In some case such as trigger_all_cpu_backtrace netpoll can wind up
generating a lot of packets in hard irq context.  My rough estimate is
perhaps 1500 packets.  That is larger than any driver tx ring, which
makes netpoll_poll_dev necessary to transmit all of the netconsole
packets immediately.  Those 1500+ packets can take up a couple megabytes
of memory if we aren't careful.  On some machines that is enough to
start depleting the polls GFP_ATOMIC can dig into, so netpoll needs to
at a minimum to be able to reuse the memory for the skbs it has
transmitted.

Today this reclamation of transmitted packets happens in
zap_completetion_queue as dev_kfree_skb_irq places all packets to be
freed on a completion queue.  netpoll then searches this queue for
packets it thinks are freeable, and frees them.  Unfortunately
the current logic netpoll uses to decided a packet is freeable
is incorrect and thus unsafe :(

The logic netpoll uses to determine if a packet is freeable is to verify
a skb does not have a destructor.  Which works most of the time.  But in
pathological cases it can report that is a packet is freeable in hard
irq context when it is not.

This set of changes adds a function skb_irq_freeable and uses that
function in zap_completion_queue to remove the bug, and in bowls
of kfree_skb in skb_release_head_state to warn if we are inappropriate
freeing a skb.

While I don't expect this will allow anything except skbs sent by
netpoll to be freed, solving the general problem rather than solving
this for just packets generated by netpoll seems like a robust way of
handling this.

Eric W. Biederman (3):
      net: Add a test to see if a skb is freeable in irq context
      netpoll: Use skb_irq_freeable to make zap_completion_queue safe.
      net: Warn when a skb is freed inappropriately in hard irq context.

 include/linux/skbuff.h | 13 +++++++++++++
 net/core/netpoll.c     |  2 +-
 net/core/skbuff.c      |  6 +++---
 3 files changed, 17 insertions(+), 4 deletions(-)

Eric