When connected mode is set, and we have connected and datagram traffic in
parallel, ipoib might crash with double free of datagram skb.

The current mechanism assumes that the order in the completion queue is
the same as the order of sent packets for all QPs. Order is kept only for
specific QP, in case of mixed UD and CM traffic we have few QPs (one UD and
few CM's) in parallel.

The problem:
----------------------------------------------------------

Transmit queue:
-----------------
UD skb pointer kept in queue itself, CM skb kept in spearate queue and
uses transmit queue as a placeholder to count the number of total
transmitted packets.

0   1   2   3   4  5  6  7  8   9  10  11 12 13 .........127
------------------------------------------------------------
NL ud1 UD2 CM1 ud3 cm2 cm3 ud4 cm4 ud5 NL NL NL ...........
------------------------------------------------------------
    ^                                  ^
   tail                               head

Completion queue (problematic scenario) - the order not the same as in
the transmit queue:

  1  2  3  4  5  6  7  8  9
------------------------------------
 ud1 CM1 UD2 ud3 cm2 cm3 ud4 cm4 ud5
------------------------------------

1. CM1 'wc' processing
   - skb freed in cm separate ring.
   - tx_tail of transmit queue increased although UD2 is not freed.
     Now driver assumes UD2 index is already freed and it could be used for
     new transmitted skb.

0   1   2   3   4  5  6  7  8   9  10  11 12 13 .........127
------------------------------------------------------------
NL NL  UD2 CM1 ud3 cm2 cm3 ud4 cm4 ud5 NL NL NL ...........
------------------------------------------------------------
        ^   ^                       ^
      (Bad)tail                    head
(Bad - Could be used for new SKB)

In this case (due to heavy load) UD2 skb pointer could be replaced by new
transmitted packet UD_NEW, as the driver assumes its free.  At this point
we will have to process two 'wc' with same index but we have only one
pointer to free.

During second attempt to free the same skb we will have NULL pointer
exception.

2. UD2 'wc' processing
   - skb freed according the index we got from 'wc', but it was already
     overwritten by mistake. So actually the skb that was released is the
     skb of the new transmitted packet and not the original one.

3. UD_NEW 'wc' processing
   - attempt to free already freed skb. NUll pointer exception.

The fix:
-----------------------------------------------------------------------

The fix is to stop using the UD ring as a placeholder for CM packets, the
cyclic ring variables tx_head and tx_tail will manage the UD tx_ring, a
new cyclic variables global_tx_head and global_tx_tail are introduced for
managing and counting the overall outstanding sent packets, then the send
queue will be stopped and waken based on these variables only.

Note that no locking is needed since global_tx_head is updated in the xmit
flow and global_tx_tail is updated in the NAPI flow only.  A previous
attempt tried to use one variable to count the outstanding sent packets,
but it did not work since xmit and NAPI flows can run at the same time and
the counter will be updated wrongly. Thus, we use the same simple cyclic
head and tail scheme that we have today for the UD tx_ring.

Fixes: 2c104ea68350 ("IB/ipoib: Get rid of the tx_outstanding variable in all modes")
Link: https://lore.kernel.org/r/20200527134705.480068-1-leon@kernel.org
Signed-off-by: Valentine Fatiev <valentinef@mellanox.com>
Signed-off-by: Alaa Hleihel <alaa@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Acked-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>