IB/core: Make device counter infrastructure dynamic

In practice, each RDMA device has a unique set of counters that the
hardware implements.  Having a central set of counters that they must
all adhere to is limiting and causes many useful counters to not be
available.

Therefore we create a dynamic counter registration infrastructure.

The driver must implement a stats structure allocation routine, in
which the driver must place the directory name it wants, a list of
names for all of the counters, an array of u64 counters themselves,
plus a few generic configuration options.

We then implement a core routine to create a sysfs file for each
of the named stats elements, and a core routine to retrieve the
stats when any of the sysfs attribute files are read.

To avoid excessive beating on the stats generation routine in the
drivers, the core code also caches the stats for a short period of
time so that someone attempting to read all of the stats in a
given device's directory will not result in a stats generation
call per file read.

Future work will attempt to standardize just the shared stats
elements, and possibly add a method to get the stats via netlink
in addition to sysfs.

Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Mark Bloch <markb@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
[ Add caching, make structure names more informative, add i40iw support,
  other significant rewrites from the original patch ]

Merge branch 'hfi1-2' into k.o/for-4.7

IB/hfi1: Fix pio map initialization

The pio map initialization function is off by 1 causing the last
kernel send context that is allocated to not get mapped into the
pio map which leads to the last kernel send context not being used
by any of the qps.

The send context reserved for VL15 is taken care of by setting the
scontext variable that is used as the index into the kernel send
context array to 1 and does not need to be accounted for in the
kernel send context counting loop as it is currently done.

Fix the kernel send context counting loop to account for all the
allocated send contexts and map all of them to the different VLs.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Jianxin Xiong <jianxin.xiong@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Correct 8051 link parameter settings

Two 8051 link settings, external device config and tuning method,
were written in the wrong location and the previous settings were
not cleared.  For both, clear the old value and write the new
value.

Fixes: 8ebd4cf1852a ("staging/rdma/hfi1: Add active and optical cable support")
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Update pkey table properly after link down or FM start

When FM is disabled, and the HFI port on the switch is
changed from MgmtAllowed=YES to MgmtAllowed=NO and the
link is bounced, FULL_MGMT_P_KEY doesn't get cleared
from the pkey table. This also occurs when the QSFP
cable is moved from a switch port with MgmtAllowed=YES
to a MgmtAllowed=NO port. Clear pkey entry properly.

Also, when the driver is loaded and the switch port is
set to MgmtAllowed=NO, FULL_MGMT_P_KEY shouldn't be added
to pkey table after FM is started. Only set FULL_MGMT_P_KEY
in the pkey table if switch port is configured to
MgmtAllowed=YES.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Sebastian Sanchez <sebastian.sanchez@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdamvt: Fix rdmavt s_ack_queue sizing

rdmavt allows the driver to specify the size of the ack queue, but
only uses it for the modify QP limit testing for setting the atomic
limit value.

The driver dependent size is now used to size the s_ack_queue ring
dynamicially.

Since the driver knows its size, the driver will use its define
for any ring size dependent code.

Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Max atomic value should be a u8

This matches the ib_qp_attr size and
avoids a extremely large value when the lower level
driver registers.

As part of the patch, the u8 ordinals are moved to the
end of the struct to reduce pahole noted excesses.

Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Fix hard lockup due to not using save/restore spin lock

Commit b9b06cb6feda
("IB/hfi1: Fix missing lock/unlock in verbs drain callback")
added a spin lock.

Unfortunately, the new lock code can be called from a base
level interrupt state, and an interrupt that can get stacked
will attempt to get the same lock.

Fix by using the flag save/restore spin lock variation.

Cc: stable@vger.kernel.org # 4.6+
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>

IB/hfi1: Add tracing support for send with invalidate opcode

Enable trace generation for packets with the "Send Last with
Invalidate" and "Send Only with Invalidate" opcodes.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1, qib: Add ieth to the packet header definitions

A new union member "ieth" (Invalidate Extended Transport Header) is
added to the packet header definition in preparation of supporting
the send with invalidate opcode.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

Merge branches 'misc-4.7-2', 'ipoib' and 'ib-router' into k.o/for-4.7

IB/hfi1: Move driver out of staging

The TODO list for the hfi1 driver was completed during 4.6. In addition
other objections raised (which are far beyond what was in the TODO list)
have been addressed as well. It is now time to remove the driver from
staging and into the drivers/infiniband sub-tree.

Reviewed-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Do not free hfi1 cdev parent structure early

The deletion of a cdev is not a fence for holding off references to the
structure. The driver attempts to delete the cdev and then proceeds to
free the parent structure, the hfi1_devdata, or dd. This can potentially
lead to a kernel panic in situations where a user has an FD for the cdev
open, and the pci device gets removed. If the user then closes the FD
there will be a NULL dereference when trying to do put on the cdev's
kobject.

Fix this by pointing the cdev's kobject.parent at a new kobject embedded
in its parent structure. Also take a reference when the device is opened
and put it back when it is closed.

Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Add trace message in user IOCTL handling

Add a trace message to HFI1s user IOCTL handling. This allows debugging
of which IOCTLs are being handled by the driver.

Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Remove write(), use ioctl() for user cmds

Remove the write() handler for user space commands now that ioctl
handling is available. User apps will need to change to use ioctl from
this point forward.

Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Add ioctl() interface for user commands

IOCTL is more suited to what user space commands need to do than the
write() interface. Add IOCTL definitions for all existing write commands
and the handling for those. The write() interface will be removed in a
follow on patch.

Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Remove unused user command

The HFI1_CMD_SDMA_STATUS_UPD command was never implemented it has no
reason to live in the driver. Remove it.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Remove snoop/diag interface

The snoop/diag interface is better served by an implementation which is
more general and usable by other drivers perhaps. Go ahead and remove
the code now and get rid of the char dev. We can put the feature back
when we have a more agreeable solution.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-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: Remove EPROM functionality from data device

Remove EPROM handling from the cdev which is used for user application
data traffic.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-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: Remove UI char device

Remove UI char device which exposes direct access to registers for user
space. This was put in to aid in debugging the hardware. We are looking
into alternatives means of providing the same functionality. This
removes another char device from HFI1's footprint.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Reviewed-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: Remove multiple device cdev

hfi1 current exports a cdev that can be used to target all of the hfi's
in the system. However there is a problem with this approach in
that the devices could be on different subnets. This is a problem that
user space can figure out and explicitly tell the driver on which device
to create a context.

Remove the multi-purpose cdev leaving a dedicated cdev for each port.
Also remove the striping capability that is dependent upon the user
choosing the multi-purpose cdev. It is now up to user space to determine
how to stripe contexts.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Remove anti-pattern in cdev init

Remove the usage of an anti-pattern goto in hfi1_cdev_init to improve
code readability.

Suggested-by: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Fix bug that blocks process on exit after port bounce

During the processing of a user SDMA request, if there was an
error before the request counter was increased, the state of
the packet queue could be updated incorrectly, causing the
counter to underflow. As the result, the process could get
stuck later since the counter could never get back to 0.

This patch adds a condition to guard the packet queue update
so that the counter is only decreased if it has been increased
before the error happens.

Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/qib: Remove unused qib_7322_intr_msgs[]

Building the qib driver with gcc version 6.1.0 raises the following
build warning:
drivers/infiniband/hw/qib/qib_iba7322.c:1311:39: warning:
'qib_7322_intr_msgs' defined but not used [-Wunused-const-variable=]
 static const struct  qib_hwerror_msgs qib_7322_intr_msgs[] = {
                                       ^~~~~~~~~~~~~~~~~~
Remove the unused qib_7322_intr_msgs[]

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Remove unnecessary comment

This comment was old, the MTU enums have been defined.

Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Fix sdma_event_names[] build warning

sdma_event_names[] is only used within CONFIG_SDMA_VERBOSITY ifdefs, so
when CONFIG_SDMA_VERBOSITY is disabled, it results in the following
0-day build warning:
>> drivers/infiniband/hw/hfi1/sdma.c:137:27: warning: 'sdma_event_names'
>> defined but not used [-Wunused-const-variable=]
    static const char * const sdma_event_names[] = {
                              ^~~~~~~~~~~~~~~~
This occurs on the following compiler:
compiler: gcc-6 (Debian 6.1.1-1) 6.1.1 20160430

For more information check:
https://lists.01.org/pipermail/kbuild-all/2016-May/020060.html

Fix this warning by defining sdma_event_name[] only within the
CONFIG_SDMA_VERBOSITY ifdefs.

Reported-by: kbuild test robot <fengguang.wu@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Use kzalloc_node

Use kzalloc_node instead of kzalloc for rdmavt memory region segment
allocation to optimize for performance on NUMA platforms.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Insure QP vmalloc variants zero memory

The usage of the various vmalloc APIs do not consistently zero memory
when allocating the swqe. Insure zeroing variants are used.

Reviewed-by: Mitko Haralanov <mitko.haralanov@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: Fix an interval RB node reference count leak

Commit e88c9271d9f8 ("IB/hfi1: Fix buffer cache corner case which
may cause corruption") introduced a bug which may cause a reference
count of a interval RB node to be leaked in the case where an SDMA
transfer from that node completes at the same time as the node is
being extended.

If a node is being extended, it is first removed from the RB tree
in order to be processed without the risk of an invalidation event
removing the node at the same time.

If a SDMA completion happens during that time, the completion handler
will fail to find the node in the RB tree and, therefore, fail to
correctly decrement its refcount. This leaves the node in the tree and
its pages pinned for the duration of the user process.

To prevent this from happening the io vector adds a reference to the
RB node, which is used during the SDMA completion instead of looking
up the node in the RB tree.

This change adds a performance improvement as a side effect by avoiding
the RB tree lookup.

Fixes: e88c9271d9f8 ("IB/hfi1: Fix buffer cache corner case which may cause corruption")
Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Harish Chegondi <harish.chegondi@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/IPoIB: Allow setting the device address

In IB networks, and specifically in IPoIB/rdmacm traffic, the device
address of an IPoIB interface is used as a means to exchange information
between nodes needed for communication.

Currently an IPoIB interface will always be created with a device
address based on its node GUID without a way to change that.

This change adds the ability to set the device address of an IPoIB
interface by value. We use the set mac address ndo to do that.

The flow should be broken down to two:
1) The GID value is already in the GID table,
   in this case the interface will be able to set carrier up.

2) The GID value is not yet in the GID table,
   in this case the interface won't try to join the multicast group
   and will wait (listen on GID_CHANGE event) until the GID is inserted.

In order to track those changes, we add a new flag:
* IPOIB_FLAG_DEV_ADDR_SET.

When set, it means the dev_addr is a based on a value in the gid
table. this bit will be cleared upon a dev_addr change triggered
by the user and set after validation.

Per IB spec the port GUID can't change if the module is loaded.
port GUID is the basis for GID at index 0 which is the basis for
the default device address of a ipoib interface.

The issue is that there are devices that don't follow the spec,
they change the port GUID while HCA is powered on, so in order
not to break userspace applications. We need to check if the
user wanted to control the device address and we assume that
if he sets the device address back to be based on GID index 0,
he no longer wishs to control it.

In order to track this, we add an additional flag:
* IPOIB_FLAG_DEV_ADDR_CTRL

When setting the device address, there is no validation of the upper
twelve bytes of the device address (flags, qpn, subnet prefix) as those
bytes are not under the control of the user.

Signed-off-by: Mark Bloch <markb@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/ipoib: Support SendOnlyFullMember MCG for SendOnly join

Check (via an SA query) if the SM supports the new option for SendOnly
multicast joins.
If the SM supports that option it will use the new join state to create
such multicast group.
If SendOnlyFullMember is supported, we wouldn't use faked FullMember state
join for SendOnly MCG, use the correct state if supported.

This check is performed at every invocation of mcast_restart task, to be
sure that the driver stays in sync with the current state of the SM.

Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Support new type of join-state for multicast

There are four types for MCG, FullMember, NonMember, SendOnlyNonMember,
and the new added type: SendOnlyFullMember.
Add support for the new SendOnlyFullMember join state.

The new type allows host to send join request as sendonly, it will cause the
group to be created but without getting packets from this multicast back to the
host.

Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Christoph Lameter <cl@linux.com>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/SA Agent: Add support for SA agent get ClassPortInfo

New SA query function to return the ClassPortInfo struct from the SA.
If the SM supports FullMemberSendOnly mode for MCG's, it sets a
capability bit in the capability_mask2 field of the response.

Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Introduce capabilitymask2 field in ClassPortInfo mad

Change struct ib_class_port_info to conform to IB Spec 1.3
That in order to get specific capability mask from ClassPortInfo mad.

>From the IB Spec, ClassPortInfo section:
        "CapabilityMask2 Bits 0-26: Additional class-specific capabilities...
         RespTimeValue the rest 5 bits"

The new struct now has one field for capabilitymask2 (previously was the
reserved field) and the resp_time field.

And it fixes up qib and srpt, use of the field repurposed to be used as
capabilitymask2:
IB/qib: Change pma_get_classportinfo
IB/srpt: Adjust the use of ib_class_port_info

Signed-off-by: Erez Shitrit <erezsh@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Hal Rosenstock <hal@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Add IP to GID netlink offload

There is an assumption that rdmacm is used only between nodes
in the same IB subnet, this why ARP resolution can be used to turn
IP to GID in rdmacm.

When dealing with IB communication between subnets this assumption
is no longer valid. ARP resolution will get us the next hop device
address and not the peer node's device address.

To solve this issue, we will check user space if it can provide the
GID of the peer node, and fail if not.

We add a sequence number to identify each request and fill in the GID
upon answer from userspace.

Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Register SA ibnl client during ib_core initialization

Move SA ibnl client registration to ib_core module init.
This will allow us to register a single client to handle
all RDMA_NL_LS operations and make it SA independent.

Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/netlink: Add a new local service operation

This commits adds a new RDMA local service operation:
- IP to GID resolution.

The client request would include the ifindex of the outgoing interface
and would place in an attribute (LS_NLA_TYPE_IPV4 or LS_NLA_TYPE_IPV6)
the destnation IP.

The local service would answer with a message that has the attribute:
- LS_NLA_TYPE_DGID - The destination GID.

Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/SA: Integrate ib_sa module into ib_core module

Consolidate ib_sa into ib_core, this commit eliminates
ib_sa.ko and makes it part of ib_core.ko

Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/MAD: Integrate ib_mad module into ib_core module

Consolidate ib_mad into ib_core, this commit eliminates
ib_mad.ko and makes it part of ib_core.ko

Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Integrate IB address resolution module into core

IB address resolution is declared as a module (ib_addr.ko) which loads
itself before IB core module (ib_core.ko).

It causes to the scenario where IB netlink which is initialized by IB
core can't be used by ib_addr.ko.

In order to solve it, we are converting ib_addr.ko to be part of
IB core module.

Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/cxgb3: device driver frees DMA memory with different size

[  598.852037] ------------[ cut here ]------------
[  598.856698] WARNING: at lib/dma-debug.c:887 check_unmap+0xf8/0x920()
[  598.863079] cxgb3 0000:01:00.0: DMA-API: device driver frees DMA memory with different size [device address=0x0000000003310000] [map size=17 bytes] [unmap size=16 bytes]
[  598.878265] Modules linked in: xprtrdma ib_isert iscsi_target_mod ib_iser libiscsi scsi_transport_iscsi ib_srpt target_core_mod ib_srp scsi_transport_srp scsi_tgt ib_ipoib rdma_ucm ib_ucm ib_uverbs ib_umad rdma_cm ib_cm iw_cm ib_sa ib_mad kvm_amd kvm ipmi_devintf ipmi_ssif dcdbas pcspkr ipmi_si sg ipmi_msghandler acpi_power_meter amd64_edac_mod shpchp edac_core sp5100_tco k10temp edac_mce_amd i2c_piix4 acpi_cpufreq nfsd auth_rpcgss nfs_acl lockd grace sunrpc ip_tables xfs libcrc32c sd_mod crc_t10dif crct10dif_generic crct10dif_common ata_generic iw_cxgb3 pata_acpi ib_core ib_addr mgag200 syscopyarea sysfillrect sysimgblt i2c_algo_bit drm_kms_helper ttm pata_atiixp drm ahci libahci serio_raw i2c_core cxgb3 libata bnx2 mdio dm_mirror dm_region_hash dm_log dm_mod
[  598.946822] CPU: 3 PID: 11820 Comm: cmtime Not tainted 3.10.0-327.el7.x86_64.debug #1
[  598.954681] Hardware name: Dell Inc. PowerEdge R415/0GXH08, BIOS 2.0.2 10/22/2012
[  598.962193]  ffff8808077479a8 000000000381a432 ffff880807747960 ffffffff81700918
[  598.969663]  ffff880807747998 ffffffff8108b6c0 ffff880807747a80 ffff8808063f55c0
[  598.977132]  ffffffff833ca850 0000000000000282 ffff88080b1bb800 ffff880807747a00
[  598.984602] Call Trace:
[  598.987062]  [<ffffffff81700918>] dump_stack+0x19/0x1b
[  598.992224]  [<ffffffff8108b6c0>] warn_slowpath_common+0x70/0xb0
[  598.998254]  [<ffffffff8108b75c>] warn_slowpath_fmt+0x5c/0x80
[  599.004033]  [<ffffffff813903b8>] check_unmap+0xf8/0x920
[  599.009369]  [<ffffffff81025959>] ? sched_clock+0x9/0x10
[  599.014702]  [<ffffffff81390cee>] debug_dma_free_coherent+0x7e/0xa0
[  599.021008]  [<ffffffffa01ece2c>] cxio_destroy_cq+0xcc/0x160 [iw_cxgb3]
[  599.027654]  [<ffffffffa01e8da0>] iwch_destroy_cq+0xf0/0x140 [iw_cxgb3]
[  599.034307]  [<ffffffffa01c4bfe>] ib_destroy_cq+0x1e/0x30 [ib_core]
[  599.040601]  [<ffffffffa04ff2d2>] ib_uverbs_close+0x302/0x4d0 [ib_uverbs]
[  599.047417]  [<ffffffff812335a2>] __fput+0x102/0x310
[  599.052401]  [<ffffffff8123388e>] ____fput+0xe/0x10
[  599.057297]  [<ffffffff810bbde4>] task_work_run+0xb4/0xe0
[  599.062719]  [<ffffffff81092a84>] do_exit+0x304/0xc60
[  599.067789]  [<ffffffff81025905>] ? native_sched_clock+0x35/0x80
[  599.073820]  [<ffffffff81025959>] ? sched_clock+0x9/0x10
[  599.079153]  [<ffffffff8170a49c>] ? _raw_spin_unlock_irq+0x2c/0x50
[  599.085358]  [<ffffffff8109346c>] do_group_exit+0x4c/0xc0
[  599.090779]  [<ffffffff810a8661>] get_signal_to_deliver+0x2e1/0x960
[  599.097071]  [<ffffffff8101c497>] do_signal+0x57/0x6e0
[  599.102229]  [<ffffffff81714bd1>] ? sysret_signal+0x5/0x4e
[  599.107738]  [<ffffffff8101cb7f>] do_notify_resume+0x5f/0xb0
[  599.113418]  [<ffffffff81714e7d>] int_signal+0x12/0x17
[  599.118576] ---[ end trace 1e4653102e7e7019 ]---
[  599.123211] Mapped at:
[  599.125577]  [<ffffffff8138ed8b>] debug_dma_alloc_coherent+0x2b/0x80
[  599.131968]  [<ffffffffa01ec862>] cxio_create_cq+0xf2/0x1f0 [iw_cxgb3]
[  599.139920]  [<ffffffffa01e9c05>] iwch_create_cq+0x105/0x4e0 [iw_cxgb3]
[  599.147895]  [<ffffffffa0500584>] create_cq.constprop.14+0x184/0x2e0 [ib_uverbs]
[  599.156649]  [<ffffffffa05027fb>] ib_uverbs_create_cq+0x10b/0x140 [ib_uverbs]

Fixes: b955150ea784 ('RDMA/cxgb3: When a user QP is marked in error, also mark the CQs in error')
Signed-off-by: Honggang Li <honli@redhat.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx5: Fire the CQ completion handler from tasklet

Previously, mlx5_ib_cq_comp was executed from interrupt context.
Under heavy load, this could cause the CPU core to be in an interrupt
context too long.
Instead of executing the handler from the interrupt context we
execute it from a much friendly tasklet context.

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

net/mlx5_core: Use tasklet for user-space CQ completion events

Previously, we've fired all our completion callbacks straight from
our ISR.

Some of those callbacks were lightweight (for example, mlx5 Ethernet
napi callbacks), but some of them did more work (for example,
the user-space RDMA stack uverbs' completion handler). Besides that,
doing more than the minimal work in ISR is generally considered wrong,
it could even lead to a hard lockup of the system. Since when a lot
of completion events are generated by the hardware, the loop over
those events could be so long, that we'll get into a hard lockup by
the system watchdog.

In order to avoid that, add a new way of invoking completion events
callbacks. In the interrupt itself, we add the CQs which receive
completion event to a per-EQ list and schedule a tasklet. In the
tasklet context we loop over all the CQs in the list and invoke the
user callback.

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

IB/core: Do not require CAP_NET_ADMIN for packet sniffing

In the Ethernet/TCP world, CAP_NET_RAW is sufficient to allow a program
to listen to all incoming packets on a specific interface, and the
higher CAP_NET_ADMIN is required to set the interface into promiscuous
mode.  We want to emulate that same basic division of privilege in the
RDMA stack, so when dealing with Raw Ethernet QPs, allow apps with
CAP_NET_RAW to listen to all incoming flows (and direct them as they see
fit in their own listen stream).  Do not require CAP_NET_ADMIN just to
listen to traffic already incoming.  Reserve CAP_NET_ADMIN if we attempt
to set promiscuous mode.

Signed-off-by: Christoph Lameter <cl@linux.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx4: Fix unaligned access in send_reply_to_slave

The problem is that the function 'send_reply_to_slave' gets the
'req_sa_mad' as a pointer whose address is only aliged to 4 bytes
but is 8 bytes in size.  This can result in unaligned access faults
on certain architectures.

Sowmini Varadhan pointed to this reply from Dave Miller that say
that memcpy should not be used to solve alignment issues:
https://lkml.org/lkml/2015/10/21/352

Optimization of memcpy to 'ldx' instruction can only happen if the
compiler knows that the size of the data we are copying is 8 bytes
and it assumes it is aligned to 8 bytes. If the compiler know the
type is not aligned to 8 it must not optimize the 8 byte copy.
Defining the data type as aligned to 4 forces the compiler to treat
all accesses as though they aren't aligned and avoids the 'ldx'
optimization.

Full credit for the idea goes to Jason Gunthorpe
<jgunthorpe@obsidianresearch.com>.

Signed-off-by: Shamir Rabinovitch <shamir.rabinovitch@oracle.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

Merge branches 'cxgb4-2', 'i40iw-2', 'ipoib', 'misc-4.7' and 'mlx5-fcs' into k.o/for-4.7

IB/mlx5: Report Scatter FCS device capability when supported

Report Scatter FCS support when the Firmware supports as well.

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

IB/mlx5: Add Scatter FCS support for Raw Packet QP

Enable Scatter FCS in the RQ context when the user passes
Scatter FCS create flag.

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

IB/core: Add Scatter FCS create flag

Raw Packet QPs that were created with Scatter FCS flag, will scatter
the FCS into the receive buffers.

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

IB/core: Add Raw Scatter FCS device capability

Raw Scatter FCS device capability is set when the NIC supports
scattering the FCS to the receive buffers of Raw Packet QPs.

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

IB/core: Add extended device capability flags

Since all the uverbs device_cap_flags are occupied, we need a place to
expose more device capabilities.

This patch adds a new 64 bit device_cap_flags_ex to expose new
device capabilities.

The lower 32 bits will be identical to the original device_cap_flags,
The upper 32 bits will be new capabilities.

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

i40iw: pass hw_stats by reference rather than by value

passing hw_stats by value requires a 280 byte copy so instead
pass it by reference is much more efficient.

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Acked-by: Chien Tin Tung <chien.tin.tung@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

i40iw: Remove unnecessary synchronize_irq() before free_irq()

Calling synchronize_irq() right before free_irq() is quite useless. On one
hand the IRQ can easily fire again before free_irq() is entered, on the
other hand free_irq() itself calls synchronize_irq() internally (in a race
condition free way), before any state associated with the IRQ is freed.

Patch was generated using the following semantic patch:
// <smpl>
@@
expression irq;
@@
-synchronize_irq(irq);
 free_irq(irq, ...);
// </smpl>

Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Acked-by: Faisal Latif <faisal.latif#intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

i40iw: constify i40iw_vf_cqp_ops structure

The i40iw_vf_cqp_ops structure is never modified, so declare it as const.

Done with the help of Coccinelle.

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx5: Add UARs write-combining and non-cached mapping

By this patch, the user space library will be able to improve
performance using appropriate ringing DoorBell method according to the
memory type it asked for.

Currently only one mapping command is allowed for UARs:
MLX5_IB_MMAP_REGULAR_PAGE. Using this mapping, the kernel maps the
UARs to write-combining (WC) if the system supports it.
If the system is not supporting WC the UARs are mapped to
non-cached(NC). In this case the user space library can't tell which
mapping is applied.
This patch adds 2 new mapping commands: MLX5_IB_MMAP_WC_PAGE and
MLX5_IB_MMAP_NC_PAGE. For these commands the kernel maps exactly as
requested and fails if it can't.

Since there is no generic way to check if the requested memory region
can be mapped as WC, driver enables conclusive WC mapping only for
x86, PowerPC and ARM which support WC for the device's memory region.

Signed-off-by: Guy Levy <guyle@mellanox.com>
Signed-off-by: Moshe Lazer <moshel@mellanox.com>
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx5: Allow mapping the free running counter on PROT_EXEC

The current mlx5 code disallows mapping the free running counter of
mlx5 based hardwares when PROT_EXEC is set.
Although this behaviour is correct, Linux does add an implicit VM_EXEC
to the vm_flags if the READ_IMPLIES_EXEC bit is set in the process
personality. This happens for example if the process stack is
executable.

This causes libmlx5 to output a warning and prevents the user from
reading the free running clock.
Executing the init segment of the hardware isn't a security risk
(at least no more than executing a process own stack), so we just
prevent writes to there.

Fixes: d69e3bcf7976 ('IB/mlx5: Mmap the HCA's core clock register to
      user-space')
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx4: Use list_for_each_entry_safe

Simplify the code in search_relocate_mgid0_group with by using
list_for_each_entry_safe().

Signed-off-by: Geliang Tang <geliangtang@163.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/SA: Use correct free function

Fixes a direct call to kfree_skb when nlmsg_free should be used.

Fixes: 2ca546b92a02 ('IB/sa: Route SA pathrecord query through netlink')
Signed-off-by: Mark Bloch <markb@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Reviewed-by: Ira Weiny <ira.weiny@intel.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Fix a potential array overrun in CMA and SA agent

Fix array overrun when going over callback table.
In declaration of callback table, the max size isn't provided and
in registration phase, it is provided.

There is potential scenario where a new operation is added
and it is not supported by current client. The acceptance of
such operation by ib_netlink will cause to array overrun.

Fixes: 809d5fc9bf65 ("infiniband: pass rdma_cm module to netlink_dump_start")
Fixes: b493d91d333e ("iwcm: common code for port mapper")
Fixes: 2ca546b92a02 ("IB/sa: Route SA pathrecord query through netlink")
Signed-off-by: Mark Bloch <markb@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/core: Remove unnecessary check in ibnl_rcv_msg

RDMA_NL_GET_OP is defined like this: (type & ((1 << 10) - 1))
which means op (defined as an int) can never be a negative number.

Fixes: b2cbae2c2487 ('RDMA: Add netlink infrastructure')
Signed-off-by: Mark Bloch <markb@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/IWPM: Fix a potential skb leak

In case ibnl_put_msg fails in send_nlmsg_done,
the function returns with -ENOMEM without freeing.

This patch fixes this behavior.

Fixes: 30dc5e63d6a5 ("RDMA/core: Add support for iWARP Port Mapper user space service")
Signed-off-by: Mark Bloch <markb@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/nes: replace custom print_hex_dump()

There is no need to duplicate a lot of code that is in the kernel library for ages.
Replace duplicating code by calling to print_hex_dump() directly.

Note that output is slightly changed:
- hex and ascii parts have just two spaces delimeter
- there is no delimeter for ascii portions
- file and line removed from prefix (they were redundant anyway since previous
  output shows same closer enough)

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Reviewed-by: Tatyana Nikolova <tatyana.e.nikolova@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/mlx4: trivial fix of spelling mistake on "argument"

fix spelling mistake, argumant -> argument

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/nes: Deinline nes_free_qp_mem, save 1072 bytes

This function compiles to 550 bytes of machine code.
Three callsites, all in nes_create_qp.

Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
CC: Faisal Latif <faisal.latif@intel.com>
CC: Doug Ledford <dledford@redhat.com>
CC: linux-rdma@vger.kernel.org
CC: linux-kernel@vger.kernel.org
Reviewed-By: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/nes: Adding queue drain functions

Adding sq and rq drain functions, which block until all
previously posted wr-s in the specified queue have completed.
A completion object is signaled to unblock the thread,
when the last cqe for the corresponding queue is processed.

Signed-off-by: Tatyana Nikolova <tatyana.e.nikolova@intel.com>
Signed-off-by: Faisal Latif <faisal.latif@intel.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Doug Ledford <dledford@redhat.com>

iwcm: Fix a sparse warning

Avoid that sparse complains about the comparison of s_addr
with INADDR_ANY.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Steve Wise <swise@opengridcomputing.com>
Cc: Faisal Latif <faisal.latif@intel.com>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/srp: Fix a debug kernel crash

Avoid that the following BUG() is triggered against a debug
kernel:

kernel BUG at include/linux/scatterlist.h:92!
RIP: 0010:[<ffffffffa0467199>]  [<ffffffffa0467199>] srp_map_idb+0x199/0x1a0 [ib_srp]
Call Trace:
 [<ffffffffa04685fa>] srp_map_data+0x84a/0x890 [ib_srp]
 [<ffffffffa0469674>] srp_queuecommand+0x1e4/0x610 [ib_srp]
 [<ffffffff813f5a5e>] scsi_dispatch_cmd+0x9e/0x180
 [<ffffffff813f8b07>] scsi_request_fn+0x477/0x610
 [<ffffffff81298ffe>] __blk_run_queue+0x2e/0x40
 [<ffffffff81299070>] blk_delay_work+0x20/0x30
 [<ffffffff81071f07>] process_one_work+0x197/0x480
 [<ffffffff81072239>] worker_thread+0x49/0x490
 [<ffffffff810787ea>] kthread+0xea/0x100
 [<ffffffff8159b632>] ret_from_fork+0x22/0x40

Fixes: f7f7aab1a5c0 ("IB/srp: Convert to new registration API")
Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Sagi Grimberg <sagi@grimberg.me>
Cc: Christoph Hellwig <hch@lst.de>
Cc: <stable@vger.kernel.org> # v4.4+
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/ipoib: Add readout of statistics using ethtool

IPoIB collects statistics of traffic including number of packets
sent/received, number of bytes transferred, and certain errors. This
patch makes these statistics available to be queried by ethtool.

Signed-off-by: Hans Westgaard Ry <hans.westgaard.ry@oracle.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Reviewed-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>
Tested-by: Yuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

infiniband/ulp/ipoib: remove pkey_mutex

The last user of pkey_mutex was removed in db84f8803759 ("IB/ipoib: Use
P_Key change event instead of P_Key polling mechanism") but the lock
remained.
This patch removes it.

Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

i40iw: pass hw_stats by reference rather than by value

passing hw_stats by value requires a 280 byte copy so instead
pass it by reference is much more efficient.

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Acked-by: Chien Tin Tung <chien.tin.tung@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

i40iw: Remove unnecessary synchronize_irq() before free_irq()

Calling synchronize_irq() right before free_irq() is quite useless. On one
hand the IRQ can easily fire again before free_irq() is entered, on the
other hand free_irq() itself calls synchronize_irq() internally (in a race
condition free way), before any state associated with the IRQ is freed.

Patch was generated using the following semantic patch:
// <smpl>
@@
expression irq;
@@
-synchronize_irq(irq);
 free_irq(irq, ...);
// </smpl>

Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Acked-by: Faisal Latif <faisal.latif#intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

i40iw: constify i40iw_vf_cqp_ops structure

The i40iw_vf_cqp_ops structure is never modified, so declare it as const.

Done with the help of Coccinelle.

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

staging/rdma/hfi1: use RCU_INIT_POINTER() when NULLing.

It is safe to use RCU_INIT_POINTER() to NULL a pointer, instead of
rcu_assign_pointer().
This results in slightly smaller/faster code.

Signed-off-by: Muhammad Falak R Wani <falakreyaz@gmail.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Change hfi1_init loop to preserve error returns

If one iteration of the loop causes an error return and a later iteration
doesn't, the later iteration causes the earlier error condition to be
lost. This could result in driver probe succeeding when it should have
failed. Therefore save off the error return in the loop itself rather than
outside the loop.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

ib_pack.h: Add opcode definition for send with invalidate

The opcode for "SEND Last with Invalidate" and "SEND Only with
Invalidate" have been defined for RC in IBA Specification Vol 1
since Release 1.2. Add the definition to the header file in
preparation of supporting these opcodes in rdmavt based drivers.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Keep SC_USER as the last send context type

SC_USER needs to be the last send context type to ensure other
send context types get their allocation when num_user_contexts
is set to a large number.

This fixes a panic when the module parameter num_user_contexts
is set to 141 and larger.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Immediately apply congestion setting MAD

The handling of the congestion setting MAD packet only
saved off the values, waiting for a congestion control
table packet before going active.  Instead, immediately
apply the values.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Correct log message strings

Remove "IB" keyword from log messages.
Correct comment for thermal sensor init function.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Jakub Pawlak <jakub.pawlak@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/rdmavt: Increase CQ callback thread priority

The priority of the send engines is higher than the CQ completion
thread potentially causing completions to be starved for very
fast interfaces.

Change the CQ kthread to match the send engine threads to minimize
this delay for ULP completion processing.

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

IB/hfi1: Fix hfi_rcvhdr tracepoint

The hfi_rcvhdr tracepoint has the ctxt and eflags switched in the
prototype of the trace event, compared to the args and usage of the
trace function. Fix this by swapping these 2 fields in the trace event
prototype.

Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Remove unnecessary header

While running perftests, there is a significant utilization of the
random number daemon. This is due to the linux/random.h header being
included in qp.c and verbs.c. However, none of the functions from this
header are being used in these files, so remove the unnecessary header.

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Jubin John <jubin.john@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Improve performance of interval RB trees

The interval RB tree management functions use handlers to
store user-specific callback for the various tree operations.
These handlers are put on a doubly-linked list. When a RB
tree function is called, the list is searched for the handler
of the particular tree.

The list which holds the handlers is modified very rarely - when
a handler is created and when a handler is removed. On the other
hand, it is searched very often. This a perfect usage scenario
for RCU.

The result is a much lower overhead of traversing the list as most
of the time no locking will be required.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Mitko Haralanov <mitko.haralanov@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Fix potential panic with sdma drained mechanism

The guard is backwards, potentially causing the SDMA client
to panic if a wait structure was not specified.

psm and verbs are not exposed to the issue, but fix the
code just to be correct.

Fixes: a545f5308b6c ("staging/rdma/hfi: fix CQ completion order issue")
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Fix pio wait counter double increment

The code unconditionlly increments the pio wait counter
making the counter inacurate and unusable.

Fixes: 14553ca11039 ("staging/rdma/hfi1: Adaptive PIO for short messages")
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Remove no-op QSFP reset code

The RESET_N bit of the ASIC_QSFPn_OE register is not used by
the hardware.  Remove code that tries to use it - it does
nothing.

Reviewed-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Correct external device configuration shift

The external device configuration was incorrectly shifted to byte 3 of
the 32 bit DC_HOST_COMM_SETTINGS instead of byte 0. This patch corrects
the shift and provides the cable capability information in byte 0.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Wait for QSFP modules to initialize

The function level reset in init_chip() and subsequent write of all 1s
to the ASIC_QSFP registers effectively resets attached active and
optical QSFP modules that pay attention to the RESET_N pin.

We subsequently try to access the QSFP management interface to qualify
and tune the channel and fabric SerDes before enough time (2 seconds
per SFF 8679 spec for QSFP28 modules) has elapsed for the module to
finish initialization. This fails and causes the failure of the channel
tuning algorithm, preventing us from bringing the link up.

This patch checks the port type prior to beginning channel and SerDes
tuning, and if found to be QSFP, watches for the QSFP initialization
complete interrupt, with a maximum timeout of 2 seconds, to allow the
initialization to complete.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

IB/hfi1: Ignore non-temperature warnings on a downed link

QSFP modules can raise an interrupt to inform us of expected conditions
while the link is down, such as RX power low. Actively ignore these
conditions when the link is down as they only add reporting noise.
Continue reporting conditions that are valid at all times, such as
temperature alarms and warnings.

Reviewed-by: Dean Luick <dean.luick@intel.com>
Signed-off-by: Easwar Hariharan <easwar.hariharan@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

iw_cxgb4: Convert a __force cast

__force casts should be avoided if there is a better alternative.
Hence modify the comparison of s_addr with INADDR_ANY such that the
__force cast is no longer necessary.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Cc: Steve Wise <swise@opengridcomputing.com>
Cc: Vipul Pandya <vipul@chelsio.com>
Acked-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: Add arp failure handlers to send_mpa_reply/reject()

These handlers when called print error message to the kernel log,
but the actual handling is done by _c4iw_free_ep() and process_timeout().

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: Always wake up waiter in c4iw_peer_abort_intr()

Currently c4iw_peer_abort_intr() does not wake up the waiter if the
endpoint state indicates we're using MPAv2 and we're currently trying to
connect. This was introduced with commit 7c0a33d61187a ("RDMA/cxgb4:
Don't wakeup threads for MPAv2")

However, this original fix is flawed because it introduces a race that
can cause a deadlock of the iwarp stack.  Here is the race:

->local side sets up an active offload connection.

->local side sends MPA_START request.

->peer sends MPA_START response.

->local side ingress cpl thread begins processing the MPA_START response,
but before it changes the state from MPA_REQ_SENT to FPDU_MODE:

->peer sends a RST which results in a ABORT_REQ_RSS.  This triggers
peer_abort_intr() which sees the state in MPA_REQ_SENT and since mpa_rev
is 2, it will avoid waking up the endpoint with -ECONNRESET, assuming the
stack will re-attempt the connection using MPAv1.

->Meanwhile, the cpl thread moves the state to FPDU_MODE and calls
c4iw_modify_rc_qp() which calls rdma_init() which sends a RI_WR/INIT WR
to firmware.  But since HW sent an abort, FW correctly drops the RI_WR/INIT
WR.

->So the cpl thread is stuck waiting for a reply and cannot process the
ABORT_REQ_RSS cpl sitting in its input queue. Thus everything comes to a
halt because no more ingress cpls are processed by the stack...

The correct fix for the issue is to always do the wake up in
c4iw_abort_intr() but reinitialize the wait object in c4iw_reconnect().

Fixes: 7c0a33d61187a ("RDMA/cxgb4: Don't wakeup threads for MPAv2")
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: Handle ret value of process_mpa_reply() in rx_data

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: atomic find and reference for listening endpoints

Add get_ep_from_stid() which will atomically find and reference the
endpoint struct if found. This avoids touch-after-free races between
threads destroying listening endpoints and the CPL processing thread
processing an incoming PASS_ACCEPT_REQ CPL.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: Handle ULP accept/reject during ABORTING

c4iw_reject() and c4iw_accept() need to handle the case where the
endpoint has timed out and is in the middle of ABORTING the connection.

Here is the flow that causes the BUG_ON() to fire on the server side:

1) offload connection setup and endpoint timer started

2) MPA_START request received from peer, CONNECT_REQUEST passed to ULP

3) endpoint timer fires, and process_timeout() aborts the connection,
this moves the endpoint state to ABORTING until HW sends up the
ABORT_RPL_RSS.

4) application exits closing the CONNECT_REQUEST cm_id.  The IWCM calls
c4iw_reject_cr() to destroy this connection request.

5) WHAMO: BUG_ON() because the state is ABORTING.

The fix is to change c4iw_reject_cr() and c4iw_accept_cr() to fail the
operation if the state is not in MPA_REQ_RCVD vs in DEAD.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: Release ep for for FPDU_MODE and MPA_REQ_RCVD in process_timeout

ARP failure may also happen when ep in FPDU_MODE and these failures need
to be handled by process_timeout(). process_timeout() also has to handle
case MPA_REQ_RCVD, setting abort to 1, leading to ep resource release.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: Free skb in case of arp failure in _c4iw_free_ep()

Arp failure for send_mpa_reply/reject() is handled by freeing the
mpa_skb in c4iw_free_ep() before releasing ep.

Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: atomically lookup ep and get a reference

There is a race between ULP threads calling c4iw_ep_disconnect() via
c4iw_modify_rc_qp() and the ingress CPL thread where the ULP thread
can free the endpoint just after the ingress CPL thread finds the ep
pointer in the tid table.  To avoid this, we now use the hwtid_idr table
for lookups instead of the LLD tid table so we can lock around insert,
remove, and lookup+get_ep to avoid the race.  The CPL handlers now will
either find the ep ptr and have a ref on it, or not find it and they
can discard the CPL.  Callers of get_ep_from_tid() will have a ref
on the ep if found, and thus must deref when they are done.
Negative advice in peer_abort_intr() need to dereference the ep.
therefore peer_abort() is scheduled to dereference the ep later.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: Handle return value of c4iw_ofld_send() in abort_arp_failure()

In abort_arp_failure(), the return value from c4iw_ofld_send() is
ignored and thus if the CPL isn't sent, the endpoint is stuck and never
gets aborted. Failure of c4iw_ofld_send() is treated as fatal error, and
the ep resources are released in a safer context through process_work().

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: in process_timeout() don't move ep state to ABORTING

Moving the state to ABORTING causes the ep to get stuck because
c4iw_ep_timeout() thinks the ABORT has already been done. So leave the
state alone and let c4iw_ep_disconnect() do the right thing given the
ep state.

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

RDMA/iw_cxgb4: handle return value of c4iw_l2t_send() and send_mpa_req()

->In act_open_rpl(), CPL_ERR_TCAM_FULL error handling branch, there is
no handling of the return value of send_fw_act_open_req().
->In send_fw_act_open_req(), there is no handling of return value of
c4iw_l2t_send(), which may cause a ep leak and won't notify upper layers
on connection establish failure.
->send_mpa_req() should act on the return from c4iw_l2t_send() and
return the error to the caller.
->In case of c4iw_l2t_send() failure in send_mpa_req(), returns without
starting the timer and not changing the ep state, which is further
handled by act_establish()
-> In act_establish()?if send_mpa_request's get_skb returns an error,
may cause an ep leak. So handle return value of send_mpa_req()

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>