEM). > > No it doesn't. > > "access_ok()" only checks the range being a valid user address range. > > It doesn't actually help at all if the worry is "what if we take a > page fault in the middle". Because it simply doesn't check those > kinds of things. > > Now, if somebody passes actual invalid ranges (ie kernel addresses or > other crazy stuff), they only have themselves to blame. The invalid > range will be noticed when actually doing the user copy, and then > you'll get EFAULT there. But there's no point in trying to figure that > out early - it's only adding overhead, and it doesn't help any normal > case.
It might be a good idea to add Documentation/what-access_ok-does_not ;-/ In addition to what you've mentioned, * access_ok() does not fault anything in; never had. * access_ok() does not verify that memory is readable/writable/there at all; never had, except for genuine 80386 and (maybe) some of the shittier 486 clones. * access_ok() does not protect you from the length being insanely large; even on i386 it can pass with length being a bit under 3Gb. If you count upon it to prevent kmalloc() complaints about insanely large allocation (yes, I've seen that excuse used), you are wrong. * on a bunch of architectures access_ok() never rejects anything, and no, that's _not_ MMU-less ones. sparc64, for example. Or s390, or parisc, etc.
