hns_roce_hw_v2.h has a direct dependency on hnae3.h due to the
inline function hns_roce_write64(), but it doesn't include this
header currently. This leads to that files including
hns_roce_hw_v2.h must also include hnae3.h to avoid compilation
errors, even if they themselves don't really rely on hnae3.h.
This doesn't make sense, hns_roce_hw_v2.h should include hnae3.h
directly.
Fixes: d3743fa94c ("RDMA/hns: Fix the chip hanging caused by sending doorbell during reset")
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20250421132750.1363348-6-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Pull rdma updates from Jason Gunthorpe:
- Usual minor updates and fixes for bnxt_re, hfi1, rxe, mana, iser,
mlx5, vmw_pvrdma, hns
- Make rxe work on tun devices
- mana gains more standard verbs as it moves toward supporting
in-kernel verbs
- DMABUF support for mana
- Fix page size calculations when memory registration exceeds 4G
- On Demand Paging support for rxe
- mlx5 support for RDMA TRANSPORT flow tables and a new ucap mechanism
to access control use of them
- Optional RDMA_TX/RX counters per QP in mlx5
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (73 commits)
IB/mad: Check available slots before posting receive WRs
RDMA/mana_ib: Fix integer overflow during queue creation
RDMA/mlx5: Fix calculation of total invalidated pages
RDMA/mlx5: Fix mlx5_poll_one() cur_qp update flow
RDMA/mlx5: Fix page_size variable overflow
RDMA/mlx5: Drop access_flags from _mlx5_mr_cache_alloc()
RDMA/mlx5: Fix cache entry update on dereg error
RDMA/mlx5: Fix MR cache initialization error flow
RDMA/mlx5: Support optional-counters binding for QPs
RDMA/mlx5: Compile fs.c regardless of INFINIBAND_USER_ACCESS config
RDMA/core: Pass port to counter bind/unbind operations
RDMA/core: Add support to optional-counters binding configuration
RDMA/core: Create and destroy rdma_counter using rdma_zalloc_drv_obj()
RDMA/mlx5: Add optional counters for RDMA_TX/RX_packets/bytes
RDMA/core: Fix use-after-free when rename device name
RDMA/bnxt_re: Support perf management counters
RDMA/rxe: Fix incorrect return value of rxe_odp_atomic_op()
RDMA/uverbs: Propagate errors from rdma_lookup_get_uobject()
RDMA/mana_ib: Handle net event for pointing to the current netdev
net: mana: Change the function signature of mana_get_primary_netdev_rcu
...
SQ params from userspace are checked in by set_user_sq_size(). But
when the check fails, the function doesn't return but instead keep
running and overwrite 'ret'. As a result, the invalid params will
not get blocked actually.
Add a return right after the failed check. Besides, although the
check result of kernel sq params will not be overwritten, to keep
coding style unified, move default_congest_type() before
set_kernel_sq_size().
Fixes: 6ec429d588 ("RDMA/hns: Support userspace configuring congestion control algorithm with QP granularity")
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20250311084857.3803665-5-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Driver runs a for-loop when allocating bt pages and mapping them with
buffer pages. When a large buffer (e.g. MR over 100GB) is being allocated,
it may require a considerable loop count. This will lead to soft lockup:
watchdog: BUG: soft lockup - CPU#27 stuck for 22s!
...
Call trace:
hem_list_alloc_mid_bt+0x124/0x394 [hns_roce_hw_v2]
hns_roce_hem_list_request+0xf8/0x160 [hns_roce_hw_v2]
hns_roce_mtr_create+0x2e4/0x360 [hns_roce_hw_v2]
alloc_mr_pbl+0xd4/0x17c [hns_roce_hw_v2]
hns_roce_reg_user_mr+0xf8/0x190 [hns_roce_hw_v2]
ib_uverbs_reg_mr+0x118/0x290
watchdog: BUG: soft lockup - CPU#35 stuck for 23s!
...
Call trace:
hns_roce_hem_list_find_mtt+0x7c/0xb0 [hns_roce_hw_v2]
mtr_map_bufs+0xc4/0x204 [hns_roce_hw_v2]
hns_roce_mtr_create+0x31c/0x3c4 [hns_roce_hw_v2]
alloc_mr_pbl+0xb0/0x160 [hns_roce_hw_v2]
hns_roce_reg_user_mr+0x108/0x1c0 [hns_roce_hw_v2]
ib_uverbs_reg_mr+0x120/0x2bc
Add a cond_resched() to fix soft lockup during these loops. In order not
to affect the allocation performance of normal-size buffer, set the loop
count of a 100GB MR as the threshold to call cond_resched().
Fixes: 38389eaa4d ("RDMA/hns: Add mtr support for mixed multihop addressing")
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20250311084857.3803665-3-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
If a QP is modified to error state and a flush CQE process is triggered,
the subsequent QP destruction mbox can still be successfully posted but
will be blocked in HW until the flush CQE process finishes. This causes
further mbox posting timeouts in driver. The blocking time is related
to QP depth. Considering an extreme case where SQ depth and RQ depth
are both 32K, the blocking time can reach about 135ms.
This patch adds a retry mechanism for mbox posting. For each try, FW
waits 15ms for HW to complete the previous mbox, otherwise return a
timeout error code to driver. Counting other time consumption in FW,
set 8 tries for mbox posting and a 5ms time gap before each retry to
increase to a sufficient timeout limit.
Fixes: 0425e3e6e0 ("RDMA/hns: Support flush cqe for hip08 in kernel space")
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20250208105930.522796-1-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Pull rdma updates from Jason Gunthorpe:
"Lighter that normal, but the now usual collection of driver fixes and
small improvements:
- Small fixes and minor improvements to cxgb4, bnxt_re, rxe, srp,
efa, cxgb4
- Update mlx4 to use the new umem APIs, avoiding direct use of
scatterlist
- Support ROCEv2 in erdma
- Remove various uncalled functions, constify bin_attribute
- Provide core infrastructure to catch netdev events and route them
to drivers, consolidating duplicated driver code
- Fix rare race condition crashes in mlx5 ODP flows"
* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma: (63 commits)
RDMA/mlx5: Fix implicit ODP use after free
RDMA/mlx5: Fix a race for an ODP MR which leads to CQE with error
RDMA/qib: Constify 'struct bin_attribute'
RDMA/hfi1: Constify 'struct bin_attribute'
RDMA/rxe: Fix the warning "__rxe_cleanup+0x12c/0x170 [rdma_rxe]"
RDMA/cxgb4: Notify rdma stack for IB_EVENT_QP_LAST_WQE_REACHED event
RDMA/bnxt_re: Allocate dev_attr information dynamically
RDMA/bnxt_re: Pass the context for ulp_irq_stop
RDMA/bnxt_re: Add support to handle DCB_CONFIG_CHANGE event
RDMA/bnxt_re: Query firmware defaults of CC params during probe
RDMA/bnxt_re: Add Async event handling support
bnxt_en: Add ULP call to notify async events
RDMA/mlx5: Fix indirect mkey ODP page count
MAINTAINERS: Update the bnxt_re maintainers
RDMA/hns: Clean up the legacy CONFIG_INFINIBAND_HNS
RDMA/rtrs: Add missing deinit() call
RDMA/efa: Align interrupt related fields to same type
RDMA/bnxt_re: Fix to drop reference to the mmap entry in case of error
RDMA/mlx5: Fix link status down event for MPV
RDMA/erdma: Support create_ah/destroy_ah in non-sleepable contexts
...
hns driver used to support hip06 and hip08 devices with
CONFIG_INFINIBAND_HNS_HIP06 and CONFIG_INFINIBAND_HNS_HIP08
respectively, which both depended on CONFIG_INFINIBAND_HNS.
But we no longer provide support for hip06 and only support
hip08 and higher since the commit in fixes line, so there is
no need to have CONFIG_INFINIBAND_HNS any more. Remove it and
only keep CONFIG_INFINIBAND_HNS_HIP08.
Fixes: 38d2208824 ("RDMA/hns: Remove support for HIP06")
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20250106111211.3945051-1-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
hns3 NIC driver can directly notify the RoCE driver about link status
events bypassing the netdev notifier. This can provide more timely
event dispatching for ULPs.
Signed-off-by: Yuyu Li <liyuyu6@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Flush CQE handler has not been called if QP state gets into errored
mode in DWQE path. So, the new added outstanding WQEs will never be
flushed.
It leads to a hung task timeout when using NFS over RDMA:
__switch_to+0x7c/0xd0
__schedule+0x350/0x750
schedule+0x50/0xf0
schedule_timeout+0x2c8/0x340
wait_for_common+0xf4/0x2b0
wait_for_completion+0x20/0x40
__ib_drain_sq+0x140/0x1d0 [ib_core]
ib_drain_sq+0x98/0xb0 [ib_core]
rpcrdma_xprt_disconnect+0x68/0x270 [rpcrdma]
xprt_rdma_close+0x20/0x60 [rpcrdma]
xprt_autoclose+0x64/0x1cc [sunrpc]
process_one_work+0x1d8/0x4e0
worker_thread+0x154/0x420
kthread+0x108/0x150
ret_from_fork+0x10/0x18
Fixes: 01584a5edc ("RDMA/hns: Add support of direct wqe")
Signed-off-by: Chengchang Tang <tangchengchang@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20241220055249.146943-5-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Due to HW limitation, the three region of WQE buffer must be mapped
and set to HW in a fixed order: SQ buffer, SGE buffer, and RQ buffer.
Currently when one region is zero-hop while the other two are not,
the zero-hop region will not be mapped. This violate the limitation
above and leads to address error.
Fixes: 38389eaa4d ("RDMA/hns: Add mtr support for mixed multihop addressing")
Signed-off-by: wenglianfa <wenglianfa@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20241220055249.146943-2-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
DIP algorithm requires a one-to-one mapping between dgid and dip_idx.
Currently a queue 'spare_idx' is used to store QPN of QPs that use
DIP algorithm. For a new dgid, use a QPN from spare_idx as dip_idx.
This method lacks a mechanism for deduplicating QPN, which may result
in different dgids sharing the same dip_idx and break the one-to-one
mapping requirement.
This patch replaces spare_idx with xarray and introduces a refcnt of
a dip_idx to indicate the number of QPs that using this dip_idx.
The state machine for dip_idx management is implemented as:
* The entry at an index in xarray is empty -- This indicates that the
corresponding dip_idx hasn't been created.
* The entry at an index in xarray is not empty but with 0 refcnt --
This indicates that the corresponding dip_idx has been created but
not used as dip_idx yet.
* The entry at an index in xarray is not empty and with non-0 refcnt --
This indicates that the corresponding dip_idx is being used by refcnt
number of DIP QPs.
Fixes: eb653eda1e ("RDMA/hns: Bugfix for incorrect association between dip_idx and dgid")
Fixes: f91696f2f0 ("RDMA/hns: Support congestion control type selection according to the FW")
Signed-off-by: Feng Fang <fangfeng4@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20241112055553.3681129-1-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The FENCE indicator in hns WQE doesn't ensure that response data from
a previous Read/Atomic operation has been written to the requester's
memory before the subsequent Send/Write operation is processed. This
may result in the subsequent Send/Write operation accessing the original
data in memory instead of the expected response data.
Unlike FENCE, the SO (Strong Order) indicator blocks the subsequent
operation until the previous response data is written to memory and a
bresp is returned. Set the SO indicator instead of FENCE to maintain
strict order.
Fixes: 9a4435375c ("IB/hns: Add driver files for hns RoCE driver")
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20241108075743.2652258-2-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
During reset, cmd to destroy resources such as qp, cq, and mr may fail,
and error logs will be printed. When a large number of resources are
destroyed, there will be lots of printings, and it may lead to a cpu
stuck.
Delete some unnecessary printings and replace other printing functions
in these paths with the ratelimited version.
Fixes: 9a4435375c ("IB/hns: Add driver files for hns RoCE driver")
Fixes: c7bcb13442 ("RDMA/hns: Add SRQ support for hip08 kernel mode")
Fixes: 70f9252158 ("RDMA/hns: Use the reserved loopback QPs to free MR before destroying MPT")
Fixes: 926a01dc00 ("RDMA/hns: Add QP operations support for hip08 SoC")
Signed-off-by: wenglianfa <wenglianfa@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20241024124000.2931869-6-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
QP needs to be modified to IB_QPS_ERROR to trigger HW flush cqe. But
when this process races with destroy qp, the destroy-qp process may
modify the QP to IB_QPS_RESET first. In this case flush cqe will fail
since it is invalid to modify qp from IB_QPS_RESET to IB_QPS_ERROR.
Add lock and bit flag to make sure pending flush cqe work is completed
first and no more new works will be added.
Fixes: ffd541d457 ("RDMA/hns: Add the workqueue framework for flush cqe handler")
Signed-off-by: wenglianfa <wenglianfa@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20241024124000.2931869-3-huangjunxian6@hisilicon.com
Reviewed-by: Zhu Yanjun <yanjun.zhu@linux.dev>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
eq_db_ci is updated only after all AEQEs are processed in the AEQ
interrupt handler, which is not timely enough and may result in
AEQ overflow. Two optimization methods are proposed:
1. Set an upper limit for AEQE processing.
2. Move time-consuming operations such as printings to the bottom
half of the interrupt.
cmd events and flush_cqe events are still fully processed in the top half
to ensure timely handling.
Fixes: a5073d6054 ("RDMA/hns: Add eq support of hip08")
Signed-off-by: wenglianfa <wenglianfa@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20241024124000.2931869-2-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
When HW is being reset, userspace should not ring doorbell otherwise
it may lead to abnormal consequence such as RAS.
Disassociate mmap pages for all uctx to prevent userspace from ringing
doorbell to HW. Since all resources will be destroyed during HW reset,
no new mmap is allowed after HW reset is completed.
Fixes: 9a4435375c ("IB/hns: Add driver files for hns RoCE driver")
Signed-off-by: Chengchang Tang <tangchengchang@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20240927103323.1897094-3-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
When allocating MTT hem, for each hop level of each hem that is being
allocated, the driver iterates the hem list to find out whether the
bt page has been allocated in this hop level. If not, allocate a new
one and splice it to the list. The time complexity is O(n^2) in worst
cases.
Currently the allocation for-loop uses 'unit' as the step size. This
actually has taken into account the reuse of last-hop-level MTT bt
pages by multiple buffer pages. Thus pages of last hop level will
never have been allocated, so there is no need to iterate the hem list
in last hop level.
Removing this unnecessary iteration can reduce the time complexity to
O(n).
Fixes: 38389eaa4d ("RDMA/hns: Add mtr support for mixed multihop addressing")
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://patch.msgid.link/20240906093444.3571619-9-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
When a large number of tasks are issued, the speed of HW processing
mbx will slow down. The standard for judging mbx timeout in the current
firmware is 30ms, and the current timeout standard for the driver is also
30ms.
Considering that firmware scheduling in multi-function scenarios takes a
certain amount of time, this will cause the driver to time out too early
and report a failure before mbx execution times out.
This patch introduces a new mechanism that can set different timeouts for
different cmds and extends the timeout of mbx to 35ms.
Fixes: a04ff739f2 ("RDMA/hns: Add command queue support for hip08 RoCE driver")
Signed-off-by: Chengchang Tang <tangchengchang@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://lore.kernel.org/r/20240710133705.896445-9-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
