linux-kernel - Re: [PATCH -next] uaccess: fix __access

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <88ff36b3-558b-9c3f-f21d-5ef05b3227c5@huawei.com>
Date:   Tue, 22 Mar 2022 20:55:56 +0800
From:   "chenjiahao (C)" <chenjiahao16@...wei.com>
To:     Arnd Bergmann <arnd@...db.de>
CC:     "Eric W . Biederman" <ebiederm@...ssion.com>,
        Sam Ravnborg <sam@...nborg.org>,
        Stafford Horne <shorne@...il.com>,
        Dinh Nguyen <dinguyen@...nel.org>,
        linux-arch <linux-arch@...r.kernel.org>,
        "Linux Kernel Mailing List" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH -next] uaccess: fix __access_ok limit setup in compat mode


在 2022/3/18 15:44, Arnd Bergmann 写道:
> On Fri, Mar 18, 2022 at 8:11 AM Chen Jiahao <chenjiahao16@...wei.com> wrote:
>> In __access_ok, TASK_SIZE_MAX is used to check if a memory access
>> is in user address space, but some cases may get omitted in compat
>> mode.
>>
>> For example, a 32-bit testcase calling pread64(fd, buf, -1, 1)
>> and running in x86-64 kernel, the obviously illegal size "-1" will
>> get ignored by __access_ok. Since from the kernel point of view,
>> 32-bit userspace 0xffffffff is within the limit of 64-bit
>> TASK_SIZE_MAX.
>>
>> Replacing the limit TASK_SIZE_MAX with TASK_SIZE in __access_ok
>> will fix the problem above.
> I don't see what problem this fixes, the choice of TASK_SIZE_MAX in
> __access_ok() is intentional, as this means we can use a compile-time
> constant as the limit, which produces better code.
>
> Any user pointer between COMPAT_TASK_SIZE and TASK_SIZE_MAX is
> not accessible by a user process but will not let user space access
> any kernel data either, which is the point of the check.
>
> In your example of using '-1' as the pointer, access_ok() returns true,
> so the kernel can go on to perform an unchecked __get_user() on
> __put_user() on 0xffffffffull, which causes page fault that is intercepted
> by the ex_table fixup.
>
> This should not result in any user visible difference, in both cases
> user process will see a -EFAULT return code from its system call.
> Are you able to come up with a test case that shows an observable
> difference in behavior?
>
>        Arnd
>
> .

Actually, this patch do comes from a testcase failure, the code is 
pasted below:

#define TMPFILE "__1234567890"
#define BUF_SIZE    1024

int main()
{
     char buf[BUF_SIZE] = {0};
     int fd;
     int ret;
     int err;

     fd = open(TMPFILE, O_CREAT | O_RDWR);
     if(-1 == fd)
     {
         perror("open");
         return 1;
     }

     ret = pread64(fd, buf, -1, 1);
     if((-1 == ret) && (EFAULT == errno))
     {
         close(fd);
         unlink(TMPFILE);
         printf("PASS\n");
         return 0;
     }
     err = errno;
     perror("pread64");
     printf("err = %d\n", err);
     close(fd);
     unlink(TMPFILE);
     printf("FAIL\n");

     return 1;
  }

The expected result is:

PASS

but the result of 32-bit testcase running in x86-64 kernel with compat 
mode is:

pread64: Success
err = 0
FAIL


In my explanation, pread64 is called with count '0xffffffffull' and 
offset '1', which might still not trigger

page fault in 64-bit kernel.


This patch uses TASK_SIZE as the addr_limit to performance a stricter 
address check and intercepts

the illegal pointer address from 32-bit userspace at a very early time. 
Which is roughly the same

address limit check as __access_ok in arch/ia64.


This is why this fixes my testcase failure above, or have I missed 
anything else?


Jiahao