IB/hfi1: Prevent kernel QP post send hard lockups
authorMike Marciniszyn <mike.marciniszyn@intel.com>
Sun, 9 Apr 2017 17:16:35 +0000 (10:16 -0700)
committerDoug Ledford <dledford@redhat.com>
Fri, 28 Apr 2017 17:48:01 +0000 (13:48 -0400)
commitb6eac931b9bb2bce4db7032c35b41e5e34ec22a5
tree9f6757a5a525248cb8ed496ad39ce8bad60c24da
parent3d591099a0a2b45a50913130f0599ab838002fc3
IB/hfi1: Prevent kernel QP post send hard lockups

The driver progress routines can call cond_resched() when
a timeslice is exhausted and irqs are enabled.

If the ULP had been holding a spin lock without disabling irqs and
the post send directly called the progress routine, the cond_resched()
could yield allowing another thread from the same ULP to deadlock
on that same lock.

Correct by replacing the current hfi1_do_send() calldown with a unique
one for post send and adding an argument to hfi1_do_send() to indicate
that the send engine is running in a thread.   If the routine is not
running in a thread, avoid calling cond_resched().

CC: <stable@vger.kernel.org> # 4.7.x-
Fixes: Commit 831464ce4b74 ("IB/hfi1: Don't call cond_resched in atomic mode when sending packets")
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/hw/hfi1/ruc.c
drivers/infiniband/hw/hfi1/verbs.c
drivers/infiniband/hw/hfi1/verbs.h