Re: [PATCH] drivers/gpu/vga: use __GFP_NOWARN for user-controlled kmalloc

[Dmitry Vyukov]

On Thu, Feb 4, 2016 at 5:59 PM, Ville SyrjÃlÃ
<ville.syrjala@xxxxxxxxxxxxxxx> wrote:
> On Thu, Feb 04, 2016 at 05:37:49PM +0100, Dmitry Vyukov wrote:
>> On Thu, Feb 4, 2016 at 5:32 PM, Ville SyrjÃlÃ
>> <ville.syrjala@xxxxxxxxxxxxxxx> wrote:
>> > On Thu, Feb 04, 2016 at 04:49:49PM +0100, Dmitry Vyukov wrote:
>> >> Size of kmalloc() in vga_arb_write() is controlled by user.
>> >> Too large kmalloc() size triggers WARNING message on console.
>> >>
>> >> Use GFP_USER | __GFP_NOWARN for this kmalloc() to not scare admins.
>> >>
>> >> Signed-off-by: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
>> >> ---
>> >> Example WARNING:
>> >>
>> >> WARNING: CPU: 2 PID: 29322 at mm/page_alloc.c:2999
>> >> __alloc_pages_nodemask+0x7d2/0x1760()
>> >> Modules linked in:
>> >> CPU: 2 PID: 29322 Comm: syz-executor Tainted: G    B  4.5.0-rc1+ #283
>> >> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
>> >>  00000000ffffffff ffff880069eff670 ffffffff8299a06d 0000000000000000
>> >>  ffff8800658a4740 ffffffff864985a0 ffff880069eff6b0 ffffffff8134fcf9
>> >>  ffffffff8166de32 ffffffff864985a0 0000000000000bb7 00000000024040c0
>> >> Call Trace:
>> >>  [<     inline     >] __dump_stack lib/dump_stack.c:15
>> >>  [<ffffffff8299a06d>] dump_stack+0x6f/0xa2 lib/dump_stack.c:50
>> >>  [<ffffffff8134fcf9>] warn_slowpath_common+0xd9/0x140 kernel/panic.c:482
>> >>  [<ffffffff8134ff29>] warn_slowpath_null+0x29/0x30 kernel/panic.c:515
>> >>  [<     inline     >] __alloc_pages_slowpath mm/page_alloc.c:2999
>> >>  [<ffffffff8166de32>] __alloc_pages_nodemask+0x7d2/0x1760 mm/page_alloc.c:3253
>> >>  [<ffffffff81745c99>] alloc_pages_current+0xe9/0x450 mm/mempolicy.c:2090
>> >>  [<     inline     >] alloc_pages include/linux/gfp.h:459
>> >>  [<ffffffff81669bb6>] alloc_kmem_pages+0x16/0x100 mm/page_alloc.c:3433
>> >>  [<ffffffff816c20af>] kmalloc_order+0x1f/0x80 mm/slab_common.c:1008
>> >>  [<ffffffff816c212f>] kmalloc_order_trace+0x1f/0x140 mm/slab_common.c:1019
>> >>  [<     inline     >] kmalloc_large include/linux/slab.h:395
>> >>  [<ffffffff81756b24>] __kmalloc+0x2f4/0x340 mm/slub.c:3557
>> >>  [<     inline     >] kmalloc include/linux/slab.h:468
>> >>  [<ffffffff832c65a4>] vga_arb_write+0xd4/0xe40 drivers/gpu/vga/vgaarb.c:926
>> >>  [<ffffffff817a9831>] do_loop_readv_writev+0x141/0x1e0 fs/read_write.c:719
>> >>  [<ffffffff817ad698>] do_readv_writev+0x5f8/0x6e0 fs/read_write.c:849
>> >>  [<ffffffff817ad8b6>] vfs_writev+0x86/0xc0 fs/read_write.c:886
>> >>  [<     inline     >] SYSC_writev fs/read_write.c:919
>> >>  [<ffffffff817b0a21>] SyS_writev+0x111/0x2b0 fs/read_write.c:911
>> >>  [<ffffffff86359636>] entry_SYSCALL_64_fastpath+0x16/0x7a
>> >> arch/x86/entry/entry_64.S:185
>> >> ---
>> >>  drivers/gpu/vga/vgaarb.c | 2 +-
>> >>  1 file changed, 1 insertion(+), 1 deletion(-)
>> >>
>> >> diff --git a/drivers/gpu/vga/vgaarb.c b/drivers/gpu/vga/vgaarb.c
>> >> index f17cb04..d73b85b 100644
>> >> --- a/drivers/gpu/vga/vgaarb.c
>> >> +++ b/drivers/gpu/vga/vgaarb.c
>> >> @@ -923,7 +923,7 @@ static ssize_t vga_arb_write(struct file *file, const char __user *buf,
>> >>       int i;
>> >>
>> >>
>> >> -     kbuf = kmalloc(count + 1, GFP_KERNEL);
>> >> +     kbuf = kmalloc(count + 1, GFP_USER | __GFP_NOWARN);
>> >
>> > I don't really see why it does this user controlled malloc in the
>> > first place. The max legth of the string it will actually handle looks
>> > well bounded, so it could just use some fixed length buffer (on stack
>> > even).
>>
>>
>> What would be the right limit on data len?
>
> From the looks of things the longest command could be the
> "target PCI:domain:bus:dev.fn" thing. Even assuming something silly like
> having 10 characters for each domain,bus,dev,fn that would still be only
> 55 bytes. So based on that even something like 64 bytes should be more
> than enough AFAICS.

David, what do you think? I can allocate char kbuf[64] on stack.


> The other thing that strikes me as bit odd in this code is that it
> just ignores whatever data is left over after it's done parsing the
> string. But it returns the full count to userspace, indicating it
> ate all of it. I guess that's fairly sane when userspace just uses a
> fixed size buffer and checks that the kernel consumed it all. But
> maybe there should be an actual check to see that there's a '\0'
> or maybe <any amount of whitespace>+'\0' after the parsed string.