Profiling has shown that sdma_lock is proving a bottleneck for
performance. The situations include:
- RDMA reads when krcvqs > 1
- post sends from multiple threads
For RDMA read the current global qib_wq mechanism runs on all CPUs
and contends for the sdma_lock when multiple RMDA read requests are
fielded on differenct CPUs. For post sends, the direct call to
qib_do_send() from multiple threads causes the contention.
Since the sdma mechanism is per port, this fix converts the existing
workqueue to a per port single thread workqueue to reduce the lock
contention in the RDMA read case, and for any other case where the QP
is scheduled via the workqueue mechanism from more than 1 CPU.
For the post send case, This patch modifies the post send code to test
for a non empty sdma engine. If the sdma is not idle the (now single
thread) workqueue will be used to trigger the send engine instead of
the direct call to qib_do_send().
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
There is a cut-and-paste typo in the function qib_pci_slot_reset()
where it prints that the "link_reset" function is called rather than
the "slot_reset" function. This makes the message misleading.
Signed-off-by: Betty Dall <betty.dall@hp.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Profiling indicates that MR validation locking is expensive. The MR
table is largely read-only and is a suitable candidate for RCU locking.
The patch uses RCU locking during validation to eliminate one
lock/unlock during that validation.
Reviewed-by: Mike Heinz <michael.william.heinz@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
A timing issue can occur where qib_mr_dereg can return -EBUSY if the
MR use count is not zero.
This can occur if the MR is de-registered while RDMA read response
packets are being progressed from the SDMA ring. The suspicion is
that the peer sent an RDMA read request, which has already been copied
across to the peer. The peer sees the completion of his request and
then communicates to the responder that the MR is not needed any
longer. The responder tries to de-register the MR, catching some
responses remaining in the SDMA ring holding the MR use count.
The code now uses a get/put paradigm to track MR use counts and
coordinates with the MR de-registration process using a completion
when the count has reached zero. A timeout on the delay is in place
to catch other EBUSY issues.
The reference count protocol is as follows:
- The return to the user counts as 1
- A reference from the lk_table or the qib_ibdev counts as 1.
- Transient I/O operations increase/decrease as necessary
A lot of code duplication has been folded into the new routines
init_qib_mregion() and deinit_qib_mregion(). Additionally, explicit
initialization of fields to zero is now handled by kzalloc().
Also, duplicated code 'while.*num_sge' that decrements reference
counts have been consolidated in qib_put_ss().
Reviewed-by: Ramkrishna Vepa <ramkrishna.vepa@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
An MR reference leak exists when handling UC RDMA writes with
immediate data because we manipulate the reference counts as if the
operation had been a send.
This patch moves the last_imm label so that the RDMA write operations
with immediate data converge at the cq building code. The copy/mr
deref code is now done correctly prior to the branch to last_imm.
Reviewed-by: Edward Mascarenhas <edward.mascarenhas@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
The dev->sgid_tbl[] array is allocated in ocrdma_alloc_resources().
It has OCRDMA_MAX_SGID elements so the test here is off by one.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Fix RQ/SRQ error CQE polling. Return error CQE to consumer for error
case which was not returned previously.
Signed-off-by: Parav Pandit <parav.pandit@emulex.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
1. Fix reporting GID table addition events.
2. Enable vlan based GID entries only when VLAN is enabled at compile
time (test CONFIG_VLAN_8021Q / CONFIG_VLAN_8021Q_MODULE).
Signed-off-by: Parav Pandit <parav.pandit@emulex.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
1. Limit the max number of WQEs per QP reported when querying the
device, so that ib_create_qp() will not fail for a QP size that the
device claimed to support due to additional headroom WQEs being
allocated.
2. Limit qp resources accepted for ib_create_qp() to the limits
reported in ib_query_device(). In kernel space, make sure that the
limits returned to the caller following qp creation also lie within
the reported device limits. For userspace, report as before, and do
adjustment in libmlx4 (so as not to break ABI).
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Sagi Grimberg <sagig@mellanox.co.il>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Commit e605b743f3 ("IB/mlx4: Increase the number of vectors (EQs)
available for ULPs") didn't handle correctly the case where there
aren't enough MSI-X vectors to increase the number of EQs, so only the
legacy EQs are allocated. This results in an attempt to memset() to
zero the EQ table which was never allocated and a kernel crash.
Fix this by checking in the teardown flow if the table of EQs was ever
allocated. Also remove some unneeded setting to zero of the EQ
related fields in struct mlx4_ib_dev.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
When using rping -c -a 0.0.0.0 with iw_cxgb4, the system crashes when
rdma_connect() is called. ip_dev_find() will return NULL, but pdev is
accessed anyway.
Checking that pdev is NULL and returning -ENODEV prevents the system
from crashing.
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@linux.vnet.ibm.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Correct queue free count math for SQ, RQ for all hardware type.
Update user-kernel ABI interface.
Signed-off-by: Parav Pandit <parav.pandit@emulex.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
We need to use a different loop index for mlx4_counter_alloc() and for
device_create_file() iterations: the mlx4_counter_alloc() loop index
is used in the error flow to free counters.
If the same loop index is used for device_create_file() and, say, the
device_create_file() loop fails on the first iteration, the allocated
counters will not be freed.
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Enable IB ULPs to use a larger portion of the device EQs (which map to
IRQs). The mlx4_ib driver follows the mlx4_core framework of the EQs
to be divided among the device ports. In this scheme, for each IB
port, the number of allocated EQs follows the number of cores, subject
to other system constraints, such as number available MSI-X vectors.
Signed-off-by: Shlomo Pongratz <shlomop@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Using kfifos for ID management was limiting the number of QPs and
preventing NP384 MPI jobs. So replace it with a simple bitmap
allocator.
Remove IDs from the IDR tables before deallocating them. This bug was
causing the BUG_ON() in insert_handle() to fire because the ID was
getting reused before being removed from the IDR table.
Signed-off-by: Vipul Pandya <vipul@chelsio.com>
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Mike Marciniszyn
Betty Dall
Dan Carpenter
Parav Pandit
Mahesh Vardhamanaiah
Roland Dreier
Sagi Grimberg
Shlomo Pongratz
Thadeu Lima de Souza Cascardo
Devendra Naga
Vipul Pandya
Jack Morgenstein