- 29 May, 2020 11 commits
-
-
Yamin Friedman authored
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.comSigned-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>
-
Yamin Friedman authored
A pre-step for adding shared CQs. Add the infrastructure to prevent shared CQ users from altering the CQ configurations. For now all cqs are marked as private (non-shared). The core driver should use the new force functions to perform resize/destroy/moderation changes that are not allowed for users of shared CQs. Link: https://lore.kernel.org/r/1590568495-101621-2-git-send-email-yaminf@mellanox.comSigned-off-by:
-
Qiushi Wu authored
kobject_init_and_add() takes reference even when it fails. If this function returns an error, kobject_put() must be called to properly clean up the memory associated with the object. Previous commit b8eb7183 ("net-sysfs: Fix reference count leak in rx|netdev_queue_add_kobject") fixed a similar problem. Link: https://lore.kernel.org/r/20200528030231.9082-1-wu000273@umn.eduSigned-off-by:
Qiushi Wu <wu000273@umn.edu> Reviewed-by:
-
Colin Ian King authored
There is a spelling mistake in an error message. Fix it. Link: https://lore.kernel.org/r/20200528110709.400935-1-colin.king@canonical.comSigned-off-by:
Colin Ian King <colin.king@canonical.com> Acked-by:
Dennis Dalessandro <dennis.dalessandro@intel.com> Signed-off-by:
-
Jason Gunthorpe authored
No reason to open code this. Link: https://lore.kernel.org/r/0-v1-0bc346e08476+585-drop_offsetofend_jgg@mellanox.comSigned-off-by:
-
Colin Ian King authored
The pointer raq is being assigned twice. Fix this by removing one of the redundant assignments. Fixes: 14ba8730 ("RDMA/hns: Remove redundant type cast for general pointers") Link: https://lore.kernel.org/r/20200528150427.420624-1-colin.king@canonical.com Addressses-Coverity: ("Evaluation order violation") Signed-off-by:
-
Mark Zhang authored
The mlx5 VF driver doesn't set QP tx port affinity because it doesn't know if the lag is active or not, since the "lag_active" works only for PF interfaces. In this case for VF interfaces only one lag is used which brings performance issue. Add a lag_tx_port_affinity CAP bit; When it is enabled and "num_lag_ports > 1", then driver always set QP tx affinity, regardless of lag state. Link: https://lore.kernel.org/r/20200527055014.355093-1-leon@kernel.orgSigned-off-by:
Mark Zhang <markz@mellanox.com> Reviewed-by:
Maor Gottlieb <maorg@mellanox.com> Signed-off-by:
-
Bart Van Assche authored
The ib_srpt driver limits max_send_sge to 16. Since that is a workaround for an mlx4 bug that has been fixed, increase max_send_sge. See also commit f95ccffc ("IB/mlx4: Use 4K pages for kernel QP's WQE buffer"). Link: https://lore.kernel.org/r/20200525172212.14413-5-bvanassche@acm.orgSigned-off-by:
Bart Van Assche <bvanassche@acm.org> Signed-off-by:
-
Bart Van Assche authored
Since srpt_post_recv() always sets num_sge to 1, reduce the max_recv_sge parameter that is used at queue pair allocation time to 1. Link: https://lore.kernel.org/r/20200525172212.14413-4-bvanassche@acm.orgSigned-off-by:
-
Bart Van Assche authored
Since the session name by itself is not sufficient to uniquely identify a queue pair, include the queue pair number. Show the ASCII channel state name instead of the numeric value. This change makes the ib_srpt debug output more consistent. Link: https://lore.kernel.org/r/20200525172212.14413-3-bvanassche@acm.orgSigned-off-by:
-
Bart Van Assche authored
Increase the flexibility of the SRP initiator driver by making the channel count configurable per target instead of only providing a kernel module parameter for configuring the channel count. Link: https://lore.kernel.org/r/20200525172212.14413-2-bvanassche@acm.orgSigned-off-by:
-
- 27 May, 2020 17 commits
-
-
Leon Romanovsky authored
After users sets the ECE option, FW will return the agreed/supported bits through an output structures of modify QP stages for regular QPs or through create QP for the DCT. Link: https://lore.kernel.org/r/20200526115440.205922-9-leon@kernel.orgReviewed-by:
-
Leon Romanovsky authored
The most common way to set ECE option will be during modify QP command in INIT2RTR, RTR2RTS and RTS2RTS stages, so update mlx5 to support it. The new bit in the comp_mask is needed to mark that kernel supports ECE and can receive data instead of "reserved" field in the struct mlx5_ib_modify_qp. Link: https://lore.kernel.org/r/20200526115440.205922-8-leon@kernel.orgReviewed-by:
-
Leon Romanovsky authored
Instead of hand crafted mlx5_qp_context and mlx5_qp_path use common MLX5_SET() macros. Link: https://lore.kernel.org/r/20200526115440.205922-7-leon@kernel.orgReviewed-by:
-
Leon Romanovsky authored
As a preparation to removal hand crafted mlx5_qp_context, convert query_qp_attr() to use proper MLX5_GET() macros. Link: https://lore.kernel.org/r/20200526115440.205922-6-leon@kernel.orgReviewed-by:
-
Leon Romanovsky authored
As a preparation to removal hand crafted mlx5_qp_context, convert counter code to use mlx5_cmd_exec_in() directly. Link: https://lore.kernel.org/r/20200526115440.205922-5-leon@kernel.orgReviewed-by:
-
Leon Romanovsky authored
Allow users to ask creation of QPs with specific ECE options. Such early set even before RDMA-CM connection is established is useful if user knows exactly which option he needs. Link: https://lore.kernel.org/r/20200526115440.205922-4-leon@kernel.orgReviewed-by:
-
Leon Romanovsky authored
Supported ECE options are returned from FW in the create_qp phase and zero means that field is not valid. Such default value allows us to reuse reserved field without worries about comp_mask. Update create QP API to return ECE options. Link: https://lore.kernel.org/r/20200526115440.205922-3-leon@kernel.orgReviewed-by:
-
Leon Romanovsky authored
IBTA declares "vendor option not supported" reject reason in REJ messages if passive side doesn't want to accept proposed ECE options. Due to the fact that ECE is managed by userspace, there is a need to let users to provide such rejected reason. Link: https://lore.kernel.org/r/20200526103304.196371-7-leon@kernel.orgSigned-off-by:
-
Leon Romanovsky authored
The rdma_accept() is called by both passive and active sides of CMID connection to mark readiness to start data transfer. For passive side, this is called explicitly, for active side, it is called implicitly while receiving REP message. Provide ECE data to rdma_accept function needed for passive side to send that REP message. Link: https://lore.kernel.org/r/20200526103304.196371-6-leon@kernel.orgSigned-off-by:
-
Leon Romanovsky authored
ECE parameters are exchanged through REQ->REP/SIDR_REP messages, this patch adds the data to provide to other side of CMID communication channel. Link: https://lore.kernel.org/r/20200526103304.196371-5-leon@kernel.orgSigned-off-by:
-
Leon Romanovsky authored
Passive side of CMID connection receives ECE request through REQ message and needs to respond with relevant REP message which will be forwarded to active side. The UCMA events interface is responsible for such communication with the user space (librdmacm). Extend it to provide ECE wire data. Link: https://lore.kernel.org/r/20200526103304.196371-4-leon@kernel.orgSigned-off-by:
-
Leon Romanovsky authored
Active side of CMID initiates connection through librdmacm's rdma_connect() and kernel's ucma_connect(). Extend UCMA interface to handle those new parameters. Link: https://lore.kernel.org/r/20200526103304.196371-3-leon@kernel.orgSigned-off-by:
-
Leon Romanovsky authored
Extend REQ (request for communications), REP (reply to request for communication), rejected reason and SIDR_REP (service ID resolution response) structures with hardware vendor ID bits according to IBTA v1.4. Link: https://lore.kernel.org/r/20200526103304.196371-2-leon@kernel.orgSigned-off-by:
-
Jason Gunthorpe authored
From the mlx5-next branch at git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux Required for dependencies in following patches * branch 'mellanox/mlx5-next': net/mlx5: Add ability to read and write ECE options net/mlx5: Add support for RDMA TX FT headers modifying net/mlx5: Move iseg access helper routines close to mlx5_core driver net/mlx5: Cleanup mlx5_ifc_fte_match_set_misc2_bits Signed-off-by:
-
Mark Zhang authored
The commit citied in the Fixes line wasn't complete and solved only part of the problems. Update the mlx5_ib to properly support MLX5_CMD_OP_INIT2INIT_QP command in the DEVX, that is required when modify the QP tx_port_affinity. Fixes: 819f7427 ("RDMA/mlx5: Add init2init as a modify command") Link: https://lore.kernel.org/r/20200527135703.482501-1-leon@kernel.orgSigned-off-by:
-
Gustavo A. R. Silva authored
Make use of the sizeof_field() helper instead of an open-coded version. Link: https://lore.kernel.org/r/20200527144152.GA22605@embeddedorSigned-off-by:
Gustavo A. R. Silva <gustavoars@kernel.org> Signed-off-by:
-
Leon Romanovsky authored
The end result of RDMA-CM ECE handshake is ECE options, which is needed to be used while configuring data QPs. Such options can come in any QP state, so add in/out fields to set and query ECE options. OUT fields: * create_qp() - default ECE options for that type of QP. * modify_qp() - enabled ECE options after QP state transition. IN fields: * create_qp() - create QP with this ECE option. * modify_qp() - requested options. For unconnected QPs, the FW will return an error if ECE is already configured with any options that not equal to previously set. Reviewed-by:
-
- 25 May, 2020 12 commits
-
-
Kamal Heib authored
can_sleep is always 0 when iboib_mcast_alloc() is called, so remove it and use GFP_ATOMIC instead of GFP_KERNEL. Link: https://lore.kernel.org/r/20200525130305.171509-1-kamalheib1@gmail.comSigned-off-by:
Kamal Heib <kamalheib1@gmail.com> Signed-off-by:
-
Potnuri Bharat Teja authored
Remove device specific debugfs entries immediately if LLD detaches a particular ULD device in case of fatal PCI errors. Link: https://lore.kernel.org/r/20200524190814.17599-1-bharat@chelsio.comSigned-off-by:
Potnuri Bharat Teja <bharat@chelsio.com> Signed-off-by:
-
Yixian Liu authored
Instead of i with the sge number of wr will make the comparision more clear, that is, when the sge number in wr is small than the maximum supported sge number in the queue, then a stop sge needed to be filled at the end of sges in wr. Link: https://lore.kernel.org/r/1590152579-32364-5-git-send-email-liweihang@huawei.comSigned-off-by:
Yixian Liu <liuyixian@huawei.com> Signed-off-by:
Weihang Li <liweihang@huawei.com> Signed-off-by:
-
Lang Cheng authored
Set hns_roce_v2_cq_set_ci to inline type and remove unnecessary next_cqe_sw_v2(). Link: https://lore.kernel.org/r/1590152579-32364-4-git-send-email-liweihang@huawei.comSigned-off-by:
Lang Cheng <chenglang@huawei.com> Signed-off-by:
-
Wenpeng Liang authored
The redundant parameters "hr_dev" need to be removed from free_kernel_wrid() and free_srq_wrid(). Link: https://lore.kernel.org/r/1590152579-32364-3-git-send-email-liweihang@huawei.comSigned-off-by:
Wenpeng Liang <liangwenpeng@huawei.com> Signed-off-by:
-
Weihang Li authored
There is no need to do a type cast on genernal pointers, they could be assigned to any type of variables. In addition, optimize initialization of some variables and adjust order of them. Link: https://lore.kernel.org/r/1590152579-32364-2-git-send-email-liweihang@huawei.comSigned-off-by:
-
Xi Wang authored
Currently, the MTR region is configed before hns_roce_mtr_map() is invoked, but in some scenarios, the region is configed at MTR creation, the caller need to store this config and call hns_roce_mtr_map() later. So optimize the usage by wrapping the MTR region config into MTR. Link: https://lore.kernel.org/r/1589982799-28728-10-git-send-email-liweihang@huawei.comSigned-off-by:
Xi Wang <wangxi11@huawei.com> Signed-off-by:
-
Xi Wang authored
Split the code related to WQE buffer configure from the QPC filling process into two functions: config_qp_sq_buf() and config_qp_rq_buf(), this will make the code more readable. Link: https://lore.kernel.org/r/1589982799-28728-9-git-send-email-liweihang@huawei.comSigned-off-by:
-
Weihang Li authored
Number of sge/eqe is always non-negative, they should be defined in type of unsigned. Link: https://lore.kernel.org/r/1589982799-28728-8-git-send-email-liweihang@huawei.comSigned-off-by:
-
Weihang Li authored
page_shift is used to calculate the page size, it's always non-negative, and should be in type of unsigned. Link: https://lore.kernel.org/r/1589982799-28728-7-git-send-email-liweihang@huawei.comSigned-off-by:
-
Xi Wang authored
Rename the function related to QP buffer to make the code more readable. Link: https://lore.kernel.org/r/1589982799-28728-6-git-send-email-liweihang@huawei.comSigned-off-by:
-
Yangyang Li authored
The codes related to assert are no longer used and need to be deleted. Link: https://lore.kernel.org/r/1589982799-28728-5-git-send-email-liweihang@huawei.comSigned-off-by:
Yangyang Li <liyangyang20@huawei.com> Signed-off-by:
-
