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>
Due to the discrete nature of the HIP08 timer unit, a requester might
finish the timeout period sooner, in elapsed real time, than its responder
does, even when both sides share the identical RNR timeout length included
in the RNR Nak packet and the responder indeed starts the timing prior to
the requester. Furthermore, if a 'providential' resend packet arrived
before the responder's timeout period expired, the responder is certainly
entitled to drop the packet silently in the light of IB protocol.
To address this problem, our team made good use of certain hardware facts:
1) The timing resolution regards the transmission arrangements is 1
microsecond, e.g. if cq_period field is set to 3, it would be
interpreted as 3 microsecond by hardware
2) A QPC field shall inform the hardware how many timing unit (ticks)
constitutes a full microsecond, which, by default, is 1000
3) It takes 14ns for the processor to handle a packet in the buffer, so
the RNR timeout length of 10ns would ensure our processing mechanism is
disabled during the entire timeout period and the packet won't be
dropped silently
To achieve (3), we permanently set the QPC field mentioned in (2) to zero
which nominally indicates every time tick is equivalent to a microsecond
in wall-clock time; now, a RNR timeout period at face value of 10 would
only last 10 ticks, which is 10ns in wall-clock time.
It's worth noting that we adapt the driver by magnifying certain
configuration parameters(cq_period, eq_period and ack_timeout)by 1000
given the user assumes the configuring timing unit to be microseconds.
Also, this particular improvisation is only deployed on HIP08 since other
hardware has already solved this issue.
Fixes: cfc85f3e4b ("RDMA/hns: Add profile support for hip08 driver")
Link: https://lore.kernel.org/r/20211209140655.49493-1-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>
When hns_roce_v2_destroy_qp() is called, the brief calling process of the
driver is as follows:
......
hns_roce_v2_destroy_qp
hns_roce_v2_qp_modify
hns_roce_cmd_mbox
hns_roce_qp_destroy
If hns_roce_cmd_mbox() detects that the hardware is being reset during the
execution of the hns_roce_cmd_mbox(), the driver will not be able to get
the return value from the hardware (the firmware cannot respond to the
driver's mailbox during the hardware reset phase).
The driver needs to wait for the hardware reset to complete before
continuing to execute hns_roce_qp_destroy(), otherwise it may happen that
the driver releases the resources but the hardware is still accessing. In
order to fix this problem, HNS RoCE needs to add a piece of code to wait
for the hardware reset to complete.
The original interface get_hw_reset_stat() is the instantaneous state of
the hardware reset, which cannot accurately reflect whether the hardware
reset is completed, so it needs to be replaced with the ae_dev_reset_cnt
interface.
The sign that the hardware reset is complete is that the return value of
the ae_dev_reset_cnt interface is greater than the original value
reset_cnt recorded by the driver.
Fixes: 6a04aed6af ("RDMA/hns: Fix the chip hanging caused by sending mailbox&CMQ during reset")
Link: https://lore.kernel.org/r/20211123142402.26936-1-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>
is_reset is used to indicate whether the hardware starts to reset. When
hns_roce_hw_v2_reset_notify_down() is called, the hardware has not yet
started to reset. If is_reset is set at this time, all mailbox operations
of resource destroy actions will be intercepted by driver. When the driver
cleans up resources, but the hardware is still accessed, the following
errors will appear:
arm-smmu-v3 arm-smmu-v3.2.auto: event 0x10 received:
arm-smmu-v3 arm-smmu-v3.2.auto: 0x0000350100000010
arm-smmu-v3 arm-smmu-v3.2.auto: 0x000002088000003f
arm-smmu-v3 arm-smmu-v3.2.auto: 0x00000000a50e0800
arm-smmu-v3 arm-smmu-v3.2.auto: 0x0000000000000000
arm-smmu-v3 arm-smmu-v3.2.auto: event 0x10 received:
arm-smmu-v3 arm-smmu-v3.2.auto: 0x0000350100000010
arm-smmu-v3 arm-smmu-v3.2.auto: 0x000002088000043e
arm-smmu-v3 arm-smmu-v3.2.auto: 0x00000000a50a0800
arm-smmu-v3 arm-smmu-v3.2.auto: 0x0000000000000000
arm-smmu-v3 arm-smmu-v3.2.auto: event 0x10 received:
arm-smmu-v3 arm-smmu-v3.2.auto: 0x0000350100000010
arm-smmu-v3 arm-smmu-v3.2.auto: 0x0000020880000436
arm-smmu-v3 arm-smmu-v3.2.auto: 0x00000000a50a0880
arm-smmu-v3 arm-smmu-v3.2.auto: 0x0000000000000000
arm-smmu-v3 arm-smmu-v3.2.auto: event 0x10 received:
arm-smmu-v3 arm-smmu-v3.2.auto: 0x0000350100000010
arm-smmu-v3 arm-smmu-v3.2.auto: 0x000002088000043a
arm-smmu-v3 arm-smmu-v3.2.auto: 0x00000000a50e0840
hns3 0000:35:00.0: INT status: CMDQ(0x0) HW errors(0x0) other(0x0)
arm-smmu-v3 arm-smmu-v3.2.auto: 0x0000000000000000
hns3 0000:35:00.0: received unknown or unhandled event of vector0
arm-smmu-v3 arm-smmu-v3.2.auto: event 0x10 received:
arm-smmu-v3 arm-smmu-v3.2.auto: 0x0000350100000010
{34}[Hardware Error]: Hardware error from APEI Generic Hardware Error Source: 7
is_reset will be set correctly in check_aedev_reset_status(), so the
setting in hns_roce_hw_v2_reset_notify_down() should be deleted.
Fixes: 726be12f5c ("RDMA/hns: Set reset flag when hw resetting")
Link: https://lore.kernel.org/r/20211123084809.37318-1-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>
Patches held over for a possible rc8.
* for-rc:
RDMA/qedr: Fix NULL deref for query_qp on the GSI QP
RDMA/hns: Modify the value of MAX_LP_MSG_LEN to meet hardware compatibility
RDMA/hns: Fix initial arm_st of CQ
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The upper limit of MAX_LP_MSG_LEN on HIP08 is 64K, and the upper limit on
HIP09 is 16K. Regardless of whether it is HIP08 or HIP09, only 16K will be
used. In order to ensure compatibility, it is unified to 16K.
Setting MAX_LP_MSG_LEN to 16K will not cause performance loss on HIP08.
Fixes: fbed9d2be2 ("RDMA/hns: Fix configuration of ack_req_freq in QPC")
Link: https://lore.kernel.org/r/20211029100537.27299-1-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>
Replacing kmalloc/kfree/dma_map_single/dma_unmap_single() with
dma_alloc_coherent/dma_free_coherent() helps to reduce code size, and
simplify the code, and coherent DMA will not clear the cache every time.
The SOC that this driver supports does not have incoherent DMA, so this
makes the code follow the DMA API properly with no performance
impact. Currently there are missing dma sync calls around the DMA
transfers.
Link: https://lore.kernel.org/r/20210926061116.282-1-caihuoqing@baidu.com
Signed-off-by: Cai Huoqing <caihuoqing@baidu.com>
Reviewed-by: Wenpeng Liang <liangwenpeng@huawei.com>
Tested-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
gcc 8.3 and 5.4 throw this:
In function 'modify_qp_init_to_rtr',
././include/linux/compiler_types.h:322:38: error: call to '__compiletime_assert_1859' declared with attribute error: FIELD_PREP: value too large for the field
_compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__)
[..]
drivers/infiniband/hw/hns/hns_roce_common.h:91:52: note: in expansion of macro 'FIELD_PREP'
*((__le32 *)ptr + (field_h) / 32) |= cpu_to_le32(FIELD_PREP( \
^~~~~~~~~~
drivers/infiniband/hw/hns/hns_roce_common.h:95:39: note: in expansion of macro '_hr_reg_write'
#define hr_reg_write(ptr, field, val) _hr_reg_write(ptr, field, val)
^~~~~~~~~~~~~
drivers/infiniband/hw/hns/hns_roce_hw_v2.c:4412:2: note: in expansion of macro 'hr_reg_write'
hr_reg_write(context, QPC_LP_PKTN_INI, lp_pktn_ini);
Because gcc has miscalculated the constantness of lp_pktn_ini:
mtu = ib_mtu_enum_to_int(ib_mtu);
if (WARN_ON(mtu < 0)) [..]
lp_pktn_ini = ilog2(MAX_LP_MSG_LEN / mtu);
Since mtu is limited to {256,512,1024,2048,4096} lp_pktn_ini is between 4
and 8 which is compatible with the 4 bit field in the FIELD_PREP.
Work around this broken compiler by adding a 'can never be true'
constraint on lp_pktn_ini's value which clears out the problem.
Fixes: f0cb411aad ("RDMA/hns: Use new interface to modify QP context")
Link: https://lore.kernel.org/r/0-v1-c773ecb137bc+11f-hns_gcc8_jgg@nvidia.com
Reported-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
From Maor Gottlieb
====================
Fix the use of nents and orig_nents in the sg table append helpers. The
nents should be used by the DMA layer to store the number of DMA mapped
sges, the orig_nents is the number of CPU sges.
Since the sg append logic doesn't always create a SGL with exactly
orig_nents entries store a total_nents as well to allow the table to be
properly free'd and reorganize the freeing logic to share across all the
use cases.
====================
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
* 'sg_nents':
RDMA: Use the sg_table directly and remove the opencoded version from umem
lib/scatterlist: Fix wrong update of orig_nents
lib/scatterlist: Provide a dedicated function to support table append
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>
