Re: [PATCH 2/2] Perform check in iov_iter_fault_in_readable() by check_readable_bytes()

[Jamie Lokier <jamie@shareable.org>]

Andi Kleen wrote:
> > ptr = mmap(0, page_size, ....);
> > ...
> > write(fd, ptr + page_size - 256, 512);
> > 
> > Write() will fail here, but it can write first 256 bytes. Previously,
> > all 512 bytes were proceeded, but last 256 bytes were zeroed, and
> > sys_write() returned 256. Not very nice too. 
> 
> Is that really something that users rely on? It looks like a seriously
> broken user program. Which one is that? (just that I can avoid it :)

A few programs set pages read-only, and rely on SIGSEGVs to trigger
mprotect() in the signal handler and thus track dirty pages.

I think the Boehm garbage collector has this option, as do some LISP
interpreters.

System calls don't trigger SIGSEGVs so they can't rely on that when
calling read().  I'm not sure how they handle that.

It would be quite nice if it were safe to call read(), get EFAULT
immediately, or a truncated read() then the next read() gets EFAULT
because it starts at a missing page boundary, and then that's a hint
for the program to consult it's data structures and do it's mprotect()
thing.

Hopefully no programs assume they can do that already, but it would be
nice if they could begin to assume it, instead of checking their data
structure in advance of every read() call.

I don't know of any program which would need the same thing with
write(), but obviously good for symmetry.

-- Jamie