Pull rdma updates from Jason Gunthorpe:
"Seveal fixes scattered across the drivers and a few new features:
- Minor updates and bug fixes to hfi1, efa, iopob, bnxt, hns
- Force disassociate the userspace FD when hns does an async reset
- bnxt new features for optimized modify QP to skip certain stayes,
CQ coalescing, better debug dumping
- mlx5 new data placement ordering feature
- Faster destruction of mlx5 devx HW objects
- Improvements to RDMA CM mad handling"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (51 commits)
RDMA/bnxt_re: Correct the sequence of device suspend
RDMA/bnxt_re: Use the default mode of congestion control
RDMA/bnxt_re: Support different traffic class
IB/cm: Rework sending DREQ when destroying a cm_id
IB/cm: Do not hold reference on cm_id unless needed
IB/cm: Explicitly mark if a response MAD is a retransmission
RDMA/mlx5: Move events notifier registration to be after device registration
RDMA/bnxt_re: Cache MSIx info to a local structure
RDMA/bnxt_re: Refurbish CQ to NQ hash calculation
RDMA/bnxt_re: Refactor NQ allocation
RDMA/bnxt_re: Fail probe early when not enough MSI-x vectors are reserved
RDMA/hns: Fix different dgids mapping to the same dip_idx
RDMA/bnxt_re: Add set_func_resources support for P5/P7 adapters
RDMA/bnxt_re: Enhance RoCE SRIOV resource configuration design
bnxt_en: Add support for RoCE sriov configuration
RDMA/hns: Fix NULL pointer derefernce in hns_roce_map_mr_sg()
RDMA/hns: Fix out-of-order issue of requester when setting FENCE
RDMA/nldev: Add IB device and net device rename events
RDMA/mlx5: Add implementation for ufile_hw_cleanup device operation
RDMA/core: Move ib_uverbs_file struct to uverbs_types.h
...
There are few use cases where CQ create and destroy
is seen before re-creating the CQ, this kind of use
case is disturbing the RR distribution and all the
active CQ getting mapped to only 2 NQ alternatively.
Fixing the CQ to NQ hash calculation by implementing
a quick load sorting mechanism under a mutex.
Using this, if the CQ was allocated and destroyed
before using it, the nq selecting algorithm still
obtains the least loaded CQ. Thus balancing the load
on NQs.
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Link: https://patch.msgid.link/1731577748-1804-4-git-send-email-selvin.xavier@broadcom.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Invalidate rkey is cpu endian and immediate data is in big endian format.
Both immediate data and invalidate the remote key returned by
HW is in little endian format.
While handling the commit in fixes tag, the difference between
immediate data and invalidate rkey endianness was not considered.
Without changes of this patch, Kernel ULP was failing while processing
inv_rkey.
dmesg log snippet -
nvme nvme0: Bogus remote invalidation for rkey 0x2000019Fix in this patch
Do endianness conversion based on completion queue entry flag.
Also, the HW completions are already converted to host endianness in
bnxt_qplib_cq_process_res_rc and bnxt_qplib_cq_process_res_ud and there
is no need to convert it again in bnxt_re_poll_cq. Modified the union to
hold the correct data type.
Fixes: 95b087f87b ("bnxt_re: Fix imm_data endianness")
Signed-off-by: Kashyap Desai <kashyap.desai@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Link: https://patch.msgid.link/1730110014-20755-1-git-send-email-selvin.xavier@broadcom.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
When variable size WQE is supported, max_qp_sges reported
is more than 6. For devices that supports variable size WQE,
the Send WQE size calculation is wrong when an an older library
that doesn't support variable size WQE is used.
Set the WQE size to 128 when static WQE is supported.
Fixes: de1d364c38 ("RDMA/bnxt_re: Add support for Variable WQE in Genp7 adapters")
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Link: https://patch.msgid.link/1725444253-13221-3-git-send-email-selvin.xavier@broadcom.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Gen P7 adapters needs to share a toggle bits information received
in kernel driver with the user space. User space needs this
info to arm the SRQ.
User space application can get this page using the
UAPI routines. Library will mmap this page and get the
toggle bits to be used in the next ARM Doorbell.
Uses a hash list to map the SRQ structure from the SRQ ID.
SRQ structure is retrieved from the hash list while the
library calls the UAPI routine to get the toggle page
mapping. Currently the full page is mapped per SRQ. This
can be optimized to enable multiple SRQs from the same
application share the same page and different offsets
in the page
Signed-off-by: Chandramohan Akula <chandramohan.akula@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Link: https://patch.msgid.link/1724945645-14989-4-git-send-email-selvin.xavier@broadcom.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Selvin Xavier says:
=============
Enable the Variable size Work Queue entry support for Gen P7
adapters. This would help in the better utilization of the queue memory
and pci bandwidth due to the smaller send queue Work entries.
=============
Based on v6.11-rc5 for dependencies.
* bnxt_re_variable_wqes: (829 commits)
RDMA/bnxt_re: Enable variable size WQEs for user space applications
RDMA/bnxt_re: Handle variable WQE support for user applications
RDMA/bnxt_re: Fix the table size for PSN/MSN entries
RDMA/bnxt_re: Get the WQE index from slot index while completing the WQEs
RDMA/bnxt_re: Add support for Variable WQE in Genp7 adapters
Linux 6.11-rc5
...
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
User library calculates the number of slots required for user applications
and it can pass that information to the driver. Driver can use this value
and update the HW directly. This mechanism is currently used only for the
newly introduced variable size WQEs.
Extend the bnxt_re_qp_req structure to pass the Send Queue slot count.
Reorganize the code to get the sq_slots before initializing the Send Queue
attributes.
Link: https://patch.msgid.link/r/1724042847-1481-5-git-send-email-selvin.xavier@broadcom.com
Signed-off-by: Hongguang Gao <hongguang.gao@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Variable size WQE means that each send Work Queue Entry to HW can use
different WQE sizes as opposed to the static WQE size on the current
devices. Set variable WQE mode for Gen P7 devices. Depth of the Queue will
be a multiple of slot which is 16 bytes. The number of slots should be a
multiple of 256 as per the HW requirement.
Initialize the Software shadow queue to hold requests equal to the number
of slots. Also, do not expose the variable size WQE capability until the
last patch in the series.
Link: https://patch.msgid.link/r/1724042847-1481-2-git-send-email-selvin.xavier@broadcom.com
Signed-off-by: Hongguang Gao <hongguang.gao@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
When map a device between servers with MLX and BCM RoCE nics, RTRS
server complain about unknown imm type, and can't map the device,
After more debug, it seems bnxt_re wrongly handle the
imm_data, this patch fixed the compat issue with MLX for us.
In off list discussion, Selvin confirmed HW is working in little endian format
and all data needs to be converted to LE while providing.
This patch fix the endianness for imm_data
Fixes: 1ac5a40479 ("RDMA/bnxt_re: Add bnxt_re RoCE driver")
Signed-off-by: Jack Wang <jinpu.wang@ionos.com>
Link: https://lore.kernel.org/r/20240710122102.37569-1-jinpu.wang@ionos.com
Acked-by: Selvin Xavier <selvin.xavier@broadcom.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
BNXT_RE_COMP_MASK_UCNTX_HW_RETX_ENABLED was introduced to share the HW
retransmit capability between driver and lib. The main difference in
implementation for HW Retransmit support is the usage of MSN table or
PSN table . When HW retrans is enabled, HW expects MSN table to be
allocated by driver/lib, else PSN table (for older adapters).
FW expose a new field which gives MSN capability. Drivers and libs can
depend on the new field instead of HW Retrasns capability. For
adapters which support HW_RETX feature, MSN table capability will be
set. For older adapters, this value will be 0(to maintain backward
compatibility with older FW).
Rename UAPI just to capture the correct name of the HW capability
that driver/library is interested in. No functional impact even if
older rdma-core is used.
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Link: https://lore.kernel.org/r/1716876697-25970-3-git-send-email-selvin.xavier@broadcom.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Gen P7 adapters needs to share a toggle bits information received
in kernel driver with the user space. User space needs this
info during the request notify call back to arm the CQ.
User space application can get this page using the
UAPI routines. Library will mmap this page and get the
toggle bits to be used in the next ARM Doorbell.
Uses a hash list to map the CQ structure from the CQ ID.
CQ structure is retrieved from the hash list while the
library calls the UAPI routine to get the toggle page
mapping. Currently the full page is mapped per CQ. This
can be optimized to enable multiple CQs from the same
application share the same page and different offsets
in the page.
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Link: https://lore.kernel.org/r/1702535484-26844-3-git-send-email-selvin.xavier@broadcom.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Pull rdma updates from Jason Gunthorpe:
"Many small changes across the subystem, some highlights:
- Usual driver cleanups in qedr, siw, erdma, hfi1, mlx4/5, irdma,
mthca, hns, and bnxt_re
- siw now works over tunnel and other netdevs with a MAC address by
removing assumptions about a MAC/GID from the connection manager
- "Doorbell Pacing" for bnxt_re - this is a best effort scheme to
allow userspace to slow down the doorbell rings if the HW gets full
- irdma egress VLAN priority, better QP/WQ sizing
- rxe bug fixes in queue draining and srq resizing
- Support more ethernet speed options in the core layer
- DMABUF support for bnxt_re
- Multi-stage MTT support for erdma to allow much bigger MR
registrations
- A irdma fix with a CVE that came in too late to go to -rc, missing
bounds checking for 0 length MRs"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (87 commits)
IB/hfi1: Reduce printing of errors during driver shut down
RDMA/hfi1: Move user SDMA system memory pinning code to its own file
RDMA/hfi1: Use list_for_each_entry() helper
RDMA/mlx5: Fix trailing */ formatting in block comment
RDMA/rxe: Fix redundant break statement in switch-case.
RDMA/efa: Fix wrong resources deallocation order
RDMA/siw: Call llist_reverse_order in siw_run_sq
RDMA/siw: Correct wrong debug message
RDMA/siw: Balance the reference of cep->kref in the error path
Revert "IB/isert: Fix incorrect release of isert connection"
RDMA/bnxt_re: Fix kernel doc errors
RDMA/irdma: Prevent zero-length STAG registration
RDMA/erdma: Implement hierarchical MTT
RDMA/erdma: Refactor the storage structure of MTT entries
RDMA/erdma: Renaming variable names and field names of struct erdma_mem
RDMA/hns: Support hns HW stats
RDMA/hns: Dump whole QP/CQ/MR resource in raw
RDMA/irdma: Add missing kernel-doc in irdma_setup_umode_qp()
RDMA/mlx4: Copy union directly
RDMA/irdma: Drop unused kernel push code
...
HW may generate completions that indicates QP is destroyed.
Driver should not be scheduling any more completion handlers
for this QP, after the QP is destroyed. Since CQs are active
during the QP destroy, driver may still schedule completion
handlers. This can cause a race where the destroy_cq and poll_cq
running simultaneously.
Snippet of kernel panic while doing bnxt_re driver load unload in loop.
This indicates a poll after the CQ is freed.
[77786.481636] Call Trace:
[77786.481640] <TASK>
[77786.481644] bnxt_re_poll_cq+0x14a/0x620 [bnxt_re]
[77786.481658] ? kvm_clock_read+0x14/0x30
[77786.481693] __ib_process_cq+0x57/0x190 [ib_core]
[77786.481728] ib_cq_poll_work+0x26/0x80 [ib_core]
[77786.481761] process_one_work+0x1e5/0x3f0
[77786.481768] worker_thread+0x50/0x3a0
[77786.481785] ? __pfx_worker_thread+0x10/0x10
[77786.481790] kthread+0xe2/0x110
[77786.481794] ? __pfx_kthread+0x10/0x10
[77786.481797] ret_from_fork+0x2c/0x50
To avoid this, complete all completion handlers before returning the
destroy QP. If free_cq is called soon after destroy_qp, IB stack
will cancel the CQ work before invoking the destroy_cq verb and
this will prevent any race mentioned.
Fixes: 1ac5a40479 ("RDMA/bnxt_re: Add bnxt_re RoCE driver")
Signed-off-by: Kashyap Desai <kashyap.desai@broadcom.com>
Signed-off-by: Selvin Xavier <selvin.xavier@broadcom.com>
Link: https://lore.kernel.org/r/1689322969-25402-2-git-send-email-selvin.xavier@broadcom.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
