Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Mon, 13 Nov 2017 11:19:38 +0100
From:   Michael Holzheu <holzheu@linux.vnet.ibm.com>
To:     Kees Cook <keescook@chromium.org>
Cc:     Tycho Andersen <tycho@docker.com>, Arnd Bergmann <arnd@arndb.de>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Heiko Carstens <heiko.carstens@de.ibm.com>,
        Martin Schwidefsky <schwidefsky@de.ibm.com>,
        Vasily Gorbik <gor@linux.vnet.ibm.com>
Subject: Re: Does CONFIG_HARDENED_USERCOPY break /dev/mem?
In-Reply-To: <CAGXu5jKrjkYCt1v80qCw4MpfzeHgSqLFZXN7FW5yUqN3XMgeiA@mail.gmail.com>
References: <20171110164529.14db25d6@TP-holzheu>
        <CAGXu5jKrjkYCt1v80qCw4MpfzeHgSqLFZXN7FW5yUqN3XMgeiA@mail.gmail.com>
Organization: IBM
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Message-Id: <20171113111938.6c222a81@TP-holzheu>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Am Fri, 10 Nov 2017 10:46:49 -0800
schrieb Kees Cook <keescook@chromium.org>:

> On Fri, Nov 10, 2017 at 7:45 AM, Michael Holzheu
> <holzheu@linux.vnet.ibm.com> wrote:
> > Hello Kees,
> >
> > When I try to run the crash tool on my s390 live system I get a kernel panic
> > when reading memory within the kernel image:
> >
> >  # uname -a
> >    Linux r3545011 4.14.0-rc8-00066-g1c9dbd4615fd #45 SMP PREEMPT Fri Nov 10 16:16:22 CET 2017 s390x s390x s390x GNU/Linux
> >  # crash /boot/vmlinux-devel /dev/mem
> >  # crash> rd 0x100000
> >
> >  usercopy: kernel memory exposure attempt detected from 0000000000100000 (<kernel text>) (8 bytes)
> >  ------------[ cut here ]------------
> >  kernel BUG at mm/usercopy.c:72!
> >  illegal operation: 0001 ilc:1 [#1] PREEMPT SMP.
> >  Modules linked in:
> >  CPU: 0 PID: 1461 Comm: crash Not tainted 4.14.0-rc8-00066-g1c9dbd4615fd-dirty #46
> >  Hardware name: IBM 2827 H66 706 (z/VM 6.3.0)
> >  task: 000000001ad10100 task.stack: 000000001df78000
> >  Krnl PSW : 0704d00180000000 000000000038165c (__check_object_size+0x164/0x1d0)
> >             R:0 T:1 IO:1 EX:1 Key:0 M:1 W:0 P:0 AS:3 CC:1 PM:0 RI:0 EA:3
> >  Krnl GPRS: 0000000012440e1d 0000000080000000 0000000000000061 00000000001cabc0
> >             00000000001cc6d6 0000000000000000 0000000000cc4ed2 0000000000001000
> >             000003ffc22fdd20 0000000000000008 0000000000100008 0000000000000001
> >             0000000000000008 0000000000100000 0000000000381658 000000001df7bc90
> >  Krnl Code: 000000000038164c: c020004a1c4a        larl    %r2,cc4ee0
> >             0000000000381652: c0e5fff2581b        brasl   %r14,1cc688
> >            #0000000000381658: a7f40001            brc     15,38165a
> >            >000000000038165c: eb42000c000c        srlg    %r4,%r2,12
> >             0000000000381662: eb32001c000c        srlg    %r3,%r2,28
> >             0000000000381668: c0110003ffff        lgfi    %r1,262143
> >             000000000038166e: ec31ff752065        clgrj   %r3,%r1,2,381558
> >             0000000000381674: a7f4ff67            brc     15,381542
> >  Call Trace:
> >  ([<0000000000381658>] __check_object_size+0x160/0x1d0)
> >   [<000000000082263a>] read_mem+0xaa/0x130.
> >   [<0000000000386182>] __vfs_read+0x42/0x168.
> >   [<000000000038632e>] vfs_read+0x86/0x140.
> >   [<0000000000386a26>] SyS_read+0x66/0xc0.
> >   [<0000000000ace6a4>] system_call+0xc4/0x2b0.
> >  INFO: lockdep is turned off.
> >  Last Breaking-Event-Address:
> >   [<0000000000381658>] __check_object_size+0x160/0x1d0
> >
> >  Kernel panic - not syncing: Fatal exception: panic_on_oops
> >
> > With CONFIG_HARDENED_USERCOPY copy_to_user() checks in __check_object_size()
> > if the source address is within the kernel image:
> >
> >  - __check_object_size() -> check_kernel_text_object():
> >
> >  /* Is this address range in the kernel text area? */
> >  static inline const char *check_kernel_text_object(const void *ptr,
> >                                                     unsigned long n)
> >  {
> >          unsigned long textlow = (unsigned long)_stext;
> >          unsigned long texthigh = (unsigned long)_etext;
> >          unsigned long textlow_linear, texthigh_linear;
> >
> >          if (overlaps(ptr, n, textlow, texthigh))
> >                  return "<kernel text>";
> >
> > When the crash tool reads from 0x100000, this check leads to the kernel BUG()
> > in drivers/char/mem.c:
> >
> >  144                 } else {
> >  145                         /*
> >  146                          * On ia64 if a page has been mapped somewhere as
> >  147                          * uncached, then it must also be accessed uncached
> >  148                          * by the kernel or data corruption may occur.
> >  149                          */
> >  150                         ptr = xlate_dev_mem_ptr(p);
> >  151                         if (!ptr)
> >  152                                 return -EFAULT;
> >  153
> >  154                         remaining = copy_to_user(buf, ptr, sz); <<<---- BUG
> >  155
> >  156                         unxlate_dev_mem_ptr(p, ptr);
> >  157                 }
> >
> > Here the reporting function in mm/usercopy.c:
> >
> >  61 static void report_usercopy(const void *ptr, unsigned long len,
> >  62                             bool to_user, const char *type)
> >  63 {
> >  64         pr_emerg("kernel memory %s attempt detected %s %p (%s) (%lu bytes)\n",
> >  65                 to_user ? "exposure" : "overwrite",
> >  66                 to_user ? "from" : "to", ptr, type ? : "unknown", len);
> >  67         /*
> >  68          * For greater effect, it would be nice to do do_group_exit(),
> >  69          * but BUG() actually hooks all the lock-breaking and per-arch
> >  70          * Oops code, so that is used here instead.
> >  71          */
> >  72         BUG();
> >  73 }
> >
> > Shouldn't we skip the kernel address check for /dev/mem - at least when
> > CONFIG_STRICT_DEVMEM is not enabled?
> 
> Some kind of better interaction is needed here, I agree. The prior
> discussions around this basically resulted in declaring that
> HARDENED_USERCOPY without STRICT_DEVMEM didn't make a lot of sense.
> i.e. HARDENED_USERCOPY should maybe require STRICT_DEVMEM, etc. Tycho
> wrote this up after some back-and-forth:
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/kees/linux.git/commit/?h=kspp/kconfig&id=ae98b44ceb338ae165a7f18f29f6244893712da3
> 
> In the end, I was still uncomfortable with it because the usercopy
> protection is wider than just the kernel text, so it seemed strange to
> force it off when not using STRICT_DEVMEM.
> 
> What's the use-case here where you've got hardened usercopy without
> strict devmem?

We use that configuration for development and test. We disabled STRICT_DEVMEM
for debugging the live system with crash. We enabled HARDENED_USERCOPY for
finding user-copy bugs.

Michael


From 1583705926978964627@xxx Fri Nov 10 18:47:47 +0000 2017
X-GM-THRID: 1583694540722198621
X-Gmail-Labels: Inbox,Category Forums,HistoricalUnread