Commit cae2a173fe ("x86: clean up/fix 'copy_in_user()' tail zeroing")
fixed the failure case tail zeroing of one special case of the x86-64
generic user-copy routine, namely when used for the user-to-user case
("copy_in_user()").
But in the process it broke an even more unusual case: using the user
copy routine for kernel-to-kernel copying.
Now, normally kernel-kernel copies are obviously done using memcpy(),
but we have a couple of special cases when we use the user-copy
functions. One is when we pass a kernel buffer to a regular user-buffer
routine, using set_fs(KERNEL_DS). That's a "normal" case, and continued
to work fine, because it never takes any faults (with the possible
exception of a silent and successful vmalloc fault).
But Jan Beulich pointed out another, very unusual, special case: when we
use the user-copy routines not because it's a path that expects a user
pointer, but for a couple of ftrace/kgdb cases that want to do a kernel
copy, but do so using "unsafe" buffers, and use the user-copy routine to
gracefully handle faults. IOW, for probe_kernel_write().
And that broke for the case of a faulting kernel destination, because we
saw the kernel destination and wanted to try to clear the tail of the
buffer. Which doesn't work, since that's what faults.
This only triggers for things like kgdb and ftrace users (eg trying
setting a breakpoint on read-only memory), but it's definitely a bug.
The fix is to not compare against the kernel address start (TASK_SIZE),
but instead use the same limits "access_ok()" uses.
Reported-and-tested-by: Jan Beulich <jbeulich@suse.com>
Cc: stable@vger.kernel.org # 4.0
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The rule for 'copy_from_user()' is that it zeroes the remaining kernel
buffer even when the copy fails halfway, just to make sure that we don't
leave uninitialized kernel memory around. Because even if we check for
errors, some kernel buffers stay around after thge copy (think page
cache).
However, the x86-64 logic for user copies uses a copy_user_generic()
function for all the cases, that set the "zerorest" flag for any fault
on the source buffer. Which meant that it didn't just try to clear the
kernel buffer after a failure in copy_from_user(), it also tried to
clear the destination user buffer for the "copy_in_user()" case.
Not only is that pointless, it also means that the clearing code has to
worry about the tail clearing taking page faults for the user buffer
case. Which is just stupid, since that case shouldn't happen in the
first place.
Get rid of the whole "zerorest" thing entirely, and instead just check
if the destination is in kernel space or not. And then just use
memset() to clear the tail of the kernel buffer if necessary.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The increment of "to" in copy_user_handle_tail() will have incremented
before a failure has been noted. This causes us to skip a byte in the
failure case.
Only do the increment when assured there is no failure.
Signed-off-by: CQ Tang <cq.tang@intel.com>
Link: http://lkml.kernel.org/r/20130318150221.8439.993.stgit@phlsvslse11.ph.intel.com
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Cc: <stable@vger.kernel.org>
When Supervisor Mode Access Prevention (SMAP) is enabled, access to
userspace from the kernel is controlled by the AC flag. To make the
performance of manipulating that flag acceptable, there are two new
instructions, STAC and CLAC, to set and clear it.
This patch adds those instructions, via alternative(), when the SMAP
feature is enabled. It also adds X86_EFLAGS_AC unconditionally to the
SYSCALL entry mask; there is simply no reason to make that one
conditional.
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Link: http://lkml.kernel.org/r/1348256595-29119-9-git-send-email-hpa@linux.intel.com
This throws away the old x86-specific functions in favor of the generic
optimized version.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
This merges the 32- and 64-bit versions of the x86 strncpy_from_user()
by just rewriting it in C rather than the ancient inline asm versions
that used lodsb/stosb and had been duplicated for (trivial) differences
between the 32-bit and 64-bit versions.
While doing that, it also speeds them up by doing the accesses a word at
a time. Finally, the new routines also properly handle the case of
hitting the end of the address space, which we have never done correctly
before (fs/namei.c has a hack around it for that reason).
Despite all these improvements, it actually removes more lines than it
adds, due to the de-duplication. Also, we no longer export (or define)
the legacy __strncpy_from_user() function (that was defined to not do
the user permission checks), since it's not actually used anywhere, and
the user address space checks are built in to the new code.
Other architecture maintainers have been notified that the old hack in
fs/namei.c will be going away in the 3.5 merge window, in case they
copied the x86 approach of being a bit cavalier about the end of the
address space.
Cc: linux-arch@vger.kernel.org
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Peter Anvin" <hpa@zytor.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The discussion about using "access_ok()" in get_user_pages_fast() (see
commit 7f8189068726492950bf1a2dcfd9b51314560abf: "x86: don't use
'access_ok()' as a range check in get_user_pages_fast()" for details and
end result), made us notice that x86-64 was really being very sloppy
about virtual address checking.
So be way more careful and straightforward about masking x86-64 virtual
addresses:
- All the VIRTUAL_MASK* variants now cover half of the address
space, it's not like we can use the full mask on a signed
integer, and the larger mask just invites mistakes when
applying it to either half of the 48-bit address space.
- /proc/kcore's kc_offset_to_vaddr() becomes a lot more
obvious when it transforms a file offset into a
(kernel-half) virtual address.
- Unify/simplify the 32-bit and 64-bit USER_DS definition to
be based on TASK_SIZE_MAX.
This cleanup and more careful/obvious user virtual address checking also
uncovered a buglet in the x86-64 implementation of strnlen_user(): it
would do an "access_ok()" check on the whole potential area, even if the
string itself was much shorter, and thus return an error even for valid
strings. Our sloppy checking had hidden this.
So this fixes 'strnlen_user()' to do this properly, the same way we
already handled user strings in 'strncpy_from_user()'. Namely by just
checking the first byte, and then relying on fault handling for the
rest. That always works, since we impose a guard page that cannot be
mapped at the end of the user space address space (and even if we
didn't, we'd have the address space hole).
Acked-by: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Hugh Dickins <hugh.dickins@tiscali.co.uk>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Impact: fix rare (but currently harmless) miscompile with certain configs and gcc versions
Hugh Dickins noticed that strncpy_from_user() was miscompiled
in some circumstances with gcc 4.3.
Thanks to Hugh's excellent analysis it was easy to track down.
Hugh writes:
> Try building an x86_64 defconfig 2.6.29-rc1 kernel tree,
> except not quite defconfig, switch CONFIG_PREEMPT_NONE=y
> and CONFIG_PREEMPT_VOLUNTARY off (because it expands a
> might_fault() there, which hides the issue): using a
> gcc 4.3.2 (I've checked both openSUSE 11.1 and Fedora 10).
>
> It generates the following:
>
> 0000000000000000 <__strncpy_from_user>:
> 0: 48 89 d1 mov %rdx,%rcx
> 3: 48 85 c9 test %rcx,%rcx
> 6: 74 0e je 16 <__strncpy_from_user+0x16>
> 8: ac lods %ds:(%rsi),%al
> 9: aa stos %al,%es:(%rdi)
> a: 84 c0 test %al,%al
> c: 74 05 je 13 <__strncpy_from_user+0x13>
> e: 48 ff c9 dec %rcx
> 11: 75 f5 jne 8 <__strncpy_from_user+0x8>
> 13: 48 29 c9 sub %rcx,%rcx
> 16: 48 89 c8 mov %rcx,%rax
> 19: c3 retq
>
> Observe that "sub %rcx,%rcx; mov %rcx,%rax", whereas gcc 4.2.1
> (and many other configs) say "sub %rcx,%rdx; mov %rdx,%rax".
> Isn't it returning 0 when it ought to be returning strlen?
The asm constraints for the strncpy_from_user() result were missing an
early clobber, which tells gcc that the last output arguments
are written before all input arguments are read.
Also add more early clobbers in the rest of the file and fix 32-bit
usercopy.c in the same way.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
[ since this API is rarely used and no in-kernel user relies on a 'len'
return value (they only rely on negative return values) this miscompile
was never noticed in the field. But it's worth fixing it nevertheless. ]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
copy_to/from_user and all its variants (except the atomic ones) can take a
page fault and perform non-trivial work like taking mmap_sem and entering
the filesyste/pagecache.
Unfortunately, this often escapes lockdep because a common pattern is to
use it to read in some arguments just set up from userspace, or write data
back to a hot buffer. In those cases, it will be unlikely for page reclaim
to get a window in to cause copy_*_user to fault.
With the new might_lock primitives, add some annotations to x86. I don't
know if I caught all possible faulting points (it's a bit of a maze, and I
didn't really look at 32-bit). But this is a starting point.
Boots and runs OK so far.
Signed-off-by: Nick Piggin <npiggin@suse.de>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Use the _ASM_EXTABLE macro from <asm/asm.h>, instead of open-coding
__ex_table entires in arch/x86/lib/usercopy_64.c.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Linus Torvalds