2 * This file is provided under a dual BSD/GPLv2 license. When using or
3 * redistributing this file, you may do so under either license.
7 * Copyright(c) 2015 Intel Corporation. All rights reserved.
8 * Copyright(c) 2017 T-Platforms. All Rights Reserved.
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of version 2 of the GNU General Public License as
12 * published by the Free Software Foundation.
16 * Copyright(c) 2015 Intel Corporation. All rights reserved.
17 * Copyright(c) 2017 T-Platforms. All Rights Reserved.
19 * Redistribution and use in source and binary forms, with or without
20 * modification, are permitted provided that the following conditions
23 * * Redistributions of source code must retain the above copyright
24 * notice, this list of conditions and the following disclaimer.
25 * * Redistributions in binary form must reproduce the above copy
26 * notice, this list of conditions and the following disclaimer in
27 * the documentation and/or other materials provided with the
29 * * Neither the name of Intel Corporation nor the names of its
30 * contributors may be used to endorse or promote products derived
31 * from this software without specific prior written permission.
33 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
34 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
35 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
36 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
37 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
38 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
39 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
40 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
41 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
42 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
43 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45 * PCIe NTB Perf Linux driver
49 * How to use this tool, by example.
51 * Assuming $DBG_DIR is something like:
52 * '/sys/kernel/debug/ntb_perf/0000:00:03.0'
53 * Suppose aside from local device there is at least one remote device
54 * connected to NTB with index 0.
55 *-----------------------------------------------------------------------------
56 * Eg: install driver with specified chunk/total orders and dma-enabled flag
58 * root@self# insmod ntb_perf.ko chunk_order=19 total_order=28 use_dma
59 *-----------------------------------------------------------------------------
60 * Eg: check NTB ports (index) and MW mapping information
62 * root@self# cat $DBG_DIR/info
63 *-----------------------------------------------------------------------------
64 * Eg: start performance test with peer (index 0) and get the test metrics
66 * root@self# echo 0 > $DBG_DIR/run
67 * root@self# cat $DBG_DIR/run
70 #include <linux/init.h>
71 #include <linux/kernel.h>
72 #include <linux/module.h>
73 #include <linux/sched.h>
74 #include <linux/wait.h>
75 #include <linux/dma-mapping.h>
76 #include <linux/dmaengine.h>
77 #include <linux/pci.h>
78 #include <linux/ktime.h>
79 #include <linux/slab.h>
80 #include <linux/delay.h>
81 #include <linux/sizes.h>
82 #include <linux/workqueue.h>
83 #include <linux/debugfs.h>
84 #include <linux/random.h>
85 #include <linux/ntb.h>
87 #define DRIVER_NAME "ntb_perf"
88 #define DRIVER_VERSION "2.0"
90 MODULE_LICENSE("Dual BSD/GPL");
91 MODULE_VERSION(DRIVER_VERSION);
92 MODULE_AUTHOR("Dave Jiang <dave.jiang@intel.com>");
93 MODULE_DESCRIPTION("PCIe NTB Performance Measurement Tool");
95 #define MAX_THREADS_CNT 32
96 #define DEF_THREADS_CNT 1
97 #define MAX_CHUNK_SIZE SZ_1M
98 #define MAX_CHUNK_ORDER 20 /* no larger than 1M */
100 #define DMA_TRIES 100
101 #define DMA_MDELAY 10
103 #define MSG_TRIES 500
104 #define MSG_UDELAY_LOW 1000
105 #define MSG_UDELAY_HIGH 2000
107 #define PERF_BUF_LEN 1024
109 static unsigned long max_mw_size;
110 module_param(max_mw_size, ulong, 0644);
111 MODULE_PARM_DESC(max_mw_size, "Upper limit of memory window size");
113 static unsigned char chunk_order = 19; /* 512K */
114 module_param(chunk_order, byte, 0644);
115 MODULE_PARM_DESC(chunk_order, "Data chunk order [2^n] to transfer");
117 static unsigned char total_order = 30; /* 1G */
118 module_param(total_order, byte, 0644);
119 MODULE_PARM_DESC(total_order, "Total data order [2^n] to transfer");
121 static bool use_dma; /* default to 0 */
122 module_param(use_dma, bool, 0644);
123 MODULE_PARM_DESC(use_dma, "Use DMA engine to measure performance");
125 /*==============================================================================
126 * Perf driver data definition
127 *==============================================================================
131 PERF_CMD_INVAL = -1,/* invalid spad command */
132 PERF_CMD_SSIZE = 0, /* send out buffer size */
133 PERF_CMD_RSIZE = 1, /* recv in buffer size */
134 PERF_CMD_SXLAT = 2, /* send in buffer xlat */
135 PERF_CMD_RXLAT = 3, /* recv out buffer xlat */
136 PERF_CMD_CLEAR = 4, /* clear allocated memory */
137 PERF_STS_DONE = 5, /* init is done */
138 PERF_STS_LNKUP = 6, /* link up state flag */
144 struct perf_ctx *perf;
148 /* Outbound MW params */
150 resource_size_t outbuf_size;
151 void __iomem *outbuf;
153 /* Inbound MW params */
154 dma_addr_t inbuf_xlat;
155 resource_size_t inbuf_size;
158 /* NTB connection setup service */
159 struct work_struct service;
162 struct completion init_comp;
164 #define to_peer_service(__work) \
165 container_of(__work, struct perf_peer, service)
168 struct perf_ctx *perf;
171 /* DMA-based test sync parameters */
173 wait_queue_head_t dma_wait;
174 struct dma_chan *dma_chan;
176 /* Data source and measured statistics */
181 struct work_struct work;
183 #define to_thread_work(__work) \
184 container_of(__work, struct perf_thread, work)
189 /* Global device index and peers descriptors */
192 struct perf_peer *peers;
194 /* Performance measuring work-threads interface */
195 unsigned long busy_flag;
196 wait_queue_head_t twait;
199 struct perf_peer *test_peer;
200 struct perf_thread threads[MAX_THREADS_CNT];
202 /* Scratchpad/Message IO operations */
203 int (*cmd_send)(struct perf_peer *peer, enum perf_cmd cmd, u64 data);
204 int (*cmd_recv)(struct perf_ctx *perf, int *pidx, enum perf_cmd *cmd,
207 struct dentry *dbgfs_dir;
211 * Scratchpads-base commands interface
213 #define PERF_SPAD_CNT(_pcnt) \
215 #define PERF_SPAD_CMD(_gidx) \
217 #define PERF_SPAD_LDATA(_gidx) \
219 #define PERF_SPAD_HDATA(_gidx) \
221 #define PERF_SPAD_NOTIFY(_gidx) \
225 * Messages-base commands interface
227 #define PERF_MSG_CNT 3
228 #define PERF_MSG_CMD 0
229 #define PERF_MSG_LDATA 1
230 #define PERF_MSG_HDATA 2
232 /*==============================================================================
233 * Static data declarations
234 *==============================================================================
237 static struct dentry *perf_dbgfs_topdir;
239 static struct workqueue_struct *perf_wq __read_mostly;
241 /*==============================================================================
242 * NTB cross-link commands execution service
243 *==============================================================================
246 static void perf_terminate_test(struct perf_ctx *perf);
248 static inline bool perf_link_is_up(struct perf_peer *peer)
252 link = ntb_link_is_up(peer->perf->ntb, NULL, NULL);
253 return !!(link & BIT_ULL_MASK(peer->pidx));
256 static int perf_spad_cmd_send(struct perf_peer *peer, enum perf_cmd cmd,
259 struct perf_ctx *perf = peer->perf;
263 dev_dbg(&perf->ntb->dev, "CMD send: %d 0x%llx\n", cmd, data);
266 * Perform predefined number of attempts before give up.
267 * We are sending the data to the port specific scratchpad, so
268 * to prevent a multi-port access race-condition. Additionally
269 * there is no need in local locking since only thread-safe
270 * service work is using this method.
272 for (try = 0; try < MSG_TRIES; try++) {
273 if (!perf_link_is_up(peer))
276 sts = ntb_peer_spad_read(perf->ntb, peer->pidx,
277 PERF_SPAD_CMD(perf->gidx));
278 if (sts != PERF_CMD_INVAL) {
279 usleep_range(MSG_UDELAY_LOW, MSG_UDELAY_HIGH);
283 ntb_peer_spad_write(perf->ntb, peer->pidx,
284 PERF_SPAD_LDATA(perf->gidx),
285 lower_32_bits(data));
286 ntb_peer_spad_write(perf->ntb, peer->pidx,
287 PERF_SPAD_HDATA(perf->gidx),
288 upper_32_bits(data));
290 ntb_peer_spad_write(perf->ntb, peer->pidx,
291 PERF_SPAD_CMD(perf->gidx),
294 ntb_peer_db_set(perf->ntb, PERF_SPAD_NOTIFY(peer->gidx));
296 dev_dbg(&perf->ntb->dev, "DB ring peer %#llx\n",
297 PERF_SPAD_NOTIFY(peer->gidx));
302 return try < MSG_TRIES ? 0 : -EAGAIN;
305 static int perf_spad_cmd_recv(struct perf_ctx *perf, int *pidx,
306 enum perf_cmd *cmd, u64 *data)
308 struct perf_peer *peer;
311 ntb_db_clear(perf->ntb, PERF_SPAD_NOTIFY(perf->gidx));
314 * We start scanning all over, since cleared DB may have been set
315 * by any peer. Yes, it makes peer with smaller index being
316 * serviced with greater priority, but it's convenient for spad
317 * and message code unification and simplicity.
319 for (*pidx = 0; *pidx < perf->pcnt; (*pidx)++) {
320 peer = &perf->peers[*pidx];
322 if (!perf_link_is_up(peer))
325 val = ntb_spad_read(perf->ntb, PERF_SPAD_CMD(peer->gidx));
326 if (val == PERF_CMD_INVAL)
331 val = ntb_spad_read(perf->ntb, PERF_SPAD_LDATA(peer->gidx));
334 val = ntb_spad_read(perf->ntb, PERF_SPAD_HDATA(peer->gidx));
335 *data |= (u64)val << 32;
337 /* Next command can be retrieved from now */
338 ntb_spad_write(perf->ntb, PERF_SPAD_CMD(peer->gidx),
341 dev_dbg(&perf->ntb->dev, "CMD recv: %d 0x%llx\n", *cmd, *data);
349 static int perf_msg_cmd_send(struct perf_peer *peer, enum perf_cmd cmd,
352 struct perf_ctx *perf = peer->perf;
356 dev_dbg(&perf->ntb->dev, "CMD send: %d 0x%llx\n", cmd, data);
359 * Perform predefined number of attempts before give up. Message
360 * registers are free of race-condition problem when accessed
361 * from different ports, so we don't need splitting registers
362 * by global device index. We also won't have local locking,
363 * since the method is used from service work only.
365 outbits = ntb_msg_outbits(perf->ntb);
366 for (try = 0; try < MSG_TRIES; try++) {
367 if (!perf_link_is_up(peer))
370 ret = ntb_msg_clear_sts(perf->ntb, outbits);
374 ntb_peer_msg_write(perf->ntb, peer->pidx, PERF_MSG_LDATA,
375 lower_32_bits(data));
377 if (ntb_msg_read_sts(perf->ntb) & outbits) {
378 usleep_range(MSG_UDELAY_LOW, MSG_UDELAY_HIGH);
382 ntb_peer_msg_write(perf->ntb, peer->pidx, PERF_MSG_HDATA,
383 upper_32_bits(data));
386 /* This call shall trigger peer message event */
387 ntb_peer_msg_write(perf->ntb, peer->pidx, PERF_MSG_CMD, cmd);
392 return try < MSG_TRIES ? 0 : -EAGAIN;
395 static int perf_msg_cmd_recv(struct perf_ctx *perf, int *pidx,
396 enum perf_cmd *cmd, u64 *data)
401 inbits = ntb_msg_inbits(perf->ntb);
403 if (hweight64(ntb_msg_read_sts(perf->ntb) & inbits) < 3)
406 val = ntb_msg_read(perf->ntb, pidx, PERF_MSG_CMD);
409 val = ntb_msg_read(perf->ntb, pidx, PERF_MSG_LDATA);
412 val = ntb_msg_read(perf->ntb, pidx, PERF_MSG_HDATA);
413 *data |= (u64)val << 32;
415 /* Next command can be retrieved from now */
416 ntb_msg_clear_sts(perf->ntb, inbits);
418 dev_dbg(&perf->ntb->dev, "CMD recv: %d 0x%llx\n", *cmd, *data);
423 static int perf_cmd_send(struct perf_peer *peer, enum perf_cmd cmd, u64 data)
425 struct perf_ctx *perf = peer->perf;
427 if (cmd == PERF_CMD_SSIZE || cmd == PERF_CMD_SXLAT)
428 return perf->cmd_send(peer, cmd, data);
430 dev_err(&perf->ntb->dev, "Send invalid command\n");
434 static int perf_cmd_exec(struct perf_peer *peer, enum perf_cmd cmd)
444 dev_err(&peer->perf->ntb->dev, "Exec invalid command\n");
448 /* No need of memory barrier, since bit ops have invernal lock */
449 set_bit(cmd, &peer->sts);
451 dev_dbg(&peer->perf->ntb->dev, "CMD exec: %d\n", cmd);
453 (void)queue_work(system_highpri_wq, &peer->service);
458 static int perf_cmd_recv(struct perf_ctx *perf)
460 struct perf_peer *peer;
464 while (!(ret = perf->cmd_recv(perf, &pidx, &cmd, &data))) {
465 peer = &perf->peers[pidx];
469 peer->inbuf_size = data;
470 return perf_cmd_exec(peer, PERF_CMD_RSIZE);
472 peer->outbuf_xlat = data;
473 return perf_cmd_exec(peer, PERF_CMD_RXLAT);
475 dev_err(&perf->ntb->dev, "Recv invalid command\n");
480 /* Return 0 if no data left to process, otherwise an error */
481 return ret == -ENODATA ? 0 : ret;
484 static void perf_link_event(void *ctx)
486 struct perf_ctx *perf = ctx;
487 struct perf_peer *peer;
491 for (pidx = 0; pidx < perf->pcnt; pidx++) {
492 peer = &perf->peers[pidx];
494 lnk_up = perf_link_is_up(peer);
497 !test_and_set_bit(PERF_STS_LNKUP, &peer->sts)) {
498 perf_cmd_exec(peer, PERF_CMD_SSIZE);
499 } else if (!lnk_up &&
500 test_and_clear_bit(PERF_STS_LNKUP, &peer->sts)) {
501 perf_cmd_exec(peer, PERF_CMD_CLEAR);
506 static void perf_db_event(void *ctx, int vec)
508 struct perf_ctx *perf = ctx;
510 dev_dbg(&perf->ntb->dev, "DB vec %d mask %#llx bits %#llx\n", vec,
511 ntb_db_vector_mask(perf->ntb, vec), ntb_db_read(perf->ntb));
513 /* Just receive all available commands */
514 (void)perf_cmd_recv(perf);
517 static void perf_msg_event(void *ctx)
519 struct perf_ctx *perf = ctx;
521 dev_dbg(&perf->ntb->dev, "Msg status bits %#llx\n",
522 ntb_msg_read_sts(perf->ntb));
524 /* Messages are only sent one-by-one */
525 (void)perf_cmd_recv(perf);
528 static const struct ntb_ctx_ops perf_ops = {
529 .link_event = perf_link_event,
530 .db_event = perf_db_event,
531 .msg_event = perf_msg_event
534 static void perf_free_outbuf(struct perf_peer *peer)
536 (void)ntb_peer_mw_clear_trans(peer->perf->ntb, peer->pidx, peer->gidx);
539 static int perf_setup_outbuf(struct perf_peer *peer)
541 struct perf_ctx *perf = peer->perf;
544 /* Outbuf size can be unaligned due to custom max_mw_size */
545 ret = ntb_peer_mw_set_trans(perf->ntb, peer->pidx, peer->gidx,
546 peer->outbuf_xlat, peer->outbuf_size);
548 dev_err(&perf->ntb->dev, "Failed to set outbuf translation\n");
552 /* Initialization is finally done */
553 set_bit(PERF_STS_DONE, &peer->sts);
554 complete_all(&peer->init_comp);
559 static void perf_free_inbuf(struct perf_peer *peer)
564 (void)ntb_mw_clear_trans(peer->perf->ntb, peer->pidx, peer->gidx);
565 dma_free_coherent(&peer->perf->ntb->pdev->dev, peer->inbuf_size,
566 peer->inbuf, peer->inbuf_xlat);
570 static int perf_setup_inbuf(struct perf_peer *peer)
572 resource_size_t xlat_align, size_align, size_max;
573 struct perf_ctx *perf = peer->perf;
576 /* Get inbound MW parameters */
577 ret = ntb_mw_get_align(perf->ntb, peer->pidx, perf->gidx,
578 &xlat_align, &size_align, &size_max);
580 dev_err(&perf->ntb->dev, "Couldn't get inbuf restrictions\n");
584 if (peer->inbuf_size > size_max) {
585 dev_err(&perf->ntb->dev, "Too big inbuf size %pa > %pa\n",
586 &peer->inbuf_size, &size_max);
590 peer->inbuf_size = round_up(peer->inbuf_size, size_align);
592 perf_free_inbuf(peer);
594 peer->inbuf = dma_alloc_coherent(&perf->ntb->pdev->dev,
595 peer->inbuf_size, &peer->inbuf_xlat,
598 dev_err(&perf->ntb->dev, "Failed to alloc inbuf of %pa\n",
602 if (!IS_ALIGNED(peer->inbuf_xlat, xlat_align)) {
604 dev_err(&perf->ntb->dev, "Unaligned inbuf allocated\n");
608 ret = ntb_mw_set_trans(perf->ntb, peer->pidx, peer->gidx,
609 peer->inbuf_xlat, peer->inbuf_size);
611 dev_err(&perf->ntb->dev, "Failed to set inbuf translation\n");
616 * We submit inbuf xlat transmission cmd for execution here to follow
617 * the code architecture, even though this method is called from service
618 * work itself so the command will be executed right after it returns.
620 (void)perf_cmd_exec(peer, PERF_CMD_SXLAT);
625 perf_free_inbuf(peer);
630 static void perf_service_work(struct work_struct *work)
632 struct perf_peer *peer = to_peer_service(work);
634 if (test_and_clear_bit(PERF_CMD_SSIZE, &peer->sts))
635 perf_cmd_send(peer, PERF_CMD_SSIZE, peer->outbuf_size);
637 if (test_and_clear_bit(PERF_CMD_RSIZE, &peer->sts))
638 perf_setup_inbuf(peer);
640 if (test_and_clear_bit(PERF_CMD_SXLAT, &peer->sts))
641 perf_cmd_send(peer, PERF_CMD_SXLAT, peer->inbuf_xlat);
643 if (test_and_clear_bit(PERF_CMD_RXLAT, &peer->sts))
644 perf_setup_outbuf(peer);
646 if (test_and_clear_bit(PERF_CMD_CLEAR, &peer->sts)) {
647 init_completion(&peer->init_comp);
648 clear_bit(PERF_STS_DONE, &peer->sts);
649 if (test_bit(0, &peer->perf->busy_flag) &&
650 peer == peer->perf->test_peer) {
651 dev_warn(&peer->perf->ntb->dev,
652 "Freeing while test on-fly\n");
653 perf_terminate_test(peer->perf);
655 perf_free_outbuf(peer);
656 perf_free_inbuf(peer);
660 static int perf_init_service(struct perf_ctx *perf)
664 if (ntb_peer_mw_count(perf->ntb) < perf->pcnt) {
665 dev_err(&perf->ntb->dev, "Not enough memory windows\n");
669 if (ntb_msg_count(perf->ntb) >= PERF_MSG_CNT) {
670 perf->cmd_send = perf_msg_cmd_send;
671 perf->cmd_recv = perf_msg_cmd_recv;
673 dev_dbg(&perf->ntb->dev, "Message service initialized\n");
678 dev_dbg(&perf->ntb->dev, "Message service unsupported\n");
680 mask = GENMASK_ULL(perf->pcnt, 0);
681 if (ntb_spad_count(perf->ntb) >= PERF_SPAD_CNT(perf->pcnt) &&
682 (ntb_db_valid_mask(perf->ntb) & mask) == mask) {
683 perf->cmd_send = perf_spad_cmd_send;
684 perf->cmd_recv = perf_spad_cmd_recv;
686 dev_dbg(&perf->ntb->dev, "Scratchpad service initialized\n");
691 dev_dbg(&perf->ntb->dev, "Scratchpad service unsupported\n");
693 dev_err(&perf->ntb->dev, "Command services unsupported\n");
698 static int perf_enable_service(struct perf_ctx *perf)
703 mask = ntb_db_valid_mask(perf->ntb);
704 (void)ntb_db_set_mask(perf->ntb, mask);
706 ret = ntb_set_ctx(perf->ntb, perf, &perf_ops);
710 if (perf->cmd_send == perf_msg_cmd_send) {
713 inbits = ntb_msg_inbits(perf->ntb);
714 outbits = ntb_msg_outbits(perf->ntb);
715 (void)ntb_msg_set_mask(perf->ntb, inbits | outbits);
717 incmd_bit = BIT_ULL(__ffs64(inbits));
718 ret = ntb_msg_clear_mask(perf->ntb, incmd_bit);
720 dev_dbg(&perf->ntb->dev, "MSG sts unmasked %#llx\n", incmd_bit);
722 scnt = ntb_spad_count(perf->ntb);
723 for (sidx = 0; sidx < scnt; sidx++)
724 ntb_spad_write(perf->ntb, sidx, PERF_CMD_INVAL);
725 incmd_bit = PERF_SPAD_NOTIFY(perf->gidx);
726 ret = ntb_db_clear_mask(perf->ntb, incmd_bit);
728 dev_dbg(&perf->ntb->dev, "DB bits unmasked %#llx\n", incmd_bit);
731 ntb_clear_ctx(perf->ntb);
735 ntb_link_enable(perf->ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
736 /* Might be not necessary */
737 ntb_link_event(perf->ntb);
742 static void perf_disable_service(struct perf_ctx *perf)
746 ntb_link_disable(perf->ntb);
748 if (perf->cmd_send == perf_msg_cmd_send) {
751 inbits = ntb_msg_inbits(perf->ntb);
752 (void)ntb_msg_set_mask(perf->ntb, inbits);
754 (void)ntb_db_set_mask(perf->ntb, PERF_SPAD_NOTIFY(perf->gidx));
757 ntb_clear_ctx(perf->ntb);
759 for (pidx = 0; pidx < perf->pcnt; pidx++)
760 perf_cmd_exec(&perf->peers[pidx], PERF_CMD_CLEAR);
762 for (pidx = 0; pidx < perf->pcnt; pidx++)
763 flush_work(&perf->peers[pidx].service);
766 /*==============================================================================
767 * Performance measuring work-thread
768 *==============================================================================
771 static void perf_dma_copy_callback(void *data)
773 struct perf_thread *pthr = data;
775 atomic_dec(&pthr->dma_sync);
776 wake_up(&pthr->dma_wait);
779 static int perf_copy_chunk(struct perf_thread *pthr,
780 void __iomem *dst, void *src, size_t len)
782 struct dma_async_tx_descriptor *tx;
783 struct dmaengine_unmap_data *unmap;
784 struct device *dma_dev;
785 int try = 0, ret = 0;
788 memcpy_toio(dst, src, len);
789 goto ret_check_tsync;
792 dma_dev = pthr->dma_chan->device->dev;
794 if (!is_dma_copy_aligned(pthr->dma_chan->device, offset_in_page(src),
795 offset_in_page(dst), len))
798 unmap = dmaengine_get_unmap_data(dma_dev, 2, GFP_NOWAIT);
803 unmap->addr[0] = dma_map_page(dma_dev, virt_to_page(src),
804 offset_in_page(src), len, DMA_TO_DEVICE);
805 if (dma_mapping_error(dma_dev, unmap->addr[0])) {
807 goto err_free_resource;
811 unmap->addr[1] = dma_map_page(dma_dev, virt_to_page(dst),
812 offset_in_page(dst), len, DMA_FROM_DEVICE);
813 if (dma_mapping_error(dma_dev, unmap->addr[1])) {
815 goto err_free_resource;
820 tx = dmaengine_prep_dma_memcpy(pthr->dma_chan, unmap->addr[1],
821 unmap->addr[0], len, DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
824 } while (!tx && (try++ < DMA_TRIES));
828 goto err_free_resource;
831 tx->callback = perf_dma_copy_callback;
832 tx->callback_param = pthr;
833 dma_set_unmap(tx, unmap);
835 ret = dma_submit_error(dmaengine_submit(tx));
837 dmaengine_unmap_put(unmap);
838 goto err_free_resource;
841 dmaengine_unmap_put(unmap);
843 atomic_inc(&pthr->dma_sync);
844 dma_async_issue_pending(pthr->dma_chan);
847 return likely(atomic_read(&pthr->perf->tsync) > 0) ? 0 : -EINTR;
850 dmaengine_unmap_put(unmap);
855 static bool perf_dma_filter(struct dma_chan *chan, void *data)
857 struct perf_ctx *perf = data;
860 node = dev_to_node(&perf->ntb->dev);
862 return node == NUMA_NO_NODE || node == dev_to_node(chan->device->dev);
865 static int perf_init_test(struct perf_thread *pthr)
867 struct perf_ctx *perf = pthr->perf;
868 dma_cap_mask_t dma_mask;
870 pthr->src = kmalloc_node(perf->test_peer->outbuf_size, GFP_KERNEL,
871 dev_to_node(&perf->ntb->dev));
875 get_random_bytes(pthr->src, perf->test_peer->outbuf_size);
880 dma_cap_zero(dma_mask);
881 dma_cap_set(DMA_MEMCPY, dma_mask);
882 pthr->dma_chan = dma_request_channel(dma_mask, perf_dma_filter, perf);
883 if (!pthr->dma_chan) {
884 dev_err(&perf->ntb->dev, "%d: Failed to get DMA channel\n",
886 atomic_dec(&perf->tsync);
887 wake_up(&perf->twait);
892 atomic_set(&pthr->dma_sync, 0);
897 static int perf_run_test(struct perf_thread *pthr)
899 struct perf_peer *peer = pthr->perf->test_peer;
900 struct perf_ctx *perf = pthr->perf;
901 void __iomem *flt_dst, *bnd_dst;
902 u64 total_size, chunk_size;
906 total_size = 1ULL << total_order;
907 chunk_size = 1ULL << chunk_order;
908 chunk_size = min_t(u64, peer->outbuf_size, chunk_size);
911 bnd_dst = peer->outbuf + peer->outbuf_size;
912 flt_dst = peer->outbuf;
914 pthr->duration = ktime_get();
916 /* Copied field is cleared on test launch stage */
917 while (pthr->copied < total_size) {
918 ret = perf_copy_chunk(pthr, flt_dst, flt_src, chunk_size);
920 dev_err(&perf->ntb->dev, "%d: Got error %d on test\n",
925 pthr->copied += chunk_size;
927 flt_dst += chunk_size;
928 flt_src += chunk_size;
929 if (flt_dst >= bnd_dst || flt_dst < peer->outbuf) {
930 flt_dst = peer->outbuf;
934 /* Give up CPU to give a chance for other threads to use it */
941 static int perf_sync_test(struct perf_thread *pthr)
943 struct perf_ctx *perf = pthr->perf;
948 wait_event(pthr->dma_wait,
949 (atomic_read(&pthr->dma_sync) == 0 ||
950 atomic_read(&perf->tsync) < 0));
952 if (atomic_read(&perf->tsync) < 0)
956 pthr->duration = ktime_sub(ktime_get(), pthr->duration);
958 dev_dbg(&perf->ntb->dev, "%d: copied %llu bytes\n",
959 pthr->tidx, pthr->copied);
961 dev_dbg(&perf->ntb->dev, "%d: lasted %llu usecs\n",
962 pthr->tidx, ktime_to_us(pthr->duration));
964 dev_dbg(&perf->ntb->dev, "%d: %llu MBytes/s\n", pthr->tidx,
965 div64_u64(pthr->copied, ktime_to_us(pthr->duration)));
970 static void perf_clear_test(struct perf_thread *pthr)
972 struct perf_ctx *perf = pthr->perf;
978 * If test finished without errors, termination isn't needed.
979 * We call it anyway just to be sure of the transfers completion.
981 (void)dmaengine_terminate_sync(pthr->dma_chan);
983 dma_release_channel(pthr->dma_chan);
986 atomic_dec(&perf->tsync);
987 wake_up(&perf->twait);
991 static void perf_thread_work(struct work_struct *work)
993 struct perf_thread *pthr = to_thread_work(work);
997 * Perform stages in compliance with use_dma flag value.
998 * Test status is changed only if error happened, otherwise
999 * status -ENODATA is kept while test is on-fly. Results
1000 * synchronization is performed only if test fininshed
1001 * without an error or interruption.
1003 ret = perf_init_test(pthr);
1009 ret = perf_run_test(pthr);
1012 goto err_clear_test;
1015 pthr->status = perf_sync_test(pthr);
1018 perf_clear_test(pthr);
1021 static int perf_set_tcnt(struct perf_ctx *perf, u8 tcnt)
1023 if (tcnt == 0 || tcnt > MAX_THREADS_CNT)
1026 if (test_and_set_bit_lock(0, &perf->busy_flag))
1031 clear_bit_unlock(0, &perf->busy_flag);
1036 static void perf_terminate_test(struct perf_ctx *perf)
1040 atomic_set(&perf->tsync, -1);
1041 wake_up(&perf->twait);
1043 for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
1044 wake_up(&perf->threads[tidx].dma_wait);
1045 cancel_work_sync(&perf->threads[tidx].work);
1049 static int perf_submit_test(struct perf_peer *peer)
1051 struct perf_ctx *perf = peer->perf;
1052 struct perf_thread *pthr;
1055 ret = wait_for_completion_interruptible(&peer->init_comp);
1059 if (test_and_set_bit_lock(0, &perf->busy_flag))
1062 perf->test_peer = peer;
1063 atomic_set(&perf->tsync, perf->tcnt);
1065 for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
1066 pthr = &perf->threads[tidx];
1068 pthr->status = -ENODATA;
1070 pthr->duration = ktime_set(0, 0);
1071 if (tidx < perf->tcnt)
1072 (void)queue_work(perf_wq, &pthr->work);
1075 ret = wait_event_interruptible(perf->twait,
1076 atomic_read(&perf->tsync) <= 0);
1077 if (ret == -ERESTARTSYS) {
1078 perf_terminate_test(perf);
1082 clear_bit_unlock(0, &perf->busy_flag);
1087 static int perf_read_stats(struct perf_ctx *perf, char *buf,
1088 size_t size, ssize_t *pos)
1090 struct perf_thread *pthr;
1093 if (test_and_set_bit_lock(0, &perf->busy_flag))
1096 (*pos) += scnprintf(buf + *pos, size - *pos,
1097 " Peer %d test statistics:\n", perf->test_peer->pidx);
1099 for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
1100 pthr = &perf->threads[tidx];
1102 if (pthr->status == -ENODATA)
1106 (*pos) += scnprintf(buf + *pos, size - *pos,
1107 "%d: error status %d\n", tidx, pthr->status);
1111 (*pos) += scnprintf(buf + *pos, size - *pos,
1112 "%d: copied %llu bytes in %llu usecs, %llu MBytes/s\n",
1113 tidx, pthr->copied, ktime_to_us(pthr->duration),
1114 div64_u64(pthr->copied, ktime_to_us(pthr->duration)));
1117 clear_bit_unlock(0, &perf->busy_flag);
1122 static void perf_init_threads(struct perf_ctx *perf)
1124 struct perf_thread *pthr;
1127 perf->tcnt = DEF_THREADS_CNT;
1128 perf->test_peer = &perf->peers[0];
1129 init_waitqueue_head(&perf->twait);
1131 for (tidx = 0; tidx < MAX_THREADS_CNT; tidx++) {
1132 pthr = &perf->threads[tidx];
1136 pthr->status = -ENODATA;
1137 init_waitqueue_head(&pthr->dma_wait);
1138 INIT_WORK(&pthr->work, perf_thread_work);
1142 static void perf_clear_threads(struct perf_ctx *perf)
1144 perf_terminate_test(perf);
1147 /*==============================================================================
1149 *==============================================================================
1152 static ssize_t perf_dbgfs_read_info(struct file *filep, char __user *ubuf,
1153 size_t size, loff_t *offp)
1155 struct perf_ctx *perf = filep->private_data;
1156 struct perf_peer *peer;
1162 buf_size = min_t(size_t, size, 0x1000U);
1164 buf = kmalloc(buf_size, GFP_KERNEL);
1168 pos += scnprintf(buf + pos, buf_size - pos,
1169 " Performance measuring tool info:\n\n");
1171 pos += scnprintf(buf + pos, buf_size - pos,
1172 "Local port %d, Global index %d\n", ntb_port_number(perf->ntb),
1174 pos += scnprintf(buf + pos, buf_size - pos, "Test status: ");
1175 if (test_bit(0, &perf->busy_flag)) {
1176 pos += scnprintf(buf + pos, buf_size - pos,
1177 "on-fly with port %d (%d)\n",
1178 ntb_peer_port_number(perf->ntb, perf->test_peer->pidx),
1179 perf->test_peer->pidx);
1181 pos += scnprintf(buf + pos, buf_size - pos, "idle\n");
1184 for (pidx = 0; pidx < perf->pcnt; pidx++) {
1185 peer = &perf->peers[pidx];
1187 pos += scnprintf(buf + pos, buf_size - pos,
1188 "Port %d (%d), Global index %d:\n",
1189 ntb_peer_port_number(perf->ntb, peer->pidx), peer->pidx,
1192 pos += scnprintf(buf + pos, buf_size - pos,
1193 "\tLink status: %s\n",
1194 test_bit(PERF_STS_LNKUP, &peer->sts) ? "up" : "down");
1196 pos += scnprintf(buf + pos, buf_size - pos,
1197 "\tOut buffer addr 0x%pK\n", peer->outbuf);
1199 pos += scnprintf(buf + pos, buf_size - pos,
1200 "\tOut buffer size %pa\n", &peer->outbuf_size);
1202 pos += scnprintf(buf + pos, buf_size - pos,
1203 "\tOut buffer xlat 0x%016llx[p]\n", peer->outbuf_xlat);
1206 pos += scnprintf(buf + pos, buf_size - pos,
1207 "\tIn buffer addr: unallocated\n");
1211 pos += scnprintf(buf + pos, buf_size - pos,
1212 "\tIn buffer addr 0x%pK\n", peer->inbuf);
1214 pos += scnprintf(buf + pos, buf_size - pos,
1215 "\tIn buffer size %pa\n", &peer->inbuf_size);
1217 pos += scnprintf(buf + pos, buf_size - pos,
1218 "\tIn buffer xlat %pad[p]\n", &peer->inbuf_xlat);
1221 ret = simple_read_from_buffer(ubuf, size, offp, buf, pos);
1227 static const struct file_operations perf_dbgfs_info = {
1228 .open = simple_open,
1229 .read = perf_dbgfs_read_info
1232 static ssize_t perf_dbgfs_read_run(struct file *filep, char __user *ubuf,
1233 size_t size, loff_t *offp)
1235 struct perf_ctx *perf = filep->private_data;
1236 ssize_t ret, pos = 0;
1239 buf = kmalloc(PERF_BUF_LEN, GFP_KERNEL);
1243 ret = perf_read_stats(perf, buf, PERF_BUF_LEN, &pos);
1247 ret = simple_read_from_buffer(ubuf, size, offp, buf, pos);
1254 static ssize_t perf_dbgfs_write_run(struct file *filep, const char __user *ubuf,
1255 size_t size, loff_t *offp)
1257 struct perf_ctx *perf = filep->private_data;
1258 struct perf_peer *peer;
1261 ret = kstrtoint_from_user(ubuf, size, 0, &pidx);
1265 if (pidx < 0 || pidx >= perf->pcnt)
1268 peer = &perf->peers[pidx];
1270 ret = perf_submit_test(peer);
1277 static const struct file_operations perf_dbgfs_run = {
1278 .open = simple_open,
1279 .read = perf_dbgfs_read_run,
1280 .write = perf_dbgfs_write_run
1283 static ssize_t perf_dbgfs_read_tcnt(struct file *filep, char __user *ubuf,
1284 size_t size, loff_t *offp)
1286 struct perf_ctx *perf = filep->private_data;
1290 pos = scnprintf(buf, sizeof(buf), "%hhu\n", perf->tcnt);
1292 return simple_read_from_buffer(ubuf, size, offp, buf, pos);
1295 static ssize_t perf_dbgfs_write_tcnt(struct file *filep,
1296 const char __user *ubuf,
1297 size_t size, loff_t *offp)
1299 struct perf_ctx *perf = filep->private_data;
1303 ret = kstrtou8_from_user(ubuf, size, 0, &val);
1307 ret = perf_set_tcnt(perf, val);
1314 static const struct file_operations perf_dbgfs_tcnt = {
1315 .open = simple_open,
1316 .read = perf_dbgfs_read_tcnt,
1317 .write = perf_dbgfs_write_tcnt
1320 static void perf_setup_dbgfs(struct perf_ctx *perf)
1322 struct pci_dev *pdev = perf->ntb->pdev;
1324 perf->dbgfs_dir = debugfs_create_dir(pci_name(pdev), perf_dbgfs_topdir);
1325 if (!perf->dbgfs_dir) {
1326 dev_warn(&perf->ntb->dev, "DebugFS unsupported\n");
1330 debugfs_create_file("info", 0600, perf->dbgfs_dir, perf,
1333 debugfs_create_file("run", 0600, perf->dbgfs_dir, perf,
1336 debugfs_create_file("threads_count", 0600, perf->dbgfs_dir, perf,
1339 /* They are made read-only for test exec safety and integrity */
1340 debugfs_create_u8("chunk_order", 0500, perf->dbgfs_dir, &chunk_order);
1342 debugfs_create_u8("total_order", 0500, perf->dbgfs_dir, &total_order);
1344 debugfs_create_bool("use_dma", 0500, perf->dbgfs_dir, &use_dma);
1347 static void perf_clear_dbgfs(struct perf_ctx *perf)
1349 debugfs_remove_recursive(perf->dbgfs_dir);
1352 /*==============================================================================
1353 * Basic driver initialization
1354 *==============================================================================
1357 static struct perf_ctx *perf_create_data(struct ntb_dev *ntb)
1359 struct perf_ctx *perf;
1361 perf = devm_kzalloc(&ntb->dev, sizeof(*perf), GFP_KERNEL);
1363 return ERR_PTR(-ENOMEM);
1365 perf->pcnt = ntb_peer_port_count(ntb);
1366 perf->peers = devm_kcalloc(&ntb->dev, perf->pcnt, sizeof(*perf->peers),
1369 return ERR_PTR(-ENOMEM);
1376 static int perf_setup_peer_mw(struct perf_peer *peer)
1378 struct perf_ctx *perf = peer->perf;
1379 phys_addr_t phys_addr;
1382 /* Get outbound MW parameters and map it */
1383 ret = ntb_peer_mw_get_addr(perf->ntb, perf->gidx, &phys_addr,
1384 &peer->outbuf_size);
1388 peer->outbuf = devm_ioremap_wc(&perf->ntb->dev, phys_addr,
1393 if (max_mw_size && peer->outbuf_size > max_mw_size) {
1394 peer->outbuf_size = max_mw_size;
1395 dev_warn(&peer->perf->ntb->dev,
1396 "Peer %d outbuf reduced to %pa\n", peer->pidx,
1397 &peer->outbuf_size);
1403 static int perf_init_peers(struct perf_ctx *perf)
1405 struct perf_peer *peer;
1406 int pidx, lport, ret;
1408 lport = ntb_port_number(perf->ntb);
1410 for (pidx = 0; pidx < perf->pcnt; pidx++) {
1411 peer = &perf->peers[pidx];
1415 if (lport < ntb_peer_port_number(perf->ntb, pidx)) {
1416 if (perf->gidx == -1)
1418 peer->gidx = pidx + 1;
1422 INIT_WORK(&peer->service, perf_service_work);
1423 init_completion(&peer->init_comp);
1425 if (perf->gidx == -1)
1429 * Hardware with only two ports may not have unique port
1430 * numbers. In this case, the gidxs should all be zero.
1432 if (perf->pcnt == 1 && ntb_port_number(perf->ntb) == 0 &&
1433 ntb_peer_port_number(perf->ntb, 0) == 0) {
1435 perf->peers[0].gidx = 0;
1438 for (pidx = 0; pidx < perf->pcnt; pidx++) {
1439 ret = perf_setup_peer_mw(&perf->peers[pidx]);
1444 dev_dbg(&perf->ntb->dev, "Global port index %d\n", perf->gidx);
1449 static int perf_probe(struct ntb_client *client, struct ntb_dev *ntb)
1451 struct perf_ctx *perf;
1454 perf = perf_create_data(ntb);
1456 return PTR_ERR(perf);
1458 ret = perf_init_peers(perf);
1462 perf_init_threads(perf);
1464 ret = perf_init_service(perf);
1468 ret = perf_enable_service(perf);
1472 perf_setup_dbgfs(perf);
1477 static void perf_remove(struct ntb_client *client, struct ntb_dev *ntb)
1479 struct perf_ctx *perf = ntb->ctx;
1481 perf_clear_dbgfs(perf);
1483 perf_disable_service(perf);
1485 perf_clear_threads(perf);
1488 static struct ntb_client perf_client = {
1490 .probe = perf_probe,
1491 .remove = perf_remove
1495 static int __init perf_init(void)
1499 if (chunk_order > MAX_CHUNK_ORDER) {
1500 chunk_order = MAX_CHUNK_ORDER;
1501 pr_info("Chunk order reduced to %hhu\n", chunk_order);
1504 if (total_order < chunk_order) {
1505 total_order = chunk_order;
1506 pr_info("Total data order reduced to %hhu\n", total_order);
1509 perf_wq = alloc_workqueue("perf_wq", WQ_UNBOUND | WQ_SYSFS, 0);
1513 if (debugfs_initialized())
1514 perf_dbgfs_topdir = debugfs_create_dir(KBUILD_MODNAME, NULL);
1516 ret = ntb_register_client(&perf_client);
1518 debugfs_remove_recursive(perf_dbgfs_topdir);
1519 destroy_workqueue(perf_wq);
1524 module_init(perf_init);
1526 static void __exit perf_exit(void)
1528 ntb_unregister_client(&perf_client);
1529 debugfs_remove_recursive(perf_dbgfs_topdir);
1530 destroy_workqueue(perf_wq);
1532 module_exit(perf_exit);