| Age | Commit message (Collapse) | Author |
|---|
|
This consolidates a bunch of repeated code patterns into a helper.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
After all the rework is done it is now possible to include single flags in
the type macros. Any user of UVERBS_ATTR_STRUCT needs to zero check data
past the end of the known struct to be correct, so make this mandatory,
and get rid of MIN_SZ_OR_ZERO as a user flag.
This changes UVERBS_ATTR_TYPE to refer to a struct of exact size with not
possibility of extension, convert the few users of UVERBS_ATTR_TYPE and
MIN_SZ_OR_ZERO to use UVERBS_ATTR_STRUCT.
The one user of UVERBS_ATTR_STRUCT without MIN_SZ_OR_ZERO is just
confused. There is some padding at the end of that struct, but userspace
always provides it with the padding. The construction doesn't test if the
padding is zero, so it is pointless. Just use UVERBS_ATTR_TYPE.
Finally, rename min_sz_or_zero to zero_trailing to better reflect what it
does and hopefully avoid such mis-uses in the future.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
This newer macro allows specifying a lower bound on the accepted size, and
has an 'unlimited' upper bound. Due to this it never checks for trailing
zeroing so it doesn't make any sense to combine it with MIN_SZ_OR_ZERO, so
drop MIN_SZ_OR_ZERO when they are used together
There were a couple of places that open coded this pattern, switch them to
use the clearer UVERBS_ATTR_MIN_SIZE for clarity.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
This bit of boilerplate isn't really necessary, we can use bitfields
instead of a flags enum and the macros can then individually initialize
them through the __VA_ARGS__ like everything else.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
Hide it inside the macros. The & is confusing and interferes with using
this as a generic DSL in later patches.
Since this also touches almost every line, also run the specs through
clang-format (with 'BinPackParameters: false') to make the maintenance
easier.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
Instead of the large set of indirecting macros, define the few needed
macros to directly instantiate the struct uverbs_oject_tree_def and
associated objects list.
This is small amount of code duplication but the readability is far
better.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
Instead of the large set of indirecting macros, define the few needed
macros to directly instantiate the struct uverbs_method_def and associated
attributes list.
This is small amount of code duplication but the readability is far
better.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
Instead of using a complex cascade of macros, just directly provide the
initializer list each of the declarations is trying to create.
Now that the macros are simplified this also reworks the uverbs_attr_spec
to be friendly to older compilers by eliminating any unnamed
structures/unions inside, and removing the duplication of some fields. The
structure size remains at 16 bytes which was the original motivation for
some of this oddness.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
Two methods are sharing the same attribute constant, but the attribute
definitions are not the same. This should not have been done, instead
split them into two attributes with the same number.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
The specs are required to operate the uverbs file, so they belong inside
the ib_uverbs_device, not inside the ib_device. The spec passed in the
ib_device is just a communication from the driver and should not be used
during runtime.
This also changes the lifetime of the spec memory to match the
ib_uverbs_device, however at this time the spec_root can still contain
driver pointers after disassociation, so it cannot be used if ib_dev is
NULL. This is preparation for another series.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
For dependencies, branch based on 'mellanox/mlx5-next' of
git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux.git
Pull Dump and fill MKEY from Leon Romanovsky:
====================
MLX5 IB HCA offers the memory key, dump_fill_mkey to increase performance,
when used in a send or receive operations.
It is used to force local HCA operations to skip the PCI bus access, while
keeping track of the processed length in the ibv_sge handling.
In this three patch series, we expose various bits in our HW spec
file (mlx5_ifc.h), move unneeded for mlx5_core FW command and export such
memory key to user space thought our mlx5-abi header file.
====================
Botched auto-merge in mlx5_ib_alloc_ucontext() resolved by hand.
* branch 'mlx5-dump-fill-mkey':
IB/mlx5: Expose dump and fill memory key
net/mlx5: Add hardware definitions for dump_fill_mkey
net/mlx5: Limit scope of dump_fill_mkey function
net/mlx5: Rate limit errors in command interface
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
MLX5 IB HCA offers the memory key, dump_fill_mkey to boost
performance, when used in a send or receive operations.
It is used to force local HCA operations to skip the PCI bus access,
while keeping track of the processed length in the ibv_sge handling.
Meaning, instead of a PCI write access the HCA leaves the target
memory untouched, and skips filling that packet section. Similar
behavior is done upon send, the HCA skips data in memory relevant
to this key and saves PCI bus access.
This functionality saves PCI read/write operations.
Signed-off-by: Yonatan Cohen <yonatanc@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Reviewed-by: Guy Levi <guyle@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
MLX5 IB HCA offers the memory key, dump_fill_mkey to boost
performance by forcing local HCA operations to skip the PCI bus
access,
This patch adds needed hardware definitions.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
mlx5_core_dump_fill_mkey() is going to be used in next
patch in IB and doesn't need to be visible to whole
mlx5_core. Move that command to mlx5_ib.
Signed-off-by: Yonatan Cohen <yonatanc@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
IB_MULTICAST_LID_BASE is defined as follows:
#define IB_MULTICAST_LID_BASE cpu_to_be16(0xC000)
Hence use be16_to_cpu() to convert it to CPU endianness. Compile-tested
only.
Fixes: af808ece5ce9 ("IB/SA: Check dlid before SA agent queries for ClassPortInfo")
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Venkata Sandeep Dhanalakota <venkata.s.dhanalakota@intel.com>
Cc: Mike Marciniszyn <mike.marciniszyn@intel.com>
Cc: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Improve uverbs_cleanup_ucontext algorithm to work properly when the
topology graph of the objects cannot be determined at compile time. This
is the case with objects created via the devx interface in mlx5.
Typically uverbs objects must be created in a strict topologically sorted
order, so that LIFO ordering will generally cause them to be freed
properly. There are only a few cases (eg memory windows) where objects can
point to things out of the strict LIFO order.
Instead of using an explicit ordering scheme where the HW destroy is not
allowed to fail, go over the list multiple times and allow the destroy
function to fail. If progress halts then a final, desperate, cleanup is
done before leaking the memory. This indicates a driver bug.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Following the removal of ib_create_flow(), adjust the code to get rid of
ib_destroy_flow() too.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
There are no kernel users of this interface so lets drop it.
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Now that all users have been converted to use the version of these APIs
that returns a gid_attr pointer we can delete the old entry points.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
While processing a path record entry in CM messages the associated GID
attribute is now also supplied.
Currently for RoCE a netdevice's net namespace pointer and ifindex are
stored in path record entry. Both of these fields of the netdev can change
anytime while processing CM messages. Additionally storing net namespace
without holding reference will lead to use-after-free crash. Therefore it
is removed. Netdevice information for RoCE is instead provided via
referenced gid attribute in ib_cm requests.
Such a design leads to a situation where the kernel can crash when the net
pointer becomes invalid. However today it is always initialized to
init_net, which cannot become invalid. In order to support processing
packets in any arbitrary namespace of the received packet, it is necessary
to avoid such conditions.
This patch removes the dependency on the net pointer and ifindex; instead
it will rely on SGID attribute which contains a pointer to netdev.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
Make the sgid_attr available along with path information to the event
consumer, this allows the consumer to keep using the same GID table entry
as the event is related to.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
Hold reference to the the sgid_attr which is used in a cm_id until the
cm_id is destroyed.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
The work completion is inspected to determine what dgid table entry was
used to receieve the packet, produces a sgid_attr that matches and sticks
it in the ah_attr.
All callers of this function are now required to release the ah_attr on
success.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
In this context the uobject is not allowed to be NULL, so type is the same
as uobject->type, and at least for IDR, id is the same as uobject->id.
FD objects should never handle the FD number outside the uAPI boundary
code.
Suggested-by: Guy Levi <guyle@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
For dependencies, branch based on 'mellanox/mlx5-next' of
git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux.git
Pull RoCE ICRC counters from Leon Romanovsky:
====================
This series exposes RoCE ICRC counter through existing RDMA hw_counters
sysfs interface.
The first patch has all HW definitions in mlx5_ifc.h file and second patch
is the actual counter implementation.
====================
* branch 'icrc-counter':
IB/mlx5: Support RoCE ICRC encapsulated error counter
net/mlx5: Add RoCE RX ICRC encapsulated counter
|
|
Add capability bit in PCAM register and RoCE ICRC error counter
to PPCNT register.
Signed-off-by: Talat Batheesh <talatb@mellanox.com>
Reviewed-by: Mark Bloch <markb@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
Move some s_flags defines out of rdmavt and into hfi1 because they are
hfi1 specific and therefore should remain in the driver instead of
bubbling up to rdmavt.
Document device specific ranges in rdmavt and remap
those in hfi1.
Reviewed-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Return the matching device EQN for a given user vector number via the
DEVX interface.
Note:
EQs are owned by the kernel and shared by all user processes.
Basically, a user CQ can point to any EQ.
The kernel doesn't enforce any such limitation today either.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Add support to register a memory with the firmware via the DEVX
interface.
The driver translates a given user address to ib_umem then it will
register the physical addresses with the firmware and get a unique id
for this registration to be used for this virtual address.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Return a device UAR index for a given user index via the DEVX interface.
Security note:
The hardware protection mechanism works like this: Each device object that
is subject to UAR doorbells (QP/SQ/CQ) gets a UAR ID (called uar_page in
the device specification manual) upon its creation. Then upon doorbell,
hardware fetches the object context for which the doorbell was rang, and
validates that the UAR through which the DB was rang matches the UAR ID
of the object.
If no match the doorbell is silently ignored by the hardware. Of
course, the user cannot ring a doorbell on a UAR that was not mapped to
it.
Now in devx, as the devx kernel does not manipulate the QP/SQ/CQ command
mailboxes (except tagging them with UID), we expose to the user its UAR
ID, so it can embed it in these objects in the expected specification
format. So the only thing the user can do is hurt itself by creating a
QP/SQ/CQ with a UAR ID other than his, and then in this case other users
may ring a doorbell on its objects.
The consequence of that will be that another user can schedule a QP/SQ
of the buggy user for execution (just insert it to the hardware schedule
queue or arm its CQ for event generation), no further harm is expected.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Add support in DEVX for modify and query commands, the required lock is
taken (i.e. READ/WRITE) by the KABI infrastructure accordingly.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Add support to create and destroy firmware objects via the DEVX
interface.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Add support to run general firmware command via the DEVX interface.
A command that works on some object (e.g. CQ, WQ, etc.) will be added
in next patches while maintaining the required object lock.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Introduce DEVX to enable direct device commands in downstream patches
from this series.
In that mode of work the firmware manages the isolation between
processes' resources and as such a DEVX user id is created and assigned
to the given user context upon allocation request.
A capability check is done to make sure that this feature is really
supported by the firmware prior to creating the DEVX user id.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Drivers that use the IOCTL API may have the ib_uverbs_file and need a
way to get the related ib_ucontext from it, this is enabled by this
patch.
Downstream patches from this series will use it.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Introduce a new macro to be used for global methods on a singleton
object.
This macros sets internally the type_attrs to be NULL as such an object
can't be created.
Downstream patches from this series will use this macro.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Sometimes the uverbs uAPI doesn't really care about the structure it gets
from user-space. All it wants to do is to allocate enough space and send
it to the hardware/provider driver. Adding a UVERBS_ATTR_MIN_SIZE that
could be used for this scenarios. We use USHRT_MAX as the kernel known
size to bypass any zero validations.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Adding UVERBS_ATTR_SPEC_F_ALLOC_AND_COPY flag to PTR_IN attributes.
By using this flag, the parse automatically allocates and copies the
user-space data. This data is accessible by using uverbs_attr_get_len
and uverbs_attr_get_alloced_ptr inline accessor functions from the
handler.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
From git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux.git
* branch 'mellanox/mlx5-next':
net/mlx5: Expose DEVX specification
net/mlx5: Prevent warns in dmesg upon firmware commands
|
|
This patch updates the mlx5_ifc structures and
command interface to support DEVX.
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
This patch replaces the ib_device_attr.max_sge with max_send_sge and
max_recv_sge. It allows ulps to take advantage of devices that have very
different send and recv sge depths. For example cxgb4 has a max_recv_sge
of 4, yet a max_send_sge of 16. Splitting out these attributes allows
much more efficient use of the SQ for cxgb4 with ulps that use the RDMA_RW
API. Consider a large RDMA WRITE that has 16 scattergather entries.
With max_sge of 4, the ulp would send 4 WRITE WRs, but with max_sge of
16, it can be done with 1 WRITE WR.
Acked-by: Sagi Grimberg <sagi@grimberg.me>
Acked-by: Christoph Hellwig <hch@lst.de>
Acked-by: Selvin Xavier <selvin.xavier@broadcom.com>
Acked-by: Shiraz Saleem <shiraz.saleem@intel.com>
Acked-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Allow users of the IDR to use the XArray lock for their own
synchronisation purposes. The IDR continues to rely on the caller to
handle locking, but this lets the caller use the lock embedded in the
IDR data structure instead of allocating their own lock.
Signed-off-by: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
If the AH has a GRH then hold a reference to the sgid_attr inside the
common struct.
If the QP is modified with an AV that includes a GRH then also hold a
reference to the sgid_attr inside the common struct.
This informs the cache that the sgid_index is in-use so long as the AH or
QP using it exists.
This also means that all drivers can access the sgid_attr directly from
the ah_attr instead of querying the cache during their UD post-send paths.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
The core code now ensures that all driver callbacks that receive an
rdma_ah_attrs will have a sgid_attr's pointer if there is a GRH present.
Drivers can use this pointer instead of calling a query function with
sgid_index. This simplifies the drivers and also avoids races where a
gid_index lookup may return different data if it is changed.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
Introduce AH attribute copy, move and replace APIs to be used by core and
provider drivers.
In CM code flow when ah attribute might be re-initialized twice while
processing incoming request, or initialized once while from path record
while sending out CM requests. Therefore use rdma_move_ah_attr API to
handle such scenarios instead of memcpy().
Provider drivers keeps a copy ah_attr during the lifetime of the ah.
Therefore, use rdma_replace_ah_attr() which conditionally release
reference to old ah_attr and holds reference to new attribute whose
referrence is released when the AH is freed.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
The sgid_attr will ultimately replace the sgid_index in the ah_attr.
This will allow for all layers to have a consistent view of what
gid table entry was selected as processing runs through all stages of the
stack.
This commit introduces the pointer and ensures it is set before calling
any driver callback that includes a struct ah_attr callback, allowing
future patches to adjust both the drivers and the callers to use
sgid_attr instead of sgid_index.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
|
|
If the gid_attr argument is NULL then the functions behave identically to
rdma_query_gid. ib_query_gid just calls ib_get_cached_gid, so everything
can be consolidated to one function.
Now that all callers either use rdma_query_gid() or ib_get_cached_gid(),
ib_query_gid() API is removed.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
These versions are functionally similar but all return gid_attrs and
related information via reference instead of via copy.
The old API is preserved, implemented as wrappers around the new, until
all callers can be converted.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
This patch introduces three APIs, rdma_get_gid_attr(),
rdma_put_gid_attr(), and rdma_hold_gid_attr() which expose the reference
counting for GID table entries to the entire stack. The kref counting is
based on the struct ib_gid_attr pointer
Later patches will convert more cache query function to return struct
ib_gid_attrs.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
|
Now that ib_gid_attr contains the GID, make use of that in the add_gid()
callback functions for the provider drivers to simplify the add_gid()
implementations.
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
|
