Expose a non standard query port via IOCTL that will be used to expose
port attributes that are specific to mlx5 devices.

The new interface receives a port number to query and returns a structure
that contains the available attributes for that port.  This will be used
to fill the gap between pure DEVX use cases and use cases where a kernel
needs to inform userspace about various kernel driver configurations that
userspace must use in order to work correctly.

Flags is used to indicate which fields are valid on return.

MLX5_IB_UAPI_QUERY_PORT_VPORT:
	The vport number of the queered port.

MLX5_IB_UAPI_QUERY_PORT_VPORT_VHCA_ID:
	The VHCA ID of the vport of the queered port.

MLX5_IB_UAPI_QUERY_PORT_VPORT_STEERING_ICM_RX:
	The vport's RX ICM address used for sw steering.

MLX5_IB_UAPI_QUERY_PORT_VPORT_STEERING_ICM_TX:
	The vport's TX ICM address used for sw steering.

MLX5_IB_UAPI_QUERY_PORT_VPORT_REG_C0:
	The metadata used to tag egress packets of the vport.

MLX5_IB_UAPI_QUERY_PORT_ESW_OWNER_VHCA_ID:
	The E-Switch owner vhca id of the vport.

Link: https://lore.kernel.org/r/6e2ef13e5a266a6c037eb0105eb1564c7bb52f23.1618743394.git.leonro@nvidia.comReviewed-by: NMaor Gottlieb <maorg@nvidia.com>
Signed-off-by: NMark Bloch <mbloch@nvidia.com>
Signed-off-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

'in6' is unused. It is just declared and filled-in.  It can be removed.

This is a left over from commit 5ea8bbfc
("mlx4: Implement IP based gids support for RoCE/SRIOV")

Link: https://lore.kernel.org/r/413dc6e2ea9d85bda1131bbd6a730b7620839eab.1618932283.git.christophe.jaillet@wanadoo.frSigned-off-by: NChristophe JAILLET <christophe.jaillet@wanadoo.fr>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

We should hold the UAPI DM type in the base struct and not the internal
mlx5 type.

Fixes: 251b9d78 ("RDMA/mlx5: Re-organize the DM code")
Link: https://lore.kernel.org/r/58dedbd5c132660f808e59166d434e2eaa6ecf7a.1618753425.git.leonro@nvidia.comSigned-off-by: NMaor Gottlieb <maorg@nvidia.com>
Signed-off-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Currently when GID is deleted, it zero out all the fields of the RoCE
address in the SET_ROCE_ADDRESS command for a specified index.

roce_version = 0 means RoCEv1 in the SET_ROCE_ADDRESS command.

This assumes that device has RoCEv1 always enabled which is not always
correct. For example Subfunction does not support RoCEv1.

Due to this assumption a previously added RoCEv2 GID is always deleted as
RoCEv1 GID. This results in a below syndrome:

   mlx5_core.sf mlx5_core.sf.4: mlx5_cmd_check:777:(pid 4256): SET_ROCE_ADDRESS(0x761) op_mod(0x0) failed, status bad parameter(0x3), syndrome (0x12822d)

Hence set the right RoCE version during GID deletion provided by the core.

Link: https://lore.kernel.org/r/d3f54129c90ca329caf438dbe31875d8ad08d91a.1618753425.git.leonro@nvidia.comSigned-off-by: NParav Pandit <parav@nvidia.com>
Signed-off-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

When i40iw_hmc_sd_one fails, chunk is freed without the deletion of chunk
entry in the PBLE info list.

Fix it by adding the chunk entry to the PBLE info list only after
successful addition of SD in i40iw_hmc_sd_one.

This fixes a static checker warning reported here:
  https://lore.kernel.org/linux-rdma/YHV4CFXzqTm23AOZ@mwanda/

Fixes: 97158301 ("i40iw: add pble resource files")
Link: https://lore.kernel.org/r/20210416002104.323-1-shiraz.saleem@intel.comReported-by: NDan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: NSindhu Devale <sindhu.devale@intel.com>
Signed-off-by: NShiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

missing qpid increment leads to skipping few qpids while allocating QP.
This eventually leads to adapter running out of qpids after establishing
fewer connections than it actually supports.
Current patch increments the qpid correctly.

Fixes: cfdda9d7 ("RDMA/cxgb4: Add driver for Chelsio T4 RNIC")
Link: https://lore.kernel.org/r/20210415151422.9139-1-bharat@chelsio.comSigned-off-by: NPotnuri Bharat Teja <bharat@chelsio.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

struct ipoib_cm_tx is defined at 245th line.  And the definition is
independent on the MACRO.  The declaration here is unnecessary. Remove it.

Link: https://lore.kernel.org/r/20210415092124.27684-1-wanjiabing@vivo.comSigned-off-by: NWan Jiabing <wanjiabing@vivo.com>
Reviewed-by: NChristoph Lameter <cl@linux.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Instead of manually messing with parent driver module reference counting
rely on export symbol mechanism to ensure that proper probe/remove chain
is performed.

Link: https://lore.kernel.org/r/20210401065715.565226-4-leon@kernel.orgSigned-off-by: NLeon Romanovsky <leonro@nvidia.com>
Acked-By: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Convert indirect probe call to its direct equivalent to create a symbol
link between RDMA and netdev modules. This will give us an ability to
remove custom module reference counting that doesn't belong to the driver.

Link: https://lore.kernel.org/r/20210401065715.565226-3-leon@kernel.orgSigned-off-by: NLeon Romanovsky <leonro@nvidia.com>
Acked-By: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

The "select" kconfig keyword provides reverse dependency, however it
doesn't check that selected symbol meets its own dependencies. Usually
"select" is used for non-visible symbols, so instead of trying to keep
dependencies in sync with BNXT ethernet driver, simply "depends on" it,
like Kconfig documentation suggest.

* CONFIG_PCI is already required by BNXT
* CONFIG_NETDEVICES and CONFIG_ETHERNET are needed to chose BNXT

Link: https://lore.kernel.org/r/20210401065715.565226-2-leon@kernel.orgSigned-off-by: NLeon Romanovsky <leonro@nvidia.com>
Acked-By: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Currently IPoIB connected mode queries the device to get the pkey table
entry during connection formation. This will increase the time taken to
form the connection, especially when limited pkeys are in use.  This gets
worse when multiple connection attempts are done in parallel.

Since ipoib interfaces are locked to a single pkey, use the pkey index
that was determined at link up time instead of searching for anything.

This improved the latency from 500ms to 1ms on an internal setup.

Link: https://lore.kernel.org/r/1618338965-16717-1-git-send-email-manjunath.b.patil@oracle.comSuggested-by: NJason Gunthorpe <jgg@nvidia.com>
Signed-off-by: NManjunath Patil <manjunath.b.patil@oracle.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

In cm_req_handler(), unify the check for RoCE and re-factor to avoid
one test.

Link: https://lore.kernel.org/r/1617705423-15570-1-git-send-email-haakon.bugge@oracle.comSuggested-by: NJason Gunthorpe <jgg@nvidia.com>
Fixes: 8f974860 ("IB/cm: Reduce dependency on gid attribute ndev check")
Fixes: 194f64a3 ("RDMA/core: Fix corrupted SL on passive side")
Signed-off-by: NHåkon Bugge <haakon.bugge@oracle.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

When "enhanced IPoIB" is enabled for CX-5 devices it requires the parent
device to be UP, otherwise the child devices won't work.

Thus add a debug message to give admin a hint when only the child
interface is UP but parent interface is not.

Link: https://lore.kernel.org/r/20210408093215.24023-1-jinpu.wang@ionos.comSigned-off-by: NJack Wang <jinpu.wang@ionos.com>
Reviewed-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Remove duplicated code for setting hem page size in PF and VF.

Link: https://lore.kernel.org/r/1617715514-29039-7-git-send-email-liweihang@huawei.comSigned-off-by: NXi Wang <wangxi11@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Introduce the VF support by adding code changes to allow VF PCI device
initialization, assgining the reserved resource of the PF to the active
VFs, setting the default abilities, applying the interruptions, resetting
and reducing the default QP/GID number to aovid exceeding the hardware
limitation.

Link: https://lore.kernel.org/r/1617715514-29039-6-git-send-email-liweihang@huawei.comSigned-off-by: NWei Xu <xuwei5@hisilicon.com>
Signed-off-by: NShengming Shu <shushengming1@huawei.com>
Signed-off-by: NXi Wang <wangxi11@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Switch parameters of all functions belong to a PF should be set including
VFs.

Link: https://lore.kernel.org/r/1617715514-29039-5-git-send-email-liweihang@huawei.comSigned-off-by: NWei Xu <xuwei5@hisilicon.com>
Signed-off-by: NShengming Shu <shushengming1@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Query the resource including EQC/SMAC/SGID from the firmware in the PF and
distribute fairly among all the functions belong to the PF.

Link: https://lore.kernel.org/r/1617715514-29039-4-git-send-email-liweihang@huawei.comSigned-off-by: NWei Xu <xuwei5@hisilicon.com>
Signed-off-by: NShengming Shu <shushengming1@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Query how many functions are supported by the PF from the FW and store it
in the hns_roce_dev structure which will be used to support the
configuration of virtual functions.

Link: https://lore.kernel.org/r/1617715514-29039-3-git-send-email-liweihang@huawei.comSigned-off-by: NWei Xu <xuwei5@hisilicon.com>
Signed-off-by: NShengming Shu <shushengming1@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Use hr_reg_write/read() to simplify codes about configuring function's
resource. And because the design of PF/VF fields is same, they can be
defined only once.

Link: https://lore.kernel.org/r/1617715514-29039-2-git-send-email-liweihang@huawei.comSigned-off-by: NXi Wang <wangxi11@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Two error messages are only different message but have common
code to generate the path string.

Link: https://lore.kernel.org/r/20210406123639.202899-4-gi-oh.kim@ionos.comSigned-off-by: NGioh Kim <gi-oh.kim@ionos.com>
Signed-off-by: NJack Wang <jinpu.wang@ionos.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

It does not help to debug if it only print error message
without any debugging information which session and connection
the error happened.

Link: https://lore.kernel.org/r/20210406123639.202899-3-gi-oh.kim@ionos.comSigned-off-by: NGioh Kim <gi-oh.kim@ionos.com>
Signed-off-by: NJack Wang <jinpu.wang@ionos.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Client prints only error value and it is not enough for debugging.

1. When client receives an error from server: the client does not only
   print the error value but also more information of server connection.

2. When client failes to send IO: the client gets an error from RDMA
   layer. It also print more information of server connection.

Link: https://lore.kernel.org/r/20210406123639.202899-2-gi-oh.kim@ionos.comSigned-off-by: NGioh Kim <gi-oh.kim@ionos.com>
Signed-off-by: NJack Wang <jinpu.wang@ionos.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Describe new multipath policy min-latency of the RTRS client.

Link: https://lore.kernel.org/r/20210407113444.150961-5-gi-oh.kim@ionos.comSigned-off-by: NGioh Kim <gi-oh.kim@ionos.com>
Signed-off-by: NJack Wang <jinpu.wang@ionos.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

It shows the latest latency that the client checked when sending the
heart-beat.

Link: https://lore.kernel.org/r/20210407113444.150961-3-gi-oh.kim@ionos.comSigned-off-by: NGioh Kim <gi-oh.kim@ionos.com>
Signed-off-by: NJack Wang <jinpu.wang@ionos.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

This patch adds new multipath policy: min-latency.  Client checks the
latency of each path when it sends the heart-beat.  And it sends IO to the
path with the minimum latency.

Link: https://lore.kernel.org/r/20210407113444.150961-2-gi-oh.kim@ionos.comSigned-off-by: NGioh Kim <gi-oh.kim@ionos.com>
Signed-off-by: NJack Wang <jinpu.wang@ionos.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Maor Gottlieb says:
====================
This series from Maor extends MEMIC to support atomic operations from the
host in addition to already supported regular read/write.
====================

* 'memic_ops':
  RDMA/mlx5: Expose UAPI to query DM
  RDMA/mlx5: Add support in MEMIC operations
  RDMA/mlx5: Add support to MODIFY_MEMIC command
  RDMA/mlx5: Re-organize the DM code
  RDMA/mlx5: Move all DM logic to separate file
  RDMA/uverbs: Make UVERBS_OBJECT_METHODS to consider line number
  net/mlx5: Add MEMIC operations related bits

Expose UAPI to query MEMIC DM, this will let user space application
that didn't allocate the DM but has access to by owning the matching
command FD to retrieve its information.

Link: https://lore.kernel.org/r/20210411122924.60230-8-leon@kernel.orgSigned-off-by: NMaor Gottlieb <maorg@nvidia.com>
Signed-off-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

MEMIC buffer, in addition to regular read and write operations, can
support atomic operations from the host.

Introduce and implement new UAPI to allocate address space for MEMIC
operations such as atomic. This includes:

1. Expose new IOCTL for request mapping of MEMIC operation.
2. Hold the operations address in a list, so same operation to same DM
   will be allocated only once.
3. Manage refcount on the mlx5_ib_dm object, so it would be keep valid
   until all addresses were unmapped.

Link: https://lore.kernel.org/r/20210411122924.60230-7-leon@kernel.orgSigned-off-by: NMaor Gottlieb <maorg@nvidia.com>
Signed-off-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Add two functions to allocate and deallocate MEMIC operations
by using the MODIFY_MEMIC command.

Link: https://lore.kernel.org/r/20210411122924.60230-6-leon@kernel.orgSigned-off-by: NMaor Gottlieb <maorg@nvidia.com>
Signed-off-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

1. Inline the checks from check_dm_type_support() into their
   respective allocation functions.
2. Fix use after free when driver fails to copy the MEMIC address to the
   user by moving the allocation code into their respective functions,
   hence we avoid the explicit call to free the DM in the error flow.
3. Split mlx5_ib_dm struct to memic and icm proper typesafety
   throughout.

Fixes: dc2316eb ("IB/mlx5: Fix device memory flows")
Link: https://lore.kernel.org/r/20210411122924.60230-5-leon@kernel.orgSigned-off-by: NMaor Gottlieb <maorg@nvidia.com>
Signed-off-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Move all device memory related code to a separate file.

Link: https://lore.kernel.org/r/20210411122924.60230-4-leon@kernel.orgSigned-off-by: NMaor Gottlieb <maorg@nvidia.com>
Signed-off-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

In order to support multiple methods declaration in the same file we
should use the line number as part of the name.

Link: https://lore.kernel.org/r/20210411122924.60230-3-leon@kernel.orgSigned-off-by: NMaor Gottlieb <maorg@nvidia.com>
Signed-off-by: NLeon Romanovsky <leonro@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Add the MEMIC operations bits and structures to the mlx5_ifc file.
Signed-off-by: NMaor Gottlieb <maorg@nvidia.com>
Signed-off-by: NLeon Romanovsky <leonro@nvidia.com>

All other users of the dummy netdevice embed the netdev in other
structures:

init_dummy_netdev(&mal->dummy_dev);
init_dummy_netdev(&eth->dummy_dev);
init_dummy_netdev(&ar->napi_dev);
init_dummy_netdev(&irq_grp->napi_ndev);
init_dummy_netdev(&wil->napi_ndev);
init_dummy_netdev(&trans_pcie->napi_dev);
init_dummy_netdev(&dev->napi_dev);
init_dummy_netdev(&bus->mux_dev);

The AIP and VNIC implementation turns that model inside out and used a
kfree() to free what appears to be a netdev struct when in reality, it is
a struct that enbodies the rx state as well as the dummy netdev used to
support napi_poll across disparate receive contexts.  The relationship is
infered by the odd allocation:

	const int netdev_size = sizeof(*dd->dummy_netdev) +
		sizeof(struct hfi1_netdev_priv);
	<snip>
	dd->dummy_netdev = kcalloc_node(1, netdev_size, GFP_KERNEL, dd->node);

Correct the issue by:
- Correctly naming the alloc and free functions
- Renaming hfi1_netdev_priv to hfi1_netdev_rx
- Replacing dd dummy_netdev with a netdev_rx pointer
- Embedding the net_device in hfi1_netdev_rx
- Moving the init_dummy_netdev to the alloc routine
- Adjusting wrappers to fit the new model

Fixes: 6991abcb ("IB/hfi1: Add functions to receive accelerated ipoib packets")
Link: https://lore.kernel.org/r/1617026056-50483-11-git-send-email-dennis.dalessandro@cornelisnetworks.comReviewed-by: NKaike Wan <kaike.wan@intel.com>
Signed-off-by: NMike Marciniszyn <mike.marciniszyn@cornelisnetworks.com>
Signed-off-by: NDennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Fix the following clang warning:

drivers/infiniband/hw/qib/qib_iba7322.c:803:19: warning: unused function 'qib_read_ureg' [-Wunused-function].

Link: https://lore.kernel.org/r/1618305063-29007-1-git-send-email-jiapeng.chong@linux.alibaba.comReported-by: NAbaci Robot <abaci@linux.alibaba.com>
Signed-off-by: NJiapeng Chong <jiapeng.chong@linux.alibaba.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

As a flush operation is implemented inside destroy_workqueue(), there is
no need to do flush operation before.

Fixes: bfcc681b ("IB/hns: Fix the bug when free mr")
Fixes: 0425e3e6 ("RDMA/hns: Support flush cqe for hip08 in kernel space")
Link: https://lore.kernel.org/r/1618305087-30799-1-git-send-email-liweihang@huawei.comSigned-off-by: NYixian Liu <liuyixian@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

A session can be removed dynamically by sysfs interface "remove_path" that
eventually calls rtrs_clt_remove_path_from_sysfs function.  The current
rtrs_clt_remove_path_from_sysfs first removes the sysfs interfaces and
frees sess->stats object. Second it removes the session from the active
list.

Therefore some functions could access non-connected session and access the
freed sess->stats object even-if they check the session status before
accessing the session.

For instance rtrs_clt_request and get_next_path_min_inflight check the
session status and try to send IO to the session.  The session status
could be changed when they are trying to send IO but they could not catch
the change and update the statistics information in sess->stats object,
and generate use-after-free problem.
(see: "RDMA/rtrs-clt: Check state of the rtrs_clt_sess before reading its
stats")

This patch changes the rtrs_clt_remove_path_from_sysfs to remove the
session from the active session list and then destroy the sysfs
interfaces.

Each function still should check the session status because closing or
error recovery paths can change the status.

Fixes: 6a98d71d ("RDMA/rtrs: client: main functionality")
Link: https://lore.kernel.org/r/20210412084002.33582-1-gi-oh.kim@ionos.comSigned-off-by: NGioh Kim <gi-oh.kim@ionos.com>
Reviewed-by: NJack Wang <jinpu.wang@ionos.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Fix to return a negative error code from the error handling case instead
of 0, as done elsewhere in this function.

Fixes: db7683d7 ("IB/srpt: Fix login-related race conditions")
Link: https://lore.kernel.org/r/20210408113132.87250-1-wangwensheng4@huawei.comReported-by: NHulk Robot <hulkci@huawei.com>
Signed-off-by: NWang Wensheng <wangwensheng4@huawei.com>
Reviewed-by: NBart Van Assche <bvanassche@acm.org>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

For some HCAs, ib_modify_qp() is an expensive operation running
virtualized.

For both the active and passive side, the QP returned by the CM has the
state set to RTS, so no need for this excess RTS -> RTS transition. With
IB Core's ability to set the RNR Retry timer, we use this interface to
shave off another ib_modify_qp().

Fixes: ec16227e ("RDS/IB: Infiniband transport")
Link: https://lore.kernel.org/r/1617216194-12890-3-git-send-email-haakon.bugge@oracle.comSigned-off-by: NHåkon Bugge <haakon.bugge@oracle.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>

Introduce the ability for kernel ULPs to adjust the minimum RNR Retry
timer. The INIT -> RTR transition executed by RDMA CM will be used for
this adjustment. This avoids an additional ib_modify_qp() call.

rdma_set_min_rnr_timer() must be called before the call to rdma_connect()
on the active side and before the call to rdma_accept() on the passive
side.

The default value of RNR Retry timer is zero, which translates to 655
ms. When the receiver is not ready to accept a send messages, it encodes
the RNR Retry timer value in the NAK. The requestor will then wait at
least the specified time value before retrying the send.

The 5-bit value to be supplied to the rdma_set_min_rnr_timer() is
documented in IBTA Table 45: "Encoding for RNR NAK Timer Field".

Link: https://lore.kernel.org/r/1617216194-12890-2-git-send-email-haakon.bugge@oracle.comSigned-off-by: NHåkon Bugge <haakon.bugge@oracle.com>
Acked-by: NJason Gunthorpe <jgg@nvidia.com>
Signed-off-by: NJason Gunthorpe <jgg@nvidia.com>