Allow a set of completion vectors to be associated with each InfiniBand
HCA port and allocate the completion vector for each session in a
round-robin fashion from the per-port set. This helps to spread the
InfiniBand interrupt workload over multiple CPU's, at least if different
InfiniBand MSI-X vectors are associated with different CPU's.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5728 d57e44dd-8a1f-0410-8b47-8ef2f437770f
When unloading the ib_srpt kernel module, instead of waiting until
all connected queue pairs have left the TimeWait state, destroy
these queue pairs immediately.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5727 d57e44dd-8a1f-0410-8b47-8ef2f437770f
The lockdep_assert_held() macro is a convenient debugging tool.
However, it is inconvenient to surround each invocation of that
macro by an #ifdef/#endif pair. Hence make it easier to use this
macro with older kernel versions.
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5573 d57e44dd-8a1f-0410-8b47-8ef2f437770f
This is allowed because all CM callback functions are invoked from thread
context.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5443 d57e44dd-8a1f-0410-8b47-8ef2f437770f
Move IB/CM members into a new struct. Report channel pointer instead of CM ID
pointer in diagnostic messages.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5442 d57e44dd-8a1f-0410-8b47-8ef2f437770f
ib_send_cm_rep() checks the connection state before sending a response.
Hence checking ch->state before calling ib_send_cm_rep() is superfluous,
so remove that check and also the locking that is no longer needed.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5441 d57e44dd-8a1f-0410-8b47-8ef2f437770f
With the SRP protocol all RDMA operations are initiated by the target.
Since no RDMA operations are initiated by the initiator, do not grant
the initiator permission to submit RDMA reads or writes to the target.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5423 d57e44dd-8a1f-0410-8b47-8ef2f437770f
Complain if srpt_abort_cmd() is called for an I/O context that is
being processed by SCST and not by the HCA.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5393 d57e44dd-8a1f-0410-8b47-8ef2f437770f
Only complain about a missing completion for I/O contexts that are
in a state where the ib_srpt driver is waiting for the HCA.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5392 d57e44dd-8a1f-0410-8b47-8ef2f437770f
Apparently with mlx4 firmware up to and including 2.30.8000 it is not
guaranteed that for a QP associated with an SRQ error completions are
generated for all pending work requests. Avoid triggering
srpt_pending_cmd_timeout() for missing error completions.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5370 d57e44dd-8a1f-0410-8b47-8ef2f437770f
Invoking srpt_zerolength_write() before the queue pair has reached
the error state is a bug, so complain loudly if that happens.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5369 d57e44dd-8a1f-0410-8b47-8ef2f437770f
The overhead of atomic_add_return() is lower than that of a
spin_lock() / spin_unlock() pair, hence switch to the former.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5367 d57e44dd-8a1f-0410-8b47-8ef2f437770f
All callers guarantee that the completion queue is empty so it is
not necessary to invoke ib_poll_cq() from inside this function.
git-svn-id: http://svn.code.sf.net/p/scst/svn/trunk@5366 d57e44dd-8a1f-0410-8b47-8ef2f437770f
