There is a potential for deadlock when allocating a struct sk_buff for
  data that needs to be written out to aoe storage.  If the data is
  being written from a dirty page in order to free that page, and if
  there are no other pages available, then deadlock may occur when a
  free page is needed for the sk_buff allocation.  This situation has
  not been observed, but it would be nice to eliminate any potential for
  deadlock under memory pressure.
  
  Because ATA over Ethernet is not fragmented by the kernel's IP code,
  the destructor member of the struct sk_buff is available to the aoe
  driver.  By using a mempool for allocating all but the first few
  sk_buffs, and by registering a destructor, we should be able to
  efficiently allocate sk_buffs without introducing any potential for
  deadlock.