SCCC (SCC Context) is a context with QP granularity that contains
information about congestion control. Dump SCCC and QPC together
to improve troubleshooting.
When dumping raw QPC with rdmatool, there will be a total of 576 bytes
data output, where the first 512 bytes is QPC and the last 64 bytes is
SCCC. When congestion control is disabled, the 64 byte SCCC will be all 0.
Example:
$rdma res show qp -jpr
[ {
"ifindex": 0,
"ifname": "hns_0",
"data": [ 67,0,0,0... 512bytes
4,0,2... 64bytes]
},...
} ]
Signed-off-by: wenglianfa <wenglianfa@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://lore.kernel.org/r/20240305055257.823513-1-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Currently, congestion control algorithm is statically configured in
FW, and all QPs use the same algorithm(except UD which has a fixed
configuration of DCQCN). This is not flexible enough.
Support userspace configuring congestion control algorithm with QP
granularity while creating QPs. If the algorithm is not specified in
userspace, use the default one.
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://lore.kernel.org/r/20240301104845.1141083-1-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Commit 27c5fd271d ("RDMA/hns: The UD mode can only be configured
with DCQCN") adds a check of congest control alorithm for UD. But
that patch causes a problem: hr_dev->caps.congest_type is global,
used by all QPs, so modifying this field to DCQCN for UD QPs causes
other QPs unable to use any other algorithm except DCQCN.
Revert the modification in commit 27c5fd271d ("RDMA/hns: The UD
mode can only be configured with DCQCN"). Add a new field cong_type
to struct hns_roce_qp and configure DCQCN for UD QPs.
Fixes: 27c5fd271d ("RDMA/hns: The UD mode can only be configured with DCQCN")
Fixes: f91696f2f0 ("RDMA/hns: Support congestion control type selection according to the FW")
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://lore.kernel.org/r/20240219061805.668170-1-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Compared with normal doorbell, using record doorbell can shorten the
process of ringing the doorbell and reduce the latency.
Add a flag HNS_ROCE_CAP_FLAG_SRQ_RECORD_DB to allow FW to
enable/disable SRQ record doorbell.
If the flag above is set, allocate the dma buffer for SRQ record
doorbell and write the buffer address into SRQC during SRQ creation.
For userspace SRQ, add a flag HNS_ROCE_RSP_SRQ_CAP_RECORD_DB to notify
userspace whether the SRQ record doorbell is enabled.
Signed-off-by: Yangyang Li <liyangyang20@huawei.com>
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://lore.kernel.org/r/20230926130026.583088-1-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The user usually configures the number of sge through the max_send_sge
parameter when creating qp, and configures the maximum size of inline data
that can be sent through max_inline_data. Inline uses sge to fill data to
send. Expect the following:
1) When the sge space cannot hold inline data, the sge space needs to be
expanded to accommodate all inline data
2) When the sge space is enough to accommodate inline data, the upper
limit of inline data can be increased so that users can send larger
inline data
Currently case one is not implemented. When the inline data is larger than
the sge space, an error of insufficient sge space occurs. This part of
the code needs to be reimplemented according to the expected rules. The
calculation method of sge num is modified to take the maximum value of
max_send_sge and the sge for max_inline_data to solve this problem.
Fixes: 05201e01be ("RDMA/hns: Refactor process of setting extended sge")
Fixes: 30b707886a ("RDMA/hns: Support inline data in extented sge space for RC")
Link: https://lore.kernel.org/r/20221108133847.2304539-3-xuhaoyue1@hisilicon.com
Signed-off-by: Luoyouming <luoyouming@huawei.com>
Signed-off-by: Haoyue Xu <xuhaoyue1@hisilicon.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
As hns_roce_init_hem_table() is always called with use_lowmem
being '1', and table->lowmem is set according to that argument,
so remove table->lowmem too.
Also, as the table->lowmem is used to indicate a dma buffer
is allocated with GFP_HIGHUSER or GFP_KERNEL, and calling
dma_alloc_coherent() with GFP_KERNEL seems like a common
pattern.
Link: https://lore.kernel.org/r/20220922123315.3732205-7-xuhaoyue1@hisilicon.com
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Haoyue Xu <xuhaoyue1@hisilicon.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The bt number of qpc_timer of HIP09 increases compared with that of HIP08.
Therefore, qpc_timer_bt_num and num_qpc_timer do not match. As a result,
the driver may fail to allocate qpc_timer. So the driver needs to uniquely
uses qpc_timer_bt_num to represent the bt number of qpc_timer.
Fixes: 0e40dc2f70 ("RDMA/hns: Add timer allocation support for hip08")
Link: https://lore.kernel.org/r/20220829105021.1427804-4-liangwenpeng@huawei.com
Signed-off-by: Yixing Liu <liuyixing1@huawei.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The MR raw restrack attributes come from the queue context maintained by
the ROCEE.
For example:
$ rdma res show mr dev hns_0 mrn 6 -dd -jp -r
[ {
"ifindex": 4,
"ifname": "hns_0",
"data": [ 1,0,0,0,2,0,0,0,0,3,0,0,0,0,2,0,0,0,0,0,32,0,0,0,2,0,0,0,
2,0,0,0,0,0,0,0 ]
} ]
Link: https://lore.kernel.org/r/20220822104455.2311053-8-liangwenpeng@huawei.com
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The QP raw restrack attributes come from the queue context maintained by
the ROCEE.
For example:
$ rdma res show qp link hns_0 -jp -dd -r
[ {
"ifindex": 4,
"ifname": "hns_0",
"data": [ 2,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,4,0,0,0,
5,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,255,156,0,0,63,156,0,0,
7,0,0,0,1,0,0,0,9,0,0,0,0,0,0,0,2,0,0,0,2,0,0,0,0,0,0,
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,63,156,0,
0,0,0,0,0 ]
} ]
Link: https://lore.kernel.org/r/20220822104455.2311053-6-liangwenpeng@huawei.com
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The CQ raw restrack attributes come from the queue context maintained by
the ROCEE.
For example:
$ rdma res show cq dev hns_0 cqn 14 -dd -jp -r
[ {
"ifindex": 4,
"ifname": "hns_0",
"data": [ 1,0,0,0,7,0,0,0,0,0,0,0,0,82,6,0,0,82,6,0,0,82,6,0,
1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,0,0,0,
6,0,0,0,0,0,0,0 ]
} ]
Link: https://lore.kernel.org/r/20220822104455.2311053-4-liangwenpeng@huawei.com
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Since ECC memory maintains a memory system immune to single-bit errors,
add support for correcting the 1bit-ECC error, which prevents a 1bit-ECC
error become an uncorrected type error. When a 1bit-ECC error happens in
the internal ram of the ROCE engine, such as the QPC table, as a 1bit-ECC
error caused by reading, the ROCE engine only corrects those 1bit ECC
errors by writing.
Link: https://lore.kernel.org/r/20220714134353.16700-6-liangwenpeng@huawei.com
Signed-off-by: Haoyue Xu <xuhaoyue1@hisilicon.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
The bt number of cqc_timer of HIP09 increases compared with that of HIP08.
Therefore, cqc_timer_bt_num and num_cqc_timer do not match. As a result,
the driver may fail to allocate cqc_timer. So the driver needs to uniquely
uses cqc_timer_bt_num to represent the bt number of cqc_timer.
Fixes: 0e40dc2f70 ("RDMA/hns: Add timer allocation support for hip08")
Link: https://lore.kernel.org/r/20220429093545.58070-1-liangwenpeng@huawei.com
Signed-off-by: Yixing Liu <liuyixing1@huawei.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
CMDQ may fail during HNS ROCEE initialization. The following is the log
when the execution fails:
hns3 0000:bd:00.2: In reset process RoCE client reinit.
hns3 0000:bd:00.2: CMDQ move tail from 840 to 839
hns3 0000:bd:00.2 hns_2: failed to set gid, ret = -11!
hns3 0000:bd:00.2: CMDQ move tail from 840 to 839
<...>
hns3 0000:bd:00.2: CMDQ move tail from 840 to 839
hns3 0000:bd:00.2: CMDQ move tail from 840 to 0
hns3 0000:bd:00.2: [cmd]token 14e mailbox 20 timeout.
hns3 0000:bd:00.2 hns_2: set HEM step 0 failed!
hns3 0000:bd:00.2 hns_2: set HEM address to HW failed!
hns3 0000:bd:00.2 hns_2: failed to alloc mtpt, ret = -16.
infiniband hns_2: Couldn't create ib_mad PD
infiniband hns_2: Couldn't open port 1
hns3 0000:bd:00.2: Reset done, RoCE client reinit finished.
However, even if ib_mad client registration failed, ib_register_device()
still returns success to the driver.
In the device initialization process, CMDQ execution fails because HW/FW
is abnormal. Therefore, if CMDQ fails, the initialization function should
set CMDQ to a fatal error state and return a failure to the caller.
Fixes: 9a4435375c ("IB/hns: Add driver files for hns RoCE driver")
Link: https://lore.kernel.org/r/20220429093104.26687-1-liangwenpeng@huawei.com
Signed-off-by: Yangyang Li <liyangyang20@huawei.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The current write wqe mechanism is to write to DDR first, and then notify
the hardware through doorbell to read the data. Direct wqe is a mechanism
to fill wqe directly into the hardware. In the case of light load, the wqe
will be filled into pcie bar space of the hardware, this will reduce one
memory access operation and therefore reduce the latency. SIMD
instructions allows cpu to write the 512 bits at one time to device
memory, thus it can be used for posting direct wqe.
Add direct wqe enable switch and address mapping.
Link: https://lore.kernel.org/r/20211207124901.42123-2-liangwenpeng@huawei.com
Signed-off-by: Yixing Liu <liuyixing1@huawei.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
dip_idx and dgid should be a one-to-one mapping relationship, but when
qp_num loops back to the start number, it may happen that two different
dgid are assiociated to the same dip_idx incorrectly.
One solution is to store the qp_num that is not assigned to dip_idx in an
array. When a dip_idx needs to be allocated to a new dgid, an spare qp_num
is extracted and assigned to dip_idx.
Fixes: f91696f2f0 ("RDMA/hns: Support congestion control type selection according to the FW")
Link: https://lore.kernel.org/r/1629884592-23424-4-git-send-email-liangwenpeng@huawei.com
Signed-off-by: Junxian Huang <huangjunxian4@hisilicon.com>
Signed-off-by: Yangyang Li <liyangyang20@huawei.com>
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
