Re: copyin()/copyout() constraints ?

On Fri, Jun 14, 2013 at 02:59:57PM -0700, Alfred Perlstein wrote:
> On 6/14/13 9:38 AM, Luigi Rizzo wrote:
> > On Fri, Jun 14, 2013 at 12:07:29PM -0400, John Baldwin wrote:
> >> On Wednesday, June 12, 2013 2:36:52 pm Alfred Perlstein wrote:
> >>> On 6/12/13 11:01 AM, Luigi Rizzo wrote:
> >>>> hi,
> >>>> is it possible to run copyin() or copyout() in one of these cases:
> >>>> 1. while holding a spinlock
> >>>> 2. while holding a regular mutex/lock
> >>>> 3. while holding a read lock (on an RWLOCK or RMLOCK)
> >>>> 4. while holding a write lock (on an RWLOCK or RMLOCK)
> >>>>
> >>>> I suspect #1 is forbidden, but am a bit unclear for the
> >>>> other cases.
> >>> No on all of the above unless the memory is wired.
> > ok i suppose i'll move to an sx lock, which i have been told
> > allows me to sleep ?
> >
> > My use case is that while i run the copyin(), and possibly take a
> > page fault, nobody destroys the destination buffer.  So i wanted
> > to hold a read lock (sx_slock() ?) in the thread doing the copy
> > (there may be several writers to different parts of the destination),
> > and a write lock (sx_xlock() ?) for the other thread which may
> > destroy the buffer.
> 
> We may be putting cart before horse, or horse into cart or something. :)
> 
> You may want to just wire the user buffer so it can't get ripped out 
> from under you.

I'll investigate, but i am not sure i can afford the cost of wiring
and unwiring every single buffer.

My application is a VALE/netmap switch interconnecting two virtual
machines, as below:

B and C are netmap buffers, and are wired (in the host)
A is an mbuf/skbuf within the guest OS (so for the
  host is not wired).

The current code is able to push 5-6 Mpps with 3 copies:
   A->B (done in userspace by a qemu thread for VM1),
   B->C (a memcpy in the kernel of the host)
   C->D (done in userspace by a qemu thread for VM2)

With "indirect buffers" in netmap/vale, i can eliminate the A->B copy,
and do A->C with a copyin in the kernel of the host.
But the per-packet budget is minuscule, and i am afraid that doing
an unconditional vslock() on each buffer is going to be too expensive
(and then i should also unwire the page ?


  +------------+  +-------------------------------+  +--------------+
  | VM1        |  |           VALE switch         |  | VM2          |
  |            |  |                               |  |              |
  |   mbuf     |  | .-----+               .-----. |  | mbuf         |
  |   .------. |  | |B    |    memcpy     |C    | |  | .-----.      |
  |   |A     +------>     +-------------->|     +----->|D    |      |
  |   |      | |  | |     |   (now)       |     | |  | |     |      |
  |   |      | |  | '-----'               '-+---' |  | |     |      |
  |   |      | |  |             copyin      ^     |  | |     |      |
  |   |      +------------------------------'     |  | |     |      |
  |   '------' |  |  (with indirect buffers)      |  | '-----'      |
  |            |  |                               |  |              |
  +------------+  +-------------------------------+  +--------------+

cheers
luigi