lkml.org 
[lkml]   [2018]   [Apr]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: usercopy whitelist woe in scsi_sense_cache
On Mon, Apr 16, 2018 at 8:12 PM, Kees Cook <keescook@chromium.org> wrote:
> With a hardware watchpoint, I've isolated the corruption to here:
>
> bfq_dispatch_request+0x2be/0x1610:
> __bfq_dispatch_request at block/bfq-iosched.c:3902
> 3900 if (rq) {
> 3901 inc_in_driver_start_rq:
> 3902 bfqd->rq_in_driver++;
> 3903 start_rq:
> 3904 rq->rq_flags |= RQF_STARTED;
> 3905 }

This state continues to appear to be "impossible".

[ 68.845979] watchpoint triggered
[ 68.846462] sense before:ffff8b8f9f6aae00 after:ffff8b8f9f6aae01
[ 68.847196] elevator before:ffff8b8f9a2c2000 after:ffff8b8f9a2c2000
[ 68.847905] elevator_data before:ffff8b8f9a2c0400 after:ffff8b8f9a2c0400
...
[ 68.856925] RIP: 0010:bfq_dispatch_request+0x99/0xad0
[ 68.857553] RSP: 0018:ffff900280c63a58 EFLAGS: 00000082
[ 68.858253] RAX: ffff8b8f9aefbe80 RBX: ffff8b8f9a2c0400 RCX: ffff8b8f9a2c0408
[ 68.859201] RDX: ffff8b8f9a2c0408 RSI: ffff900280c63b34 RDI: 0000000000000001
[ 68.860147] RBP: 0000000000000000 R08: 0000000f00000204 R09: 0000000000000000
[ 68.861122] R10: ffff900280c63af0 R11: 0000000000000040 R12: ffff8b8f9aefbe00
[ 68.862089] R13: ffff8b8f9a221950 R14: 0000000000000000 R15: ffff8b8f9a2c0770

Here we can see that sense buffer has, as we've seen, been
incremented. However, the "before" values for elevator and
elevator_data still match their expected values. As such, this should
be "impossible", since:

static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
{
struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
...
if (rq) {
inc_in_driver_start_rq:
bfqd->rq_in_driver++;
start_rq:
rq->rq_flags |= RQF_STARTED;
}
exit:
return rq;
}

For rq_in_driver++ to touch sense, bfqd must be equal to scsi_request
- 12 bytes (rq_in_driver is 60 byte offset from struct bfq_data, and
sense is 48 bytes offset from struct scsi_request).

The above bfq_dispatch_request+0x99/0xad0 is still
__bfq_dispatch_request at block/bfq-iosched.c:3902, just with KASAN
removed. 0x99 is 153 decimal:

(gdb) disass bfq_dispatch_request
Dump of assembler code for function bfq_dispatch_request:
...
0xffffffff8134b2ad <+141>: test %rax,%rax
0xffffffff8134b2b0 <+144>: je 0xffffffff8134b2bd
<bfq_dispatch_request+157>
0xffffffff8134b2b2 <+146>: addl $0x1,0x100(%rax)
0xffffffff8134b2b9 <+153>: addl $0x1,0x3c(%rbx)
0xffffffff8134b2bd <+157>: orl $0x2,0x18(%r12)
0xffffffff8134b2c3 <+163>: test %ebp,%ebp
0xffffffff8134b2c5 <+165>: je 0xffffffff8134b2ce
<bfq_dispatch_request+174>
0xffffffff8134b2c7 <+167>: mov 0x108(%r14),%rax
0xffffffff8134b2ce <+174>: mov %r15,%rdi
0xffffffff8134b2d1 <+177>: callq 0xffffffff81706f90 <_raw_spin_unlock_irq>

Just as a sanity-check, at +157 %r12 should be rq, rq_flags is 0x18
offset from, $0x2 is RQF_STARTED, so that maps to "rq->rq_flags |=
RQF_STARTED", the next C statement. I don't know what +146 is, though?
An increment of something 256 bytes offset? There's a lot of inline
fun and reordering happening here, so I'm ignoring that for the
moment.

So at +153 %rbx should be bfqd (i.e. elevator_data), but this is the
tripped instruction. The watchpoint dump shows RBX as ffff8b8f9a2c0400
... which matches elevator_data.

So, what can this be? Some sort of cache issue? By the time the
watchpoint handler captures the register information, it's already
masked the problem? I'm stumped again. :(

-Kees

--
Kees Cook
Pixel Security

\
 
 \ /
  Last update: 2018-04-17 22:04    [W:0.067 / U:1.120 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site