Re: [bug report] io_uring: add support for fixed wait regions

[Jens Axboe <axboe@kernel.dk>]

On 10/30/24 7:58 AM, Dan Carpenter wrote:
> On Wed, Oct 30, 2024 at 07:22:49AM -0600, Jens Axboe wrote:
>> On 10/30/24 5:40 AM, Dan Carpenter wrote:
>>> Hello Jens Axboe,
>>>
>>> Commit 4b926ab18279 ("io_uring: add support for fixed wait regions")
>>> from Oct 22, 2024 (linux-next), leads to the following Smatch static
>>> checker warning:
>>>
>>> 	io_uring/register.c:616 io_register_cqwait_reg()
>>> 	warn: was expecting a 64 bit value instead of '~(~(((1) << 12) - 1))'
>>>
>>> io_uring/register.c
>>>     594 static int io_register_cqwait_reg(struct io_ring_ctx *ctx, void __user *uarg)
>>>     595 {
>>>     596         struct io_uring_cqwait_reg_arg arg;
>>>     597         struct io_uring_reg_wait *reg;
>>>     598         struct page **pages;
>>>     599         unsigned long len;
>>>     600         int nr_pages, poff;
>>>     601         int ret;
>>>     602 
>>>     603         if (ctx->cq_wait_page || ctx->cq_wait_arg)
>>>     604                 return -EBUSY;
>>>     605         if (copy_from_user(&arg, uarg, sizeof(arg)))
>>>     606                 return -EFAULT;
>>>     607         if (!arg.nr_entries || arg.flags)
>>>     608                 return -EINVAL;
>>>     609         if (arg.struct_size != sizeof(*reg))
>>>     610                 return -EINVAL;
>>>     611         if (check_mul_overflow(arg.struct_size, arg.nr_entries, &len))
>>>     612                 return -EOVERFLOW;
>>>     613         if (len > PAGE_SIZE)
>>>     614                 return -EINVAL;
>>>     615         /* offset + len must fit within a page, and must be reg_wait aligned */
>>> --> 616         poff = arg.user_addr & ~PAGE_MASK;
>>>
>>> This is a harmless thing, but on 32 bit systems you can put whatever you want in
>>> the high 32 bits of arg.user_addr and it won't affect anything.
>>
>> That is certainly true, it'll get masked away. I suspect this kind of
>> thing is everywhere, though? What do you suggest?
> 
> The way that I normally see these warnings is with code like "if
> (u64flags & ~mask)" where only the first 3 bits of u64flags are used.
> It's not normally a real life bug.  Normally fix them the warning, but
> I have 174 old warnings from before I started complaining about them.
> 
> Maybe:
> 
>         if (U32_MAX >= SIZE_MAX && arg.user_addr > SIZE_MAX)
> 		return -EINVAL;
> 
> This code works fine as-is, but eventually I want this code to trigger
> a couple more static checker warnings.  It's so suspicious because
> we're truncating user data then re-using the same untruncated variable
> again.

Agree, that's the part I don't like. It's fine masking off for offset,
but the later passing in directly is wonky. It'll get truncated to
unsigned long at that point though, so won't _actually_ be passed in.
Hence I'm a bit dubious that this really needs fixing. Yeah the app can
put garbage in the upper 32-bits, but it's never going to be seen.
Should we catch and -EINVAL on that? Potentially, at least it can't hurt
to do so.

-- 
Jens Axboe