[RFC v11][PATCH 03/13] General infrastructure for checkpoint restart

Linus Torvalds torvalds at linux-foundation.org
Tue Dec 16 11:28:39 PST 2008



On Tue, 16 Dec 2008, Mike Waychison wrote:
> 
> set_fs(fs) here

Btw, this all is an excellent example of why people should try to aim for 
small functions and use lots of them.

It's often _way_ more readable to do

	static inline int __some_fn(...)
	{
		.. do the real work here ..
	}

	int some_fn(...)
	{
		int retval;

		prepare();
		retval = __some_fn(..)
		finish();

		return retval;
	}

where "prepare/finish" can be about locking, or set_fs(), or allocation 
and de-allocation of temporary buffers, or any number of things like that.

With set_fs() in particular, the wrapper function also tends to be the 
perfect place to change a regular (kernel) pointer into a user pointer. 
IOW, it's the place to make sparse happy, where you can do things like

	uptr = (__force void __user *)ptr;

and comment on the fact that the forced user pointer cast is valid only 
because of the set_fs().

Because it looks like the code isn't sparse-clean.

Btw, I also think that code like this is bogus:

	nwrite = file->f_op->write(file, addr, nleft, &file->f_pos);

because you're not supposed to pass in the raw file->f_pos to that 
function. It's fundamentally thread-unsafe. I realize that maybe you don't 
care, but the thing is, you're supposed to do

	loff_t pos = file->pos;
	..
	nwrite = file->f_op->write(file, addr, nleft, &pos);
	..
	file->f_pos = pos;

and in fact preferably use "file_pos_read()" and "file_pos_write()" (but 
we've never exposed them outside of fs/read_write.c, so I guess we should 
do that).

And yes, I realize that some code does take the address of f_pos directly 
(splice, nfsctl, others), and I realize that it works, but it's still bad 
form. Please don't add more of them.

			Linus


More information about the Containers mailing list