[lkml]   [2019]   [Oct]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] Convert filldir[64]() from __put_user() to unsafe_put_user()
On Mon, Oct 07, 2019 at 09:09:14PM -0700, Linus Torvalds wrote:

> > 1) cross-architecture user_access_begin_dont_use(): on everything
> > except x86 it's empty, on x86 - __uaccess_begin_nospec().
> No, just do a proper range check, and use user_access_begin()
> Stop trying to optimize that range check away. It's a couple of fast
> instructions.
> The only ones who don't want the range check are the actual kernel
> copy ones, but they don't want the user_access_begin() either.

Not at the first step. Sure, in the end we want exactly that, and
we want it ASAP. However, the main reason it grows into a tangled
mess that would be over the top for this cycle is the impacts in
arseload of places all over arch/*.

That way we can untangle those. The initial segment that would
allow to use raw_copy_to_user() cleanly in readdir.c
could be done with provably zero impact on anything in arch/*
outside of arch/x86 usercopy-related code.

Moreover, it will be fairly small. And after that the rest can
be done in any order, independent from each other. I want to
kill __copy_... completely, and I believe we'll be able to do
just that in a cycle or two.

Once that is done, the helper disappears along with __copy_...().
And it will be documented as a temporary kludge, don't use
anywhere else, no matter what from the very beginning. For
all the couple of cycles it'll take.

I'm serious about getting rid of __copy_...() in that timeframe.
There's not that much left.

The reason I don't want to do a blanket search-and-replace turning
them all into copy_...() is simply that their use is a good indicator
of code in need of serious beating^Wamount of careful review.

And hell, we might end up doing just that on case-by-case basis.
Often enough we will, by what I'd seen there...

Again, this kluge is only a splitup aid - by the end of the series
it's gone. All it allows is to keep it easier to review.

Note, BTW, that bits and pieces converting a given pointless use
of __copy_...() to copy_...() can be reordered freely at any point
of the sequence - I've already got several. _Some_ of (5) will
be conversions a-la readdir.c one and that has to follow (4), but
most of it won't be like that.

> > void *copy_mount_options(const void __user * data)
> > {
> > unsigned offs, size;
> > char *copy;
> >
> > if (!data)
> > return NULL;
> >
> > copy = kmalloc(PAGE_SIZE, GFP_KERNEL);
> > if (!copy)
> > return ERR_PTR(-ENOMEM);
> >
> > offs = (unsigned long)untagged_addr(data) & (PAGE_SIZE - 1);
> >
> > if (copy_from_user(copy, data, PAGE_SIZE - offs)) {
> > kfree(copy);
> > return ERR_PTR(-EFAULT);
> > }
> > if (offs) {
> > if (copy_from_user(copy, data + PAGE_SIZE - offs, offs))
> > memset(copy + PAGE_SIZE - offs, 0, offs);
> > }
> > return copy;
> > }
> >
> > on the theory that any fault halfway through a page means a race with
> > munmap/mprotect/etc. and we can just pretend we'd lost the race entirely.
> > And to hell with exact_copy_from_user(), byte-by-byte copying, etc.
> Looks reasonable.

OK... BTW, do you agree that the use of access_ok() in
drivers/tty/n_hdlc.c:n_hdlc_tty_read() is wrong? It's used as an early
cutoff, so we don't bother waiting if user has passed an obviously bogus
address. copy_to_user() is used for actual copying there...

There are other places following that pattern and IMO they are all
wrong. Another variety is a half-arsed filter trying to prevent warnings
from too large (and user-controllable) kmalloc() of buffer we'll be
copying to. Which is worth very little, since kmalloc() will scream and
fail well before access_ok() limits will trip. Those need explicit capping
of the size, IMO...

 \ /
  Last update: 2019-10-08 06:58    [W:0.121 / U:6.456 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site