The rdmavt QP has fields that are both needed for the control and data
path. Such mixed declaration caused to the very specific allocation flow
with kzalloc_node and SGE list embedded into the struct rvt_qp.
This patch separates QP creation to two: regular memory allocation for the
control path and specific code for the SGE list, while the access to the
later is performed through derefenced pointer.
Such pointer and its context are expected to be in the cache, so
performance difference is expected to be negligible, if any exists.
Link: https://lore.kernel.org/r/f66c1e20ccefba0db3c69c58ca9c897f062b4d1c.1627040189.git.leonro@nvidia.com
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
It is possible for the primary IPoIB network device associated with any
RDMA device to fail to join certain multicast groups preventing IPv6
neighbor discovery and possibly other network ULPs from working
correctly. The IPv4 broadcast group is not affected as the IPoIB network
device handles joining that multicast group directly.
This is because the primary IPoIB network device uses the pkey at ndex 0
in the associated RDMA device's pkey table. Anytime the pkey value of
index 0 changes, the primary IPoIB network device automatically modifies
it's broadcast address (i.e. /sys/class/net/[ib0]/broadcast), since the
broadcast address includes the pkey value, and then bounces carrier. This
includes initial pkey assignment, such as when the pkey at index 0
transitions from the opa default of invalid (0x0000) to some value such as
the OPA default pkey for Virtual Fabric 0: 0x8001 or when the fabric
manager is restarted with a configuration change causing the pkey at index
0 to change. Many network ULPs are not sensitive to the carrier bounce and
are not expecting the broadcast address to change including the linux IPv6
stack. This problem does not affect IPoIB child network devices as their
pkey value is constant for all time.
To mitigate this issue, change the default pkey in at index 0 to 0x8001 to
cover the predominant case and avoid issues as ipoib comes up and the FM
sweeps.
At some point, ipoib multicast support should automatically fix
non-broadcast addresses as it does with the primary broadcast address.
Fixes: 7724105686 ("IB/hfi1: add driver files")
Link: https://lore.kernel.org/r/20210715160445.142451.47651.stgit@awfm-01.cornelisnetworks.com
Suggested-by: Josh Collier <josh.d.collier@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@cornelisnetworks.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Leon Romanovsky says:
====================
Add ConnectX DCS offload support
This patchset from Lior adds support of DCI stream channel (DCS) support.
DCS is an offload to SW load balancing of DC initiator work requests.
A single DC QP initiator (DCI) can be connected to only one target at the
time and can't start new connection until the previous work request is
completed.
This limitation causes to delays when the initiator process needs to
transfer data to multiple targets at the same time.
====================
* branch 'mlx5_dcs':
RDMA/mlx5: Add DCS offload support
RDMA/mlx5: Separate DCI QP creation logic
net/mlx5: Add DCS caps & fields support
DCS is an offload to SW load balancing of DC initiator work requests.
A single DCI can be connected to only one target at the time and can't
start new connection until the previous work request is completed. This
limitation will cause to delay when the initiator process needs to
transfer data to multiple targets at the same time. The SW solution is to
use a process that handling and spreading the work request on many DCIs
according to destinations.
This feature is an offload to this process and coming to reduce the load
from the CPU and improve the performance.
Link: https://lore.kernel.org/r/491c2c2afdb5b07de7f03eab3f93cf0704549dbc.1624258894.git.leonro@nvidia.com
Reviewed-by: Meir Lichtinger <meirl@nvidia.com>
Signed-off-by: Lior Nahmanson <liorna@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This fields will be needed when adding a support for DCS offload
max_dci_stream_channels - maximum DCI stream channels supported per DCI.
max_dci_errored_streams - maximum DCI error stream channels
supported per DCI before a DCI move to error state.
Signed-off-by: Lior Nahmanson <liorna@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Currently the ICRC is generated as a u32 type and then forced to a __be32
and stored into the ICRC field in the packet. The actual type of the ICRC
is __be32. This patch replaces u32 by __be32 and eliminates the casts.
The computation is exactly the same as the original but the types are more
consistent.
Link: https://lore.kernel.org/r/20210707040040.15434-10-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
ib_query_port() calls device->ops.query_port() to get the port
attributes. The method of querying is device driver specific. The same
function calls device->ops.query_gid() to get the GID and extract the
subnet_prefix (gid_prefix).
The GID and subnet_prefix are stored in a cache. But they do not get
read from the cache if the device is an Infiniband device. The
following change takes advantage of the cached subnet_prefix.
Testing with RDBMS has shown a significant improvement in performance
with this change.
Link: https://lore.kernel.org/r/20210712122625.1147-4-anand.a.khoje@oracle.com
Signed-off-by: Anand Khoje <anand.a.khoje@oracle.com>
Signed-off-by: Haakon Bugge <haakon.bugge@oracle.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
kmap() is being deprecated and will break uses of device dax after PKS
protection is introduced.[1]
The use of kmap() in siw_tx_hdt() is all thread local therefore
kmap_local_page() is a sufficient replacement and will work with pgmap
protected pages when those are implemented.
siw_tx_hdt() tracks pages used in a page_array. It uses that array to
unmap pages which were mapped on function exit. Not all entries in the
array are mapped and this is tracked in kmap_mask.
kunmap_local() takes a mapped address rather than a page. Alter
siw_unmap_pages() to take the iov array to reuse the iov_base address of
each mapping. Use PAGE_MASK to get the proper address for kunmap_local().
kmap_local_page() mappings are tracked in a stack and must be unmapped in
the opposite order they were mapped in. Because segments are mapped into
the page array in increasing index order, modify siw_unmap_pages() to
unmap pages in decreasing order.
Use kmap_local_page() instead of kmap() to map pages in the page_array.
[1] https://lore.kernel.org/lkml/20201009195033.3208459-59-ira.weiny@intel.com/
Link: https://lore.kernel.org/r/20210624174814.2822896-1-ira.weiny@intel.com
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Bernard Metzler <bmt@zurich.ibm.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
kmap() is being deprecated and will break uses of device dax after PKS
protection is introduced.[1]
These uses of kmap() in the SIW driver are thread local. Therefore
kmap_local_page() is sufficient to use and will work with pgmap protected
pages when those are implemnted.
There is one more use of kmap() in this driver which is split into its own
patch because kmap_local_page() has strict ordering rules and the use of
the kmap_mask over multiple segments must be handled carefully.
Therefore, that conversion is handled in a stand alone patch.
Use kmap_local_page() instead of kmap() in the 'easy' cases.
[1] https://lore.kernel.org/lkml/20201009195033.3208459-59-ira.weiny@intel.com/
Link: https://lore.kernel.org/r/20210622061422.2633501-4-ira.weiny@intel.com
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
On idle session, because we do not do signal for heartbeat, it will
overflow the send queue after sometime.
To avoid that, we need to enable the signal for heartbeat. To do that, add
a new member signal_interval in rtrs_path, which will set min of
queue_depth and SERVICE_CON_QUEUE_DEPTH, and track it for both heartbeat
and IO, so the sq queue full accounting is correct.
Fixes: b38041d50a ("RDMA/rtrs: Do not signal for heatbeat")
Link: https://lore.kernel.org/r/20210712060750.16494-4-jinpu.wang@ionos.com
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Reviewed-by: Aleksei Marov <aleksei.marov@ionos.com>
Reviewed-by: Gioh Kim <gi-oh.kim@ionos.com>
Reviewed-by: Md Haris Iqbal <haris.iqbal@ionos.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
