Rename the qp context related functions and adjusts the code location to
distinguish between the qp context and the entire qp.

Link: https://lore.kernel.org/r/1582526258-13825-3-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@mellanox.com>

Wrap the duplicate code in hip08 and hip06 qp destruction process as
hns_roce_qp_destroy() to simply the qp destroy flow.

Link: https://lore.kernel.org/r/1582526258-13825-2-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@mellanox.com>

There are two paths to update qp producer index into hardware now, one
path is doorbell in post verbs (send and recv), the another is mailbox in
modify qp verb which is called by flush process. This will lead the
hardware to be broken to correctly generate flush cqe.  With stopping
doorbell update and holding qp spinlock in modify qp during flush process,
the problem can be solved.

Fixes: 0425e3e6 ("RDMA/hns: Support flush cqe for hip08 in kernel space")
Link: https://lore.kernel.org/r/1582367158-27030-3-git-send-email-liuyixian@huawei.comSigned-off-by: NYixian Liu <liuyixian@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

As now we already have flush framework, using it instead of current flush
process for qp error in asynchronized interrupt (aeq).

Link: https://lore.kernel.org/r/1582367158-27030-2-git-send-email-liuyixian@huawei.comSigned-off-by: NYixian Liu <liuyixian@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Set revisions that equal to or higher than HIP08_B as default to maintain
backward compatibility.

Link: https://lore.kernel.org/r/1582363039-10714-1-git-send-email-liweihang@huawei.comSigned-off-by: NLang Cheng <chenglang@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The default version provided by "ethtool -i" it the correct way
to identify Driver version. There is no need to overwrite it.

Link: https://lore.kernel.org/r/20200220071239.231800-3-leon@kernel.orgSigned-off-by: NLeon Romanovsky <leonro@mellanox.com>
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

There is no need to set driver version in in-tree kernel code.

Link: https://lore.kernel.org/r/20200220071239.231800-2-leon@kernel.orgSigned-off-by: NLeon Romanovsky <leonro@mellanox.com>
Reviewed-by: NDennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The rdma_cm must be used single threaded.

This appears to be a bug in the design, as it does have lots of locking
that seems like it should allow concurrency. However, when it is all said
and done every single place that uses the cma_exch() scheme is broken, and
all the unlocked reads from the ucma of the cm_id data are wrong too.

syzkaller has been finding endless bugs related to this.

Fixing this in any elegant way is some enormous amount of work. Take a
very big hammer and put a mutex around everything to do with the
ucma_context at the top of every syscall.

Fixes: 75216638 ("RDMA/cma: Export rdma cm interface to userspace")
Link: https://lore.kernel.org/r/20200218210432.GA31966@ziepe.ca
Reported-by: syzbot+adb15cf8c2798e4e0db4@syzkaller.appspotmail.com
Reported-by: syzbot+e5579222b6a3edd96522@syzkaller.appspotmail.com
Reported-by: syzbot+4b628fcc748474003457@syzkaller.appspotmail.com
Reported-by: syzbot+29ee8f76017ce6cf03da@syzkaller.appspotmail.com
Reported-by: syzbot+6956235342b7317ec564@syzkaller.appspotmail.com
Reported-by: syzbot+b358909d8d01556b790b@syzkaller.appspotmail.com
Reported-by: syzbot+6b46b135602a3f3ac99e@syzkaller.appspotmail.com
Reported-by: syzbot+8458d13b13562abf6b77@syzkaller.appspotmail.com
Reported-by: syzbot+bd034f3fdc0402e942ed@syzkaller.appspotmail.com
Reported-by: syzbot+c92378b32760a4eef756@syzkaller.appspotmail.com
Reported-by: syzbot+68b44a1597636e0b342c@syzkaller.appspotmail.com
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Make sure to set the active_{speed, width} attributes to avoid reporting
the same values regardless of the underlying device.

Fixes: 303ae1cd ("rdma/siw: application interface")
Link: https://lore.kernel.org/r/20200218095911.26614-1-kamalheib1@gmail.comSigned-off-by: NKamal Heib <kamalheib1@gmail.com>
Tested-by: NBernard Metzler <bmt@zurich.ibm.com>
Reviewed-by: NBernard Metzler <bmt@zurich.ibm.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Add it

Fixes: 0c4dcd60 ("RDMA/bnxt_re: Refactor hardware queue memory allocation")
Signed-off-by: NStephen Rothwell <sfr@canb.auug.org.au>
Reviewed-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Replacing the dev_err/dbg/warn with ibdev_err/dbg/warn. In the IB device
provider driver these functions are recommended to use.

Currently qplib layer function calls has not been replaced due to
unavailability of ib_device pointer at that layer.

Link: https://lore.kernel.org/r/1581786665-23705-9-git-send-email-devesh.sharma@broadcom.comSigned-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Moving all the fast path doorbell functions at one place under
qplib_res.h. To pass doorbell record information a new structure
bnxt_qplib_db_info has been introduced.  Every roce object holds an
instance of this structure and doorbell information is initialized during
resource creation.

When DB is rung only the current queue index is read from hardware ring
and rest of the data is taken from pre-initialized dbinfo structure.

Link: https://lore.kernel.org/r/1581786665-23705-8-git-send-email-devesh.sharma@broadcom.comSigned-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Cleaning up the notification queue data structures and management
code. The CQ and SRQ event handlers have been type defined instead of
in-place declaration. NQ doorbell register descriptor has been added in
base NQ structure.  The nq->vector has been renamed to nq->msix_vec.

Link: https://lore.kernel.org/r/1581786665-23705-7-git-send-email-devesh.sharma@broadcom.comSigned-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Refactoring the command queue (rcfw) management code. A new data-structure
is introduced to describe the bar register.  each object which deals with
mmio space should have a descriptor structure. This structure specifically
hold DB register information.  Thus, slow path creq structure now hold a
bar register descriptor.

Further cleanup the rcfw structure to introduce the command queue context
and command response event queue context structures. Rest of the rcfw
related code has been touched to incorporate these three structures.

Link: https://lore.kernel.org/r/1581786665-23705-6-git-send-email-devesh.sharma@broadcom.comSigned-off-by: NNaresh Kumar PBS <nareshkumar.pbs@broadcom.com>
Signed-off-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Introducing a new attribute structure to reduce the long list of arguments
passed in bnxt_re_net_ring_alloc() function.

The caller of bnxt_re_net_ring_alloc should fill in the list of attributes
in bnxt_re_ring_attr structure and then pass the pointer to the function.

Link: https://lore.kernel.org/r/1581786665-23705-5-git-send-email-devesh.sharma@broadcom.comSigned-off-by: NNaresh Kumar PBS <nareshkumar.pbs@broadcom.com>
Signed-off-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

At top level there are three major data structure addition.  viz
bnxt_qplib_hwq_attr, bnxt_qplib_sg_info and bnxt_qplib_tqm_ctx

Intorduction of first data structure reduces the arguments list to
bnxt_re_alloc_init_hwq() function. There are changes all over the driver
code to incorporate this new structure. The caller needs to fill the
attribute data structure and pass to this function.

The second data structure is to pass memory region description
viz. sghead, page_size and page_shift. There are changes all over the
driver code to initialize bnxt_re_sg_info data structure. The new data
structure helps to reduce the argument list of __alloc_pbl() function
call.

Till now the TQM rings related members were not collected under any
specific data-structure making it hard to manage. The third data
sctructure bnxt_qplib_tqm_ctx is added to refactor the TQM queue
allocation and initialization.

Link: https://lore.kernel.org/r/1581786665-23705-4-git-send-email-devesh.sharma@broadcom.comSigned-off-by: NNaresh Kumar PBS <nareshkumar.pbs@broadcom.com>
Signed-off-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The chip_ctx member in bnxt_re_dev structure is now a pointer to struct
bnxt_qplib_chip_ctx. Since the member type has changed there are changes
in rest of the code wherever dev->chip_ctx is used.

Link: https://lore.kernel.org/r/1581786665-23705-3-git-send-email-devesh.sharma@broadcom.comSigned-off-by: NNaresh Kumar PBS <nareshkumar.pbs@broadcom.com>
Signed-off-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Restructuring the bnxt_re_create_qp function. Listing below the major
changes:
 - Monolithic central part of create_qp where attributes are initialized
   is now enclosed in one function and this new function has few more
   sub-functions.
 - Top level qp limit checking code moved to a function.
 - GSI QP creation and GSI Shadow qp creation code is handled in a sub
   function.

Link: https://lore.kernel.org/r/1581786665-23705-2-git-send-email-devesh.sharma@broadcom.comSigned-off-by: NNaresh Kumar PBS <nareshkumar.pbs@broadcom.com>
Signed-off-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: NDevesh Sharma <devesh.sharma@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The current codebase makes use of the zero-length array language
extension to the C90 standard, but the preferred mechanism to declare
variable-length types such as these ones is a flexible array member[1][2],
introduced in C99:

struct foo {
        int stuff;
        struct boo array[];
};

By making use of the mechanism above, we will get a compiler warning
in case the flexible array does not occur last in the structure, which
will help us prevent some kind of undefined behavior bugs from being
inadvertently introduced[3] to the codebase from now on.

Also, notice that, dynamic memory allocations won't be affected by
this change:

"Flexible array members have incomplete type, and so the sizeof operator
may not be applied. As a quirk of the original implementation of
zero-length arrays, sizeof evaluates to zero."[1]

This issue was found with the help of Coccinelle.

[1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html
[2] https://github.com/KSPP/linux/issues/21
[3] commit 76497732 ("cxgb3/l2t: Fix undefined behaviour")

Link: https://lore.kernel.org/r/20200213010425.GA13068@embeddedor.comSigned-off-by: NGustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com> # added a few more

Prevent uninitialized fields when new fields are added, and make code look
simpler.

Link: https://lore.kernel.org/r/1582162471-50361-1-git-send-email-liweihang@huawei.comSigned-off-by: NLang Cheng <chenglang@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

There is a spelling mistake in a dev_err error message. Fix it.

Link: https://lore.kernel.org/r/20200214003338.6573-1-colin.king@canonical.comSigned-off-by: NColin Ian King <colin.king@canonical.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

From the comment above the definition of the roundup_pow_of_two() macro:

     The result is undefined when n == 0.

Hence only pass positive values to roundup_pow_of_two(). This patch fixes
the following UBSAN complaint:

  UBSAN: Undefined behaviour in ./include/linux/log2.h:57:13
  shift exponent 64 is too large for 64-bit type 'long unsigned int'
  Call Trace:
   dump_stack+0xa5/0xe6
   ubsan_epilogue+0x9/0x26
   __ubsan_handle_shift_out_of_bounds.cold+0x4c/0xf9
   rxe_qp_from_attr.cold+0x37/0x5d [rdma_rxe]
   rxe_modify_qp+0x59/0x70 [rdma_rxe]
   _ib_modify_qp+0x5aa/0x7c0 [ib_core]
   ib_modify_qp+0x3b/0x50 [ib_core]
   cma_modify_qp_rtr+0x234/0x260 [rdma_cm]
   __rdma_accept+0x1a7/0x650 [rdma_cm]
   nvmet_rdma_cm_handler+0x1286/0x14cd [nvmet_rdma]
   cma_cm_event_handler+0x6b/0x330 [rdma_cm]
   cma_ib_req_handler+0xe60/0x22d0 [rdma_cm]
   cm_process_work+0x30/0x140 [ib_cm]
   cm_req_handler+0x11f4/0x1cd0 [ib_cm]
   cm_work_handler+0xb8/0x344e [ib_cm]
   process_one_work+0x569/0xb60
   worker_thread+0x7a/0x5d0
   kthread+0x1e6/0x210
   ret_from_fork+0x24/0x30

Link: https://lore.kernel.org/r/20200217205714.26937-1-bvanassche@acm.org
Fixes: 8700e3e7 ("Soft RoCE driver")
Signed-off-by: NBart Van Assche <bvanassche@acm.org>
Reviewed-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

bnxt_re HW maintains a GID table with only a single entry for the two
duplicate GID entries (v1 and v2). Driver needs to map stack gid index to
the HW table gid index.  Use the new API rdma_read_gid_hw_context () to
retrieve the HW GID context to get the HW table index.

Link: https://lore.kernel.org/r/1582107594-5180-3-git-send-email-selvin.xavier@broadcom.comSigned-off-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Adding a helper function to retrieve the driver gid context from the gid
attr.

Link: https://lore.kernel.org/r/1582107594-5180-2-git-send-email-selvin.xavier@broadcom.comSuggested-by: NJason Gunthorpe <jgg@mellanox.com>
Signed-off-by: NSelvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Don't use an atomic as a refcount.

Link: https://lore.kernel.org/r/20200218191657.GA29724@ziepe.caReviewed-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

This function accepts a udata but does nothing with it, and is never
passed a !NULL udata. Rename it to ib_create_qp which was the only caller
and remove the udata.

Link: https://lore.kernel.org/r/20200213191911.GA9898@ziepe.caReviewed-by: NMax Gurtovoy <maxg@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

HiP08 RoCE hardware lacks ability(a known hardware problem) to flush
outstanding WQEs if QP state gets into errored mode for some reason.  To
overcome this hardware problem and as a workaround, when QP is detected to
be in errored state during various legs like post send, post receive
etc[1], flush needs to be performed from the driver.

The earlier patch[1] sent to solve the hardware limitation explained in
the cover-letter had a bug in the software flushing leg. It acquired mutex
while modifying QP state to errored state and while conveying it to the
hardware using the mailbox. This caused leg to sleep while holding
spin-lock and caused crash.

Suggested Solution: we have proposed to defer the flushing of the QP in
the Errored state using the workqueue to get around with the limitation of
our hardware.

This patch specifically adds the calls to the flush handler from where
parts of the code like post_send/post_recv etc. when the QP state gets
into the errored mode.

[1] https://patchwork.kernel.org/patch/10534271/

Link: https://lore.kernel.org/r/1580983005-13899-3-git-send-email-liuyixian@huawei.comSigned-off-by: NYixian Liu <liuyixian@huawei.com>
Reviewed-by: NSalil Mehta <salil.mehta@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

HiP08 RoCE hardware lacks ability(a known hardware problem) to flush
outstanding WQEs if QP state gets into errored mode for some reason.  To
overcome this hardware problem and as a workaround, when QP is detected to
be in errored state during various legs like post send, post receive etc
[1], flush needs to be performed from the driver.

The earlier patch[1] sent to solve the hardware limitation explained in
the cover-letter had a bug in the software flushing leg. It acquired mutex
while modifying QP state to errored state and while conveying it to the
hardware using the mailbox. This caused leg to sleep while holding
spin-lock and caused crash.

Suggested Solution:
we have proposed to defer the flushing of the QP in the Errored state
using the workqueue to get around with the limitation of our hardware.

This patch adds the framework of the workqueue and the flush handler
function.

[1] https://patchwork.kernel.org/patch/10534271/

Link: https://lore.kernel.org/r/1580983005-13899-2-git-send-email-liuyixian@huawei.comSigned-off-by: NYixian Liu <liuyixian@huawei.com>
Reviewed-by: NSalil Mehta <salil.mehta@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

For memory regions registered with IB_ACCESS_RELAXED_ORDERING will be dma
mapped with the DMA_ATTR_WEAK_ORDERING.

This will allow reads and writes to the mapping to be weakly ordered, such
change can enhance performance on some supporting architectures.

Link: https://lore.kernel.org/r/20200212073559.684139-1-leon@kernel.orgSigned-off-by: NMichael Guralnik <michaelgur@mellanox.com>
Signed-off-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

IBTA declares QPN as 24bits, mask input to ensure that kernel
doesn't get higher bits and ensure by adding WANR_ONCE() that
other CM users do the same.

Link: https://lore.kernel.org/r/20200212072635.682689-2-leon@kernel.orgSigned-off-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Make sure to set the active_mtu attribute to avoid report the following
invalid value:

$ ibv_devinfo -d siw0 | grep active_mtu
			active_mtu:		invalid MTU (0)

Fixes: 303ae1cd ("rdma/siw: application interface")
Link: https://lore.kernel.org/r/20200205081354.30438-1-kamalheib1@gmail.comSigned-off-by: NKamal Heib <kamalheib1@gmail.com>
Reviewed-by: NGal Pressman <galpress@amazon.com>
Reviewed-by: NBernard Metzler <bmt@zurich.ibm.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The in_dev_for_each_ifa_rtnl() iterator in i40iw_add_ipv4_addr requires
that the rtnl lock be held. But the rtnl_trylock/unlock scheme in this
function does not guarantee it.

Replace the rtnl locking with an RCU lookup using
in_dev_for_each_ifa_rcu()

Fixes: 8e06af71 ("i40iw: add main, hdr, status")
Link: https://lore.kernel.org/r/20200204223840.2151-1-shiraz.saleem@intel.comSigned-off-by: NShiraz Saleem <shiraz.saleem@intel.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

The eqe has a private multi-hop addressing implementation, but there is
already a set of interfaces in the hns driver that can achieve this.

So, simplify the eqe buffer allocation process by using the mtr interface
and remove large amount of repeated logic.

Link: https://lore.kernel.org/r/20200126145835.11368-1-liweihang@huawei.comSigned-off-by: NXi Wang <wangxi11@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Some magic numbers are hard to understand, so replace them with macros or
add some comments for them.

Link: https://lore.kernel.org/r/20200126145504.9700-1-liweihang@huawei.comSigned-off-by: NLang Cheng <chenglang@huawei.com>
Signed-off-by: NYixian Liu <liuyixian@huawei.com>
Signed-off-by: NWenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: NYixing Liu <liuyixing1@huawei.com>
Signed-off-by: NWeihang Li <liweihang@huawei.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Use a refcount_t for atomics being used as a refcount.

Link: https://lore.kernel.org/r/20200126142652.104803-8-leon@kernel.orgSigned-off-by: NParav Pandit <parav@mellanox.com>
Signed-off-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Helper functions which increment/decrement reference count of a
structure read better when they are named with the get/put suffix.

Hence, rename cma_ref/deref_id() to cma_id_get/put().  Also use
cma_get_id() wrapper to find the balancing put() calls.

Link: https://lore.kernel.org/r/20200126142652.104803-7-leon@kernel.orgSigned-off-by: NParav Pandit <parav@mellanox.com>
Signed-off-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Use the refcount variant to capture the reference counting of the cma
device structure.

Link: https://lore.kernel.org/r/20200126142652.104803-6-leon@kernel.orgSigned-off-by: NParav Pandit <parav@mellanox.com>
Signed-off-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Helper functions which increment/decrement reference count of the
structure read better when they are named with the get/put suffix.

Hence, rename cma_ref/deref_dev() to cma_dev_get/put().

Link: https://lore.kernel.org/r/20200126142652.104803-5-leon@kernel.orgSigned-off-by: NParav Pandit <parav@mellanox.com>
Signed-off-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

Use RDMA device port iterator to avoid open coding.

Link: https://lore.kernel.org/r/20200126142652.104803-4-leon@kernel.orgSigned-off-by: NParav Pandit <parav@mellanox.com>
Signed-off-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>

To avoid errors, with attaching ownership of work item and its cm_id
refcount which is decremented in work handler, tie them up in single
helper function. Also avoid code duplication.

Link: https://lore.kernel.org/r/20200126142652.104803-3-leon@kernel.orgSigned-off-by: NParav Pandit <parav@mellanox.com>
Signed-off-by: NLeon Romanovsky <leonro@mellanox.com>
Signed-off-by: NJason Gunthorpe <jgg@mellanox.com>