* Tetsuo Handa <penguin-ker...@i-love.sakura.ne.jp> wrote:
> > - We should probably separate out a third 'fatal error' variant: for > > example if copying to user-space generates a page fault, then we > > clearly should not pretend that all is fine and return a short read > > even if we made some progress, a -EFAULT is more informative, as we > > might have corrupted (overran) some user buffer on the failed copy as > > well, and ran off the end into the first unmapped user area. > > Is it possible that copy_from_user() corrupts user buffer in a way that > userspace > cannot retry when kernel responded with "there was a short write"? It seems > that > these functions are difficult to return appropriate errors... In the cleanest implementation I believe should be done is to differentiate between 'kernel side errors' and 'user side errors': - 'kernel side errors' are conditions that relate to the layout of kernel memory: some areas might not be readable, and there might be cachability restrictions - or the kernel ran out of memory. In this case it's not user-space's "fault" that they ran into an error and returning a partial read might improve the whole read process, as user-space can decide to continue reading at the last offset that was read - and would also be able to extract all information that can be extracted. - 'user side errors' on the other hand are conditions that are probably a user-space bug: such as trying to read() too much data into a too small buffer, running off the end of it and potentially generating a -EFAULT. In this case the kernel should not return a short read, but escalate the error immediately - bugs are easier to find the earlier the condition is reported. So this is why I think it would make sense to have two error labels: "failure" and "fatal_failure". The "failure" case would return a partial read if possible (and an error if not), the "fatal_failure" would return an error immediately. This is probably a tad over-engineered, but since you asked ... ;-) Thanks, Ingo