linux/drivers/infiniband/core/cq.c

509 lines
12 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0-only
/*
* Copyright (c) 2015 HGST, a Western Digital Company.
*/
#include <linux/module.h>
#include <linux/err.h>
#include <linux/slab.h>
#include <rdma/ib_verbs.h>
RDMA/core: Introduce shared CQ pool API 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>
2020-05-27 08:34:53 +00:00
#include "core_priv.h"
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
#include <trace/events/rdma_core.h>
RDMA/core: Introduce shared CQ pool API 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>
2020-05-27 08:34:53 +00:00
/* Max size for shared CQ, may require tuning */
#define IB_MAX_SHARED_CQ_SZ 4096U
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
/* # of WCs to poll for with a single call to ib_poll_cq */
#define IB_POLL_BATCH 16
#define IB_POLL_BATCH_DIRECT 8
/* # of WCs to iterate over before yielding */
#define IB_POLL_BUDGET_IRQ 256
#define IB_POLL_BUDGET_WORKQUEUE 65536
#define IB_POLL_FLAGS \
(IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS)
static const struct dim_cq_moder
rdma_dim_prof[RDMA_DIM_PARAMS_NUM_PROFILES] = {
{1, 0, 1, 0},
{1, 0, 4, 0},
{2, 0, 4, 0},
{2, 0, 8, 0},
{4, 0, 8, 0},
{16, 0, 8, 0},
{16, 0, 16, 0},
{32, 0, 16, 0},
{32, 0, 32, 0},
};
RDMA/core: Provide RDMA DIM support for ULPs Added the interface in the infiniband driver that applies the rdma_dim adaptive moderation. There is now a special function for allocating an ib_cq that uses rdma_dim. Performance improvement (ConnectX-5 100GbE, x86) running FIO benchmark over NVMf between two equal end-hosts with 56 cores across a Mellanox switch using null_blk device: READS without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.8GiB/s | 7.7M | 1401 usec | 2442 usec 4k | 7.0GiB/s | 1.8M | 4817 usec | 6587 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.6GiB/s | 7.5M | 1434 usec | 2474 usec 4k | 6.3GiB/s | 1.6M | 938 usec | 1221 usec 64k | 10.7GiB/s| 175k | 8979 usec | 12780 usec IO READS with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 4GiB/s | 8.2M | 816 usec | 889 usec 4k | 10.1GiB/s| 2.65M| 3359 usec | 5080 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.9GiB/s | 8.1M | 799 usec | 922 usec 4k | 9.6GiB/s | 2.5M | 717 usec | 1004 usec 64k | 10.7GiB/s| 176k | 8586 usec | 12256 usec The rdma_dim algorithm was designed to measure the effectiveness of moderation on the flow in a general way and thus should be appropriate for all RDMA storage protocols. rdma_dim is configured to be the default option based on performance improvement seen after extensive tests. Signed-off-by: Yamin Friedman <yaminf@mellanox.com> Reviewed-by: Max Gurtovoy <maxg@mellanox.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Saeed Mahameed <saeedm@mellanox.com> Signed-off-by: Leon Romanovsky <leonro@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-07-08 10:59:03 +00:00
static void ib_cq_rdma_dim_work(struct work_struct *w)
{
struct dim *dim = container_of(w, struct dim, work);
struct ib_cq *cq = dim->priv;
u16 usec = rdma_dim_prof[dim->profile_ix].usec;
u16 comps = rdma_dim_prof[dim->profile_ix].comps;
dim->state = DIM_START_MEASURE;
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
trace_cq_modify(cq, comps, usec);
RDMA/core: Provide RDMA DIM support for ULPs Added the interface in the infiniband driver that applies the rdma_dim adaptive moderation. There is now a special function for allocating an ib_cq that uses rdma_dim. Performance improvement (ConnectX-5 100GbE, x86) running FIO benchmark over NVMf between two equal end-hosts with 56 cores across a Mellanox switch using null_blk device: READS without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.8GiB/s | 7.7M | 1401 usec | 2442 usec 4k | 7.0GiB/s | 1.8M | 4817 usec | 6587 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.6GiB/s | 7.5M | 1434 usec | 2474 usec 4k | 6.3GiB/s | 1.6M | 938 usec | 1221 usec 64k | 10.7GiB/s| 175k | 8979 usec | 12780 usec IO READS with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 4GiB/s | 8.2M | 816 usec | 889 usec 4k | 10.1GiB/s| 2.65M| 3359 usec | 5080 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.9GiB/s | 8.1M | 799 usec | 922 usec 4k | 9.6GiB/s | 2.5M | 717 usec | 1004 usec 64k | 10.7GiB/s| 176k | 8586 usec | 12256 usec The rdma_dim algorithm was designed to measure the effectiveness of moderation on the flow in a general way and thus should be appropriate for all RDMA storage protocols. rdma_dim is configured to be the default option based on performance improvement seen after extensive tests. Signed-off-by: Yamin Friedman <yaminf@mellanox.com> Reviewed-by: Max Gurtovoy <maxg@mellanox.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Saeed Mahameed <saeedm@mellanox.com> Signed-off-by: Leon Romanovsky <leonro@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-07-08 10:59:03 +00:00
cq->device->ops.modify_cq(cq, comps, usec);
}
static void rdma_dim_init(struct ib_cq *cq)
{
struct dim *dim;
if (!cq->device->ops.modify_cq || !cq->device->use_cq_dim ||
cq->poll_ctx == IB_POLL_DIRECT)
return;
dim = kzalloc(sizeof(struct dim), GFP_KERNEL);
if (!dim)
return;
dim->state = DIM_START_MEASURE;
dim->tune_state = DIM_GOING_RIGHT;
dim->profile_ix = RDMA_DIM_START_PROFILE;
dim->priv = cq;
cq->dim = dim;
INIT_WORK(&dim->work, ib_cq_rdma_dim_work);
}
static void rdma_dim_destroy(struct ib_cq *cq)
{
if (!cq->dim)
return;
cancel_work_sync(&cq->dim->work);
kfree(cq->dim);
}
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
static int __poll_cq(struct ib_cq *cq, int num_entries, struct ib_wc *wc)
{
int rc;
rc = ib_poll_cq(cq, num_entries, wc);
trace_cq_poll(cq, num_entries, rc);
return rc;
}
static int __ib_process_cq(struct ib_cq *cq, int budget, struct ib_wc *wcs,
int batch)
{
int i, n, completed = 0;
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
trace_cq_process(cq);
/*
* budget might be (-1) if the caller does not
* want to bound this call, thus we need unsigned
* minimum here.
*/
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
while ((n = __poll_cq(cq, min_t(u32, batch,
budget - completed), wcs)) > 0) {
for (i = 0; i < n; i++) {
struct ib_wc *wc = &wcs[i];
if (wc->wr_cqe)
wc->wr_cqe->done(cq, wc);
else
WARN_ON_ONCE(wc->status == IB_WC_SUCCESS);
}
completed += n;
if (n != batch || (budget != -1 && completed >= budget))
break;
}
return completed;
}
/**
* ib_process_cq_direct - process a CQ in caller context
* @cq: CQ to process
* @budget: number of CQEs to poll for
*
* This function is used to process all outstanding CQ entries.
* It does not offload CQ processing to a different context and does
* not ask for completion interrupts from the HCA.
* Using direct processing on CQ with non IB_POLL_DIRECT type may trigger
* concurrent processing.
*
* Note: do not pass -1 as %budget unless it is guaranteed that the number
* of completions that will be processed is small.
*/
int ib_process_cq_direct(struct ib_cq *cq, int budget)
{
struct ib_wc wcs[IB_POLL_BATCH_DIRECT];
return __ib_process_cq(cq, budget, wcs, IB_POLL_BATCH_DIRECT);
}
EXPORT_SYMBOL(ib_process_cq_direct);
static void ib_cq_completion_direct(struct ib_cq *cq, void *private)
{
WARN_ONCE(1, "got unsolicited completion for CQ 0x%p\n", cq);
}
static int ib_poll_handler(struct irq_poll *iop, int budget)
{
struct ib_cq *cq = container_of(iop, struct ib_cq, iop);
RDMA/core: Provide RDMA DIM support for ULPs Added the interface in the infiniband driver that applies the rdma_dim adaptive moderation. There is now a special function for allocating an ib_cq that uses rdma_dim. Performance improvement (ConnectX-5 100GbE, x86) running FIO benchmark over NVMf between two equal end-hosts with 56 cores across a Mellanox switch using null_blk device: READS without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.8GiB/s | 7.7M | 1401 usec | 2442 usec 4k | 7.0GiB/s | 1.8M | 4817 usec | 6587 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.6GiB/s | 7.5M | 1434 usec | 2474 usec 4k | 6.3GiB/s | 1.6M | 938 usec | 1221 usec 64k | 10.7GiB/s| 175k | 8979 usec | 12780 usec IO READS with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 4GiB/s | 8.2M | 816 usec | 889 usec 4k | 10.1GiB/s| 2.65M| 3359 usec | 5080 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.9GiB/s | 8.1M | 799 usec | 922 usec 4k | 9.6GiB/s | 2.5M | 717 usec | 1004 usec 64k | 10.7GiB/s| 176k | 8586 usec | 12256 usec The rdma_dim algorithm was designed to measure the effectiveness of moderation on the flow in a general way and thus should be appropriate for all RDMA storage protocols. rdma_dim is configured to be the default option based on performance improvement seen after extensive tests. Signed-off-by: Yamin Friedman <yaminf@mellanox.com> Reviewed-by: Max Gurtovoy <maxg@mellanox.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Saeed Mahameed <saeedm@mellanox.com> Signed-off-by: Leon Romanovsky <leonro@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-07-08 10:59:03 +00:00
struct dim *dim = cq->dim;
int completed;
completed = __ib_process_cq(cq, budget, cq->wc, IB_POLL_BATCH);
if (completed < budget) {
irq_poll_complete(&cq->iop);
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
if (ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0) {
trace_cq_reschedule(cq);
irq_poll_sched(&cq->iop);
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
}
}
RDMA/core: Provide RDMA DIM support for ULPs Added the interface in the infiniband driver that applies the rdma_dim adaptive moderation. There is now a special function for allocating an ib_cq that uses rdma_dim. Performance improvement (ConnectX-5 100GbE, x86) running FIO benchmark over NVMf between two equal end-hosts with 56 cores across a Mellanox switch using null_blk device: READS without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.8GiB/s | 7.7M | 1401 usec | 2442 usec 4k | 7.0GiB/s | 1.8M | 4817 usec | 6587 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.6GiB/s | 7.5M | 1434 usec | 2474 usec 4k | 6.3GiB/s | 1.6M | 938 usec | 1221 usec 64k | 10.7GiB/s| 175k | 8979 usec | 12780 usec IO READS with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 4GiB/s | 8.2M | 816 usec | 889 usec 4k | 10.1GiB/s| 2.65M| 3359 usec | 5080 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.9GiB/s | 8.1M | 799 usec | 922 usec 4k | 9.6GiB/s | 2.5M | 717 usec | 1004 usec 64k | 10.7GiB/s| 176k | 8586 usec | 12256 usec The rdma_dim algorithm was designed to measure the effectiveness of moderation on the flow in a general way and thus should be appropriate for all RDMA storage protocols. rdma_dim is configured to be the default option based on performance improvement seen after extensive tests. Signed-off-by: Yamin Friedman <yaminf@mellanox.com> Reviewed-by: Max Gurtovoy <maxg@mellanox.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Saeed Mahameed <saeedm@mellanox.com> Signed-off-by: Leon Romanovsky <leonro@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-07-08 10:59:03 +00:00
if (dim)
rdma_dim(dim, completed);
return completed;
}
static void ib_cq_completion_softirq(struct ib_cq *cq, void *private)
{
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
trace_cq_schedule(cq);
irq_poll_sched(&cq->iop);
}
static void ib_cq_poll_work(struct work_struct *work)
{
struct ib_cq *cq = container_of(work, struct ib_cq, work);
int completed;
completed = __ib_process_cq(cq, IB_POLL_BUDGET_WORKQUEUE, cq->wc,
IB_POLL_BATCH);
if (completed >= IB_POLL_BUDGET_WORKQUEUE ||
ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0)
queue_work(cq->comp_wq, &cq->work);
RDMA/core: Provide RDMA DIM support for ULPs Added the interface in the infiniband driver that applies the rdma_dim adaptive moderation. There is now a special function for allocating an ib_cq that uses rdma_dim. Performance improvement (ConnectX-5 100GbE, x86) running FIO benchmark over NVMf between two equal end-hosts with 56 cores across a Mellanox switch using null_blk device: READS without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.8GiB/s | 7.7M | 1401 usec | 2442 usec 4k | 7.0GiB/s | 1.8M | 4817 usec | 6587 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.6GiB/s | 7.5M | 1434 usec | 2474 usec 4k | 6.3GiB/s | 1.6M | 938 usec | 1221 usec 64k | 10.7GiB/s| 175k | 8979 usec | 12780 usec IO READS with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 4GiB/s | 8.2M | 816 usec | 889 usec 4k | 10.1GiB/s| 2.65M| 3359 usec | 5080 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.9GiB/s | 8.1M | 799 usec | 922 usec 4k | 9.6GiB/s | 2.5M | 717 usec | 1004 usec 64k | 10.7GiB/s| 176k | 8586 usec | 12256 usec The rdma_dim algorithm was designed to measure the effectiveness of moderation on the flow in a general way and thus should be appropriate for all RDMA storage protocols. rdma_dim is configured to be the default option based on performance improvement seen after extensive tests. Signed-off-by: Yamin Friedman <yaminf@mellanox.com> Reviewed-by: Max Gurtovoy <maxg@mellanox.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Saeed Mahameed <saeedm@mellanox.com> Signed-off-by: Leon Romanovsky <leonro@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-07-08 10:59:03 +00:00
else if (cq->dim)
rdma_dim(cq->dim, completed);
}
static void ib_cq_completion_workqueue(struct ib_cq *cq, void *private)
{
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
trace_cq_schedule(cq);
queue_work(cq->comp_wq, &cq->work);
}
/**
* __ib_alloc_cq - allocate a completion queue
* @dev: device to allocate the CQ for
* @private: driver private data, accessible from cq->cq_context
* @nr_cqe: number of CQEs to allocate
* @comp_vector: HCA completion vectors for this CQ
* @poll_ctx: context to poll the CQ from.
* @caller: module owner name.
*
* This is the proper interface to allocate a CQ for in-kernel users. A
* CQ allocated with this interface will automatically be polled from the
* specified context. The ULP must use wr->wr_cqe instead of wr->wr_id
* to use this CQ abstraction.
*/
struct ib_cq *__ib_alloc_cq(struct ib_device *dev, void *private, int nr_cqe,
int comp_vector, enum ib_poll_context poll_ctx,
const char *caller)
{
struct ib_cq_init_attr cq_attr = {
.cqe = nr_cqe,
.comp_vector = comp_vector,
};
struct ib_cq *cq;
int ret = -ENOMEM;
cq = rdma_zalloc_drv_obj(dev, ib_cq);
if (!cq)
return ERR_PTR(ret);
cq->device = dev;
cq->cq_context = private;
cq->poll_ctx = poll_ctx;
atomic_set(&cq->usecnt, 0);
RDMA/core: Introduce shared CQ pool API 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>
2020-05-27 08:34:53 +00:00
cq->comp_vector = comp_vector;
cq->wc = kmalloc_array(IB_POLL_BATCH, sizeof(*cq->wc), GFP_KERNEL);
if (!cq->wc)
goto out_free_cq;
rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ);
rdma_restrack_set_name(&cq->res, caller);
ret = dev->ops.create_cq(cq, &cq_attr, NULL);
if (ret)
goto out_free_wc;
RDMA/core: Provide RDMA DIM support for ULPs Added the interface in the infiniband driver that applies the rdma_dim adaptive moderation. There is now a special function for allocating an ib_cq that uses rdma_dim. Performance improvement (ConnectX-5 100GbE, x86) running FIO benchmark over NVMf between two equal end-hosts with 56 cores across a Mellanox switch using null_blk device: READS without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.8GiB/s | 7.7M | 1401 usec | 2442 usec 4k | 7.0GiB/s | 1.8M | 4817 usec | 6587 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES without DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.6GiB/s | 7.5M | 1434 usec | 2474 usec 4k | 6.3GiB/s | 1.6M | 938 usec | 1221 usec 64k | 10.7GiB/s| 175k | 8979 usec | 12780 usec IO READS with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 4GiB/s | 8.2M | 816 usec | 889 usec 4k | 10.1GiB/s| 2.65M| 3359 usec | 5080 usec 64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec IO WRITES with DIM: blk size | BW | IOPS | 99th percentile latency | 99.99th latency 512B | 3.9GiB/s | 8.1M | 799 usec | 922 usec 4k | 9.6GiB/s | 2.5M | 717 usec | 1004 usec 64k | 10.7GiB/s| 176k | 8586 usec | 12256 usec The rdma_dim algorithm was designed to measure the effectiveness of moderation on the flow in a general way and thus should be appropriate for all RDMA storage protocols. rdma_dim is configured to be the default option based on performance improvement seen after extensive tests. Signed-off-by: Yamin Friedman <yaminf@mellanox.com> Reviewed-by: Max Gurtovoy <maxg@mellanox.com> Reviewed-by: Sagi Grimberg <sagi@grimberg.me> Signed-off-by: Saeed Mahameed <saeedm@mellanox.com> Signed-off-by: Leon Romanovsky <leonro@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-07-08 10:59:03 +00:00
rdma_dim_init(cq);
switch (cq->poll_ctx) {
case IB_POLL_DIRECT:
cq->comp_handler = ib_cq_completion_direct;
break;
case IB_POLL_SOFTIRQ:
cq->comp_handler = ib_cq_completion_softirq;
irq_poll_init(&cq->iop, IB_POLL_BUDGET_IRQ, ib_poll_handler);
ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
break;
case IB_POLL_WORKQUEUE:
case IB_POLL_UNBOUND_WORKQUEUE:
cq->comp_handler = ib_cq_completion_workqueue;
INIT_WORK(&cq->work, ib_cq_poll_work);
ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
cq->comp_wq = (cq->poll_ctx == IB_POLL_WORKQUEUE) ?
ib_comp_wq : ib_comp_unbound_wq;
break;
default:
ret = -EINVAL;
goto out_destroy_cq;
}
rdma_restrack_add(&cq->res);
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
trace_cq_alloc(cq, nr_cqe, comp_vector, poll_ctx);
return cq;
out_destroy_cq:
rdma_dim_destroy(cq);
cq->device->ops.destroy_cq(cq, NULL);
out_free_wc:
rdma_restrack_put(&cq->res);
kfree(cq->wc);
out_free_cq:
kfree(cq);
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
trace_cq_alloc_error(nr_cqe, comp_vector, poll_ctx, ret);
return ERR_PTR(ret);
}
EXPORT_SYMBOL(__ib_alloc_cq);
/**
* __ib_alloc_cq_any - allocate a completion queue
* @dev: device to allocate the CQ for
* @private: driver private data, accessible from cq->cq_context
* @nr_cqe: number of CQEs to allocate
* @poll_ctx: context to poll the CQ from
* @caller: module owner name
*
* Attempt to spread ULP Completion Queues over each device's interrupt
* vectors. A simple best-effort mechanism is used.
*/
struct ib_cq *__ib_alloc_cq_any(struct ib_device *dev, void *private,
int nr_cqe, enum ib_poll_context poll_ctx,
const char *caller)
{
static atomic_t counter;
int comp_vector = 0;
if (dev->num_comp_vectors > 1)
comp_vector =
atomic_inc_return(&counter) %
min_t(int, dev->num_comp_vectors, num_online_cpus());
return __ib_alloc_cq(dev, private, nr_cqe, comp_vector, poll_ctx,
caller);
}
EXPORT_SYMBOL(__ib_alloc_cq_any);
/**
* ib_free_cq - free a completion queue
* @cq: completion queue to free.
*/
void ib_free_cq(struct ib_cq *cq)
{
int ret;
if (WARN_ON_ONCE(atomic_read(&cq->usecnt)))
return;
RDMA/core: Introduce shared CQ pool API 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>
2020-05-27 08:34:53 +00:00
if (WARN_ON_ONCE(cq->cqe_used))
return;
switch (cq->poll_ctx) {
case IB_POLL_DIRECT:
break;
case IB_POLL_SOFTIRQ:
irq_poll_disable(&cq->iop);
break;
case IB_POLL_WORKQUEUE:
case IB_POLL_UNBOUND_WORKQUEUE:
cancel_work_sync(&cq->work);
break;
default:
WARN_ON_ONCE(1);
}
rdma_dim_destroy(cq);
RDMA/core: Trace points for diagnosing completion queue issues Sample trace events: kworker/u29:0-300 [007] 120.042217: cq_alloc: cq.id=4 nr_cqe=161 comp_vector=2 poll_ctx=WORKQUEUE <idle>-0 [002] 120.056292: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.056402: cq_process: cq.id=4 wake-up took 109 [us] from interrupt kworker/2:1H-482 [002] 120.056407: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067503: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067537: cq_process: cq.id=4 wake-up took 34 [us] from interrupt kworker/2:1H-482 [002] 120.067541: cq_poll: cq.id=4 requested 16, returned 1 <idle>-0 [002] 120.067657: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 120.067672: cq_process: cq.id=4 wake-up took 15 [us] from interrupt kworker/2:1H-482 [002] 120.067674: cq_poll: cq.id=4 requested 16, returned 1 ... systemd-1 [002] 122.392653: cq_schedule: cq.id=4 kworker/2:1H-482 [002] 122.392688: cq_process: cq.id=4 wake-up took 35 [us] from interrupt kworker/2:1H-482 [002] 122.392693: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392836: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.392970: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393083: cq_poll: cq.id=4 requested 16, returned 16 kworker/2:1H-482 [002] 122.393195: cq_poll: cq.id=4 requested 16, returned 3 Several features to note in this output: - The WCE count and context type are reported at allocation time - The CPU and kworker for each CQ is evident - The CQ's restracker ID is tagged on each trace event - CQ poll scheduling latency is measured - Details about how often single completions occur versus multiple completions are evident - The cost of the ULP's completion handler is recorded Link: https://lore.kernel.org/r/20191218201815.30584.3481.stgit@manet.1015granger.net Signed-off-by: Chuck Lever <chuck.lever@oracle.com> Reviewed-by: Parav Pandit <parav@mellanox.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-12-18 20:18:15 +00:00
trace_cq_free(cq);
ret = cq->device->ops.destroy_cq(cq, NULL);
WARN_ONCE(ret, "Destroy of kernel CQ shouldn't fail");
rdma_restrack_del(&cq->res);
kfree(cq->wc);
kfree(cq);
}
EXPORT_SYMBOL(ib_free_cq);
RDMA/core: Introduce shared CQ pool API 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>
2020-05-27 08:34:53 +00:00
RDMA/core: Clean up cq pool mechanism The CQ pool mechanism had two problems: 1. The CQ pool lists were uninitialized in the device registration error flow. As a result, all the list pointers remained NULL. This caused the kernel to crash (in procedure ib_cq_pool_destroy) when that error flow was taken (and unregister called). The stack trace snippet: BUG: kernel NULL pointer dereference, address: 0000000000000000 #PF: supervisor read access in kernel mode #PF: error_code(0×0000) ? not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI . . . RIP: 0010:ib_cq_pool_destroy+0x1b/0×70 [ib_core] . . . Call Trace: disable_device+0x9f/0×130 [ib_core] __ib_unregister_device+0x35/0×90 [ib_core] ib_register_device+0x529/0×610 [ib_core] __mlx5_ib_add+0x3a/0×70 [mlx5_ib] mlx5_add_device+0x87/0×1c0 [mlx5_core] mlx5_register_interface+0x74/0xc0 [mlx5_core] do_one_initcall+0x4b/0×1f4 do_init_module+0x5a/0×223 load_module+0x1938/0×1d40 2. At device unregister, when cleaning up the cq pool, the cq's in the pool lists were freed, but the cq entries were left in the list. The fix for the first issue is to initialize the cq pool lists when the ib_device structure is allocated for a new device (in procedure _ib_alloc_device). The fix for the second problem is to delete cq entries from the pool lists when cleaning up the cq pool. In addition, procedure ib_cq_pool_destroy() is renamed to the more appropriate name ib_cq_pool_cleanup(). Fixes: 4aa1615268a8 ("RDMA/core: Fix ordering of CQ pool destruction") Link: https://lore.kernel.org/r/20201208073545.9723-2-leon@kernel.org Suggested-by: Jason Gunthorpe <jgg@nvidia.com> Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il> Signed-off-by: Leon Romanovsky <leonro@nvidia.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
2020-12-08 07:35:43 +00:00
void ib_cq_pool_cleanup(struct ib_device *dev)
RDMA/core: Introduce shared CQ pool API 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>
2020-05-27 08:34:53 +00:00
{
struct ib_cq *cq, *n;
unsigned int i;
for (i = 0; i < ARRAY_SIZE(dev->cq_pools); i++) {
list_for_each_entry_safe(cq, n, &dev->cq_pools[i],
pool_entry) {
WARN_ON(cq->cqe_used);
RDMA/core: Clean up cq pool mechanism The CQ pool mechanism had two problems: 1. The CQ pool lists were uninitialized in the device registration error flow. As a result, all the list pointers remained NULL. This caused the kernel to crash (in procedure ib_cq_pool_destroy) when that error flow was taken (and unregister called). The stack trace snippet: BUG: kernel NULL pointer dereference, address: 0000000000000000 #PF: supervisor read access in kernel mode #PF: error_code(0×0000) ? not-present page PGD 0 P4D 0 Oops: 0000 [#1] SMP PTI . . . RIP: 0010:ib_cq_pool_destroy+0x1b/0×70 [ib_core] . . . Call Trace: disable_device+0x9f/0×130 [ib_core] __ib_unregister_device+0x35/0×90 [ib_core] ib_register_device+0x529/0×610 [ib_core] __mlx5_ib_add+0x3a/0×70 [mlx5_ib] mlx5_add_device+0x87/0×1c0 [mlx5_core] mlx5_register_interface+0x74/0xc0 [mlx5_core] do_one_initcall+0x4b/0×1f4 do_init_module+0x5a/0×223 load_module+0x1938/0×1d40 2. At device unregister, when cleaning up the cq pool, the cq's in the pool lists were freed, but the cq entries were left in the list. The fix for the first issue is to initialize the cq pool lists when the ib_device structure is allocated for a new device (in procedure _ib_alloc_device). The fix for the second problem is to delete cq entries from the pool lists when cleaning up the cq pool. In addition, procedure ib_cq_pool_destroy() is renamed to the more appropriate name ib_cq_pool_cleanup(). Fixes: 4aa1615268a8 ("RDMA/core: Fix ordering of CQ pool destruction") Link: https://lore.kernel.org/r/20201208073545.9723-2-leon@kernel.org Suggested-by: Jason Gunthorpe <jgg@nvidia.com> Signed-off-by: Jack Morgenstein <jackm@dev.mellanox.co.il> Signed-off-by: Leon Romanovsky <leonro@nvidia.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
2020-12-08 07:35:43 +00:00
list_del(&cq->pool_entry);
RDMA/core: Introduce shared CQ pool API 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>
2020-05-27 08:34:53 +00:00
cq->shared = false;
ib_free_cq(cq);
}
}
}
static int ib_alloc_cqs(struct ib_device *dev, unsigned int nr_cqes,
enum ib_poll_context poll_ctx)
{
LIST_HEAD(tmp_list);
unsigned int nr_cqs, i;
struct ib_cq *cq, *n;
RDMA/core: Introduce shared CQ pool API 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>
2020-05-27 08:34:53 +00:00
int ret;
if (poll_ctx > IB_POLL_LAST_POOL_TYPE) {
WARN_ON_ONCE(poll_ctx > IB_POLL_LAST_POOL_TYPE);
return -EINVAL;
}
/*
* Allocate at least as many CQEs as requested, and otherwise
* a reasonable batch size so that we can share CQs between
* multiple users instead of allocating a larger number of CQs.
*/
nr_cqes = min_t(unsigned int, dev->attrs.max_cqe,
max(nr_cqes, IB_MAX_SHARED_CQ_SZ));
nr_cqs = min_t(unsigned int, dev->num_comp_vectors, num_online_cpus());
for (i = 0; i < nr_cqs; i++) {
cq = ib_alloc_cq(dev, NULL, nr_cqes, i, poll_ctx);
if (IS_ERR(cq)) {
ret = PTR_ERR(cq);
goto out_free_cqs;
}
cq->shared = true;
list_add_tail(&cq->pool_entry, &tmp_list);
}
spin_lock_irq(&dev->cq_pools_lock);
list_splice(&tmp_list, &dev->cq_pools[poll_ctx]);
spin_unlock_irq(&dev->cq_pools_lock);
return 0;
out_free_cqs:
list_for_each_entry_safe(cq, n, &tmp_list, pool_entry) {
RDMA/core: Introduce shared CQ pool API 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>
2020-05-27 08:34:53 +00:00
cq->shared = false;
ib_free_cq(cq);
}
return ret;
}
/**
* ib_cq_pool_get() - Find the least used completion queue that matches
* a given cpu hint (or least used for wild card affinity) and fits
* nr_cqe.
* @dev: rdma device
* @nr_cqe: number of needed cqe entries
* @comp_vector_hint: completion vector hint (-1) for the driver to assign
* a comp vector based on internal counter
* @poll_ctx: cq polling context
*
* Finds a cq that satisfies @comp_vector_hint and @nr_cqe requirements and
* claim entries in it for us. In case there is no available cq, allocate
* a new cq with the requirements and add it to the device pool.
* IB_POLL_DIRECT cannot be used for shared cqs so it is not a valid value
* for @poll_ctx.
*/
struct ib_cq *ib_cq_pool_get(struct ib_device *dev, unsigned int nr_cqe,
int comp_vector_hint,
enum ib_poll_context poll_ctx)
{
static unsigned int default_comp_vector;
unsigned int vector, num_comp_vectors;
struct ib_cq *cq, *found = NULL;
int ret;
if (poll_ctx > IB_POLL_LAST_POOL_TYPE) {
WARN_ON_ONCE(poll_ctx > IB_POLL_LAST_POOL_TYPE);
return ERR_PTR(-EINVAL);
}
num_comp_vectors =
min_t(unsigned int, dev->num_comp_vectors, num_online_cpus());
/* Project the affinty to the device completion vector range */
if (comp_vector_hint < 0) {
comp_vector_hint =
(READ_ONCE(default_comp_vector) + 1) % num_comp_vectors;
WRITE_ONCE(default_comp_vector, comp_vector_hint);
}
vector = comp_vector_hint % num_comp_vectors;
/*
* Find the least used CQ with correct affinity and
* enough free CQ entries
*/
while (!found) {
spin_lock_irq(&dev->cq_pools_lock);
list_for_each_entry(cq, &dev->cq_pools[poll_ctx],
pool_entry) {
/*
* Check to see if we have found a CQ with the
* correct completion vector
*/
if (vector != cq->comp_vector)
continue;
if (cq->cqe_used + nr_cqe > cq->cqe)
continue;
found = cq;
break;
}
if (found) {
found->cqe_used += nr_cqe;
spin_unlock_irq(&dev->cq_pools_lock);
return found;
}
spin_unlock_irq(&dev->cq_pools_lock);
/*
* Didn't find a match or ran out of CQs in the device
* pool, allocate a new array of CQs.
*/
ret = ib_alloc_cqs(dev, nr_cqe, poll_ctx);
if (ret)
return ERR_PTR(ret);
}
return found;
}
EXPORT_SYMBOL(ib_cq_pool_get);
/**
* ib_cq_pool_put - Return a CQ taken from a shared pool.
* @cq: The CQ to return.
* @nr_cqe: The max number of cqes that the user had requested.
*/
void ib_cq_pool_put(struct ib_cq *cq, unsigned int nr_cqe)
{
if (WARN_ON_ONCE(nr_cqe > cq->cqe_used))
return;
spin_lock_irq(&cq->device->cq_pools_lock);
cq->cqe_used -= nr_cqe;
spin_unlock_irq(&cq->device->cq_pools_lock);
}
EXPORT_SYMBOL(ib_cq_pool_put);