IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
QP's with the same SRQ, working on different CQs and running in parallel
on different CPUs could lead to a race when maintaining the SRQ consumer
count, and leads to FW running out of SRQs. Update the consumer
atomically. Make sure the wqe_prod is updated after the sge_prod due to
FW requirements.
Fixes: 3491c9e799fb ("qedr: Add support for kernel mode SRQ's")
Link: https://lore.kernel.org/r/20200708195526.31040-1-ybason@marvell.com
Signed-off-by: Michal Kalderon <mkalderon@marvell.com>
Signed-off-by: Yuval Basson <ybason@marvell.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Current iSER target code allocates MR pool budget based on queue size.
Since there is no handshake between iSER initiator and target on max IO
size, we'll set the iSER target to support upto 16MiB IO operations and
allocate the correct number of RDMA ctxs according to the factor of MR's
per IO operation. This would guarantee sufficient size of the MR pool for
the required IO queue depth and IO size.
Link: https://lore.kernel.org/r/20200708091908.162263-1-maxg@mellanox.com
Reported-by: Krishnamraju Eraparaju <krishna2@chelsio.com>
Tested-by: Krishnamraju Eraparaju <krishna2@chelsio.com>
Signed-off-by: Max Gurtovoy <maxg@mellanox.com>
Acked-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
When using action drop dest_type was never assigned to any value. Add
initialization of dest_type to -1 since 0 is valid.
Fixes: f29de9eee782 ("RDMA/mlx5: Add support for drop action in DV steering")
Link: https://lore.kernel.org/r/20200707110259.882276-1-leon@kernel.org
Signed-off-by: Daria Velikovsky <daria@mellanox.com>
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Instead of returning IB_LINK_LAYER_ETHERNET from rxe_link_layer, return it
directly from get_link_layer callback and remove rxe_link_layer().
Fixes: 8700e3e7c485 ("Soft RoCE driver")
Link: https://lore.kernel.org/r/20200705104313.283034-5-kamalheib1@gmail.com
Signed-off-by: Kamal Heib <kamalheib1@gmail.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The return value from rxe_mem_init_dma() is always 0 - change it to be
void and fix the callers accordingly.
Fixes: 8700e3e7c485 ("Soft RoCE driver")
Link: https://lore.kernel.org/r/20200705104313.283034-4-kamalheib1@gmail.com
Signed-off-by: Kamal Heib <kamalheib1@gmail.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The return value from rxe_init_port_param() is always 0 - change it to be
void.
Fixes: 8700e3e7c485 ("Soft RoCE driver")
Link: https://lore.kernel.org/r/20200705104313.283034-3-kamalheib1@gmail.com
Signed-off-by: Kamal Heib <kamalheib1@gmail.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Both pkey_tbl_len and gid_tbl_len are set in rxe_init_port_param() - so no
need to check if they aren't set.
Fixes: 8700e3e7c485 ("Soft RoCE driver")
Link: https://lore.kernel.org/r/20200705104313.283034-2-kamalheib1@gmail.com
Signed-off-by: Kamal Heib <kamalheib1@gmail.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
ROCE uses "VA % buf_page_size" to caclulate the offset in the PBL's first
page, the actual PA corresponding to the MR's VA is equal to MR's PA plus
this offset. The first PA in PBL has already been aligned to PAGE_SIZE
after calling ib_umem_get(), but the MR's VA may not. If the buf_page_size
is smaller than the PAGE_SIZE, this will lead the HW to access the wrong
memory because the offset is smaller than expected.
Fixes: 9b2cf76c9f05 ("RDMA/hns: Optimize PBL buffer allocation process")
Link: https://lore.kernel.org/r/1594726935-45666-1-git-send-email-liweihang@huawei.com
Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The RoCE Engine will schedule to another QP after one has sent
(2 ^ lp_pktn_ini) packets. lp_pktn_ini is set in QPC and should be
calculated from 2 factors:
1. current MTU as a integer
2. the RoCE Engine's maximum slice length 64KB
But the driver use MTU as a enum ib_mtu and the max inline capability, the
lp_pktn_ini will be much bigger than expected which may cause traffic of
some QPs to never get scheduled.
Fixes: b713128de7a1 ("RDMA/hns: Adjust lp_pktn_ini dynamically")
Link: https://lore.kernel.org/r/1594726138-49294-1-git-send-email-liweihang@huawei.com
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In manual mode allow bind user QPs with different pids to same counter,
since this is allowed in auto mode.
Bind kernel QPs and user QPs to the same counter are not allowed.
Fixes: 1bd8e0a9d0fd ("RDMA/counter: Allow manual mode configuration support")
Link: https://lore.kernel.org/r/20200702082933.424537-4-leon@kernel.org
Signed-off-by: Mark Zhang <markz@mellanox.com>
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In auto mode only bind user QPs to a dynamic counter, since this feature
is mainly used for system statistic and diagnostic purpose, while there's
no need to counter kernel QPs so far.
Fixes: 99fa331dc862 ("RDMA/counter: Add "auto" configuration mode support")
Link: https://lore.kernel.org/r/20200702082933.424537-3-leon@kernel.org
Signed-off-by: Mark Zhang <markz@mellanox.com>
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
With the "PID" category QPs have same PID will be bound to same counter;
If this category is not set then QPs have different PIDs will be bound
to same counter.
This is implemented for 2 reasons:
1. The counter is a limited resource, while there may be dozens of
applications, each of which creates several types of QPs, which means
it may doesn't have enough counter.
2. The system administrator needs all QPs created by all applications
with same type bound to one counter.
The counter name and PID is only make sense when "PID" category are
configured.
This category can also be used in combine with others, e.g. QP type.
Link: https://lore.kernel.org/r/20200702082933.424537-2-leon@kernel.org
Signed-off-by: Mark Zhang <markz@mellanox.com>
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Some released FW versions mistakenly don't set the capability that 50G per
lane link-modes are supported for VFs (ptys_extended_ethernet capability
bit).
Use PTYS.ext_eth_proto_capability instead, as this indication is always
accurate. If PTYS.ext_eth_proto_capability is valid
(has a non-zero value) conclude that the HCA supports 50G per lane.
Otherwise, conclude that the HCA doesn't support 50G per lane.
Fixes: 08e8676f1607 ("IB/mlx5: Add support for 50Gbps per lane link modes")
Link: https://lore.kernel.org/r/20200707110612.882962-3-leon@kernel.org
Signed-off-by: Aya Levin <ayal@mellanox.com>
Reviewed-by: Eran Ben Elisha <eranbe@mellanox.com>
Reviewed-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Move the initialization of the vendor_part_id to be before calling
ib_register_device(), this is needed because the query_device() callback
is called from the context of ib_register_device() before initializing the
vendor_part_id, so the reported value is wrong.
Fixes: bdcf26bf9b3a ("rdma/siw: network and RDMA core interface")
Link: https://lore.kernel.org/r/20200707130931.444724-1-kamalheib1@gmail.com
Signed-off-by: Kamal Heib <kamalheib1@gmail.com>
Reviewed-by: Bernard Metzler <bmt@zurich.ibm.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Merge them into their callers, usually the only thing the caller did was
to call the one function, so this is clearer.
Link: https://lore.kernel.org/r/20200702081809.423482-7-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Move DEVX initialization and cleanup flows to the devx.c instead of having
almost empty functions in main.c
Link: https://lore.kernel.org/r/20200702081809.423482-6-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Move flow steering logic to be in separate file and rename flow.c to be
fs.c because it is better describe the content.
Link: https://lore.kernel.org/r/20200702081809.423482-5-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
There are number of counters types supported in mlx5_ib: HW counters,
congestion counters, Q-counters and flow counters. Almost all supporting
code was placed in main.c that made almost impossible to maintain the code
anymore. Let's create separate code namespace for the counters to easy
future generalization effort.
Link: https://lore.kernel.org/r/20200702081809.423482-4-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The restrack code has separate .c, so move callbacks initialization to
that file to improve code locality.
Link: https://lore.kernel.org/r/20200702081809.423482-3-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The mlx5_ib_enable_driver() is local function and doesn't need to be
shared in mlx5_ib, so change it's signature to have static keyword in it.
Link: https://lore.kernel.org/r/20200702081809.423482-2-leon@kernel.org
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
If hns ROCEE is set to level-0 addressing, the length of the entire buffer
can be used as the page size. The driver needn't to split the buffer into
small units because all pages are continuous.
Link: https://lore.kernel.org/r/1593525696-12570-1-git-send-email-liweihang@huawei.com
Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In the loopback tests, the following call trace occurs.
Call Trace:
__rxe_do_task+0x1a/0x30 [rdma_rxe]
rxe_qp_destroy+0x61/0xa0 [rdma_rxe]
rxe_destroy_qp+0x20/0x60 [rdma_rxe]
ib_destroy_qp_user+0xcc/0x220 [ib_core]
uverbs_free_qp+0x3c/0xc0 [ib_uverbs]
destroy_hw_idr_uobject+0x24/0x70 [ib_uverbs]
uverbs_destroy_uobject+0x43/0x1b0 [ib_uverbs]
uobj_destroy+0x41/0x70 [ib_uverbs]
__uobj_get_destroy+0x39/0x70 [ib_uverbs]
ib_uverbs_destroy_qp+0x88/0xc0 [ib_uverbs]
ib_uverbs_handler_UVERBS_METHOD_INVOKE_WRITE+0xb9/0xf0 [ib_uverbs]
ib_uverbs_cmd_verbs+0xb16/0xc30 [ib_uverbs]
The root cause is that the actual RDMA connection is not created in the
loopback tests and the rxe_match_dgid will fail randomly.
To fix this call trace which appear in the loopback tests, skip check of
the dgid.
Fixes: 8700e3e7c485 ("Soft RoCE driver")
Link: https://lore.kernel.org/r/20200630123605.446959-1-leon@kernel.org
Signed-off-by: Zhu Yanjun <yanjunz@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Update the code to allocate and free ib_xrcd structure in the
ib_core instead of inside drivers.
Link: https://lore.kernel.org/r/20200630101855.368895-4-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Expose UAPI to query MR, this will let user space application that
didn't allocate the MR but has access to by owning the matching command
FD to retrieve its information.
Link: https://lore.kernel.org/r/20200630093916.332097-8-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Introduce UAPI to query PD attributes, this can be used to retrieve PD
attributes by having the PD handle of the created one and owning the
command FD for the ucontxet.
Link: https://lore.kernel.org/r/20200630093916.332097-7-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Implement the query ucontext functionality by returning the original
ucontext data as part of an extra mlx5 attribute that holds the driver
UAPI response.
Link: https://lore.kernel.org/r/20200630093916.332097-6-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Refactor mlx5_ib_alloc_ucontext() to set its response fields in a
cleaner way.
It includes,
- Move the relevant code to a self contained function.
- Calculate the response length once and drop redundant code all around.
- Reuse previously set ucontext fields once preparing the response.
The self contained function will be used in next patch as part of
implementing the query ucontext functionality.
Link: https://lore.kernel.org/r/20200630093916.332097-5-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Expose UAPI to query ucontext, this will let user space application that
didn't allocate the ucontext but has access to by owning the matching
command FD to retrieve the ucontext information.
Link: https://lore.kernel.org/r/20200630093916.332097-4-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Set IOVA on IB MR in uverbs layer to let all drivers have it, this
includes both reg/rereg MR flows.
As part of this change cleaned-up this setting from the drivers that
already did it by themselves in their user flows.
Fixes: e6f0330106f4 ("mlx4_ib: set user mr attributes in struct ib_mr")
Link: https://lore.kernel.org/r/20200630093916.332097-3-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Enable CQ ioctl commands by default, this functionality is fully mature
to be used over ioctl, no reason to maintain any more the EXP KCONFIG
entry to enable it.
Link: https://lore.kernel.org/r/20200630093916.332097-2-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Replace the mutex with read write semaphore and use xarray instead of
linked list for XRC target QPs. This will give faster XRC target
lookup. In addition, when QP is closed, don't insert it back to the xarray
if the destroy command failed.
Link: https://lore.kernel.org/r/20200706122716.647338-4-leon@kernel.org
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
ib_alloc_xrcd() already does the required initialization, so move the
uverbs to call it and save code duplication, while cleaning the function
argument lists of that function.
Link: https://lore.kernel.org/r/20200706122716.647338-3-leon@kernel.org
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The mlx5_ib creates XRC domain and uses for creating internal SRQ.
However all that is needed is XRCD number and not full blown ib_xrcd
objects.
Update the code to get and store the number only.
Link: https://lore.kernel.org/r/20200706122716.647338-2-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Allocating an MR flow can only be initiated by kernel users, and not from
userspace so a udata parameter is redundant.
Link: https://lore.kernel.org/r/20200706120343.10816-4-galpress@amazon.com
Signed-off-by: Gal Pressman <galpress@amazon.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Allocating an MR flow can only be initiated by kernel users, and not from
userspace. As a result, the udata parameter is always being passed as
NULL. Rename ib_alloc_mr_user function to ib_alloc_mr and remove the udata
parameter.
Link: https://lore.kernel.org/r/20200706120343.10816-3-galpress@amazon.com
Signed-off-by: Gal Pressman <galpress@amazon.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The common kernel pattern is to check for error, not success. Flip the if
statement accordingly and keep the main flow unindented.
Link: https://lore.kernel.org/r/20200706120343.10816-2-galpress@amazon.com
Signed-off-by: Gal Pressman <galpress@amazon.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Michael Guralnik says:
====================
This series handles IPoIB child interface creation with setting
interface's HW address.
In current implementation, lladdr requested by user is ignored and
overwritten. Child interface gets the same GID as the parent interface and
a QP number which is assigned by the underlying drivers.
In this series we fix this behavior so that user's requested address is
assigned to the newly created interface.
As specific QP number request is not supported for all vendors, QP number
requested by user will still be overwritten when this is not supported.
Behavior of creation of child interfaces through the sysfs mechanism or
without specifying a requested address, stays the same.
====================
Based on the mlx5-next branch at
git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux
due to dependencies.
* branch 'mlx5_ipoib_qpn':
RDMA/ipoib: Handle user-supplied address when creating child
net/mlx5: Enable QP number request when creating IPoIB underlay QP
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Use the address supplied by user when creating a child interface.
Previously, the address requested by the user was ignored and overridden
with parent's GID and the random QP number assigned to the child.
Link: https://lore.kernel.org/r/20200623110105.1225750-3-leon@kernel.org
Signed-off-by: Michael Guralnik <michaelgur@mellanox.com>
Reviewed-by: Feras Daoud <ferasda@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
For debugging purpose it will be easier to understand if prefetch works
okay if it has its own counter. Introduce ODP prefetch counter and count
per MR the total number of prefetched pages.
In addition remove comment which is not relevant anymore and anyway not in
the correct place.
Link: https://lore.kernel.org/r/20200621104147.53795-1-leon@kernel.org
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
There is a race condition where ib_nl_make_request() inserts the request
data into the linked list but the timer in ib_nl_request_timeout() can see
it and destroy it before ib_nl_send_msg() is done touching it. This could
happen, for instance, if there is a long delay allocating memory during
nlmsg_new()
This causes a use-after-free in the send_mad() thread:
[<ffffffffa02f43cb>] ? ib_pack+0x17b/0x240 [ib_core]
[ <ffffffffa032aef1>] ib_sa_path_rec_get+0x181/0x200 [ib_sa]
[<ffffffffa0379db0>] rdma_resolve_route+0x3c0/0x8d0 [rdma_cm]
[<ffffffffa0374450>] ? cma_bind_port+0xa0/0xa0 [rdma_cm]
[<ffffffffa040f850>] ? rds_rdma_cm_event_handler_cmn+0x850/0x850 [rds_rdma]
[<ffffffffa040f22c>] rds_rdma_cm_event_handler_cmn+0x22c/0x850 [rds_rdma]
[<ffffffffa040f860>] rds_rdma_cm_event_handler+0x10/0x20 [rds_rdma]
[<ffffffffa037778e>] addr_handler+0x9e/0x140 [rdma_cm]
[<ffffffffa026cdb4>] process_req+0x134/0x190 [ib_addr]
[<ffffffff810a02f9>] process_one_work+0x169/0x4a0
[<ffffffff810a0b2b>] worker_thread+0x5b/0x560
[<ffffffff810a0ad0>] ? flush_delayed_work+0x50/0x50
[<ffffffff810a68fb>] kthread+0xcb/0xf0
[<ffffffff816ec49a>] ? __schedule+0x24a/0x810
[<ffffffff816ec49a>] ? __schedule+0x24a/0x810
[<ffffffff810a6830>] ? kthread_create_on_node+0x180/0x180
[<ffffffff816f25a7>] ret_from_fork+0x47/0x90
[<ffffffff810a6830>] ? kthread_create_on_node+0x180/0x180
The ownership rule is once the request is on the list, ownership transfers
to the list and the local thread can't touch it any more, just like for
the normal MAD case in send_mad().
Thus, instead of adding before send and then trying to delete after on
errors, move the entire thing under the spinlock so that the send and
update of the lists are atomic to the conurrent threads. Lightly reoganize
things so spinlock safe memory allocations are done in the final NL send
path and the rest of the setup work is done before and outside the lock.
Fixes: 3ebd2fd0d011 ("IB/sa: Put netlink request into the request list before sending")
Link: https://lore.kernel.org/r/1592964789-14533-1-git-send-email-divya.indi@oracle.com
Signed-off-by: Divya Indi <divya.indi@oracle.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
