Potential double fetch in sg_scsi_ioctl()

Potential double fetch in sg_scsi_ioctl()

[Alexander Potapenko]

Hi Christof, Jens,

We've found a double-fetch in sg_scsi_ioctl() using a prototype tool
(see the report below).

Turns out that sg_scsi_ioctl() reads the first byte of sic->data
twice: first when getting the opcode
(https://elixir.bootlin.com/linux/latest/source/block/scsi_ioctl.c#L439),
then when reading the command of the size calculated from that opcode
(https://elixir.bootlin.com/linux/latest/source/block/scsi_ioctl.c#L464).

At this point opcode and req->cmd[0] may mismatch.
The opcode is then used to determine rq->timeout and req->retries,
whereas req->cmd[0] is used by the underlying device drivers.
Not sure invalid timeout or retries is a big deal, but since the
command length also depends on the opcode, it is possible to trick the
kernel into using the remnants of the previous command by first
announcing a short command and then changing the opcode to a longer
one.

I've noticed that three years ago Meng Xu has reported the very same
bug already: https://patchwork.kernel.org/project/linux-block/patch/1505834638-37142-1-git-send-email-mengxu.gatech@gmail.com/
Was there any followup to that patch?

Alex

Double fetch report follows:

BUG: multi-read in sg_scsi_ioctl, syscall __x64_sys_ioctl

==================================================================
======= First Address Range Stack =======
 df_save_stack+0x33/0x70 lib/df-detection.c:202
 add_address+0x2cb/0x370 lib/df-detection.c:48
 sg_scsi_ioctl+0x160/0x560 block/scsi_ioctl.c:439
 sg_ioctl_common+0xdb6/0x1110 drivers/scsi/sg.c:1109
 sg_ioctl+0x49/0xa0 drivers/scsi/sg.c:1163
 vfs_ioctl fs/ioctl.c:48 [inline]
 __do_sys_ioctl fs/ioctl.c:753 [inline]
 __se_sys_ioctl fs/ioctl.c:739 [inline]
 __x64_sys_ioctl+0xfc/0x140 fs/ioctl.c:739
 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46
 entry_SYSCALL_64_after_hwframe+0x44/0xa9
======= Second Address Range Stack =======
 df_save_stack+0x33/0x70 lib/df-detection.c:202
 add_address+0x2cb/0x370 lib/df-detection.c:48
 _copy_from_user+0x93/0xe0 lib/usercopy.c:17
 copy_from_user include/linux/uaccess.h:167 [inline]
 sg_scsi_ioctl+0x21a/0x560 block/scsi_ioctl.c:464
 sg_ioctl_common+0xdb6/0x1110 drivers/scsi/sg.c:1109
 sg_ioctl+0x49/0xa0 drivers/scsi/sg.c:1163
 vfs_ioctl fs/ioctl.c:48 [inline]
 __do_sys_ioctl fs/ioctl.c:753 [inline]
 __se_sys_ioctl fs/ioctl.c:739 [inline]
 __x64_sys_ioctl+0xfc/0x140 fs/ioctl.c:739
 do_syscall_64+0x2d/0x70 arch/x86/entry/common.c:46
 entry_SYSCALL_64_after_hwframe+0x44/0xa9
syscall number 16  System Call: __x64_sys_ioctl+0x0/0x140 fs/ioctl.c:800
First 0000000020000188 len 1 Caller sg_ioctl_common+0xdb6/0x1110
drivers/scsi/sg.c:1109
Second 0000000020000188 len 10 Caller copy_from_user
include/linux/uaccess.h:167 [inline]
Second 0000000020000188 len 10 Caller sg_scsi_ioctl+0x21a/0x560
block/scsi_ioctl.c:464

CPU: 0 PID: 6266 Comm: syz-executor.0 Not tainted 5.9.0-rc4+ #55
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-1 04/01/2014
==================================================================




-- 
Alexander Potapenko
Software Engineer

Google Germany GmbH
Erika-Mann-Straße, 33
80636 München

Geschäftsführer: Paul Manicle, Halimah DeLaine Prado
Registergericht und -nummer: Hamburg, HRB 86891
Sitz der Gesellschaft: Hamburg

Re: Potential double fetch in sg_scsi_ioctl()

[Jens Axboe]

On 11/23/20 9:15 AM, Alexander Potapenko wrote:
> Hi Christof, Jens,
> 
> We've found a double-fetch in sg_scsi_ioctl() using a prototype tool
> (see the report below).
> 
> Turns out that sg_scsi_ioctl() reads the first byte of sic->data
> twice: first when getting the opcode
> (https://elixir.bootlin.com/linux/latest/source/block/scsi_ioctl.c#L439),
> then when reading the command of the size calculated from that opcode
> (https://elixir.bootlin.com/linux/latest/source/block/scsi_ioctl.c#L464).
> 
> At this point opcode and req->cmd[0] may mismatch.
> The opcode is then used to determine rq->timeout and req->retries,
> whereas req->cmd[0] is used by the underlying device drivers.
> Not sure invalid timeout or retries is a big deal, but since the
> command length also depends on the opcode, it is possible to trick the
> kernel into using the remnants of the previous command by first
> announcing a short command and then changing the opcode to a longer
> one.
> 
> I've noticed that three years ago Meng Xu has reported the very same
> bug already: https://patchwork.kernel.org/project/linux-block/patch/1505834638-37142-1-git-send-email-mengxu.gatech@gmail.com/
> Was there any followup to that patch?

Doesn't look like it - Christoph made a suggestion, and then the
original reporter didn't follow up. FWIW, I do agree that just copying
it once is a better idea than copying twice and then copying the opcode
again.

-- 
Jens Axboe