Re: UBSAN: array-index-out-of-bounds in arch_uprobe_analyze_insn

Re: UBSAN: array-index-out-of-bounds in arch_uprobe_analyze_insn

[syzbot]

syzbot has bisected this issue to:

commit 4b2bd5fec007a4fd3fc82474b9199af25013de4c
Author: John Stultz <john.stultz@linaro.org>
Date:   Sat Oct 8 00:02:33 2016 +0000

    proc: fix timerslack_ns CAP_SYS_NICE check when adjusting self

bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=1697348d900000
start commit:   325d0eab Merge branch 'akpm' (patches from Andrew)
git tree:       upstream
final oops:     https://syzkaller.appspot.com/x/report.txt?x=1597348d900000
console output: https://syzkaller.appspot.com/x/log.txt?x=1197348d900000
kernel config:  https://syzkaller.appspot.com/x/.config?x=b12e84189082991c
dashboard link: https://syzkaller.appspot.com/bug?extid=9b64b619f10f19d19a7c
syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=1573a8ad900000
C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=164ee6c5900000

Reported-by: syzbot+9b64b619f10f19d19a7c@syzkaller.appspotmail.com
Fixes: 4b2bd5fec007 ("proc: fix timerslack_ns CAP_SYS_NICE check when adjusting self")

For information about bisection process see: https://goo.gl/tpsmEJ#bisection

Re: UBSAN: array-index-out-of-bounds in arch_uprobe_analyze_insn

[Kees Cook]

Hi,

There appears to be a problem with prefix counting for the instruction
decoder. It looks like insn_get_prefixes() isn't keeping "nb" and "nbytes"
in sync correctly:

        while (inat_is_legacy_prefix(attr)) {
                /* Skip if same prefix */
                for (i = 0; i < nb; i++)
                        if (prefixes->bytes[i] == b)
                                goto found;
                if (nb == 4)
                        /* Invalid instruction */
                        break;
                prefixes->bytes[nb++] = b;
		...
found:
                prefixes->nbytes++;
                insn->next_byte++;
                lb = b;
                b = peek_next(insn_byte_t, insn);
                attr = inat_get_opcode_attribute(b);
        }

(nbytes is incremented on repeated prefixes, but "nb" isn't)

However, it looks like nbytes is used as an offset:

static inline int insn_offset_rex_prefix(struct insn *insn)
{
        return insn->prefixes.nbytes;
}
static inline int insn_offset_vex_prefix(struct insn *insn)
{
        return insn_offset_rex_prefix(insn) + insn->rex_prefix.nbytes;
}

Which means everything that iterates over prefixes.bytes[] is buggy,
since they may be trying to read past the end of the array:

$ git grep -A3 -E '< .*prefixes(\.|->)nbytes'
boot/compressed/sev-es.c:       for (i = 0; i < insn->prefixes.nbytes; i++) {
boot/compressed/sev-es.c-               insn_byte_t p =
insn->prefixes.bytes[i];
boot/compressed/sev-es.c-
boot/compressed/sev-es.c-               if (p == 0xf2 || p == 0xf3)
--
kernel/uprobes.c:       for (i = 0; i < insn->prefixes.nbytes; i++) {
kernel/uprobes.c-               insn_attr_t attr;
kernel/uprobes.c-
kernel/uprobes.c-               attr = inat_get_opcode_attribute(insn->prefixes.bytes[i]);
--
kernel/uprobes.c:       for (i = 0; i < insn->prefixes.nbytes; i++) {
kernel/uprobes.c-               if (insn->prefixes.bytes[i] == 0x66)
kernel/uprobes.c-                       return -ENOTSUPP;
kernel/uprobes.c-       }
--
lib/insn-eval.c:        for (i = 0; i < insn->prefixes.nbytes; i++) {
lib/insn-eval.c-                insn_byte_t p = insn->prefixes.bytes[i];
lib/insn-eval.c-
lib/insn-eval.c-                if (p == 0xf2 || p == 0xf3)
--
lib/insn-eval.c:        for (i = 0; i < insn->prefixes.nbytes; i++) {
lib/insn-eval.c-                insn_attr_t attr;
lib/insn-eval.c-
lib/insn-eval.c-                attr = inat_get_opcode_attribute(insn->prefixes.bytes[i]);

I don't see a clear way to fix this.

-Kees

On Mon, Sep 21, 2020 at 09:20:07PM -0700, syzbot wrote:
> syzbot has bisected this issue to:
> 
> commit 4b2bd5fec007a4fd3fc82474b9199af25013de4c
> Author: John Stultz <john.stultz@linaro.org>
> Date:   Sat Oct 8 00:02:33 2016 +0000
> 
>     proc: fix timerslack_ns CAP_SYS_NICE check when adjusting self
> 
> bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=1697348d900000
> start commit:   325d0eab Merge branch 'akpm' (patches from Andrew)
> git tree:       upstream
> final oops:     https://syzkaller.appspot.com/x/report.txt?x=1597348d900000
> console output: https://syzkaller.appspot.com/x/log.txt?x=1197348d900000
> kernel config:  https://syzkaller.appspot.com/x/.config?x=b12e84189082991c
> dashboard link: https://syzkaller.appspot.com/bug?extid=9b64b619f10f19d19a7c
> syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=1573a8ad900000
> C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=164ee6c5900000
> 
> Reported-by: syzbot+9b64b619f10f19d19a7c@syzkaller.appspotmail.com
> Fixes: 4b2bd5fec007 ("proc: fix timerslack_ns CAP_SYS_NICE check when adjusting self")
> 
> For information about bisection process see: https://goo.gl/tpsmEJ#bisection

-- 
Kees Cook

Re: UBSAN: array-index-out-of-bounds in arch_uprobe_analyze_insn

[Masami Hiramatsu]

Hi Kees,

On Tue, 1 Dec 2020 16:48:55 -0800
Kees Cook <keescook@chromium.org> wrote:

> Hi,
> 
> There appears to be a problem with prefix counting for the instruction
> decoder. It looks like insn_get_prefixes() isn't keeping "nb" and "nbytes"
> in sync correctly:
> 
>         while (inat_is_legacy_prefix(attr)) {
>                 /* Skip if same prefix */
>                 for (i = 0; i < nb; i++)
>                         if (prefixes->bytes[i] == b)
>                                 goto found;
>                 if (nb == 4)
>                         /* Invalid instruction */
>                         break;
>                 prefixes->bytes[nb++] = b;
> 		...
> found:
>                 prefixes->nbytes++;
>                 insn->next_byte++;
>                 lb = b;
>                 b = peek_next(insn_byte_t, insn);
>                 attr = inat_get_opcode_attribute(b);
>         }
> 
> (nbytes is incremented on repeated prefixes, but "nb" isn't)
> 
> However, it looks like nbytes is used as an offset:
> 
> static inline int insn_offset_rex_prefix(struct insn *insn)
> {
>         return insn->prefixes.nbytes;
> }
> static inline int insn_offset_vex_prefix(struct insn *insn)
> {
>         return insn_offset_rex_prefix(insn) + insn->rex_prefix.nbytes;
> }

Yes, it is designed to do that. nbytes counts how many bytes the prefix is,
and nb is how many bytes of the prefix->bytes consumed.
Since the legacy prefix can be repeated and more than 4 (bytes), we can not
store all of those.

> 
> Which means everything that iterates over prefixes.bytes[] is buggy,
> since they may be trying to read past the end of the array:

Good catch! All following usage are wrong...

> 
> $ git grep -A3 -E '< .*prefixes(\.|->)nbytes'
> boot/compressed/sev-es.c:       for (i = 0; i < insn->prefixes.nbytes; i++) {
> boot/compressed/sev-es.c-               insn_byte_t p =
> insn->prefixes.bytes[i];
> boot/compressed/sev-es.c-
> boot/compressed/sev-es.c-               if (p == 0xf2 || p == 0xf3)
> --
> kernel/uprobes.c:       for (i = 0; i < insn->prefixes.nbytes; i++) {
> kernel/uprobes.c-               insn_attr_t attr;
> kernel/uprobes.c-
> kernel/uprobes.c-               attr = inat_get_opcode_attribute(insn->prefixes.bytes[i]);
> --
> kernel/uprobes.c:       for (i = 0; i < insn->prefixes.nbytes; i++) {
> kernel/uprobes.c-               if (insn->prefixes.bytes[i] == 0x66)
> kernel/uprobes.c-                       return -ENOTSUPP;
> kernel/uprobes.c-       }
> --
> lib/insn-eval.c:        for (i = 0; i < insn->prefixes.nbytes; i++) {
> lib/insn-eval.c-                insn_byte_t p = insn->prefixes.bytes[i];
> lib/insn-eval.c-
> lib/insn-eval.c-                if (p == 0xf2 || p == 0xf3)
> --
> lib/insn-eval.c:        for (i = 0; i < insn->prefixes.nbytes; i++) {
> lib/insn-eval.c-                insn_attr_t attr;
> lib/insn-eval.c-
> lib/insn-eval.c-                attr = inat_get_opcode_attribute(insn->prefixes.bytes[i]);
> 
> I don't see a clear way to fix this.

For the loop, we can check the insn.prefixes.bytes[i] == 0 since
it is initialized by 0 and 0x0 is not a prefix like this.

for (i = 0; insn->prefixes.bytes[i] && i < 4; i++) {
...
}

Thank you,

> 
> -Kees
> 
> On Mon, Sep 21, 2020 at 09:20:07PM -0700, syzbot wrote:
> > syzbot has bisected this issue to:
> > 
> > commit 4b2bd5fec007a4fd3fc82474b9199af25013de4c
> > Author: John Stultz <john.stultz@linaro.org>
> > Date:   Sat Oct 8 00:02:33 2016 +0000
> > 
> >     proc: fix timerslack_ns CAP_SYS_NICE check when adjusting self
> > 
> > bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=1697348d900000
> > start commit:   325d0eab Merge branch 'akpm' (patches from Andrew)
> > git tree:       upstream
> > final oops:     https://syzkaller.appspot.com/x/report.txt?x=1597348d900000
> > console output: https://syzkaller.appspot.com/x/log.txt?x=1197348d900000
> > kernel config:  https://syzkaller.appspot.com/x/.config?x=b12e84189082991c
> > dashboard link: https://syzkaller.appspot.com/bug?extid=9b64b619f10f19d19a7c
> > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=1573a8ad900000
> > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=164ee6c5900000
> > 
> > Reported-by: syzbot+9b64b619f10f19d19a7c@syzkaller.appspotmail.com
> > Fixes: 4b2bd5fec007 ("proc: fix timerslack_ns CAP_SYS_NICE check when adjusting self")
> > 
> > For information about bisection process see: https://goo.gl/tpsmEJ#bisection
> 
> -- 
> Kees Cook


-- 
Masami Hiramatsu <mhiramat@kernel.org>