Merge branch 'rdma-netlink' into k.o/merge-test

Conflicts:
include/rdma/ib_verbs.h - Modified a function signature adjacent
to a newly added function signature from a previous merge

Signed-off-by: Doug Ledford <dledford@redhat.com>

Merge branches '32bit_lid' and 'irq_affinity' into k.o/merge-test

Conflicts:
drivers/infiniband/hw/mlx5/main.c - Both add new code
include/rdma/ib_verbs.h - Both add new code

Signed-off-by: Doug Ledford <dledford@redhat.com>

Merge tag 'rdma-next-2017-08-10' of git://git./linux/kernel/git/leon/linux-rdma into rdma-netlink

RDMA netlink infrastructure v2

Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/netlink: Export node_type

Add ability to get node_type for RDAM netlink users.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA/netlink: Provide port state and physical link state

Add port state and physical link state to the users of RDMA netlink.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA/netlink: Export LID mask control (LMC)

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA/netink: Export lids and sm_lids

According to the IB specification, the LID and SM_LID
are 16-bit wide, but to support OmniPath users, export
it as 32-bit value from the beginning.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA/netlink: Advertise IB subnet prefix

Add IB subnet prefix to the port properties exported
by RDMA netlink.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA/netlink: Export node_guid and sys_image_guid

Add Node GUID and system image GUID to the device properties
exported by RDMA netlink, to be used by RDMAtool.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA/netlink: Export FW version

Add FW version to the device properties exported
by RDMA netlink, to be used by RDMAtool.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA: Simplify get firmware interface

There is a need to forward FW version to user space
application through RDMA netlink. In order to make it safe, there
is need to declare nla_policy and limit the size of FW string.

The new define IB_FW_VERSION_NAME_MAX will limit the size of
FW version string. That define was chosen to be equal to
ETHTOOL_FWVERS_LEN, because many drivers anyway are limited
by that value indirectly.

The introduction of this define allows us to remove the string size
from get_fw_str function signature.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA/netlink: Expose device and port capability masks

The port capability mask is exposed to user space via sysfs interface,
while device capabilities are available for verbs only.

This patch provides those capabilities through netlink interface.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Implement nldev port doit callback

Provide ability to get specific to device and port information.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Add nldev port dumpit implementation

This patch implements the query interface to get all
ports data for the specific device.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Add nldev device doit implementation

Provide ability to query specific device.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Implement nldev device dumpit calback

This patch adds the ability to return all available devices
together with their properties.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Add nldev initialization flows

Add nldev init and exit flows to the RDMA/core.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Add netlink device definitions to UAPI

Introduce new defines to rdma_netlink.h, so the RDMA configuration tool
will be able to communicate with RDMA subsystem by using the shared defines.

The addition of new client (NLDEV) revealed the fact that we exposed by
mistake the RDMA_NL_I40IW define which is not backed by any RDMA netlink
by now and it won't be exposed in the future too. So this patch reuses
the value and deletes the old defines.

The NLDEV operates with objects. The struct ib_device has two straightforward
objects: device itself and ports of that device.

This brings us to propose the following commands to work on those objects:
 * RDMA_NLDEV_CMD_{GET,SET,NEW,DEL} - works on ib_device itself
 * RDMA_NLDEV_CMD_PORT_{GET,SET,NEW,DEL} - works on ports of specific ib_device

Those commands receive/return the device index (RDMA_NLDEV_ATTR_DEV_INDEX)
and port index (RDMA_NLDEV_ATTR_PORT_INDEX). For device object accesses,
the RDMA_NLDEV_ATTR_PORT_INDEX will return the maximum number of ports
for specific ib_device and for port access the actual port index.

The port index starts from 1 to follow RDMA/core internal semantics and
the sysfs exposed knobs.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Update copyright

Add Mellanox to the copyright header.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Convert LS to doit callback

RDMA_NL_LS protocol is actually does not dump anything,
but sets data and it should be handled by doit callback.

This patch actually converts RDMA_NL_LS to doit callback, while
preserving IWCM and RDMA_CM flows through netlink_dump_start().

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Reduce indirection access to cb_table

Introduce intermediate variable to store access to fields
of cb_table.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Add and implement doit netlink callback

The .doit callback is used by netlink core to differentiate
between get and set operations. Common convention is to use
that call for command operations like (SET, ADD, e.t.c.) and/or
access without NLF_M_DUMP flag.

This commit adds proper declaration and implementation
to RDMA netlink.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/core: Add and expose static device index

This patch adds static device index in similar fashion to
already available in netdev world (struct net->ifindex).

In downstream patches, the RDMA nelink will use this idx-to-ib_device
conversion, so as part of this commit, we are exposing the translation
function to be visible for IB/core users.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA/core: Add iterator over ib_devices

The coming nldev needs to iterate over all IB devices in the system
and in order to not expose the ib_devices list outside the devices.c,
it is necessary to provide function iterator.

Current version is written explicitly for nldev callback to avoid
over-engineering at this stage, but it can be easily extended for
other types.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Rename netlink callback struct

The RDMA netlink client infrastructure was removed and made obsolete.
The old infrastructure defined struct ibnl_client_cbs. Now that all
uses of this have been updated to the new infrastructure, rename the
struct to be compliant with the current stack naming standards:
struct rdma_nl_cbs.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Simplify and rename ibnl_chk_listeners

Make ibnl_chk_listeners function to be one line by removing
unneeded comparison.

Rename that function to be complaint to other functions in RDMA netlink.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Rename and remove redundant parameter from ibnl_multicast

The pointer to netlink header was not used in the ibnl_multicast
function, so let's remove it and simplify the function
signature.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Rename and remove redundant parameter from ibnl_unicast*

Netlink message header is not needed for unicast reply, hence remove it.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Simplify the put_msg and put_attr

Reuse standard macros to cancel the netlink message
in case of error.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/netlink: Add flag to consolidate common handling

Add ability to provide flags to control RDMA netlink callbacks
and convert addr.c and sa_query.c to be first users of such
infrastructure. It allows to move their CAP_NET_ADMIN checks
into netlink core.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>

RDMA/iwcm: Remove extra EXPORT_SYMBOLS

The iwcm exports functions which are not used outside of ib_core.
This patch simply removes these EXPORT_SYMBOLS.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Chien Tin Tung <chien.tin.tung@intel.com>

RDMA/iwcm: Remove useless check of netlink client validity

RDMA netlink implementation guarantees that supplied
client number is in allowed range.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Chien Tin Tung <chien.tin.tung@intel.com>

RDMA/netlink: Avoid double pass for RDMA netlink messages

The standard netlink_rcv_skb function skips messages without
NLM_F_REQUEST flag in it, while SA netlink client issues them.

In commit bc10ed7d3d19 ("IB/core: Add rdma netlink helper functions")
the local function was introduced to allow such messages.

This led to double pass for every incoming message.

In this patch, we unify that local implementation and netlink_rcv_skb
functions, so there will be no need for double pass anymore.

As a outcome, this combined function gained more strict check
for NLM_F_REQUEST flag and it is now allowed for SA pathquery
client only.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA/netlink: Remove redundant owner option for netlink callbacks

Owner field is not needed to be set because netlink is part of ib_core
which will be unloaded last after all other modules are unloaded.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

RDMA/netlink: Remove netlink clients infrastructure

RDMA netlink has a complicated infrastructure for dynamically
registering and de-registering netlink clients to the NETLINK_RDMA
group. The complicated portion of this code is not widely used because
2 of the 3 current clients are statically compiled together with
netlink.c. The infrastructure, therefore, is deemed overkill.

Refactor the code to eliminate the dynamically added clients. Now all
clients are pre-registered in a client array at compile time, and at run
time they merely check-in with the infrastructure to pass their callback
table for inclusion in the pre-sized client array.

This also allows for future cleanups and removal of unneeded code in the
iwcm* netlink handler.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Chien Tin Tung <chien.tin.tung@intel.com>

RDMA/core: Add wait/retry version of ibnl_unicast

Add a wait/retry version of ibnl_unicast, ibnl_unicast_wait,
and modify ibnl_unicast to not wait/retry.  This eliminates
the undesirable wait for future users of ibnl_unicast.

Change Portmapper calls originating from kernel to user-space
to use ibnl_unicast_wait and take advantage of the wait/retry
logic in netlink_unicast.

Signed-off-by: Mustafa Ismail <mustafa.ismail@intel.com>
Signed-off-by: Chien Tin Tung <chien.tin.tung@intel.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>

nvme-rdma: use intelligent affinity based queue mappings

Use the generic block layer affinity mapping helper. Also,
limit nr_hw_queues to the rdma device number of irq vectors
as we don't really need more.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Doug Ledford <dledford@redhat.com>

block: Add rdma affinity based queue mapping helper

Like pci and virtio, we add a rdma helper for affinity
spreading. This achieves optimal mq affinity assignments
according to the underlying rdma device affinity maps.

Reviewed-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Doug Ledford <dledford@redhat.com>

mlx5: support ->get_vector_affinity

Simply refer to the generic affinity mask helper.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Acked-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/core: expose affinity mappings per completion vector

This will allow ULPs to intelligently locate threads based
on completion vector cpu affinity mappings. In case the
driver does not expose a get_vector_affinity callout, return
NULL so the caller can maintain a fallback logic.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Håkon Bugge <haakon.bugge@oracle.com>
Acked-by: Doug Ledford <dledford@redhat.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Doug Ledford <dledford@redhat.com>

mlx5: move affinity hints assignments to generic code

generic api takes care of spreading affinity similar to
what mlx5 open coded (and even handles better asymmetric
configurations). Ask the generic API to spread affinity
for us, and feed him pre_vectors that do not participate
in affinity settings (which is an improvement to what we
had before).

The affinity assignments should match what mlx5 tried to
do earlier but now we do not set affinity to async, cmd
and pages dedicated vectors.

Also, remove mlx5e_get_cpu and introduce mlx5e_get_node
(used for allocation purposes) and mlx5_get_vector_affinity
(for indirection table construction) as they provide the needed
information. Luckily, we have generic helpers to get cpumask
and node given a irq vector. mlx5_get_vector_affinity will
be used by mlx5_ib in a subsequent patch.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Acked-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Doug Ledford <dledford@redhat.com>

mlx5e: don't assume anything on the irq affinity mappings of the device

mlx5e currently assumes that irq affinity is really spread first
irq vectors across device home node cpus, with the new generic affinity
mappings this is no longer the case, hence mlxe should not rely on
this anymore.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Doug Ledford <dledford@redhat.com>

mlx5: convert to generic pci_alloc_irq_vectors

Now that we have a generic code to allocate an array
of irq vectors and even correctly spread their affinity,
correctly handle cpu hotplug events and more, were much
better off using it.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Acked-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/CM: Set appropriate slid and dlid when handling CM request

If extended LIDs are being used, a connection request contains
OPA GIDs in them. Extract the lids from the OPA gids and populate
slid/dlid fields in the path records that are created when handling
a connection request.

Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Don Hiatt <don.hiatt@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/CM: Create appropriate path records when handling CM request

When handling an incoming conection request, ib_cm creates
either an IB or an OPA path record based on the gid field
in the request.

Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Don Hiatt <don.hiatt@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/CM: Add OPA Path record support to CM

Add OPA path record support to the Connection Manager.

Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Change wc.slid from 16 to 32 bits

slid field in struct ib_wc is increased to 32 bits.
This enables core components to use larger LIDs if needed.
The user ABI is unchanged and return 16 bit values when queried.

Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Change port_attr.sm_lid from 16 to 32 bits

sm_lid field in struct ib_port_attr is increased to 32 bits. This
enables core components to use larger LIDs if needed.
The user ABI is unchanged and return 16 bit values when queried.

Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Change port_attr.lid size from 16 to 32 bits

lid field in struct ib_port_attr is increased to 32 bits. This enables core
components to use larger LIDs if needed.
The user ABI is unchanged and return 16 bit values when queried.

Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mad: Change slid in RMPP recv from 16 to 32 bits

MAD RMPP contains slid field which is 16 bits in
length, increase it to 32 bits.

Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/IPoIB: Increase local_lid to 32 bits

IPoIB contains local_lid field which is 16 bits in
length, increase it to 32 bits.

Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/srpt: Increase lid and sm_lid to 32 bits

srpt contains lid and sm_lid fields which are 16 bits in
length, increase them to 32 bits.

Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Don Hiatt <don.hiatt@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Convert ah_attr from OPA to IB when copying to user

OPA address handle atttibutes that have 32 bit LIDs would have to
be converted to IB address handle attribute with the LID field
programmed in the GID before copying to user space.

Signed-off-by: Dasaratharaman Chandramouli <dasaratharaman.chandramouli@intel.com>
Reviewed-by: Don Hiatt <don.hiatt@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Always perform offline transition

Always initiate an offline transition request
when a link down occurs. The firmware will
use this request to confirm that the driver
has seen the link down message. A host version
is set to indicate this driver behavior to the
firmware.

Reviewed-by: Dean Luick <dean.luick@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>

IB/hfi1: Prevent link down request double queuing

When link interrupts occur, multiple link down requests
could be queued up when only one is needed. This could get
the hfi1 out of sync with its link partner during LNI.

Only allow one link down request to be queued at any one time.

Reviewed-by: Dean Luick <dean.luick@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>

IB/hfi1: Create workqueue for link events

Currently, link down interrupts queue link entries
on a workqueue intended for sending events only.
Create a workqueue for queuing link events.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/{rdmavt, hfi1, qib}: Fix panic with post receive and SGE compression

The server side of qperf panics as follows:

[242446.336860] IP: report_bug+0x64/0x10
[242446.341031] PGD 1c0c067
[242446.341032] P4D 1c0c067
[242446.343951] PUD 1c0d063
[242446.346870] PMD 8587ea067
[242446.349788] PTE 800000083e14016
[242446.352901]
[242446.358352] Oops: 0003 [#1] SM
[242446.437919] CPU: 1 PID: 7442 Comm: irq/92-hfi1_0 k Not tainted 4.12.0-mam-asm #1
[242446.446365] Hardware name: Intel Corporation S2600WT2/S2600WT2, BIOS SE5C610.86B.01.01.0018.C4.072020161249 07/20/201
[242446.458397] task: ffff8808392d2b80 task.stack: ffffc9000664000
[242446.465097] RIP: 0010:report_bug+0x64/0x10
[242446.469859] RSP: 0018:ffffc900066439c0 EFLAGS: 0001000
[242446.475784] RAX: ffffffffa06647e4 RBX: ffffffffa06461e1 RCX: 000000000000000
[242446.483840] RDX: 0000000000000907 RSI: ffffffffa0675040 RDI: ffffffffffff740
[242446.491897] RBP: ffffc900066439e0 R08: 0000000000000001 R09: 000000000000025
[242446.499953] R10: ffffffff81a253df R11: 0000000000000133 R12: ffffc90006643b3
[242446.508010] R13: ffffffffa065bbf0 R14: 00000000000001e5 R15: 000000000000000
[242446.516067] FS:  0000000000000000(0000) GS:ffff88085f640000(0000) knlGS:000000000000000
[242446.525191] CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003
[242446.531698] CR2: ffffffffa06647ee CR3: 0000000001c09000 CR4: 00000000001406e
[242446.539756] Call Trace
[242446.542582]  fixup_bug+0x2c/0x5
[242446.546277]  do_trap+0x12b/0x18
[242446.549972]  do_error_trap+0x89/0x11
[242446.554171]  ? hfi1_copy_sge+0x271/0x2b0 [hfi1
[242446.559324]  ? ttwu_do_wakeup+0x1e/0x14
[242446.563795]  ? ttwu_do_activate+0x77/0x8
[242446.568363]  do_invalid_op+0x20/0x3
[242446.572448]  invalid_op+0x1e/0x3
[242446.576247] RIP: 0010:hfi1_copy_sge+0x271/0x2b0 [hfi1
[242446.582075] RSP: 0018:ffffc90006643be8 EFLAGS: 0001004
[242446.587999] RAX: 0000000000000000 RBX: ffff88083e0fa240 RCX: 000000000000000
[242446.596058] RDX: 0000000000000000 RSI: ffff880842508000 RDI: ffff88083e0fa24
[242446.604116] RBP: ffffc90006643c28 R08: 0000000000000000 R09: 000000000000000
[242446.612172] R10: ffffc90009473640 R11: 0000000000000133 R12: 000000000000000
[242446.620228] R13: 0000000000000000 R14: 0000000000002000 R15: ffff88084250800
[242446.628293]  ? hfi1_copy_sge+0x1a1/0x2b0 [hfi1
[242446.633449]  hfi1_rc_rcv+0x3da/0x1270 [hfi1
[242446.638312]  ? sc_buffer_alloc+0x113/0x150 [hfi1
[242446.643662]  hfi1_ib_rcv+0x1c9/0x2e0 [hfi1
[242446.648428]  process_receive_ib+0x19a/0x270 [hfi1
[242446.653866]  ? process_rcv_qp_work+0xd2/0x160 [hfi1
[242446.659505]  handle_receive_interrupt_nodma_rtail+0x184/0x2e0 [hfi1
[242446.666693]  ? irq_finalize_oneshot+0x100/0x10
[242446.671846]  receive_context_thread+0x1b/0x140 [hfi1
[242446.677576]  irq_thread_fn+0x1e/0x4
[242446.681659]  irq_thread+0x13c/0x1b
[242446.685646]  ? irq_forced_thread_fn+0x60/0x6
[242446.690604]  kthread+0x112/0x15
[242446.694298]  ? irq_thread_check_affinity+0xe0/0xe
[242446.699738]  ? kthread_park+0x60/0x6
[242446.703919]  ? do_syscall_64+0x67/0x15
[242446.708292]  ret_from_fork+0x25/0x3
[242446.712374] Code: 63 78 04 44 0f b7 70 08 41 89 d0 4c 8d 2c 38 41 83 e0 01 f6 c2 02 74 17 66 45 85 c0 74 11 f6 c2 04 b9 01 00 00 00 75 bb 83 ca 04 <66> 89 50 0a 66 45 85 c0 74 52 0f b6 48 0b 41 0f b7 f6 4d 89 e0
[242446.733527] RIP: report_bug+0x64/0x100 RSP: ffffc900066439c
[242446.739935] CR2: ffffffffa06647e
[242446.743763] ---[ end trace 0e90a20d0aa494f7 ]--

The root cause is that the qib/hfi1 post receive call to rvt_lkey_ok()
doesn't interpret the new return value from rvt_lkey_ok() properly
leading to an mr reference count underrun.

Additionally, remove an unused argument in rvt_sge_adjacent()
aw well as an unneeded incr local in rvt_post_one_wr().

Fixes: Commit 14fe13fcd3af ("IB/rdmavt: Compress adjacent SGEs in rvt_lkey_ok()")
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>

IB/hfi1: Disambiguate corruption and uninitialized error cases

The error messages when checksum validation of the platform
configuration fields populated into the ASIC scratch registers fails are
ambiguous. Disambiguate them.

Reviewed-by: Jakub Byczkowski <jakub.byczkowski@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Jan Sokolowski <jan.sokolowski@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Only set fd pointer when base context is completely initialized

The allocate_ctxt() function adds the context to the fd data structure.
Since the context is not completely initialized, this can cause confusion
as to whether the context is valid or not.

Move the fd reference from allocate_ctxt() to setup_base_ctxt().
Update the necessary functions to be aware of this move.

Reviewed-by: Sebastian Sanchez <sebastian.sanchez@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>

IB/hfi1: Do not enable disabled port on cable insert

Fix issue where a disabled port can be enabled by
inserting a cable. The port should be explicitly
enabled instead.

Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Reviewed-by: Jakub Byczkowski <jakub.byczkowski@intel.com>
Signed-off-by: Jan Sokolowski <jan.sokolowski@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Harden state transition to Armed and Active

There is a window that allows other threads to read state of
'host_link_state' as a new, before the hardware actual state is set.
This patch closes the window by indicating a new state only after
hardware transition is complete.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Alex Estrin <alex.estrin@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Split copy_to_user data copy for better security

A copy_to_user() call assumes that two members of a data structure
are sequential.  Since this may not always be true, separate the copies
to ensure a safe copy.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@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>

IB/hfi1: Verify port data VLs credits on transition to Armed

There is a window where the FM can read the buffer control table
and decide not to program buffers. When a port goes down, the code
clears the table and if it is not programmed, posted SDMA descriptors
will never complete due to no buffer credits.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Alex Estrin <alex.estrin@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Move saving PCI values to a separate function

During PCIe initialization some registers' values from
PCI config space are saved in order to restore them later
(i.e. after reset). Restoring those value is done by a
function called restore_pci_variables, while saving them
is put directly into function hfi1_pcie_ddinit.
Move saving values to a separate function in the image
of restoring functionality.

Reviewed-by: Jakub Byczkowski <jakub.byczkowski@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Bartlomiej Dudek <bartlomiej.dudek@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Fix initialization failure for debug firmware

Loading debug signed firmware fails if started immediately after
failed attempt to load production firmware. A short delay is
required so add about a 100us delay after RSA check failure.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Easwar Hariharan <easwar.hariharan@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Jakub Byczkowski <jakub.byczkowski@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Fix code consistency for if/else blocks in chip.c

Code structure is not consistent for if/else blocks and break
instructions in set_link_state for case HLS_UP_INIT. Physical
state uses break in case of an error and if/else blocks for
logical use cases. These blocks should be implemented consistently.

Reviewed-by: Jakub Byczkowski <jakub.byczkowski@intel.com>
Signed-off-by: Jan Sokolowski <jan.sokolowski@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Send MAD traps until repressed

A trap should be sent to the FM until the FM sends a repress message.
This is in line with the IBTA 13.4.9.

Add the ability to resend traps until a repress message is received.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Michael N. Henry <michael.n.henry@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>

IB/hfi1: Pass the context pointer rather than the index

The hfi1_rcvctrl() function receives an index which it then converts
to an rcd.  Since most functions have the rcd, use that instead.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Sebastian Sanchez <sebastian.sanchez@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>

IB/hfi1: Use context pointer rather than context index

The hfi1_<set|clear>_ctxt_<j|p>key functions take a context index and
look up the context based on that index.

Since the context index is being retrieved from the context, this
doesn't seem optimal.

Pass the context pointer for use, rather than the context index.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@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>

IB/hfi1: Size rcd array index correctly and consistently

The array index for the rcd array is sized several different ways
throughout the code.

Use the user interface size (u16) as the standard size and update the
necessary code to reflect this.

u16 is large enough for the largest amount of supported contexts.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@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>

IB/hfi1: Remove unused user context data members

Several data members of the user context have become unused over time.
Cleaning them up.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
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>

IB/hfi1: Assign context does not clean up file descriptor correctly on error

In the error path for context allocation, the file descriptor pointer
should not point to a context when an error occurs.

Clean up the appropriate references on error.

Fixes: Commit 62239fc6e5545b2e59f83dfbc5db231a81f37a45 ("IB/hfi1: Clean up on context initialization failure")
Reviewed-by: Ira Weiny <ira.weiny@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>

IB/hfi1: Serve the most starved iowait entry first

When an egress resource(SDMA descriptors, pio credits) is not available,
a sending thread will be put on the resource's wait queue. When the
resource becomes available again, up to a fixed number of sending threads
can be awakened sequentially and removed from the wait queue, depending
on the number of waiting threads and the number of free resources. Since
each awakened sending thread will send as many packets as possible, it
is highly likely that the first sending thread will consume all the
egress resources. Subsequently, it will be put back to the end of the wait
queue. Depending on the timing when the later sending threads wake up,
they may not be able to send any packet and be again put back to the end
of the wait queue sequentially, right behind the first sending thread.
This starvation cycle continues until some sending threads exceed their
retry limit and consequently fail.

This patch fixes the issue by two simple approaches:
(1) Any starved sending thread will be put to the head of the wait queue
while a served sending thread will be put to the tail;
(2) The most starved sending thread will be served first.

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>

IB/hfi1: Fix bar0 mapping to use write combining

When the debugpat kernel boot flag is turned on the following
traces are printed:

[ 1884.793168] x86/PAT: Overlap at 0x90000000-0x92000000
[ 1884.803510] x86/PAT: reserve_memtype added [mem 0x91200000-0x9127ffff],
track uncached-minus, req write-combining, ret uncached-minus
[ 1884.818167] hfi1 0000:05:00.0: hfi1_0: WC Remapped RcvArray:
ffffc9000a980000

The ioremap_wc() clearly is not returning a write combining mapping due
to an overlap where the RcvArray is mapped in a uncached mapping prior
to creating the proposed write combining mapping.

The patch replaces the single base register for uncached CSRs that
used to overlap the RcvArray with two mappings.   One, kregbase1, from the
bar0 up to the RcvArray and another, kregbase2, from the end of the
RcvArray to the pio send buffer space.  A new dd field, base2_start,
is used to convert the zero-based offset in the CSR routines to the
correct kregbase1/kregbase2 mapping.  A single direct write of the
RcvArray CSRs is replaced with hfi1_put_tid() to insure correct access
using the new disjoint mapping.

Additionally, the kregend field is deleted since it is only ever written.

patdebug now shows the RcvArray as write combining:
[   35.688990] x86/PAT: reserve_memtype added [mem 0x91200000-0x9127ffff],
track write-combining, req write-combining, ret write-combining

To insulate from any potential issues with write combining, all
writeq are now flushed in hfi1_put_tid() and rcv_array_wc_fill().

Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@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>

IB/hfi1: Check return values from PCI config API calls

Ensure that return values from kernel PCI config access
API calls in HFI driver are checked and react properly if
they are not expected (i.e. not successful).

Reviewed-by: Jakub Byczkowski <jakub.byczkowski@intel.com>
Signed-off-by: Bartlomiej Dudek <bartlomiej.dudek@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hns: include linux/interrupt.h

I ran into this build error on linux-next:

drivers/infiniband/hw/hns/hns_roce_eq.c:477:8: error: unknown type name 'irqreturn_t'
 static irqreturn_t hns_roce_msi_x_interrupt(int irq, void *eq_ptr)
        ^~~~~~~~~~~
drivers/infiniband/hw/hns/hns_roce_eq.c: In function 'hns_roce_msi_x_interrupt':
drivers/infiniband/hw/hns/hns_roce_eq.c:485:9: error: implicit declaration of function 'IRQ_RETVAL'; did you mean 'BPF_RVAL'? [-Werror=implicit-function-declaration]
  return IRQ_RETVAL(int_work);

I have bisected this to a seemingly unrelated change that happened
to remove some indirect header inclusions. Simply including the
required header explicitly fixes the build failure.

Fixes: 09c7570480f7 ("xfrm: remove flow cache")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/hns: fix build regression

The 0day build system flags implicit includes as errors.  A patch from
Matan Barak to allow hns_roce, an aarch64 specific RDMA driver, to be
built on other arches, but it resulted in build regressions.  The
problem is that hns_roce_device.h needs a definition for __raw_writeq
but did not have an include to provide it.  Add <linux/io.h> as an
include to resolve the issue.

Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hns: fix semicolon.cocci warnings

drivers/infiniband/hw/hns/hns_roce_eq.c:295:3-4: Unneeded semicolon

 Remove unneeded semicolon.

Generated by: scripts/coccinelle/misc/semicolon.cocci

Fixes: 7d1b6a678e0b ("IB/hns: Support compile test for hns RoCE driver")
CC: Matan Barak <matanb@mellanox.com>
Signed-off-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hns: fix returnvar.cocci warnings

drivers/infiniband/hw/hns/hns_roce_hw_v1.c:2026:5-8: Unneeded variable: "ret". Return "0" on line 2046

 Remove unneeded variable used to store return value.

Generated by: scripts/coccinelle/misc/returnvar.cocci

Fixes: 7d1b6a678e0b ("IB/hns: Support compile test for hns RoCE driver")
CC: Matan Barak <matanb@mellanox.com>
Signed-off-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hns: fix boolreturn.cocci warnings

drivers/infiniband/hw/hns/hns_roce_qp.c:802:9-10: WARNING: return of 0/1 in function 'hns_roce_wq_overflow' with return type bool

 Return statements in functions returning bool should use
 true/false instead of 1/0.
Generated by: scripts/coccinelle/misc/boolreturn.cocci

Fixes: 7d1b6a678e0b ("IB/hns: Support compile test for hns RoCE driver")
CC: Matan Barak <matanb@mellanox.com>
Signed-off-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hns: Support compile test for hns RoCE driver

Compiling the hns RoCE driver requires ARM architecture.
In order to simplify development of IB/core, support
compile test. Add the necessary includes for that too.

Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/cma: Fix default RoCE type setting

The initial patch for changing the stack to use RoCEv2 GIDs by default
set the CMA_PREFERRED_ROCE_GID_TYPE to an incorrect value.  Instead of
an absolute value, we needed to set the right bit in a bitmask.  Correct
the default setting so we use RoCEv2 by default.

Fixes: 63a5f483af0e (IB/cma: Set default gid type to RoCEv2)
Signed-off-by: Doug Ledford <dledford@redhat.com>

Merge branch 'misc' into k.o/for-next

RDMA/qedr: notify user application of supported WIDs

The number of supported WIDs, if they are supported at all, can be
limited due to resources. Notifying the user space application the
number of available WIDs allows it to utilize them correctly.

Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/qedr: notify user application if DPM is supported

Direct Packet Mode support may be disabled, e.g, due to limited
resources. Notifying the user application prevents wasting cycles
on attempting to send these kind of packets.

Signed-off-by: Ram Amrani <Ram.Amrani@cavium.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

Merge branches 'rxe' and 'mlx' into k.o/for-next

net/mlx5: fix spelling mistake: "alloated" -> "allocated"

Trivial fix to spelling mistake in mlx5_ib_dbg message

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Add receiving queue info to qp_stats

This patch adds qp->s_ack_queue indices and size to qp_stats printout.
This information will provide information about the receiving side.

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>

IB/mlx4: Expose RSS capabilities

As a final step to support RSS feature, expose the RSS capabilities in
query device verb.

It includes:
- Max rwq indirection tables.
- Max rwq indirection table size.
- Supported qp types.

Signed-off-by: Guy Levi <guyle@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx4: Add support for RSS QP

Add support to work with a RSS QP by using an indirection table object
upon QP creation. Other related QP verbs (e.g. modify/destroy/query) were
updated as well for that QP mode.

Notes:
- The RX hash properties are supplied as driver private data.
- The RSS QP port is used on the associated WQs in its indirection
  table. Applying different ports during WQ life time is not allowed.
- The expected RSS QP flow is: create, modify(RST->INIT),
  modify(RST->RTR), destroy.

Signed-off-by: Guy Levi <guyle@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx4: Add support for WQ indirection table related verbs

To enable RSS functionality the IB indirection table object (i.e.
ib_rwq_ind_table) should be used.
This patch implements the related verbs as of create and destroy an
indirection table.

In downstream patches the indirection table will be used as part of RSS
QP creation.

Signed-off-by: Guy Levi <guyle@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx4: Add support for WQ related verbs

Support create/modify/destroy WQ related verbs.

The base IB object to enable RSS functionality is a WQ (i.e. ib_wq).
This patch implements the related WQ verbs as of create, modify and
destroy.

In downstream patches the WQ will be used as part of an indirection
table (i.e. ib_rwq_ind_table) to enable RSS QP creation.

Notes:
ConnectX-3 hardware requires consecutive WQNs list as receive descriptor
queues for the RSS QP. Hence, the driver manages consecutive ranges lists
per context which the user must respect.
Destroying the WQ does not return its WQN back to its range for
reusing. However, destroying all WQs from the same range releases the
range and in turn releases its WQNs for reusing.

Since the WQ object is not a natural object in the hardware, the driver
implements the WQ by the hardware QP.

As such, the WQ inherits its port from its RSS QP parent upon its
RST->INIT transition and by that time its state is applied to the
hardware.

Signed-off-by: Guy Levi <guyle@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

(IB, net)/mlx4: Add resource utilization support

Adding visibility of resource usage of QPs, CQs and counters used by
virtual functions. This feature will be used to give the PF administrator
more data while debugging VF status. Usage info was added to ALLOC_RES
command, to notify the PF if the resource which is being reserved or
allocated for the VF will be used by kernel driver or by user verbs.

Updated reservation and allocation functions of QP, CQ and counter with
additional usage parameter.

Signed-off-by: Moshe Shemesh <moshe@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx4: Add inline-receive support

When inline-receive is enabled, the HCA may write received
data into the receive WQE.

Inline-receive is enabled by setting its matching bit in
the QP context and each single-packet message with payload
not exceeding the receive WQE size will be delivered to
the WQE.

The completion report will indicate that the payload was placed to the WQE.

It includes:
1) Return maximum supported size of inline-receive by the hardware
in query_device vendor's data part.
2) Enable the feature when requested by the vendor data input.

Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx5: Expose extended error counters

This patch adds below requester and responder side error counters,
which will be exposed by hardware counters interface and are supported
as part of query Q counters command extension.

 +---------------------------+-------------------------------------+
 |      Name                 |           Description               |
 |---------------------------+-------------------------------------|
 |resp_local_length_error    | Number of times responder detected  |
 |                           | local length errors                 |
 |---------------------------+-------------------------------------|
 |resp_cqe_error             | Number of CQEs completed with error |
 |                           | at responder                        |
 |---------------------------+-------------------------------------|
 |req_cqe_error              | Number of CQEs completed with error |
 |                           | at requester                        |
 |---------------------------+-------------------------------------|
 |req_remote_invalid_request | Number of times requester detected  |
 |                           | remote invalid request error        |
 |---------------------------+-------------------------------------|
 |req_remote_access_error    | Number of times requester detected  |
 |                           | remote access error                 |
 |---------------------------+-------------------------------------|
 |resp_remote_access_error   | Number of times responder detected  |
 |                           | remote access error                 |
 |---------------------------+-------------------------------------|
 |resp_cqe_flush_error       | Number of CQEs completed with       |
 |                           | flushed with error at responder     |
 |---------------------------+-------------------------------------|
 |req_cqe_flush_error        | Number of CQEs completed with       |
 |                           | flushed with error at requester     |
 +---------------------------+-------------------------------------+

Signed-off-by: Parav Pandit <parav@mellanox.com>
Reviewed-by: Daniel Jurgens <danielj@mellanox.com>
Reviewed-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx5: Fix existence check for extended address vector

The extended address vector is the highest bit in be32 variable,
but it was compared with the lowest. This patch fixes the endianness
of that check and removes already declared define.

Fixes: 17d2f88f92ce ("IB/mlx5: Add ODP atomics support")
Reviewed-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx5: Fix cached MR allocation flow

When we have a miss in one order of the mkey cache, we try to get
an mkey from a higher order.

We still need to check that the higher order can be used with UMR
before using it. Otherwise, we will get an mkey with 0 entries and
the post send operation that is used to fill it will complete with
the following error:

mlx5_0:dump_cqe:275:(pid 0): dump error cqe
00000000 00000000 00000000 00000000
00000000 00000000 00000000 00000000
00000000 0f007806 25000025 49ce59d2

Fixes: 49780d42dfc9 ("IB/mlx5: Expose MR cache for mlx5_ib")
Cc: <stable@vger.kernel.org> # v4.10+
Signed-off-by: Majd Dibbiny <majd@mellanox.com>
Reviewed-by: Ilya Lesokhin <ilyal@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx5: Report RX checksum capabilities for IPoIB

Report RX checksum capabilities when 'ipoib_enhanced_offloads'
capabilities are set and support checksum.

Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

net/mlx5: Report enhanced capabilities for IPoIB

Report 'ipoib_enhanced_offloads' capabilities from
the core layer, it will be used in the next patch from this series.

Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx5: Add multicast flow steering support for underlay QP

In order to add multicast flow steering support, there is need
to block the attaching of mcg flow for underlay QP, recognize
multicast IB_FLOW_SPEC_IPV4 based on its IP and enable
creating/destroying flow for IB layer.

Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>