On Tue, Jan 16, 2018 at 10:28 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> On Tue, Jan 16, 2018 at 08:30:17PM -0800, Dan Williams wrote:
>> On Tue, Jan 16, 2018 at 2:23 PM, Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
>> > On Sat, Jan 13, 2018 at 11:33 AM, Linus Torvalds
>> [..]
>> > I'll respin this set along those lines, and drop the ifence bits.
>>
>> So now I'm not so sure. Yes, get_user_{1,2,4,8} can mask the pointer
>> with the address limit result, but this doesn't work for the
>> access_ok() + __get_user() case. We can either change the access_ok()
>> calling convention to return a properly masked pointer to be used in
>> subsequent calls to __get_user(), or go with lfence on every
>> __get_user call. There seem to be several drivers that open code
>> copy_from_user() with __get_user loops, so the 'fence every
>> __get_user' approach might have noticeable overhead. On the other hand
>> the access_ok conversion, while it could be scripted with coccinelle,
>> is ~300 sites (VERIFY_READ), if you're concerned about having
>> something small to merge for 4.15.
>>
>> I think the access_ok() conversion to return a speculation sanitized
>> pointer or NULL is the way to go unless I'm missing something simpler.
>> Other ideas?
>
> What masked pointer?
The pointer value that is masked under speculation.
diff --git a/arch/x86/lib/getuser.S b/arch/x86/lib/getuser.S
index c97d935a29e8..4c378b485399 100644
--- a/arch/x86/lib/getuser.S
+++ b/arch/x86/lib/getuser.S
@@ -40,6 +40,8 @@ ENTRY(__get_user_1)
mov PER_CPU_VAR(current_task), %_ASM_DX
cmp TASK_addr_limit(%_ASM_DX),%_ASM_AX
jae bad_get_user
+ sbb %_ASM_DX,%_ASM_DX
+ and %_ASM_DX,%_ASM_AX
ASM_STAC
1: movzbl (%_ASM_AX),%edx
xor %eax,%eax
...i.e %_ASM_AX is guaranteed to be zero if userspace tries to cause
speculation with an address above the limit. The proposal is make
access_ok do that same masking so we never speculate on pointers from
userspace aimed at kernel memory.
> access_ok() exists for other architectures as well,
I'd modify those as well...
> and the fewer callers remain outside of arch/*, the better.
>
> Anything that open-codes copy_from_user() that way is *ALREADY* fucked if
> it cares about the overhead - recent x86 boxen will have slowdown from
> hell on stac()/clac() pairs. Anything like that on a hot path is already
> deep in trouble and needs to be found and fixed. What drivers would those
> be?
So I took a closer look and the pattern is not copy_from_user it's
more like __get_user + write-to-hardware loops. If the performance is
already expected to be bad for those then perhaps an lfence each loop
iteration won't be much worse. It's still a waste because the lfence
is only needed once after the access_ok.
> We don't have that many __get_user() users left outside of arch/*
> anymore...
