Re: [linux-pm] [PATCH 2/9] PM: suspend_block: Add driver to access suspend blockers from user-space

["Arve Hjønnevåg" <arve@android.com>]

2010/4/26 Alan Stern <stern@rowland.harvard.edu>:
> On Sun, 25 Apr 2010, Arve Hjønnevåg wrote:
>
>> >> > >> > +User-space API
>> >> > >> > +==============
>> >> > >> > +
>> >> > >> > +To create a suspend_blocker from user-space, open the suspend_blocker device:
>> >> > >> > +    fd = open("/dev/suspend_blocker", O_RDWR | O_CLOEXEC);
>> >> > >> > +then call:
>> >> > >> > +    ioctl(fd, SUSPEND_BLOCKER_IOCTL_INIT(strlen(name)), name);
>> >> > >>
>> >> > >>
>> >> > >> This seems like very wrong idea -- it uses different ioctl number for
>> >> > >> each length AFAICT.
>> >> > >
>> >> > > How about specifying the name by an ordinary write() call instead of
>> >> > > by an ioctl()?
>> >> > >
>> >> >
>> >> > I prefer using ioctls. We have three operations at the moment. Init,
>> >> > block and unblock. If we do init with write but block and unblock
>> >> > using ioctls, it would be pretty strange. Specifying a command and
>> >>
>> >> Why would it be "strange"?
>> >
>> > Why indeed?  Using write() is the natural way to pass a data buffer
>> > into the kernel, especially a variable-length buffer.
>> >
>> > Mixing ioctl() and write() might seem strange at first, but it has
>> > plenty of precedent.  Consider adjusting the settings for a serial
>> > port, for example.
>> >
>> That sound like to opposite situation to me. It uses ioctls for setup
>> and read/write access the data stream.
>
> Or you could say that it uses ioctls to send commands and read/write to
> pass data.  That's pretty much what you would be doing.
>
> Let's put it another way: You find it excessively strange to do init
> using write and to send commands via ioctl, but you don't find it
> strange to abuse the ioctl number code for passing the string length?
> The old saying about motes and beams applies here...
>
> If you insist on using ioctl for init, you should use the standard
> convention for passing variable-length data.  The userspace program
> sets up a fixed-size buffer containing a pointer to the name and the
> name's length, and it passes the buffer's address as the ioctl
> argument.

Are you sure that is the standard? I searched for ioctls with NAME in
their name and only found one that passed the name that way. The rest
used fixed length string buffers, or passed the buffersize to _IOC
like I do. For instance, input.h has ioctls to read string and
bitmasks where user space specify the buffer size as an argument to
the ioctl macro. These pass data from the kernel to user space, but I
don't passing a string length is any worse than passing a buffer size.

-- 
Arve Hjønnevåg