net: ipa: fix TX queue race

Jakub Kicinski pointed out a race condition in ipa_start_xmit() in a
recently-accepted series of patches:
  https://lore.kernel.org/netdev/20210812195035.2816276-1-elder@linaro.org/
We are stopping the modem TX queue in that function if the power
state is not active.  We restart the TX queue again once hardware
resume is complete.

  TX path                       Power Management
  -------                       ----------------
  pm_runtime_get(); no power    Start resume
  Stop TX queue                      ...
  pm_runtime_put()              Resume complete
  return NETDEV_TX_BUSY         Start TX queue

  pm_runtime_get()
  Power present, transmit
  pm_runtime_put()              (auto-suspend)

The issue is that the power management (resume) activity and the
network transmit activity can occur concurrently, and there's a
chance the queue will be stopped *after* it has been started again.

  TX path                       Power Management
  -------                       ----------------
                                Resume underway
  pm_runtime_get(); no power         ...
                                Resume complete
                                Start TX queue
  Stop TX queue       <-- No more transmits after this
  pm_runtime_put()
  return NETDEV_TX_BUSY

We address this using a STARTED flag to indicate when the TX queue
has been started from the resume path, and a spinlock to make the
flag and queue updates happen atomically.

  TX path                       Power Management
  -------                       ----------------
                                Resume underway
  pm_runtime_get(); no power    Resume complete
                                start TX queue     \
  If STARTED flag is *not* set:                     > atomic
      Stop TX queue             set STARTED flag   /
  pm_runtime_put()
  return NETDEV_TX_BUSY

A second flag is used to address a different race that involves
another path requesting power.

  TX path            Other path              Power Management
  -------            ----------              ----------------
                     pm_runtime_get_sync()   Resume
                                             Start TX queue   \ atomic
                                             Set STARTED flag /
                     (do its thing)
                     pm_runtime_put()
                                             (auto-suspend)
  pm_runtime_get()                           Mark delayed resume
  STARTED *is* set, so
    do *not* stop TX queue  <-- Queue should be stopped here
  pm_runtime_put()
  return NETDEV_TX_BUSY                      Suspend done, resume
                                             Resume complete
  pm_runtime_get()
  Stop TX queue
    (STARTED is *not* set)                   Start TX queue   \ atomic
  pm_runtime_put()                           Set STARTED flag /
  return NETDEV_TX_BUSY

So a STOPPED flag is set in the transmit path when it has stopped
the TX queue, and this pair of operations is also protected by the
spinlock.  The resume path only restarts the TX queue if the STOPPED
flag is set.  This case isn't a major problem, but it avoids the
"non-trivial amount of useless work" done by the networking stack
when NETDEV_TX_BUSY is returned.

Fixes: 6b51f802

 ("net: ipa: ensure hardware has power in ipa_start_xmit()")
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>