Dma-buf based memory region requires one extra parameter and is processed
quite differently. Adding a separate method allows clean separation from
regular memory regions.
Link: https://lore.kernel.org/r/1608067636-98073-3-git-send-email-jianxin.xiong@intel.com
Signed-off-by: Jianxin Xiong <jianxin.xiong@intel.com>
Reviewed-by: Sean Hefty <sean.hefty@intel.com>
Acked-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Acked-by: Christian Koenig <christian.koenig@amd.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Fixes the following W=1 kernel build warning(s):
drivers/infiniband/core/device.c:1896: warning: Function parameter or member 'ibdev' not described in 'ib_get_client_nl_info'
drivers/infiniband/core/device.c:1896: warning: Function parameter or member 'client_name' not described in 'ib_get_client_nl_info'
drivers/infiniband/core/device.c:1896: warning: Function parameter or member 'res' not described in 'ib_get_client_nl_info'
drivers/infiniband/core/device.c:2328: warning: Function parameter or member 'nldev_cb' not described in 'ib_enum_all_devs'
drivers/infiniband/core/device.c:2328: warning: Function parameter or member 'skb' not described in 'ib_enum_all_devs'
Link: https://lore.kernel.org/r/20210118223929.512175-3-lee.jones@linaro.org
Cc: Doug Ledford <dledford@redhat.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: linux-rdma@vger.kernel.org
Signed-off-by: Lee Jones <lee.jones@linaro.org>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
In procedure ib_register_device, procedure kobject_uevent is called
(advertising that the device is ready for userspace usage) even when
device_enable_and_get() returned an error.
As a result, various RDMA modules attempted to register for the device
even while the device driver was preparing to unregister the device.
Fix this by advertising the device availability only after enabling the
device succeeds.
Fixes: e7a5b4aafd ("RDMA/device: Don't fire uevent before device is fully initialized")
Link: https://lore.kernel.org/r/20201208073545.9723-3-leon@kernel.org
Suggested-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The CQ pool mechanism had two problems:
1. The CQ pool lists were uninitialized in the device registration error
flow. As a result, all the list pointers remained NULL. This caused
the kernel to crash (in procedure ib_cq_pool_destroy) when that error
flow was taken (and unregister called). The stack trace snippet:
BUG: kernel NULL pointer dereference, address: 0000000000000000
#PF: supervisor read access in kernel mode
#PF: error_code(0×0000) ? not-present page
PGD 0 P4D 0
Oops: 0000 [#1] SMP PTI
. . .
RIP: 0010:ib_cq_pool_destroy+0x1b/0×70 [ib_core]
. . .
Call Trace:
disable_device+0x9f/0×130 [ib_core]
__ib_unregister_device+0x35/0×90 [ib_core]
ib_register_device+0x529/0×610 [ib_core]
__mlx5_ib_add+0x3a/0×70 [mlx5_ib]
mlx5_add_device+0x87/0×1c0 [mlx5_core]
mlx5_register_interface+0x74/0xc0 [mlx5_core]
do_one_initcall+0x4b/0×1f4
do_init_module+0x5a/0×223
load_module+0x1938/0×1d40
2. At device unregister, when cleaning up the cq pool, the cq's in the
pool lists were freed, but the cq entries were left in the list.
The fix for the first issue is to initialize the cq pool lists when the
ib_device structure is allocated for a new device (in procedure
_ib_alloc_device).
The fix for the second problem is to delete cq entries from the pool lists
when cleaning up the cq pool.
In addition, procedure ib_cq_pool_destroy() is renamed to the more
appropriate name ib_cq_pool_cleanup().
Fixes: 4aa1615268 ("RDMA/core: Fix ordering of CQ pool destruction")
Link: https://lore.kernel.org/r/20201208073545.9723-2-leon@kernel.org
Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Use the ib_dma_* helpers to skip the DMA translation instead. This
removes the last user if dma_virt_ops and keeps the weird layering
violation inside the RDMA core instead of burderning the DMA mapping
subsystems with it. This also means the software RDMA drivers now don't
have to mess with DMA parameters that are not relevant to them at all, and
that in the future we can use PCI P2P transfers even for software RDMA, as
there is no first fake layer of DMA mapping that the P2P DMA support.
Link: https://lore.kernel.org/r/20201106181941.1878556-8-hch@lst.de
Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Mike Marciniszyn <mike.marciniszyn@cornelisnetworks.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Drivers that need a uverbs AH should instead set the create_user_ah() op
similar to reg_user_mr(). MODIFY_AH and QUERY_AH cmds were never
implemented so are just deleted.
Link: https://lore.kernel.org/r/11-v1-caa70ba3d1ab+1436e-ucmd_mask_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
No driver sets it, and the core code sets a maximum mask, simply remove
it.
Disabled operations are now handled either by having a NULL ops pointer,
or by having the common driver callbacks check for unsupported extended
attributes.
Link: https://lore.kernel.org/r/9-v1-caa70ba3d1ab+1436e-ucmd_mask_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Each driver should check that the QP attrs create_flags is supported.
Unfortuantely when create_flags was added to the QP attrs the drivers were
not updated. uverbs_ex_cmd_mask was used to block it - even though kernel
drivers use these flags too.
Check that flags is zero in all drivers that don't use it, remove
IB_USER_VERBS_EX_CMD_CREATE_QP from uverbs_ex_cmd_mask. Fix the error code
to be EOPNOTSUPP.
Link: https://lore.kernel.org/r/8-v1-caa70ba3d1ab+1436e-ucmd_mask_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Each driver should check that the CQ attrs is supported. Unfortuantely
when flags was added to the CQ attrs the drivers were not updated,
uverbs_ex_cmd_mask was used to block it. This was missed when create CQ
was converted to ioctl, so non-zero flags could have been passed into
drivers.
Check that flags is zero in all drivers that don't use it, remove
IB_USER_VERBS_EX_CMD_CREATE_CQ from uverbs_ex_cmd_mask.
Fixes: 41b2a71fc8 ("IB/uverbs: Move ioctl path of create_cq and destroy_cq to a new file")
Link: https://lore.kernel.org/r/7-v1-caa70ba3d1ab+1436e-ucmd_mask_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Each driver should check that it can support the provided attr_mask during
modify_qp. IB_USER_VERBS_EX_CMD_MODIFY_QP was being used to block
modify_qp_ex because the driver didn't check RATE_LIMIT.
Link: https://lore.kernel.org/r/6-v1-caa70ba3d1ab+1436e-ucmd_mask_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
uverbs was blocking srq_types the driver doesn't support based on the
CREATE_XSRQ cmd_mask. Fix all drivers to check for supported srq_types
during create_srq and move CREATE_XSRQ to the core code.
Link: https://lore.kernel.org/r/5-v1-caa70ba3d1ab+1436e-ucmd_mask_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
These functions all depend on the driver providing a specific op:
- REREG_MR is rereg_user_mr(). bnxt_re set this without providing the op
- ATTACH/DEATCH_MCAST is attach_mcast()/detach_mcast(). usnic set this
without providing the op
- OPEN_QP doesn't involve the driver but requires a XRCD. qedr provides
xrcd but forgot to set it, usnic doesn't provide XRCD but set it anyhow.
- OPEN/CLOSE_XRCD are the ops alloc_xrcd()/dealloc_xrcd()
- CREATE_SRQ/DESTROY_SRQ are the ops create_srq()/destroy_srq()
- QUERY/MODIFY_SRQ is op query_srq()/modify_srq(). hns sets this but
sometimes supplies a NULL op.
- RESIZE_CQ is op resize_cq(). bnxt_re sets this boes doesn't supply an op
- ALLOC/DEALLOC_MW is alloc_mw()/dealloc_mw(). cxgb4 provided an
(now deleted) implementation but no userspace
All drivers were checked that no drivers provide the op without also
setting uverbs_cmd_mask so this should have no functional change.
Link: https://lore.kernel.org/r/4-v1-caa70ba3d1ab+1436e-ucmd_mask_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Since a while now the uverbs layer checks if the driver implements a
function before allowing the ucmd to proceed. This largely obsoletes the
cmd_mask stuff, but there is some tricky bits in drivers preventing it
from being removed.
Remove the easy elements of uverbs_ex_cmd_mask by pre-setting them in the
core code. These are triggered soley based on the related ops function
pointer.
query_device_ex is not triggered based on an op, but all drivers already
implement something compatible with the extension, so enable it globally
too.
Link: https://lore.kernel.org/r/2-v1-caa70ba3d1ab+1436e-ucmd_mask_jgg@nvidia.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The code in setup_dma_device has become rather convoluted, move all of
this to the drivers. Drives now pass in a DMA capable struct device which
will be used to setup DMA, or drivers must fully configure the ibdev for
DMA and pass in NULL.
Other than setting the masks in rvt all drivers were doing this already
anyhow.
mthca, mlx4 and mlx5 were already setting up maximum DMA segment size for
DMA based on their hardweare limits in:
__mthca_init_one()
dma_set_max_seg_size (1G)
__mlx4_init_one()
dma_set_max_seg_size (1G)
mlx5_pci_init()
set_dma_caps()
dma_set_max_seg_size (2G)
Other non software drivers (except usnic) were extended to UINT_MAX [1, 2]
instead of 2G as was before.
[1] https://lore.kernel.org/linux-rdma/20200924114940.GE9475@nvidia.com/
[2] https://lore.kernel.org/linux-rdma/20200924114940.GE9475@nvidia.com/
Link: https://lore.kernel.org/r/20201008082752.275846-1-leon@kernel.org
Link: https://lore.kernel.org/r/6b2ed339933d066622d5715903870676d8cc523a.1602590106.git.mchehab+huawei@kernel.org
Suggested-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Mauro Carvalho Chehab <mchehab+huawei@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
From Maor Gottlieb says:
====================
This series extends __sg_alloc_table_from_pages to allow chaining of new
pages to an already initialized SG table.
This allows for drivers to utilize the optimization of merging contiguous
pages without a need to pre allocate all the pages and hold them in a very
large temporary buffer prior to the call to SG table initialization.
The last patch changes the Infiniband core to use the new API. It removes
duplicate functionality from the code and benefits from the optimization
of allocating dynamic SG table from pages.
In huge pages system of 2MB page size, without this change, the SG table
would contain x512 SG entries.
====================
* branch 'dynamic_sg':
RDMA/umem: Move to allocate SG table from pages
lib/scatterlist: Add support in dynamic allocation of SG table from pages
tools/testing/scatterlist: Show errors in human readable form
tools/testing/scatterlist: Rejuvenate bit-rotten test
Move allocation and destruction of memory windows under ib_core
responsibility and clean drivers to ensure that no updates to MW
ib_core structures are done in driver layer.
Link: https://lore.kernel.org/r/20200902081623.746359-2-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
rxe will hold a refcount on the IB device as long as CQ objects exist,
this causes destruction of a rxe device to hang if the CQ pool has any
cached CQs since they are being destroyed after the refcount must go to
zero.
Treat the CQ pool like a client and create/destroy it before/after all
other clients. No users of CQ pool can exist past a client remove call.
Link: https://lore.kernel.org/r/e8a240aa-9e9b-3dca-062f-9130b787f29b@acm.org
Fixes: c7ff819aef ("RDMA/core: Introduce shared CQ pool API")
Tested-by: Bart Van Assche <bvanassche@acm.org>
Tested-by: Yi Zhang <yi.zhang@redhat.com>
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Smaller set of RDMA updates. A smaller number of 'big topics' with the
majority of changes being driver updates.
- Driver updates for hfi1, rxe, mlx5, hns, qedr, usnic, bnxt_re
- Removal of dead or redundant code across the drivers
- RAW resource tracker dumps to include a device specific data blob for
device objects to aide device debugging
- Further advance the IOCTL interface, remove the ability to turn it off.
Add QUERY_CONTEXT, QUERY_MR, and QUERY_PD commands
- Remove stubs related to devices with no pkey table
- A shared CQ scheme to allow multiple ULPs to share the CQ rings of a
device to give higher performance
- Several more static checker, syzkaller and rare crashers fixed
-----BEGIN PGP SIGNATURE-----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=0z2L
-----END PGP SIGNATURE-----
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma
Pull rdma updates from Jason Gunthorpe:
"A quiet cycle after the larger 5.8 effort. Substantially cleanup and
driver work with a few smaller features this time.
- Driver updates for hfi1, rxe, mlx5, hns, qedr, usnic, bnxt_re
- Removal of dead or redundant code across the drivers
- RAW resource tracker dumps to include a device specific data blob
for device objects to aide device debugging
- Further advance the IOCTL interface, remove the ability to turn it
off. Add QUERY_CONTEXT, QUERY_MR, and QUERY_PD commands
- Remove stubs related to devices with no pkey table
- A shared CQ scheme to allow multiple ULPs to share the CQ rings of
a device to give higher performance
- Several more static checker, syzkaller and rare crashers fixed"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (121 commits)
RDMA/mlx5: Fix flow destination setting for RDMA TX flow table
RDMA/rxe: Remove pkey table
RDMA/umem: Add a schedule point in ib_umem_get()
RDMA/hns: Fix the unneeded process when getting a general type of CQE error
RDMA/hns: Fix error during modify qp RTS2RTS
RDMA/hns: Delete unnecessary memset when allocating VF resource
RDMA/hns: Remove redundant parameters in set_rc_wqe()
RDMA/hns: Remove support for HIP08_A
RDMA/hns: Refactor hns_roce_v2_set_hem()
RDMA/hns: Remove redundant hardware opcode definitions
RDMA/netlink: Remove CAP_NET_RAW check when dump a raw QP
RDMA/include: Replace license text with SPDX tags
RDMA/rtrs: remove WQ_MEM_RECLAIM for rtrs_wq
RDMA/rtrs-clt: add an additional random 8 seconds before reconnecting
RDMA/cma: Execute rdma_cm destruction from a handler properly
RDMA/cma: Remove unneeded locking for req paths
RDMA/cma: Using the standard locking pattern when delivering the removal event
RDMA/cma: Simplify DEVICE_REMOVAL for internal_id
RDMA/efa: Add EFA 0xefa1 PCI ID
RDMA/efa: User/kernel compatibility handshake mechanism
...
The query_pkey() isn't mandatory for the iwarp providers, so remove this
requirement from the RDMA core.
Link: https://lore.kernel.org/r/20200714183414.61069-4-kamalheib1@gmail.com
Signed-off-by: Kamal Heib <kamalheib1@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Avoid the overhead of the dma ops support for tiny builds that only
use the direct mapping.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Alexey Kardashevskiy <aik@ozlabs.ru>
Reviewed-by: Alexey Kardashevskiy <aik@ozlabs.ru>
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 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>
ib_unregister_device_queued() can only be used by drivers using the new
dealloc_device callback flow, and it has a safety WARN_ON to ensure
drivers are using it properly.
However, if unregister and register are raced there is a special
destruction path that maintains the uniform error handling semantic of
'caller does ib_dealloc_device() on failure'. This requires disabling the
dealloc_device callback which triggers the WARN_ON.
Instead of using NULL to disable the callback use a special function
pointer so the WARN_ON does not trigger.
Fixes: d0899892ed ("RDMA/device: Provide APIs from the core code to help unregistration")
Link: https://lore.kernel.org/r/0-v1-a36d512e0a99+762-syz_dealloc_driver_jgg@nvidia.com
Reported-by: syzbot+4088ed905e4ae2b0e13b@syzkaller.appspotmail.com
Suggested-by: Hillf Danton <hdanton@sina.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Add support to get resource dump in raw format. It enable drivers to
return the entire device specific QP/CQ/MR context without a need from the
driver to set each field separately.
The raw query returns only the device specific data, general data is still
returned by using the existing queries.
Example:
$ rdma res show mr dev mlx5_1 mrn 2 -r -j
[{"ifindex":7,"ifname":"mlx5_1",
"data":[0,4,255,254,0,0,0,0,0,0,0,0,16,28,0,216,...]}]
Link: https://lore.kernel.org/r/20200623113043.1228482-9-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>
In order to avoid double multiplexing of the resource when it is a cm id,
add a dedicated callback function. In addition remove fill_res_entry which
is not used anymore.
Link: https://lore.kernel.org/r/20200623113043.1228482-8-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>
In order to avoid double multiplexing of the resource when it is a QP, add
a dedicated callback function.
Link: https://lore.kernel.org/r/20200623113043.1228482-7-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>
In order to avoid double multiplexing of the resource when it is a CQ, add
a dedicated callback function.
Link: https://lore.kernel.org/r/20200623113043.1228482-6-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>
In order to avoid double multiplexing of the resource when it is a MR, add
a dedicated callback function.
Link: https://lore.kernel.org/r/20200623113043.1228482-5-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>
Allow a ULP to ask the core to provide a completion queue based on a
least-used search on a per-device CQ pools. The device CQ pools grow in a
lazy fashion when more CQs are requested.
This feature reduces the amount of interrupts when using many QPs. Using
shared CQs allows for more effcient completion handling. It also reduces
the amount of overhead needed for CQ contexts.
Test setup:
Intel(R) Xeon(R) Platinum 8176M CPU @ 2.10GHz servers.
Running NVMeoF 4KB read IOs over ConnectX-5EX across Spectrum switch.
TX-depth = 32. The patch was applied in the nvme driver on both the target
and initiator. Four controllers are accessed from each core. In the
current test case we have exposed sixteen NVMe namespaces using four
different subsystems (four namespaces per subsystem) from one NVM port.
Each controller allocated X queues (RDMA QPs) and attached to Y CQs.
Before this series we had X == Y, i.e for four controllers we've created
total of 4X QPs and 4X CQs. In the shared case, we've created 4X QPs and
only X CQs which means that we have four controllers that share a
completion queue per core. Until fourteen cores there is no significant
change in performance and the number of interrupts per second is less than
a million in the current case.
==================================================
|Cores|Current KIOPs |Shared KIOPs |improvement|
|-----|---------------|--------------|-----------|
|14 |2332 |2723 |16.7% |
|-----|---------------|--------------|-----------|
|20 |2086 |2712 |30% |
|-----|---------------|--------------|-----------|
|28 |1971 |2669 |35.4% |
|=================================================
|Cores|Current avg lat|Shared avg lat|improvement|
|-----|---------------|--------------|-----------|
|14 |767us |657us |14.3% |
|-----|---------------|--------------|-----------|
|20 |1225us |943us |23% |
|-----|---------------|--------------|-----------|
|28 |1816us |1341us |26.1% |
========================================================
|Cores|Current interrupts|Shared interrupts|improvement|
|-----|------------------|-----------------|-----------|
|14 |1.6M/sec |0.4M/sec |72% |
|-----|------------------|-----------------|-----------|
|20 |2.8M/sec |0.6M/sec |72.4% |
|-----|------------------|-----------------|-----------|
|28 |2.9M/sec |0.8M/sec |63.4% |
====================================================================
|Cores|Current 99.99th PCTL lat|Shared 99.99th PCTL lat|improvement|
|-----|------------------------|-----------------------|-----------|
|14 |67ms |6ms |90.9% |
|-----|------------------------|-----------------------|-----------|
|20 |5ms |6ms |-10% |
|-----|------------------------|-----------------------|-----------|
|28 |8.7ms |6ms |25.9% |
|===================================================================
Performance improvement with sixteen disks (sixteen CQs per core) is
comparable.
Link: https://lore.kernel.org/r/1590568495-101621-3-git-send-email-yaminf@mellanox.com
Signed-off-by: Yamin Friedman <yaminf@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
When a client is added it isn't allowed to fail, but all the client's have
various failure paths within their add routines.
This creates the very fringe condition where the client was added, failed
during add and didn't set the client_data. The core code will then still
call other client_data centric ops like remove(), rename(), get_nl_info(),
and get_net_dev_by_params() with NULL client_data - which is confusing and
unexpected.
If the add() callback fails, then do not call any more client ops for the
device, even remove.
Remove all the now redundant checks for NULL client_data in ops callbacks.
Update all the add() callbacks to return error codes
appropriately. EOPNOTSUPP is used for cases where the ULP does not support
the ib_device - eg because it only works with IB.
Link: https://lore.kernel.org/r/20200421172440.387069-1-leon@kernel.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Acked-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This lock is used to protect the qp->open_list linked list. As a side
effect it seems to also globally serialize the qp event_handler, but it
isn't clear if that is a deliberate design.
Link: https://lore.kernel.org/r/20191212113024.336702-5-leon@kernel.org
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Currently when the low level driver notifies Pkey, GID, and port change
events they are notified to the registered handlers in the order they are
registered.
IB core and other ULPs such as IPoIB are interested in GID, LID, Pkey
change events.
Since all GID queries done by ULPs are serviced by IB core, and the IB
core deferes cache updates to a work queue, it is possible for other
clients to see stale cache data when they handle their own events.
For example, the below call tree shows how ipoib will call
rdma_query_gid() concurrently with the update to the cache sitting in the
WQ.
mlx5_ib_handle_event()
ib_dispatch_event()
ib_cache_event()
queue_work() -> slow cache update
[..]
ipoib_event()
queue_work()
[..]
work handler
ipoib_ib_dev_flush_light()
__ipoib_ib_dev_flush()
ipoib_dev_addr_changed_valid()
rdma_query_gid() <- Returns old GID, cache not updated.
Move all the event dispatch to a work queue so that the cache update is
always done before any clients are notified.
Fixes: f35faa4ba9 ("IB/core: Simplify ib_query_gid to always refer to cache")
Link: https://lore.kernel.org/r/20191212113024.336702-3-leon@kernel.org
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This is another round of bug fixing and cleanup. This time the focus is on
the driver pattern to use mmu notifiers to monitor a VA range. This code
is lifted out of many drivers and hmm_mirror directly into the
mmu_notifier core and written using the best ideas from all the driver
implementations.
This removes many bugs from the drivers and has a very pleasing
diffstat. More drivers can still be converted, but that is for another
cycle.
- A shared branch with RDMA reworking the RDMA ODP implementation
- New mmu_interval_notifier API. This is focused on the use case of
monitoring a VA and simplifies the process for drivers
- A common seq-count locking scheme built into the mmu_interval_notifier
API usable by drivers that call get_user_pages() or hmm_range_fault()
with the VA range
- Conversion of mlx5 ODP, hfi1, radeon, nouveau, AMD GPU, and Xen GntDev
drivers to the new API. This deletes a lot of wonky driver code.
- Two improvements for hmm_range_fault(), from testing done by Ralph
-----BEGIN PGP SIGNATURE-----
iQIzBAABCgAdFiEEfB7FMLh+8QxL+6i3OG33FX4gmxoFAl3cCjQACgkQOG33FX4g
mxpp8xAAiR9iOdT28m/tx1GF31XludrMhRZVIiz0vmCIxIiAkWekWEfAEVm9PDnh
wdrxTJohSs+B65AK3sfToOM3AIuNCuFVWmbbHI5qmOO76vaSvcZa905Z++pNsawO
Bn8mgRCprYoFHcxWLvTvnA5U0g1S2BSSOwBSZI43CbEnVvHjYAR6MnvRqfGMk+NF
bf8fTk/x+fl0DCemhynlBLuJkogzoE2Hgl0yPY5bFna4PktOxdpa1yPaQsiqZ7e6
2s2NtM3pbMBJk0W42q5BU+aPhiqfxFFszasPSLBduXrD2xDsG76HJdHj5VydKmfL
nelG4BvqJozXTEZWvTEePYhCqaZ41eJZ7Asw8BXtmacVqE5mDlTXo/Zdgbz7yEOR
mI5MVyjD5rauZJldUOWXbwrPoWVFRvboauehiSgqvxvT9HvlFp9GKObSuu4gubBQ
mzxs4t48tPhA7bswLmw0/pETSogFuVDfaB7hsyY0gi8EwxMFMpw2qFypm1PEEF+C
BuUxCSShzvNKrraNe5PWaNNFd3AzIwAOWJHE+poH4bCoXQVr5nA+rq2gnHkdY5vq
/xrBCyxkf0U05YoFGYembPVCInMehzp9Xjy8V+SueSvCg2/TYwGDCgGfsbe9dNOP
Bc40JpS7BDn5w9nyLUJmOx7jfruNV6kx1QslA7NDDrB/rzOlsEc=
=Hj8a
-----END PGP SIGNATURE-----
Merge tag 'for-linus-hmm' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma
Pull hmm updates from Jason Gunthorpe:
"This is another round of bug fixing and cleanup. This time the focus
is on the driver pattern to use mmu notifiers to monitor a VA range.
This code is lifted out of many drivers and hmm_mirror directly into
the mmu_notifier core and written using the best ideas from all the
driver implementations.
This removes many bugs from the drivers and has a very pleasing
diffstat. More drivers can still be converted, but that is for another
cycle.
- A shared branch with RDMA reworking the RDMA ODP implementation
- New mmu_interval_notifier API. This is focused on the use case of
monitoring a VA and simplifies the process for drivers
- A common seq-count locking scheme built into the
mmu_interval_notifier API usable by drivers that call
get_user_pages() or hmm_range_fault() with the VA range
- Conversion of mlx5 ODP, hfi1, radeon, nouveau, AMD GPU, and Xen
GntDev drivers to the new API. This deletes a lot of wonky driver
code.
- Two improvements for hmm_range_fault(), from testing done by Ralph"
* tag 'for-linus-hmm' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
mm/hmm: remove hmm_range_dma_map and hmm_range_dma_unmap
mm/hmm: make full use of walk_page_range()
xen/gntdev: use mmu_interval_notifier_insert
mm/hmm: remove hmm_mirror and related
drm/amdgpu: Use mmu_interval_notifier instead of hmm_mirror
drm/amdgpu: Use mmu_interval_insert instead of hmm_mirror
drm/amdgpu: Call find_vma under mmap_sem
nouveau: use mmu_interval_notifier instead of hmm_mirror
nouveau: use mmu_notifier directly for invalidate_range_start
drm/radeon: use mmu_interval_notifier_insert
RDMA/hfi1: Use mmu_interval_notifier_insert for user_exp_rcv
RDMA/odp: Use mmu_interval_notifier_insert()
mm/hmm: define the pre-processor related parts of hmm.h even if disabled
mm/hmm: allow hmm_range to be used with a mmu_interval_notifier or hmm_mirror
mm/mmu_notifier: add an interval tree notifier
mm/mmu_notifier: define the header pre-processor parts even if disabled
mm/hmm: allow snapshot of the special zero page
Mainly a collection of smaller of driver updates this cycle.
- Various driver updates and bug fixes for siw, bnxt_re, hns, qedr,
iw_cxgb4, vmw_pvrdma, mlx5
- Improvements in SRPT from working with iWarp
- SRIOV VF support for bnxt_re
- Skeleton kernel-doc files for drivers/infiniband
- User visible counters for events related to ODP
- Common code for tracking of mmap lifetimes so that drivers can link HW
object liftime to a VMA
- ODP bug fixes and rework
- RDMA READ support for efa
- Removal of the very old cxgb3 driver
-----BEGIN PGP SIGNATURE-----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=UV2u
-----END PGP SIGNATURE-----
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma
Pull rdma updates from Jason Gunthorpe:
"Again another fairly quiet cycle with few notable core code changes
and the usual variety of driver bug fixes and small improvements.
- Various driver updates and bug fixes for siw, bnxt_re, hns, qedr,
iw_cxgb4, vmw_pvrdma, mlx5
- Improvements in SRPT from working with iWarp
- SRIOV VF support for bnxt_re
- Skeleton kernel-doc files for drivers/infiniband
- User visible counters for events related to ODP
- Common code for tracking of mmap lifetimes so that drivers can link
HW object liftime to a VMA
- ODP bug fixes and rework
- RDMA READ support for efa
- Removal of the very old cxgb3 driver"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (168 commits)
RDMA/hns: Delete unnecessary callback functions for cq
RDMA/hns: Rename the functions used inside creating cq
RDMA/hns: Redefine the member of hns_roce_cq struct
RDMA/hns: Redefine interfaces used in creating cq
RDMA/efa: Expose RDMA read related attributes
RDMA/efa: Support remote read access in MR registration
RDMA/efa: Store network attributes in device attributes
IB/hfi1: remove redundant assignment to variable ret
RDMA/bnxt_re: Fix missing le16_to_cpu
RDMA/bnxt_re: Fix stat push into dma buffer on gen p5 devices
RDMA/bnxt_re: Fix chip number validation Broadcom's Gen P5 series
RDMA/bnxt_re: Fix Kconfig indentation
IB/mlx5: Implement callbacks for getting VFs GUID attributes
IB/ipoib: Add ndo operation for getting VFs GUID attributes
IB/core: Add interfaces to get VF node and port GUIDs
net/core: Add support for getting VF GUIDs
RDMA/qedr: Fix null-pointer dereference when calling rdma_user_mmap_get_offset
RDMA/cm: Use refcount_t type for refcount variable
IB/mlx5: Support extended number of strides for Striding RQ
IB/mlx4: Update HW GID table while adding vlan GID
...
Danit Goldberg says:
====================
This series extends RTNETLINK to provide IB port and node GUIDs, which
were configured for Infiniband VFs.
The functionality to set VF GUIDs already existed for a long time, and
here we are adding the missing "get" so that netlink will be symmetric and
various cloud orchestration tools will be able to manage such VFs more
naturally.
The iproute2 was extended too to present those GUIDs.
- ip link show <device>
For example:
- ip link set ib4 vf 0 node_guid 22:44:33:00:33:11:00:33
- ip link set ib4 vf 0 port_guid 10:21:33:12:00:11:22:10
- ip link show ib4
ib4: <BROADCAST,MULTICAST> mtu 4092 qdisc noop state DOWN mode DEFAULT group default qlen 256
link/infiniband 00:00:0a:2d:fe:80:00:00:00:00:00:00:ec:0d:9a:03:00:44:36:8d brd 00:ff:ff:ff:ff:12:40:1b:ff:ff:00:00:00:00:00:00:ff:ff:ff:ff
vf 0 link/infiniband 00:00:0a:2d:fe:80:00:00:00:00:00:00:ec:0d:9a:03:00:44:36:8d brd 00:ff:ff:ff:ff:12:40:1b:ff:ff:00:00:00:00:00:00:ff:ff:ff:ff,
spoof checking off, NODE_GUID 22:44:33:00:33:11:00:33, PORT_GUID 10:21:33:12:00:11:22:10, link-state disable, trust off, query_rss off
====================
Based on the mlx5-next branch from
git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux for
dependencies
* branch 'ib-guids': (35 commits)
IB/mlx5: Implement callbacks for getting VFs GUID attributes
IB/ipoib: Add ndo operation for getting VFs GUID attributes
IB/core: Add interfaces to get VF node and port GUIDs
net/core: Add support for getting VF GUIDs
net/mlx5: Add new chain for netfilter flow table offload
net/mlx5: Refactor creating fast path prio chains
net/mlx5: Accumulate levels for chains prio namespaces
net/mlx5: Define fdb tc levels per prio
net/mlx5: Rename FDB_* tc related defines to FDB_TC_* defines
net/mlx5: Simplify fdb chain and prio eswitch defines
IB/mlx5: Load profile according to RoCE enablement state
IB/mlx5: Rename profile and init methods
net/mlx5: Handle "enable_roce" devlink param
net/mlx5: Document flow_steering_mode devlink param
devlink: Add new "enable_roce" generic device param
net/mlx5: fix spelling mistake "metdata" -> "metadata"
net/mlx5: fix kvfree of uninitialized pointer spec
IB/mlx5: Introduce and use mlx5_core_is_vf()
net/mlx5: E-switch, Enable metadata on own vport
net/mlx5: Refactor ingress acl configuration
...
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Replace the internal interval tree based mmu notifier with the new common
mmu_interval_notifier_insert() API. This removes a lot of code and fixes a
deadlock that can be triggered in ODP:
zap_page_range()
mmu_notifier_invalidate_range_start()
[..]
ib_umem_notifier_invalidate_range_start()
down_read(&per_mm->umem_rwsem)
unmap_single_vma()
[..]
__split_huge_page_pmd()
mmu_notifier_invalidate_range_start()
[..]
ib_umem_notifier_invalidate_range_start()
down_read(&per_mm->umem_rwsem) // DEADLOCK
mmu_notifier_invalidate_range_end()
up_read(&per_mm->umem_rwsem)
mmu_notifier_invalidate_range_end()
up_read(&per_mm->umem_rwsem)
The umem_rwsem is held across the range_start/end as the ODP algorithm for
invalidate_range_end cannot tolerate changes to the interval
tree. However, due to the nested invalidation regions the second
down_read() can deadlock if there are competing writers. The new core code
provides an alternative scheme to solve this problem.
Fixes: ca748c39ea ("RDMA/umem: Get rid of per_mm->notifier_count")
Link: https://lore.kernel.org/r/20191112202231.3856-6-jgg@ziepe.ca
Tested-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Provide ability to get node and port GUIDs of VFs to be symmetrical
to already existing set option.
Signed-off-by: Danit Goldberg <danitg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Improve return code from ib_modify_port() by doing the following:
- Use "-EOPNOTSUPP" instead "-ENOSYS" which is the proper return code
- Allow only fake IB_PORT_CM_SUP manipulation for RoCE providers that
didn't implement the modify_port callback, otherwise return
"-EOPNOTSUPP"
Fixes: 61e0962d52 ("IB: Avoid ib_modify_port() failure for RoCE devices")
Link: https://lore.kernel.org/r/20191028155931.1114-2-kamalheib1@gmail.com
Signed-off-by: Kamal Heib <kamalheib1@gmail.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Create some common API's for adding entries to a xa_mmap. Searching for
an entry and freeing one.
The general approach is copied from the EFA driver and improved to be more
general and do more to help the drivers. Integration with the core allows
a reference counted scheme with a free function so that the driver can
know when its mmaps are all gone.
This significant new functionality will be helpful for drivers to have the
correct lifetime model for mmap objects.
Link: https://lore.kernel.org/r/20191030094417.16866-3-michal.kalderon@marvell.com
Signed-off-by: Ariel Elior <ariel.elior@marvell.com>
Signed-off-by: Michal Kalderon <michal.kalderon@marvell.com>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
-----BEGIN PGP SIGNATURE-----
iQFSBAABCAA8FiEEq68RxlopcLEwq+PEeb4+QwBBGIYFAl210Z8eHHRvcnZhbGRz
QGxpbnV4LWZvdW5kYXRpb24ub3JnAAoJEHm+PkMAQRiGv+kIAKRpO7EuDokQL4qp
hxEEaCMJA1T055EMlNU6FVAq/ZbmapzreUyNYiRMpPWKGTWNMkhIcZQfysYeGZz5
y/KRxAiVxlcB+3v3yRmoZd/XoQmhgvJmqD4zhaGI2Utonow4f/SGSEFFZqqs9WND
4HJROjZHgQ4JBxg9Z+QMo0FxbV/DCZpEOUq51N9WJywyyDRb18zotE83stpU+pE2
fjqT7mk0NLrnYXuDRAbFC1Aau9ed4H6LlwLmxaqxq/Pt5Rz7wIKwKL9HIT4Dm/0a
qpani6phhHWL7MwUpa2wkEonFCD03rJFl3DUVJo64Ijh4up5D/jyXQ+GKV2P4WKJ
275Rb5Q=
=WiZZ
-----END PGP SIGNATURE-----
Merge tag 'v5.4-rc5' into rdma.git for-next
Linux 5.4-rc5
For dependencies in the next patches
Conflict resolved by keeping the delete of the unlock.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The dma_set_max_seg_size() call in setup_dma_device() does not have any
effect since device->dev.dma_parms is NULL. Fix this by initializing
device->dev.dma_parms first.
Link: https://lore.kernel.org/r/20191025225830.257535-5-bvanassche@acm.org
Fixes: d10bcf947a ("RDMA/umem: Combine contiguous PAGE_SIZE regions in SGEs")
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
When rdmacm module is not loaded, and when netlink message is received to
get char device info, it results into a deadlock due to recursive locking
of rdma_nl_mutex with the below call sequence.
[..]
rdma_nl_rcv()
mutex_lock()
[..]
rdma_nl_rcv_msg()
ib_get_client_nl_info()
request_module()
iw_cm_init()
rdma_nl_register()
mutex_lock(); <- Deadlock, acquiring mutex again
Due to above call sequence, following call trace and deadlock is observed.
kernel: __mutex_lock+0x35e/0x860
kernel: ? __mutex_lock+0x129/0x860
kernel: ? rdma_nl_register+0x1a/0x90 [ib_core]
kernel: rdma_nl_register+0x1a/0x90 [ib_core]
kernel: ? 0xffffffffc029b000
kernel: iw_cm_init+0x34/0x1000 [iw_cm]
kernel: do_one_initcall+0x67/0x2d4
kernel: ? kmem_cache_alloc_trace+0x1ec/0x2a0
kernel: do_init_module+0x5a/0x223
kernel: load_module+0x1998/0x1e10
kernel: ? __symbol_put+0x60/0x60
kernel: __do_sys_finit_module+0x94/0xe0
kernel: do_syscall_64+0x5a/0x270
kernel: entry_SYSCALL_64_after_hwframe+0x49/0xbe
process stack trace:
[<0>] __request_module+0x1c9/0x460
[<0>] ib_get_client_nl_info+0x5e/0xb0 [ib_core]
[<0>] nldev_get_chardev+0x1ac/0x320 [ib_core]
[<0>] rdma_nl_rcv_msg+0xeb/0x1d0 [ib_core]
[<0>] rdma_nl_rcv+0xcd/0x120 [ib_core]
[<0>] netlink_unicast+0x179/0x220
[<0>] netlink_sendmsg+0x2f6/0x3f0
[<0>] sock_sendmsg+0x30/0x40
[<0>] ___sys_sendmsg+0x27a/0x290
[<0>] __sys_sendmsg+0x58/0xa0
[<0>] do_syscall_64+0x5a/0x270
[<0>] entry_SYSCALL_64_after_hwframe+0x49/0xbe
To overcome this deadlock and to allow multiple netlink messages to
progress in parallel, following scheme is implemented.
1. Split the lock protecting the cb_table into a per-index lock, and make
it a rwlock. This lock is used to ensure no callbacks are running after
unregistration returns. Since a module will not be registered once it
is already running callbacks, this avoids the deadlock.
2. Use smp_store_release() to update the cb_table during registration so
that no lock is required. This avoids lockdep problems with thinking
all the rwsems are the same lock class.
Fixes: 0e2d00eb6f ("RDMA: Add NLDEV_GET_CHARDEV to allow char dev discovery and autoload")
Link: https://lore.kernel.org/r/20191015080733.18625-1-leon@kernel.org
Signed-off-by: Parav Pandit <parav@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Jianxin Xiong