Re: [PATCH v3] SUNRPC: set desired file system root before connecting local transports

[ebiederm@xmission.com (Eric W. Biederman)]

"J. Bruce Fields" <bfields@fieldses.org> writes:

> On Tue, Oct 09, 2012 at 01:20:48PM -0700, Eric W. Biederman wrote:
>> "Myklebust, Trond" <Trond.Myklebust@netapp.com> writes:
>> 
>> > On Tue, 2012-10-09 at 15:35 -0400, J. Bruce Fields wrote:
>> >> Cc'ing Eric since I seem to recall he suggested doing it this way?
>> 
>> Yes.  On second look setting fs->root won't work. We need to change fs.
>> The problem is that by default all kernel threads share fs so changing
>> fs->root will have non-local consequences.
>
> Oh, huh.  And we can't "unshare" it somehow?

I don't fully understand how nfs uses kernel threads and work queues.
My general understanding is work queues reuse their kernel threads
between different users.  So it is mostly a don't pollute your
environment thing.  If there was a dedicated kernel thread for each
environment this would be trivial.

What I was suggesting here is changing task->fs instead of
task->fs.root.  That should just require task_lock().

> Or, previously you suggested:
>
> 	- introduce sockaddr_fd that can be applied to AF_UNIX sockets,
> 	  and teach unix_bind and unix_connect how to deal with a second
> 	  type of sockaddr, AT_FD:
> 	  struct sockaddr_fd { short fd_family; short pad; int fd; }
>
> 	- introduce sockaddr_unix_at that takes a directory file
> 	  descriptor as well as a unix path, and teach unix_bind and
> 	  unix_connect to deal with a second sockaddr type, AF_UNIX_AT:
> 	  struct sockaddr_unix_at { short family; short pad; int dfd; char path[102]; }
>
> Any other options?

I am still half hoping we don't have to change the userspace API/ABI.
There is sanity checking on that path that no one seems interested in to
solve this problem.

This is a weird issue as we are dealing with both the vfs and the
networking stack.  Fundamentally we need to change task->fs.root or
we need to capitialize on the openat functionality in the kernel, so
that we don't create mountains of special cases to support this.

I think swapping task->fs instead of task->fs.root is effecitely the
same complexity.

>> I very much believe we want if at all possible to perform a local
>> modification.
>> 
>> Changing fs isn't all that different from what devtmpfs is doing.
>
> Sorry, I don't know much about devtmpfs, are you suggesting it as a
> model?  What exactly should we look at?

Roughly all I meant was that devtmpsfsd is a kernel thread that runs
with an unshared fs struct.  Although I admit devtmpfsd is for all
practical purposes a userspace daemon that just happens to run in kernel
space.

Eric