xref: /openbmc/linux/drivers/ntb/test/ntb_perf.c (revision 781095f903f398148cd0b646d3984234a715f29e)
1 /*
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.
4  *
5  *   GPL LICENSE SUMMARY
6  *
7  *   Copyright(c) 2015 Intel Corporation. All rights reserved.
8  *
9  *   This program is free software; you can redistribute it and/or modify
10  *   it under the terms of version 2 of the GNU General Public License as
11  *   published by the Free Software Foundation.
12  *
13  *   BSD LICENSE
14  *
15  *   Copyright(c) 2015 Intel Corporation. All rights reserved.
16  *
17  *   Redistribution and use in source and binary forms, with or without
18  *   modification, are permitted provided that the following conditions
19  *   are met:
20  *
21  *     * Redistributions of source code must retain the above copyright
22  *       notice, this list of conditions and the following disclaimer.
23  *     * Redistributions in binary form must reproduce the above copy
24  *       notice, this list of conditions and the following disclaimer in
25  *       the documentation and/or other materials provided with the
26  *       distribution.
27  *     * Neither the name of Intel Corporation nor the names of its
28  *       contributors may be used to endorse or promote products derived
29  *       from this software without specific prior written permission.
30  *
31  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
32  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
33  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
34  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
35  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
36  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
37  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
38  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
39  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
40  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
41  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
42  *
43  *   PCIe NTB Perf Linux driver
44  */
45 
46 #include <linux/init.h>
47 #include <linux/kernel.h>
48 #include <linux/module.h>
49 #include <linux/kthread.h>
50 #include <linux/time.h>
51 #include <linux/timer.h>
52 #include <linux/dma-mapping.h>
53 #include <linux/pci.h>
54 #include <linux/slab.h>
55 #include <linux/spinlock.h>
56 #include <linux/debugfs.h>
57 #include <linux/dmaengine.h>
58 #include <linux/delay.h>
59 #include <linux/sizes.h>
60 #include <linux/ntb.h>
61 
62 #define DRIVER_NAME		"ntb_perf"
63 #define DRIVER_DESCRIPTION	"PCIe NTB Performance Measurement Tool"
64 
65 #define DRIVER_LICENSE		"Dual BSD/GPL"
66 #define DRIVER_VERSION		"1.0"
67 #define DRIVER_AUTHOR		"Dave Jiang <dave.jiang@intel.com>"
68 
69 #define PERF_LINK_DOWN_TIMEOUT	10
70 #define PERF_VERSION		0xffff0001
71 #define MAX_THREADS		32
72 #define MAX_TEST_SIZE		SZ_1M
73 #define MAX_SRCS		32
74 #define DMA_OUT_RESOURCE_TO	50
75 #define DMA_RETRIES		20
76 #define SZ_4G			(1ULL << 32)
77 #define MAX_SEG_ORDER		20 /* no larger than 1M for kmalloc buffer */
78 
79 MODULE_LICENSE(DRIVER_LICENSE);
80 MODULE_VERSION(DRIVER_VERSION);
81 MODULE_AUTHOR(DRIVER_AUTHOR);
82 MODULE_DESCRIPTION(DRIVER_DESCRIPTION);
83 
84 static struct dentry *perf_debugfs_dir;
85 
86 static unsigned int seg_order = 19; /* 512K */
87 module_param(seg_order, uint, 0644);
88 MODULE_PARM_DESC(seg_order, "size order [n^2] of buffer segment for testing");
89 
90 static unsigned int run_order = 32; /* 4G */
91 module_param(run_order, uint, 0644);
92 MODULE_PARM_DESC(run_order, "size order [n^2] of total data to transfer");
93 
94 static bool use_dma; /* default to 0 */
95 module_param(use_dma, bool, 0644);
96 MODULE_PARM_DESC(use_dma, "Using DMA engine to measure performance");
97 
98 struct perf_mw {
99 	phys_addr_t	phys_addr;
100 	resource_size_t	phys_size;
101 	resource_size_t	xlat_align;
102 	resource_size_t	xlat_align_size;
103 	void __iomem	*vbase;
104 	size_t		xlat_size;
105 	size_t		buf_size;
106 	void		*virt_addr;
107 	dma_addr_t	dma_addr;
108 };
109 
110 struct perf_ctx;
111 
112 struct pthr_ctx {
113 	struct task_struct	*thread;
114 	struct perf_ctx		*perf;
115 	atomic_t		dma_sync;
116 	struct dma_chan		*dma_chan;
117 	int			dma_prep_err;
118 	int			src_idx;
119 	void			*srcs[MAX_SRCS];
120 };
121 
122 struct perf_ctx {
123 	struct ntb_dev		*ntb;
124 	spinlock_t		db_lock;
125 	struct perf_mw		mw;
126 	bool			link_is_up;
127 	struct work_struct	link_cleanup;
128 	struct delayed_work	link_work;
129 	struct dentry		*debugfs_node_dir;
130 	struct dentry		*debugfs_run;
131 	struct dentry		*debugfs_threads;
132 	u8			perf_threads;
133 	bool			run;
134 	struct pthr_ctx		pthr_ctx[MAX_THREADS];
135 	atomic_t		tsync;
136 };
137 
138 enum {
139 	VERSION = 0,
140 	MW_SZ_HIGH,
141 	MW_SZ_LOW,
142 	SPAD_MSG,
143 	SPAD_ACK,
144 	MAX_SPAD
145 };
146 
147 static void perf_link_event(void *ctx)
148 {
149 	struct perf_ctx *perf = ctx;
150 
151 	if (ntb_link_is_up(perf->ntb, NULL, NULL) == 1)
152 		schedule_delayed_work(&perf->link_work, 2*HZ);
153 	else
154 		schedule_work(&perf->link_cleanup);
155 }
156 
157 static void perf_db_event(void *ctx, int vec)
158 {
159 	struct perf_ctx *perf = ctx;
160 	u64 db_bits, db_mask;
161 
162 	db_mask = ntb_db_vector_mask(perf->ntb, vec);
163 	db_bits = ntb_db_read(perf->ntb);
164 
165 	dev_dbg(&perf->ntb->dev, "doorbell vec %d mask %#llx bits %#llx\n",
166 		vec, db_mask, db_bits);
167 }
168 
169 static const struct ntb_ctx_ops perf_ops = {
170 	.link_event = perf_link_event,
171 	.db_event = perf_db_event,
172 };
173 
174 static void perf_copy_callback(void *data)
175 {
176 	struct pthr_ctx *pctx = data;
177 
178 	atomic_dec(&pctx->dma_sync);
179 }
180 
181 static ssize_t perf_copy(struct pthr_ctx *pctx, char *dst,
182 			 char *src, size_t size)
183 {
184 	struct perf_ctx *perf = pctx->perf;
185 	struct dma_async_tx_descriptor *txd;
186 	struct dma_chan *chan = pctx->dma_chan;
187 	struct dma_device *device;
188 	struct dmaengine_unmap_data *unmap;
189 	dma_cookie_t cookie;
190 	size_t src_off, dst_off;
191 	struct perf_mw *mw = &perf->mw;
192 	u64 vbase, dst_vaddr;
193 	dma_addr_t dst_phys;
194 	int retries = 0;
195 
196 	if (!use_dma) {
197 		memcpy_toio(dst, src, size);
198 		return size;
199 	}
200 
201 	if (!chan) {
202 		dev_err(&perf->ntb->dev, "DMA engine does not exist\n");
203 		return -EINVAL;
204 	}
205 
206 	device = chan->device;
207 	src_off = (size_t)src & ~PAGE_MASK;
208 	dst_off = (size_t)dst & ~PAGE_MASK;
209 
210 	if (!is_dma_copy_aligned(device, src_off, dst_off, size))
211 		return -ENODEV;
212 
213 	vbase = (u64)(u64 *)mw->vbase;
214 	dst_vaddr = (u64)(u64 *)dst;
215 	dst_phys = mw->phys_addr + (dst_vaddr - vbase);
216 
217 	unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
218 	if (!unmap)
219 		return -ENOMEM;
220 
221 	unmap->len = size;
222 	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(src),
223 				      src_off, size, DMA_TO_DEVICE);
224 	if (dma_mapping_error(device->dev, unmap->addr[0]))
225 		goto err_get_unmap;
226 
227 	unmap->to_cnt = 1;
228 
229 	do {
230 		txd = device->device_prep_dma_memcpy(chan, dst_phys,
231 						     unmap->addr[0],
232 						     size, DMA_PREP_INTERRUPT);
233 		if (!txd) {
234 			set_current_state(TASK_INTERRUPTIBLE);
235 			schedule_timeout(DMA_OUT_RESOURCE_TO);
236 		}
237 	} while (!txd && (++retries < DMA_RETRIES));
238 
239 	if (!txd) {
240 		pctx->dma_prep_err++;
241 		goto err_get_unmap;
242 	}
243 
244 	txd->callback = perf_copy_callback;
245 	txd->callback_param = pctx;
246 	dma_set_unmap(txd, unmap);
247 
248 	cookie = dmaengine_submit(txd);
249 	if (dma_submit_error(cookie))
250 		goto err_set_unmap;
251 
252 	atomic_inc(&pctx->dma_sync);
253 	dma_async_issue_pending(chan);
254 
255 	return size;
256 
257 err_set_unmap:
258 	dmaengine_unmap_put(unmap);
259 err_get_unmap:
260 	dmaengine_unmap_put(unmap);
261 	return 0;
262 }
263 
264 static int perf_move_data(struct pthr_ctx *pctx, char *dst, char *src,
265 			  u64 buf_size, u64 win_size, u64 total)
266 {
267 	int chunks, total_chunks, i;
268 	int copied_chunks = 0;
269 	u64 copied = 0, result;
270 	char *tmp = dst;
271 	u64 perf, diff_us;
272 	ktime_t kstart, kstop, kdiff;
273 
274 	chunks = div64_u64(win_size, buf_size);
275 	total_chunks = div64_u64(total, buf_size);
276 	kstart = ktime_get();
277 
278 	for (i = 0; i < total_chunks; i++) {
279 		result = perf_copy(pctx, tmp, src, buf_size);
280 		copied += result;
281 		copied_chunks++;
282 		if (copied_chunks == chunks) {
283 			tmp = dst;
284 			copied_chunks = 0;
285 		} else
286 			tmp += buf_size;
287 
288 		/* Probably should schedule every 4GB to prevent soft hang. */
289 		if (((copied % SZ_4G) == 0) && !use_dma) {
290 			set_current_state(TASK_INTERRUPTIBLE);
291 			schedule_timeout(1);
292 		}
293 	}
294 
295 	if (use_dma) {
296 		pr_info("%s: All DMA descriptors submitted\n", current->comm);
297 		while (atomic_read(&pctx->dma_sync) != 0)
298 			msleep(20);
299 	}
300 
301 	kstop = ktime_get();
302 	kdiff = ktime_sub(kstop, kstart);
303 	diff_us = ktime_to_us(kdiff);
304 
305 	pr_info("%s: copied %llu bytes\n", current->comm, copied);
306 
307 	pr_info("%s: lasted %llu usecs\n", current->comm, diff_us);
308 
309 	perf = div64_u64(copied, diff_us);
310 
311 	pr_info("%s: MBytes/s: %llu\n", current->comm, perf);
312 
313 	return 0;
314 }
315 
316 static bool perf_dma_filter_fn(struct dma_chan *chan, void *node)
317 {
318 	return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
319 }
320 
321 static int ntb_perf_thread(void *data)
322 {
323 	struct pthr_ctx *pctx = data;
324 	struct perf_ctx *perf = pctx->perf;
325 	struct pci_dev *pdev = perf->ntb->pdev;
326 	struct perf_mw *mw = &perf->mw;
327 	char *dst;
328 	u64 win_size, buf_size, total;
329 	void *src;
330 	int rc, node, i;
331 	struct dma_chan *dma_chan = NULL;
332 
333 	pr_info("kthread %s starting...\n", current->comm);
334 
335 	node = dev_to_node(&pdev->dev);
336 
337 	if (use_dma && !pctx->dma_chan) {
338 		dma_cap_mask_t dma_mask;
339 
340 		dma_cap_zero(dma_mask);
341 		dma_cap_set(DMA_MEMCPY, dma_mask);
342 		dma_chan = dma_request_channel(dma_mask, perf_dma_filter_fn,
343 					       (void *)(unsigned long)node);
344 		if (!dma_chan) {
345 			pr_warn("%s: cannot acquire DMA channel, quitting\n",
346 				current->comm);
347 			return -ENODEV;
348 		}
349 		pctx->dma_chan = dma_chan;
350 	}
351 
352 	for (i = 0; i < MAX_SRCS; i++) {
353 		pctx->srcs[i] = kmalloc_node(MAX_TEST_SIZE, GFP_KERNEL, node);
354 		if (!pctx->srcs[i]) {
355 			rc = -ENOMEM;
356 			goto err;
357 		}
358 	}
359 
360 	win_size = mw->phys_size;
361 	buf_size = 1ULL << seg_order;
362 	total = 1ULL << run_order;
363 
364 	if (buf_size > MAX_TEST_SIZE)
365 		buf_size = MAX_TEST_SIZE;
366 
367 	dst = (char *)mw->vbase;
368 
369 	atomic_inc(&perf->tsync);
370 	while (atomic_read(&perf->tsync) != perf->perf_threads)
371 		schedule();
372 
373 	src = pctx->srcs[pctx->src_idx];
374 	pctx->src_idx = (pctx->src_idx + 1) & (MAX_SRCS - 1);
375 
376 	rc = perf_move_data(pctx, dst, src, buf_size, win_size, total);
377 
378 	atomic_dec(&perf->tsync);
379 
380 	if (rc < 0) {
381 		pr_err("%s: failed\n", current->comm);
382 		rc = -ENXIO;
383 		goto err;
384 	}
385 
386 	for (i = 0; i < MAX_SRCS; i++) {
387 		kfree(pctx->srcs[i]);
388 		pctx->srcs[i] = NULL;
389 	}
390 
391 	return 0;
392 
393 err:
394 	for (i = 0; i < MAX_SRCS; i++) {
395 		kfree(pctx->srcs[i]);
396 		pctx->srcs[i] = NULL;
397 	}
398 
399 	if (dma_chan) {
400 		dma_release_channel(dma_chan);
401 		pctx->dma_chan = NULL;
402 	}
403 
404 	return rc;
405 }
406 
407 static void perf_free_mw(struct perf_ctx *perf)
408 {
409 	struct perf_mw *mw = &perf->mw;
410 	struct pci_dev *pdev = perf->ntb->pdev;
411 
412 	if (!mw->virt_addr)
413 		return;
414 
415 	ntb_mw_clear_trans(perf->ntb, 0);
416 	dma_free_coherent(&pdev->dev, mw->buf_size,
417 			  mw->virt_addr, mw->dma_addr);
418 	mw->xlat_size = 0;
419 	mw->buf_size = 0;
420 	mw->virt_addr = NULL;
421 }
422 
423 static int perf_set_mw(struct perf_ctx *perf, resource_size_t size)
424 {
425 	struct perf_mw *mw = &perf->mw;
426 	size_t xlat_size, buf_size;
427 
428 	if (!size)
429 		return -EINVAL;
430 
431 	xlat_size = round_up(size, mw->xlat_align_size);
432 	buf_size = round_up(size, mw->xlat_align);
433 
434 	if (mw->xlat_size == xlat_size)
435 		return 0;
436 
437 	if (mw->buf_size)
438 		perf_free_mw(perf);
439 
440 	mw->xlat_size = xlat_size;
441 	mw->buf_size = buf_size;
442 
443 	mw->virt_addr = dma_alloc_coherent(&perf->ntb->pdev->dev, buf_size,
444 					   &mw->dma_addr, GFP_KERNEL);
445 	if (!mw->virt_addr) {
446 		mw->xlat_size = 0;
447 		mw->buf_size = 0;
448 	}
449 
450 	return 0;
451 }
452 
453 static void perf_link_work(struct work_struct *work)
454 {
455 	struct perf_ctx *perf =
456 		container_of(work, struct perf_ctx, link_work.work);
457 	struct ntb_dev *ndev = perf->ntb;
458 	struct pci_dev *pdev = ndev->pdev;
459 	u32 val;
460 	u64 size;
461 	int rc;
462 
463 	dev_dbg(&perf->ntb->pdev->dev, "%s called\n", __func__);
464 
465 	size = perf->mw.phys_size;
466 	ntb_peer_spad_write(ndev, MW_SZ_HIGH, upper_32_bits(size));
467 	ntb_peer_spad_write(ndev, MW_SZ_LOW, lower_32_bits(size));
468 	ntb_peer_spad_write(ndev, VERSION, PERF_VERSION);
469 
470 	/* now read what peer wrote */
471 	val = ntb_spad_read(ndev, VERSION);
472 	if (val != PERF_VERSION) {
473 		dev_dbg(&pdev->dev, "Remote version = %#x\n", val);
474 		goto out;
475 	}
476 
477 	val = ntb_spad_read(ndev, MW_SZ_HIGH);
478 	size = (u64)val << 32;
479 
480 	val = ntb_spad_read(ndev, MW_SZ_LOW);
481 	size |= val;
482 
483 	dev_dbg(&pdev->dev, "Remote MW size = %#llx\n", size);
484 
485 	rc = perf_set_mw(perf, size);
486 	if (rc)
487 		goto out1;
488 
489 	perf->link_is_up = true;
490 
491 	return;
492 
493 out1:
494 	perf_free_mw(perf);
495 
496 out:
497 	if (ntb_link_is_up(ndev, NULL, NULL) == 1)
498 		schedule_delayed_work(&perf->link_work,
499 				      msecs_to_jiffies(PERF_LINK_DOWN_TIMEOUT));
500 }
501 
502 static void perf_link_cleanup(struct work_struct *work)
503 {
504 	struct perf_ctx *perf = container_of(work,
505 					     struct perf_ctx,
506 					     link_cleanup);
507 
508 	dev_dbg(&perf->ntb->pdev->dev, "%s called\n", __func__);
509 
510 	if (!perf->link_is_up)
511 		cancel_delayed_work_sync(&perf->link_work);
512 }
513 
514 static int perf_setup_mw(struct ntb_dev *ntb, struct perf_ctx *perf)
515 {
516 	struct perf_mw *mw;
517 	int rc;
518 
519 	mw = &perf->mw;
520 
521 	rc = ntb_mw_get_range(ntb, 0, &mw->phys_addr, &mw->phys_size,
522 			      &mw->xlat_align, &mw->xlat_align_size);
523 	if (rc)
524 		return rc;
525 
526 	perf->mw.vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
527 	if (!mw->vbase)
528 		return -ENOMEM;
529 
530 	return 0;
531 }
532 
533 static ssize_t debugfs_run_read(struct file *filp, char __user *ubuf,
534 				size_t count, loff_t *offp)
535 {
536 	struct perf_ctx *perf = filp->private_data;
537 	char *buf;
538 	ssize_t ret, out_offset;
539 
540 	if (!perf)
541 		return 0;
542 
543 	buf = kmalloc(64, GFP_KERNEL);
544 	out_offset = snprintf(buf, 64, "%d\n", perf->run);
545 	ret = simple_read_from_buffer(ubuf, count, offp, buf, out_offset);
546 	kfree(buf);
547 
548 	return ret;
549 }
550 
551 static ssize_t debugfs_run_write(struct file *filp, const char __user *ubuf,
552 				 size_t count, loff_t *offp)
553 {
554 	struct perf_ctx *perf = filp->private_data;
555 	int node, i;
556 
557 	if (!perf->link_is_up)
558 		return 0;
559 
560 	if (perf->perf_threads == 0)
561 		return 0;
562 
563 	if (atomic_read(&perf->tsync) == 0)
564 		perf->run = false;
565 
566 	if (perf->run) {
567 		/* lets stop the threads */
568 		perf->run = false;
569 		for (i = 0; i < MAX_THREADS; i++) {
570 			if (perf->pthr_ctx[i].thread) {
571 				kthread_stop(perf->pthr_ctx[i].thread);
572 				perf->pthr_ctx[i].thread = NULL;
573 			} else
574 				break;
575 		}
576 	} else {
577 		perf->run = true;
578 
579 		if (perf->perf_threads > MAX_THREADS) {
580 			perf->perf_threads = MAX_THREADS;
581 			pr_info("Reset total threads to: %u\n", MAX_THREADS);
582 		}
583 
584 		/* no greater than 1M */
585 		if (seg_order > MAX_SEG_ORDER) {
586 			seg_order = MAX_SEG_ORDER;
587 			pr_info("Fix seg_order to %u\n", seg_order);
588 		}
589 
590 		if (run_order < seg_order) {
591 			run_order = seg_order;
592 			pr_info("Fix run_order to %u\n", run_order);
593 		}
594 
595 		node = dev_to_node(&perf->ntb->pdev->dev);
596 		/* launch kernel thread */
597 		for (i = 0; i < perf->perf_threads; i++) {
598 			struct pthr_ctx *pctx;
599 
600 			pctx = &perf->pthr_ctx[i];
601 			atomic_set(&pctx->dma_sync, 0);
602 			pctx->perf = perf;
603 			pctx->thread =
604 				kthread_create_on_node(ntb_perf_thread,
605 						       (void *)pctx,
606 						       node, "ntb_perf %d", i);
607 			if (pctx->thread)
608 				wake_up_process(pctx->thread);
609 			else {
610 				perf->run = false;
611 				for (i = 0; i < MAX_THREADS; i++) {
612 					if (pctx->thread) {
613 						kthread_stop(pctx->thread);
614 						pctx->thread = NULL;
615 					}
616 				}
617 			}
618 
619 			if (perf->run == false)
620 				return -ENXIO;
621 		}
622 
623 	}
624 
625 	return count;
626 }
627 
628 static const struct file_operations ntb_perf_debugfs_run = {
629 	.owner = THIS_MODULE,
630 	.open = simple_open,
631 	.read = debugfs_run_read,
632 	.write = debugfs_run_write,
633 };
634 
635 static int perf_debugfs_setup(struct perf_ctx *perf)
636 {
637 	struct pci_dev *pdev = perf->ntb->pdev;
638 
639 	if (!debugfs_initialized())
640 		return -ENODEV;
641 
642 	if (!perf_debugfs_dir) {
643 		perf_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
644 		if (!perf_debugfs_dir)
645 			return -ENODEV;
646 	}
647 
648 	perf->debugfs_node_dir = debugfs_create_dir(pci_name(pdev),
649 						    perf_debugfs_dir);
650 	if (!perf->debugfs_node_dir)
651 		return -ENODEV;
652 
653 	perf->debugfs_run = debugfs_create_file("run", S_IRUSR | S_IWUSR,
654 						perf->debugfs_node_dir, perf,
655 						&ntb_perf_debugfs_run);
656 	if (!perf->debugfs_run)
657 		return -ENODEV;
658 
659 	perf->debugfs_threads = debugfs_create_u8("threads", S_IRUSR | S_IWUSR,
660 						  perf->debugfs_node_dir,
661 						  &perf->perf_threads);
662 	if (!perf->debugfs_threads)
663 		return -ENODEV;
664 
665 	return 0;
666 }
667 
668 static int perf_probe(struct ntb_client *client, struct ntb_dev *ntb)
669 {
670 	struct pci_dev *pdev = ntb->pdev;
671 	struct perf_ctx *perf;
672 	int node;
673 	int rc = 0;
674 
675 	node = dev_to_node(&pdev->dev);
676 
677 	perf = kzalloc_node(sizeof(*perf), GFP_KERNEL, node);
678 	if (!perf) {
679 		rc = -ENOMEM;
680 		goto err_perf;
681 	}
682 
683 	perf->ntb = ntb;
684 	perf->perf_threads = 1;
685 	atomic_set(&perf->tsync, 0);
686 	perf->run = false;
687 	spin_lock_init(&perf->db_lock);
688 	perf_setup_mw(ntb, perf);
689 	INIT_DELAYED_WORK(&perf->link_work, perf_link_work);
690 	INIT_WORK(&perf->link_cleanup, perf_link_cleanup);
691 
692 	rc = ntb_set_ctx(ntb, perf, &perf_ops);
693 	if (rc)
694 		goto err_ctx;
695 
696 	perf->link_is_up = false;
697 	ntb_link_enable(ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
698 	ntb_link_event(ntb);
699 
700 	rc = perf_debugfs_setup(perf);
701 	if (rc)
702 		goto err_ctx;
703 
704 	return 0;
705 
706 err_ctx:
707 	cancel_delayed_work_sync(&perf->link_work);
708 	cancel_work_sync(&perf->link_cleanup);
709 	kfree(perf);
710 err_perf:
711 	return rc;
712 }
713 
714 static void perf_remove(struct ntb_client *client, struct ntb_dev *ntb)
715 {
716 	struct perf_ctx *perf = ntb->ctx;
717 	int i;
718 
719 	dev_dbg(&perf->ntb->dev, "%s called\n", __func__);
720 
721 	cancel_delayed_work_sync(&perf->link_work);
722 	cancel_work_sync(&perf->link_cleanup);
723 
724 	ntb_clear_ctx(ntb);
725 	ntb_link_disable(ntb);
726 
727 	debugfs_remove_recursive(perf_debugfs_dir);
728 	perf_debugfs_dir = NULL;
729 
730 	if (use_dma) {
731 		for (i = 0; i < MAX_THREADS; i++) {
732 			struct pthr_ctx *pctx = &perf->pthr_ctx[i];
733 
734 			if (pctx->dma_chan)
735 				dma_release_channel(pctx->dma_chan);
736 		}
737 	}
738 
739 	kfree(perf);
740 }
741 
742 static struct ntb_client perf_client = {
743 	.ops = {
744 		.probe = perf_probe,
745 		.remove = perf_remove,
746 	},
747 };
748 module_ntb_client(perf_client);
749