There are mis-match at counting inflight IO after changing the multipath
policy.
For example, we started fio test with round-robin policy and then we
changed the policy to min-inflight. IOs created under the RR policy is
finished under the min-inflight policy and inflight counter only
decreased. So the counter would be negative value. And also we started
fio test with min-inflight policy and changed the policy to the
round-robin. IOs created under the min-inflight policy increased the
inflight IO counter but the inflight IO counter was not decreased because
the policy was the round-robin when IO was finished.
So it should count IOs only if the IO is created under the min-inflight
policy. It should not care the policy when the IO is finished.
This patch adds a field mp_policy in struct rtrs_clt_io_req and stores the
multipath policy when an object of rtrs_clt_io_req is created. Then
rtrs-clt checks the mp_policy of only struct rtrs_clt_io_req instead of
the struct rtrs_clt.
Link: https://lore.kernel.org/r/20210806112112.124313-6-haris.iqbal@ionos.com
Signed-off-by: Gioh Kim <gi-oh.kim@ionos.com>
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Reviewed-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Md Haris Iqbal <haris.iqbal@ionos.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The IO performance test with fio after swapping the likely and unlikely
macros in all if-statement shows no difference. They do not help for the
performance of rtrs.
Thanks to Haakon Bugge for the test scenario.
The fio test did random read on 32 rnbd devices and 64 processes.
Test environment:
- Intel(R) Xeon(R) Gold 6130 CPU @ 2.10GHz
- 376G memory
- kernel version: 5.4.86
- gcc version: gcc (Debian 8.3.0-6) 8.3.0
- Infiniband controller: Mellanox Technologies MT27800 Family [ConnectX-5]
Test result:
- before swapping: IOPS=829k, BW=3239MiB/s
- after swapping: IOPS=829k, BW=3238MiB/s
- remove all (un)likely: IOPS=829k, BW=3238MiB/s
Link: https://lore.kernel.org/r/20210806112112.124313-5-haris.iqbal@ionos.com
Signed-off-by: Gioh Kim <gi-oh.kim@ionos.com>
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Reviewed-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Md Haris Iqbal <haris.iqbal@ionos.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Saeed Mahameed says:
====================
This pulls mlx5-next branch which includes patches already reviewed on
net-next and rdma mailing lists.
1) mlx5 single E-Switch FDB for lag
2) IB/mlx5: Rename is_apu_thread_cq function to is_apu_cq
3) Add DCS caps & fields support
We need this in net-next as multiple features are dependent on the
single FDB feature.
====================
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
* mellanox/mlx5-next:
net/mlx5: Lag, Create shared FDB when in switchdev mode
net/mlx5: E-Switch, add logic to enable shared FDB
net/mlx5: Lag, move lag destruction to a workqueue
net/mlx5: Lag, properly lock eswitch if needed
net/mlx5: Add send to vport rules on paired device
net/mlx5: E-Switch, Add event callback for representors
net/mlx5e: Use shared mappings for restoring from metadata
net/mlx5e: Add an option to create a shared mapping
net/mlx5: E-Switch, set flow source for send to uplink rule
RDMA/mlx5: Add shared FDB support
{net, RDMA}/mlx5: Extend send to vport rules
RDMA/mlx5: Fill port info based on the relevant eswitch
net/mlx5: Lag, add initial logic for shared FDB
net/mlx5: Return mdev from eswitch
IB/mlx5: Rename is_apu_thread_cq function to is_apu_cq
Qedr code is tightly coupled with existing both INIT transitions. Here,
during first INIT transition all variables are reset and the RESET state
is checked in post_recv() before any posting.
Commit dc70f7c3ed ("RDMA/cma: Remove unnecessary INIT->INIT transition")
exposed this bug.
So moving variables reset to qedr_set_common_qp_params() and also avoid
RESET state check for post_recv().
Link: https://lore.kernel.org/r/20210811051650.14914-1-pkushwaha@marvell.com
Signed-off-by: Michal Kalderon <mkalderon@marvell.com>
Signed-off-by: Ariel Elior <aelior@marvell.com>
Signed-off-by: Shai Malin <smalin@marvell.com>
Signed-off-by: Prabhakar Kushwaha <pkushwaha@marvell.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
If both eswitches are in switchdev mode and the uplink representors
are enslaved to the same bond device create a shared FDB configuration.
When moving to shared FDB mode not only the hardware needs be configured
but the RDMA driver needs to reconfigure itself.
When such change is done, unload the RDMA devices, configure the hardware
and load the RDMA representors.
When destroying the lag (can happen if a PCI function is unbinded,
driver is unloaded or by just removing a netdev from the bond) make sure
to restore the system to the previous state only if possible.
For example, if a PCI function is unbinded there is no need to load the
representors as the device is going away.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Shared FDB allows to direct traffic from all the vports in the HCA to a
single eswitch. In order to do that three things are needed.
1) Point the ingress ACL of the slave uplink to that of the master.
With this, wire traffic from both uplinks will reach the same eswitch
with the same metadata where a single steering rule can catch traffic
from both ports.
2) Set the FDB root flow table of the slave's eswitch to that of the
master. As this flow table can change dynamically make sure to
sync it on any set root flow table FDB command.
This will make sure traffic from SFs, VFs, ECPFs and PFs reach the
master eswitch.
3) Split wire traffic at the eswitch manager egress ACL so that it's
directed to the native eswitch manager. We only treat wire traffic
from both ports the same at the eswitch level. If such traffic wasn't
handled in the eswitch it needs to reach the right representor to be
processed by software. For example LACP packets should *always*
reach the right uplink representor for correct operation.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
If a netdev is removed from the lag the lag should be destroyed.
With downstream patches this might trigger a reconfiguration of
representors on a different eswitch and such we don't have the proper
locking to so from this path. Move the destruction to be done by the
workqueue.
As the destruction won't affect the netdev side it okay to do so.
The RDMA side will be reconfigured and it already coded to handle such
reconfiguration.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Currently when doing hardware lag we check the eswitch mode
but as this isn't done under a lock the check isn't valid.
As the code needs to sync between two different devices an extra
care is needed.
- When going to change eswitch mode, if hardware lag is active destroy it.
- While changing eswitch modes block any hardware bond creation.
- Delay handling bonding events until there are no mode changes in
progress.
- When attaching a new mdev to lag, block until there is no mode change
in progress. In order for the mode change to finish the interface lock
will have to be taken. Release the lock and sleep for 100ms to
allow forward progress. As this is a very rare condition (can happen if
the user unbinds and binds a PCI function while also changing eswitch
mode of the other PCI function) it has no real world impact.
As taking multiple eswitch mode locks is now required lockdep will
complain about a possible deadlock. Register a key per eswitch to make
lockdep happy.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
When two mlx5 devices are paired in switchdev mode, always offload the
send-to-vport rule to the peer E-Switch. This allows to abstract
the logic when this is really necessary (single FDB) and combine
the logic of both cases into one.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
This callback will allow to notify representors about relevant events
when in OFFLOADS mode. In downstream patches, this will be used to notify
about PAIR/UNPAIR devcom events.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
FTEs are added with mapped metadata which is saved per eswitch.
When uplink reps are bonded and we are in a single FDB mode,
we could fail to find metadata which was stored on one eswitch mapping
but not the other or with a different id.
To resolve this issue use shared mapping between eswitch ports.
We do not have any conflict using a single mapping, for a type,
between the ports.
Signed-off-by: Roi Dayan <roid@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Set the flow source param to local vport for the uplink rep
send-to-vport rule.
This will comply with the recent changes in SW steering that
use the flow source as an indication for the rule type - rx or tx.
Since the uplink send-to-vport rule is forwarding traffic to the wire
it has to indicate that it is an sx rule and can't use the any port
value in the flow source.
Signed-off-by: Ariel Levkovich <lariel@nvidia.com>
Reviewed-by: Roi Dayan <roid@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Shared FDB allows to create a single RDMA device that holds representors
from both eswitches. As shared FDB is only active when both uplink
representors are enslaved there is a single RDMA port that represents
both uplinks.
The number of ports is the number of vports on both eswitches minus one
as we only need 1 port for both uplinks.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
In shared FDB there is only one eswitch which is active and it receives
traffic from all representors and all vports in the HCA.
While the Ethernet representor will always reside on its native PF
the IB representor will not. Extend send to vport rule creation to
support such flows. Need to account for source vport that sends the
traffic (on which the representors resides) and the target eswitch
the traffic which reach.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
In shared FDB a single RDMA device can have representors that are
connected to two different eswitches. Use the right eswitch when
preparing the response to userspace.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
As shared FDB requires changes in two subsystems first expose the needed
core functions so the RDMA side can be changed.
mlx5_lag_is_master(): return true if a given mlx5 device is the lag master.
mlx5_lag_is_shared_fdb(): Returns true if the lag mode is shared FDB.
mlx5_lag_get_peer_mdev(): Return the peer mdev in lag.
The mentioned functions will be used by downstream patches in order
to add support for shared FDB for the RDMA side.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Export a function so users can retrieve the mellanox device that manages
the eswitch from the eswitch device.
Signed-off-by: Mark Bloch <mbloch@nvidia.com>
Reviewed-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
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>
