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.com
Signed-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>
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.com
Signed-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>
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 b8eb718348 ("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.edu
Signed-off-by: Qiushi Wu <wu000273@umn.edu>
Reviewed-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
There is a spelling mistake in an error message. Fix it.
Link: https://lore.kernel.org/r/20200528110709.400935-1-colin.king@canonical.com
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Acked-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The pointer raq is being assigned twice. Fix this by removing one of the
redundant assignments.
Fixes: 14ba87304b ("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: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Signed-off-by: Mark Zhang <markz@mellanox.com>
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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 f95ccffc71 ("IB/mlx4: Use 4K pages for kernel QP's WQE buffer").
Link: https://lore.kernel.org/r/20200525172212.14413-5-bvanassche@acm.org
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Signed-off-by: Bart Van Assche <bvanassche@acm.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Reviewed-by: Mark Zhang <markz@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Reviewed-by: Mark Zhang <markz@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Reviewed-by: Mark Zhang <markz@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Reviewed-by: Mark Zhang <markz@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Reviewed-by: Mark Zhang <markz@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Reviewed-by: Mark Zhang <markz@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Reviewed-by: Mark Zhang <markz@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.org
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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: Jason Gunthorpe <jgg@mellanox.com>
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: 819f7427ba ("RDMA/mlx5: Add init2init as a modify command")
Link: https://lore.kernel.org/r/20200527135703.482501-1-leon@kernel.org
Signed-off-by: Mark Zhang <markz@mellanox.com>
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Make use of the sizeof_field() helper instead of an open-coded version.
Link: https://lore.kernel.org/r/20200527144152.GA22605@embeddedor
Signed-off-by: Gustavo A. R. Silva <gustavoars@kernel.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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: Mark Zhang <markz@mellanox.com>
Reviewed-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
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.com
Signed-off-by: Kamal Heib <kamalheib1@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.com
Signed-off-by: Potnuri Bharat Teja <bharat@chelsio.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.com
Signed-off-by: Yixian Liu <liuyixian@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.com
Signed-off-by: Wenpeng Liang <liangwenpeng@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.com
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.com
Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
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.com
Signed-off-by: Xi Wang <wangxi11@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
It's easier to understand and maintain enable flags of cq using a single
field in type of u32 than defining a field for every flags in the
structure hns_roce_cq, and we can add new flags for features more
conveniently in the future.
Link: https://lore.kernel.org/r/1589982799-28728-3-git-send-email-liweihang@huawei.com
Signed-off-by: Lang Cheng <chenglang@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The hardware can truncate PI/CI when posting or polling, the driver does
not need to do truncation. Therefore keep the software's PI/CI consistent
with it in the hardware.
Link: https://lore.kernel.org/r/1589982799-28728-2-git-send-email-liweihang@huawei.com
Signed-off-by: Lang Cheng <chenglang@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The macros do_each_path/while_each_path lead to a smatch warning:
drivers/infiniband/ulp/rtrs/rtrs-clt.c:1196 rtrs_clt_failover_req() warn: inconsistent indenting
drivers/infiniband/ulp/rtrs/rtrs-clt.c:2890 rtrs_clt_request() warn: inconsistent indenting
Also checkpatch complains:
ERROR: Macros with multiple statements should be enclosed in a do - while loop
The macros are used only in two places: for a normal IO path and for the
failover path triggered after errors.
Get rid of the macros and just use a for loop iterating over the list of
paths in both places. It is easier to read and also less lines of code.
Fixes: 6a98d71dae ("RDMA/rtrs: client: main functionality")
Link: https://lore.kernel.org/r/20200522053924.528980-1-danil.kipnis@cloud.ionos.com
Reported-by: kbuild test robot <lkp@intel.com>
Signed-off-by: Danil Kipnis <danil.kipnis@cloud.ionos.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The rtrs_sess structure has already been extracted above from the
rtrs_srv_sess structure. Use that to avoid redundant dereferencing.
Fixes: 9cb8374804 ("RDMA/rtrs: server: main functionality")
Link: https://lore.kernel.org/r/20200522082833.1480551-1-haris.phnx@gmail.com
Signed-off-by: Md Haris Iqbal <haris.phnx@gmail.com>
Acked-by: Danil Kipnis <danil.kipnis@cloud.ionos.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
module_is_live function is only defined when CONFIG_MODULES is enabled.
Use try_module_get instead to check whether the module is being removed.
When module unload and manuall unmapping is happening in parallel, we can
try removing the symlink twice: rnbd_client_exit
vs. rnbd_clt_unmap_dev_store.
This is probably not the best way to deal with this race in general, but
for now this fixes the compilation issue when CONFIG_MODULES is disabled
and has no functional impact. Regression tests passed.
Fixes: 1eb54f8f5d ("block/rnbd: client: sysfs interface functions")
Link: https://lore.kernel.org/r/20200521185909.457245-1-danil.kipnis@cloud.ionos.com
Reported-by: Randy Dunlap <rdunlap@infradead.org>
Suggested-by: Guoqing Jiang <guoqing.jiang@cloud.ionos.com>
Signed-off-by: Danil Kipnis <danil.kipnis@cloud.ionos.com>
Acked-by: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The allocated ports structure in never freed. The free function should be
called by release_cma_ports_group, but the group is never released since
we don't remove its default group.
Remove default groups when device group is deleted.
Fixes: 045959db65 ("IB/cma: Add configfs for rdma_cm")
Link: https://lore.kernel.org/r/20200521072650.567908-1-leon@kernel.org
Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The alloc_sess() function returns error pointers, it never returns NULL.
Fixes: f7a7a5c228 ("block/rnbd: client: main functionality")
Link: https://lore.kernel.org/r/20200519120347.GD42765@mwanda
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Jack Wang <jinpu.wang@cloud.ionos.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Introduce create/destroy QP commands over the ioctl interface to let it
be extended to get an asynchronous event FD.
Link: https://lore.kernel.org/r/20200519072711.257271-8-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Introduce create/destroy WQ commands over the ioctl interface to let it
be extended to get an asynchronous event FD.
Link: https://lore.kernel.org/r/20200519072711.257271-7-leon@kernel.org
Signed-off-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Yamin Friedman