The following sequence:
* Change queue pair state into IB_QPS_ERR.
* Post a work request on the queue pair.
Triggers the following race condition in the rdma_rxe driver:
* rxe_qp_error() triggers an asynchronous call of rxe_completer(), the function
that examines the QP send queue.
* rxe_post_send() posts a work request on the QP send queue.
If rxe_completer() runs prior to rxe_post_send(), it will drain the send
queue and the driver will assume no further action is necessary.
However, once we post the send to the send queue, because the queue is
in error, no send completion will ever happen and the send will get
stuck. In order to process the send, we need to make sure that
rxe_completer() gets run after a send is posted to a queue pair in an
error state. This patch ensures that happens.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Moni Shoua <monis@mellanox.com>
Cc: <stable@vger.kernel.org> # v4.8
Signed-off-by: Doug Ledford <dledford@redhat.com>
The initial assignment to mdev is redundant as mdev is re-assigned
later and the first assigned value is never read. Remove this
redundant assignment.
Cleans up clang warning:
drivers/infiniband/hw/mlx5/main.c:359:24: warning: Value stored
to 'mdev' during its initialization is never read
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Acked-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
In xmit, there is a skb_clone. This function copies the struct sk_buff.
And some parameters are changed to the new skb. Then the new skb is sent
while the old skb is freed.
While the function skb_clone is removed, the parameter changes are made on
the old skb, then the old skb is sent. It can also work well.
The following tests are made.
server client
--------- ---------
|1.1.1.1|<----rxe-channel--->|1.1.1.2|
--------- ---------
On server: rping -s -a 1.1.1.1 -v -C 1000 -S 512
On client: rping -c -a 1.1.1.1 -v -C 1000 -S 512
The kernel config CONFIG_DEBUG_KMEMLEAK is enabled on both server
and client.
This test runs for several hours. There is no memory leak and the whole
system can work well.
As the above network, the following tests are made.
Server: ibv_rc_pingpong -d rxe0 -g 1
Client: ibv_rc_pingpong -d rxe0 -g 1 1.1.1.1
The result on Server.
Before:
8192000 bytes in 0.88 seconds = 74.36 Mbit/sec
1000 iters in 0.88 seconds = 881.30 usec/iter
After:
8192000 bytes in 0.81 seconds = 81.15 Mbit/sec
1000 iters in 0.81 seconds = 807.62 usec/iter
The throughput is enhanced and the latency is reduced.
CC: Srinivas Eeda <srinivas.eeda@oracle.com>
CC: Joe Jin <joe.jin@oracle.com>
CC: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: Zhu Yanjun <yanjun.zhu@oracle.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The variable recv_sockets is only used in the file rxe_net.c. So
it is better to add static type to it.
CC: Srinivas Eeda <srinivas.eeda@oracle.com>
CC: Joe Jin <joe.jin@oracle.com>
CC: Junxiao Bi <junxiao.bi@oracle.com>
Signed-off-by: Zhu Yanjun <yanjun.zhu@oracle.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Merging in 12 patch series from Bart that required changes in the
current for-rc branch in order to apply cleanly.
Signed-off-by: Doug Ledford <dledford@redhat.com>
Since all I/O context state changes are already serialized, it is
not necessary to protect I/O context state changes with the I/O
context spinlock. Hence remove that spinlock.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
It is not necessary to obtain ioctx->spinlock when reading the ioctx
state. Since after removal of this locking only a single line remains,
inline the srpt_get_cmd_state() function.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Introduce a function for converting a GUID into an ASCII string. This
patch does not change any functionality.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Disabling an SRP target port causes the state of all QPs associated
with a port to be changed into IB_QPS_ERR. Avoid that this causes
one error message per I/O context to be reported.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
At least when running the ib_srpt driver on top of the rdma_rxe
driver it is easy to trigger a zero-length write completion in
the CH_DISCONNECTED state. Hence make the message that reports
this less noisy.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Make the ib_srpt driver use the IPv6 format for GIDs in log messages
to improve consistency of this driver with other RDMA kernel drivers.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Verify whether port numbers are in the expected range before using
these as an array index. Complain if a port number is out of range.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Since the SRQ event message is only useful for debugging purposes,
reduce its severity from "informational" to "debug".
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Rename rsp_size into max_rsp_size and SRPT_RQ_SIZE into MAX_SRPT_RQ_SIZE.
The new names better reflect the role of this member variable and constant.
Since the prefix "srp_" is superfluous in the context of the function
that creates an RDMA channel, rename srp_sq_size into sq_size.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This patch avoids that the following command reports any warnings:
scripts/kernel-doc -none drivers/infiniband/ulp/srpt/ib_srpt.h
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Avoid that warnings about missing parameter descriptions are reported
when building with W=1.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
When operating in dual port RoCE mode FW doesn't support steering for
raw QPs on the slave port. They still work on the master port, but
the user has no way of knowing which port is the master. The
capability is reported per device, not per port.
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Route performance query MADs to the correct mlx5_core_dev when using
dual port RoCE mode.
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
When in dual port mode setting a RoCE GID for any port flows through the
master ports mlx5_core_dev. Provide an interface to set the port when
sending this command.
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Update the counter interface for multiple ports. Some counter sets
always comes from the primary device.
Port specific counters should be accessed per mlx5_core_dev not always
through the IB master mdev.
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
When there are multiple ports for single IB(RoCE) device, support
debugfs entries to be available for each port.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Port operations must be routed to their native mlx5_core_dev. A
multiport RoCE device registers itself as having 2 ports even before a
2nd port is affiliated. If an unaffilated port is queried use capability
information from the master port, these values are the same.
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Because mlx5_ib_event can be called from atomic context move event
handling onto a workqueue. A mutex lock is required to get the IB device
for slave ports, so move event processing onto a work queue. When an IB
event is received, check if the mlx5_core_dev is a slave port, if so
attempt to get the IB device it's affiliated with. If found process the
event for that device, otherwise return.
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
When mlx5_ib_add is called determine if the mlx5 core device being
added is capable of dual port RoCE operation. If it is, determine
whether it is a master device or a slave device using the
num_vhca_ports and affiliate_nic_vport_criteria capabilities.
If the device is a slave, attempt to find a master device to affiliate it
with. Devices that can be affiliated will share a system image guid. If
none are found place it on a list of unaffiliated ports. If a master is
found bind the port to it by configuring the port affiliation in the NIC
vport context.
Similarly when mlx5_ib_remove is called determine the port type. If it's
a slave port, unaffiliate it from the master device, otherwise just
remove it from the unaffiliated port list.
The IB device is registered as a multiport device, even if a 2nd port is
not available for affiliation. When the 2nd port is affiliated later the
GID cache must be refreshed in order to get the default GIDs for the 2nd
port in the cache. Export roce_rescan_device to provide a mechanism to
refresh the cache after a new port is bound.
In a multiport configuration all IB object (QP, MR, PD, etc) related
commands should flow through the master mlx5_core_dev, other commands
must be sent to the slave port mlx5_core_mdev, an interface is provide
to get the correct mdev for non IB object commands.
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
When multiple RoCE ports are supported registration for events on
multiple netdevs is required. Refactor the event registration and
handling to support multiple ports.
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Remove use of the num_ports general capability throughout. The number of
ports will be variable in the future, and reported in a different way.
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
It always returns 0. Change return type to void.
Signed-off-by: Daniel Jurgens <danielj@mellanox.com>
Reviewed-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
A DC Target (DCT) QP is represented in the hardware as a unique object.
This object is created by CREATE_DCT command and destroyed by DESTROY_DCT
command. However, in the driver we describe it as a QP.
The hardware command that creates a DCT needs parameters that the verb
create_qp() does not provide. Those remaining parameters are provided
with the call to the verb modify_qp(). Therefore we delay the actual
creation of a DCT in the hardware until the stage of modify_qp() to RTR.
A support for query_qp() was added as well. It uses QUERY_DCT command to
retrieve the applicable fields.
Signed-off-by: Moni Shoua <monis@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
DC Initiator (DCI) QP is represented like any other QP in the hardware.
However, like any other transport QP there are attributes and settings
that are special to DCI QP and needs specific attention and care.
Make necessary changes to configure DCI QP.
Signed-off-by: Moni Shoua <monis@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The QP type IB_QPT_DRIVER doesn't describe the transport or the service
that the QP provides but those are known only to the hardware driver.
The actual type of the QP is stored in the hardware driver context (i.e.
mlx5_qp) under the field qp_sub_type.
Take the real QP type and any extra data that is required to create the QP
from the driver channel and modify the QP initial attributes before continuing
with create_qp().
Downstream patches from this series will add support for both DCI and
DCT driver QPs.
Signed-off-by: Moni Shoua <monis@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Reporting that a device doesn't support RoCE seems like a valuable piece
of information to have when trying to determine why a driver is not binding
to a device. Better to report this at info log level instead of requiring
a user to enable all debug messages in the driver.
Signed-off-by: Jonathan Toppins <jtoppins@redhat.com>
Acked-By: Devesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The cases for FDR/EDR signalling speed, were missing in
ib_rate_to_mult and mult_to_ib_rate giving wrong return values when
drivers convert static rate to/from inter-packet-delay.
Signed-off-by: Hans Westgaard Ry <hans.westgaard.ry@oracle.com>
Reviewed-by: Håkon Bugge <haakon.bugge@oracle.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
i40iw_wait_pe_ready is not called in an interrupt handler
nor holding a spinlock.
The function mdelay in it can be replaced with msleep,
to reduce busy wait.
Signed-off-by: Jia-Ju Bai <baijiaju1990@gmail.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Acked-by: Shiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This patch adds static trace for RNRNAK timer. Currently the output from
hrtimer static trace only shows the addresses of hrtimers in the system
and there is no easy way to correlate an RNRNAK timer with its entries in
the hrtimer trace. This patch adds the correlation among a QP, its RNRNAK
timer, and its entries in the hrtimer trace. This correlation will be
enormously helpful when debugging RNRNAK related issues. In addition, this
patch cleans up rvt_stop_rnr_timer() to be void while here.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The get_unit_name() function crafts a string based on the device name
and the device unit number. It then stores this in a static variable.
This has concurrency issues as can be seen with this log:
hfi1 0000:02:00.0: hfi1_1: read_idle_message: read idle message 0x203
hfi1 0000:01:00.0: hfi1_1: read_idle_message: read idle message 0x203
The PCI device ID (0000:02:00.0 vs. 0000:01:00.0) is correct for the
message, but the device string hfi1_1 is incorrect (it should be
hfi1_0 for the second log message).
Remove get_unit_name() function.
Instead, use the rvt accessor rvt_get_ibdev_name() to get the IB name
string.
Clean up any hfi1_early_xx calls that can now use the new path.
QIB has the same (qib_get_unit_name()) issue. Updating as necessary.
Remove qib_get_unit_name() function.
Update log message that has redundant device name.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
CQ allocation does not ensure that completion queue entries
and the completion queue structure are allocated on the correct
numa node.
Fix by allocating the rvt_cq and kernel CQ entries on the device node,
leaving the user CQ entries on the default local node. Also ensure
CQ resizes use the correct allocator when extending a CQ.
Reviewed-by: Sebastian Sanchez <sebastian.sanchez@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>
When an 8051 command times out, the entire DC block is restarted. During
the restart, the host interface version bit is set, which calls
do_8051_command() recursively. The host version bit needs to be set
before the link moves into polling, so the host version bit can be set
in set_local_link_attributes() instead. Thus, the 8051 command functions
can be simplied as a non-locking version (dd->dc8051_lock) of those
functions are no longer needed.
Fixes: 9be6a5d788 ("IB/hfi1: Prevent LNI out of sync by resetting host interface version")
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Normal receive queue allocation ensures that kernel receive queues
are allocated on the local numa node. Shared receive queues
do not behave the same way.
Ensure that kernel shared receive queues are allocated on the device
local node.
Reviewed-by: Sebastian Sanchez <sebastian.sanchez@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>
rdmavt has a down call to client drivers to retrieve a crafted card
name.
This name should be the IB defined name.
Rather than craft the name each time it is needed, simply retrieve
the IB allocated name from the IB device.
Update the function name to reflect its application.
Clean up driver code to match this change.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Currently the HFI and QIB drivers allow the IB core to assign a unit
number to the driver name string.
If multiple devices exist in a system, there is a possibility that the
device unit number and the IB core number will be mismatched.
Fix by using the driver defined unit number to generate the device
name.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
When the rdmavt's RNRNAK timer is fired, it tries to cancel the timer by
calling hrtimer_try_to_cancel(), which always returns -1 because the timer
is currently running. This patch removes this useless call.
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Make sure that the initiator port GUID is stored in ch->ini_guid.
Note: when initiating a connection sgid and dgid members in struct
sa_path_rec represent the source and destination GIDs. When accepting
a connection however sgid represents the destination GID and dgid the
source GID.
Fixes: commit 2bce1a6d22 ("IB/srpt: Accept GUIDs as port names")
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Unconditional locks/list and ODP srcu initialization should be done in
the INIT stage. Remove those from the CAPS stage and move them to the
proper stage.
Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The loopback stage only initializes a lock, move it to be in
the CAPS initialization phase and get rid loopback step completely.
Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Now that we have a stage just for hardware counters, move all relevant
initialization logic into one place.
Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Now that we have a stage just for ODP, move all relevant
initialization logic into one place.
Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Now that we have a stage just for RoCE/ETH, move all relevant
initialization logic into one place.
Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
