eady for more work, but the
  network driver still has all of the pre-allocated skbs.  In that
  case, the new patch just calls alloc_skb, as you'd expect.

  We don't want to get carried away, though.  We try not to do
  excessive allocation in the write path, so we cap the number of skbs
  we dynamically allocate.

  Probably calling it a "dynamic pool" is misleading.  We were already
  trying to use a small fixed-size set of pre-allocated skbs before
  this patch, and this patch just provides a little headroom (with a
  ceiling, though) to accomodate network drivers that hang onto skbs,
  by allocating when needed.  The d->skbpool_hd list of allocated skbs
  is necessary so that we can free them later.

  We didn't notice the need for this headroom until AoE targets got
  fast enough.

Alternatives

  If the network layer never did a put_page on the pages in the bio's
  we get from the block layer, then it would be possible for us to
  hand skbs to the network layer and forget about them, allowing the
  network layer to free skbs itself (and thereby calling our own
  skb->destructor callback function if we needed that).  In that case
  we could get rid of the pre-allocated skbs and also the
  d->skbpool_hd, instead just calling alloc_skb every time we wanted
  to transmit a packet.  The slab allocator would effectively maintain
  the list of skbs.

  Besides a loss of CPU cache locality, the main concern with that
  approach the danger that it would increase the likelihood of
  deadlock when VM is trying to free pages by writing dirty data from
  the page cache through the aoe driver out to persistent storage on
  an AoE device.  Right now we have a situation where we have
  pre-allocation that corresponds to how much we use, which seems
  ideal.

  Of course, there's still the separate issue of receiving the packets
  that tell us that a write has successfully completed on the AoE
  target.  When memory is low and VM is using AoE to flush dirty data
  to free up pages, it would be perfect if there were a way for us to
  register a fast callback that could recognize write command
  completion responses.  But I don't think the current problems with
  the receive side of the situation are a justification for
  exacerbating the problem on the transmit side.

Signed-off-by: Ed L. Cashin <ecashin@coraid.com>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
P]`ː<x