Re: [PATCH v2 3/5] IB/hfi1: Add ioctl() interface for user commands

From: ira.weiny
Date: Fri May 13 2016 - 16:54:10 EST


On Thu, May 12, 2016 at 01:40:06PM -0600, Jason Gunthorpe wrote:
> On Thu, May 12, 2016 at 03:27:27PM -0400, Dennis Dalessandro wrote:
>
> > >>+static inline int check_ioctl_access(unsigned int cmd, unsigned long arg)
> > >>+{
> > >>+ int read_cmd, write_cmd, read_ok, write_ok;
> > >>+
> > >>+ read_cmd = _IOC_DIR(cmd) & _IOC_READ;
> > >>+ write_cmd = _IOC_DIR(cmd) & _IOC_WRITE;
> > >>+ write_ok = access_ok(VERIFY_WRITE, (void __user *)arg, _IOC_SIZE(cmd));
> > >>+ read_ok = access_ok(VERIFY_READ, (void __user *)arg, _IOC_SIZE(cmd));
> > >>+
> > >>+ if ((read_cmd && !write_ok) || (write_cmd && !read_ok))
> > >>+ return -EFAULT;
> > >
> > >This seems kind of goofy, didn't Ira say this is performance senstive?
>
> Well, calling access_ok twice when only once is typically needed is
> certainly not performant. Typically this check is done at every access
> via get_user/put_user/copy_to/from_user - why is it being hoisted like
> this?
>
>
> > >Driver shouldn't be open coding __get_user like that, IMHO.
> >
> > Can you explain what you mean here? We should not use __get_user()?
>
> Generally speaking, yes. Use get_user() that includes the correct
> access_ok. Unless there is a good reason to avoid it, the standard API
> should be used.

I know this code was refactored while we were still submitting patches to Greg
KH back in Nov/Dec. Part of this was cleaning up branch on error rather than
success. Hence the check for access at the top of the function and early
return.

At that time I _thought_ there were multiple __get_users in some of the
operations so a single common access_ok would speed those up. However, I don't
see that happening any longer, so either I don't remember correctly, or we have
made this cleaner.

As it stands now I think you are correct we could use get_user and
copy_to/from_user.

Ira