set_fs(KERNEL_DS) vs iovec

* set_fs(KERNEL_DS) vs iovec
@ 2018-05-22  5:08 Benjamin Herrenschmidt
  2018-05-22 10:20 ` Christoph Hellwig
  0 siblings, 1 reply; 3+ messages in thread
From: Benjamin Herrenschmidt @ 2018-05-22  5:08 UTC (permalink / raw)
  To: Al Viro, Greg Kroah-Hartman, Greg Kroah-Hartman; +Cc: linux-kernel

Hi guys !

I was helping with a small driver when I stumbled upon a comment from a
reviwer pointing to an old lwn article talking about deprecating set_fs
due to security concerns:

https://lwn.net/Articles/722267/

Now, this is a very simple driver running on a small/slow ARM SoC,
which reads from a FIFO and writes into a destination buffer.

It provides 2 interfaces, a userspace one (read syscall) and an in-
kernel one since some other kernel drivers use it as a transport.

My existing implementation uses the good old construct of doing
put_user() in the low level FIFO pumping code, and have the in-kernel
API do:

     old_fs = get_fs();
     set_fs(KERNEL_DS);
     rc = __sbefifo_submit(sbefifo, command, cmd_len,
                           response, resp_len);
     set_fs(old_fs);

Which is simple, turns into fairly efficient code on that simple
device, and doesn't seem to have security issues...

That said, following the advice in that article, I tried to look at the
iovec stuff and noticed:

 - The APIs are almost entirely undocumented (or did I look in the
wrong place ?)

 - The code in lib/iov_iter.c is rather ... unreadable

 - It's also significantly more complex, thus would probably result in
a slower driver (remember: small SoC). It's quite overkill for my
simple use case.

 - There are very few users, set_fs(KERNEL_DS) is still the most common
method used by drivers.

Hence my question: Is is still acceptable these days to use
set_fs(KERNEL_DS) for simple cases like this ? Or is it really
deprecated and all new users should use the iovec's ?

Cheers,
Ben.

^ permalink raw reply	[flat|nested] 3+ messages in thread