xref: /openbmc/linux/drivers/ntb/test/ntb_perf.c (revision 90a53e44)
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 #include <linux/mutex.h>
62 
63 #define DRIVER_NAME		"ntb_perf"
64 #define DRIVER_DESCRIPTION	"PCIe NTB Performance Measurement Tool"
65 
66 #define DRIVER_LICENSE		"Dual BSD/GPL"
67 #define DRIVER_VERSION		"1.0"
68 #define DRIVER_AUTHOR		"Dave Jiang <dave.jiang@intel.com>"
69 
70 #define PERF_LINK_DOWN_TIMEOUT	10
71 #define PERF_VERSION		0xffff0001
72 #define MAX_THREADS		32
73 #define MAX_TEST_SIZE		SZ_1M
74 #define MAX_SRCS		32
75 #define DMA_OUT_RESOURCE_TO	msecs_to_jiffies(50)
76 #define DMA_RETRIES		20
77 #define SZ_4G			(1ULL << 32)
78 #define MAX_SEG_ORDER		20 /* no larger than 1M for kmalloc buffer */
79 #define PIDX			NTB_DEF_PEER_IDX
80 
81 MODULE_LICENSE(DRIVER_LICENSE);
82 MODULE_VERSION(DRIVER_VERSION);
83 MODULE_AUTHOR(DRIVER_AUTHOR);
84 MODULE_DESCRIPTION(DRIVER_DESCRIPTION);
85 
86 static struct dentry *perf_debugfs_dir;
87 
88 static unsigned long max_mw_size;
89 module_param(max_mw_size, ulong, 0644);
90 MODULE_PARM_DESC(max_mw_size, "Limit size of large memory windows");
91 
92 static unsigned int seg_order = 19; /* 512K */
93 module_param(seg_order, uint, 0644);
94 MODULE_PARM_DESC(seg_order, "size order [2^n] of buffer segment for testing");
95 
96 static unsigned int run_order = 32; /* 4G */
97 module_param(run_order, uint, 0644);
98 MODULE_PARM_DESC(run_order, "size order [2^n] of total data to transfer");
99 
100 static bool use_dma; /* default to 0 */
101 module_param(use_dma, bool, 0644);
102 MODULE_PARM_DESC(use_dma, "Using DMA engine to measure performance");
103 
104 static bool on_node = true; /* default to 1 */
105 module_param(on_node, bool, 0644);
106 MODULE_PARM_DESC(on_node, "Run threads only on NTB device node (default: true)");
107 
108 struct perf_mw {
109 	phys_addr_t	phys_addr;
110 	resource_size_t	phys_size;
111 	resource_size_t	xlat_align;
112 	resource_size_t	xlat_align_size;
113 	void __iomem	*vbase;
114 	size_t		xlat_size;
115 	size_t		buf_size;
116 	void		*virt_addr;
117 	dma_addr_t	dma_addr;
118 };
119 
120 struct perf_ctx;
121 
122 struct pthr_ctx {
123 	struct task_struct	*thread;
124 	struct perf_ctx		*perf;
125 	atomic_t		dma_sync;
126 	struct dma_chan		*dma_chan;
127 	int			dma_prep_err;
128 	int			src_idx;
129 	void			*srcs[MAX_SRCS];
130 	wait_queue_head_t       *wq;
131 	int			status;
132 	u64			copied;
133 	u64			diff_us;
134 };
135 
136 struct perf_ctx {
137 	struct ntb_dev		*ntb;
138 	spinlock_t		db_lock;
139 	struct perf_mw		mw;
140 	bool			link_is_up;
141 	struct delayed_work	link_work;
142 	wait_queue_head_t	link_wq;
143 	u8			perf_threads;
144 	/* mutex ensures only one set of threads run at once */
145 	struct mutex		run_mutex;
146 	struct pthr_ctx		pthr_ctx[MAX_THREADS];
147 	atomic_t		tsync;
148 	atomic_t                tdone;
149 };
150 
151 enum {
152 	VERSION = 0,
153 	MW_SZ_HIGH,
154 	MW_SZ_LOW,
155 	MAX_SPAD
156 };
157 
158 static void perf_link_event(void *ctx)
159 {
160 	struct perf_ctx *perf = ctx;
161 
162 	if (ntb_link_is_up(perf->ntb, NULL, NULL) == 1) {
163 		schedule_delayed_work(&perf->link_work, 2*HZ);
164 	} else {
165 		dev_dbg(&perf->ntb->pdev->dev, "link down\n");
166 
167 		if (!perf->link_is_up)
168 			cancel_delayed_work_sync(&perf->link_work);
169 
170 		perf->link_is_up = false;
171 	}
172 }
173 
174 static void perf_db_event(void *ctx, int vec)
175 {
176 	struct perf_ctx *perf = ctx;
177 	u64 db_bits, db_mask;
178 
179 	db_mask = ntb_db_vector_mask(perf->ntb, vec);
180 	db_bits = ntb_db_read(perf->ntb);
181 
182 	dev_dbg(&perf->ntb->dev, "doorbell vec %d mask %#llx bits %#llx\n",
183 		vec, db_mask, db_bits);
184 }
185 
186 static const struct ntb_ctx_ops perf_ops = {
187 	.link_event = perf_link_event,
188 	.db_event = perf_db_event,
189 };
190 
191 static void perf_copy_callback(void *data)
192 {
193 	struct pthr_ctx *pctx = data;
194 
195 	atomic_dec(&pctx->dma_sync);
196 }
197 
198 static ssize_t perf_copy(struct pthr_ctx *pctx, char __iomem *dst,
199 			 char *src, size_t size)
200 {
201 	struct perf_ctx *perf = pctx->perf;
202 	struct dma_async_tx_descriptor *txd;
203 	struct dma_chan *chan = pctx->dma_chan;
204 	struct dma_device *device;
205 	struct dmaengine_unmap_data *unmap;
206 	dma_cookie_t cookie;
207 	size_t src_off, dst_off;
208 	struct perf_mw *mw = &perf->mw;
209 	void __iomem *vbase;
210 	void __iomem *dst_vaddr;
211 	dma_addr_t dst_phys;
212 	int retries = 0;
213 
214 	if (!use_dma) {
215 		memcpy_toio(dst, src, size);
216 		return size;
217 	}
218 
219 	if (!chan) {
220 		dev_err(&perf->ntb->dev, "DMA engine does not exist\n");
221 		return -EINVAL;
222 	}
223 
224 	device = chan->device;
225 	src_off = (uintptr_t)src & ~PAGE_MASK;
226 	dst_off = (uintptr_t __force)dst & ~PAGE_MASK;
227 
228 	if (!is_dma_copy_aligned(device, src_off, dst_off, size))
229 		return -ENODEV;
230 
231 	vbase = mw->vbase;
232 	dst_vaddr = dst;
233 	dst_phys = mw->phys_addr + (dst_vaddr - vbase);
234 
235 	unmap = dmaengine_get_unmap_data(device->dev, 1, GFP_NOWAIT);
236 	if (!unmap)
237 		return -ENOMEM;
238 
239 	unmap->len = size;
240 	unmap->addr[0] = dma_map_page(device->dev, virt_to_page(src),
241 				      src_off, size, DMA_TO_DEVICE);
242 	if (dma_mapping_error(device->dev, unmap->addr[0]))
243 		goto err_get_unmap;
244 
245 	unmap->to_cnt = 1;
246 
247 	do {
248 		txd = device->device_prep_dma_memcpy(chan, dst_phys,
249 						     unmap->addr[0],
250 						     size, DMA_PREP_INTERRUPT);
251 		if (!txd) {
252 			set_current_state(TASK_INTERRUPTIBLE);
253 			schedule_timeout(DMA_OUT_RESOURCE_TO);
254 		}
255 	} while (!txd && (++retries < DMA_RETRIES));
256 
257 	if (!txd) {
258 		pctx->dma_prep_err++;
259 		goto err_get_unmap;
260 	}
261 
262 	txd->callback = perf_copy_callback;
263 	txd->callback_param = pctx;
264 	dma_set_unmap(txd, unmap);
265 
266 	cookie = dmaengine_submit(txd);
267 	if (dma_submit_error(cookie))
268 		goto err_set_unmap;
269 
270 	dmaengine_unmap_put(unmap);
271 
272 	atomic_inc(&pctx->dma_sync);
273 	dma_async_issue_pending(chan);
274 
275 	return size;
276 
277 err_set_unmap:
278 	dmaengine_unmap_put(unmap);
279 err_get_unmap:
280 	dmaengine_unmap_put(unmap);
281 	return 0;
282 }
283 
284 static int perf_move_data(struct pthr_ctx *pctx, char __iomem *dst, char *src,
285 			  u64 buf_size, u64 win_size, u64 total)
286 {
287 	int chunks, total_chunks, i;
288 	int copied_chunks = 0;
289 	u64 copied = 0, result;
290 	char __iomem *tmp = dst;
291 	u64 perf, diff_us;
292 	ktime_t kstart, kstop, kdiff;
293 	unsigned long last_sleep = jiffies;
294 
295 	chunks = div64_u64(win_size, buf_size);
296 	total_chunks = div64_u64(total, buf_size);
297 	kstart = ktime_get();
298 
299 	for (i = 0; i < total_chunks; i++) {
300 		result = perf_copy(pctx, tmp, src, buf_size);
301 		copied += result;
302 		copied_chunks++;
303 		if (copied_chunks == chunks) {
304 			tmp = dst;
305 			copied_chunks = 0;
306 		} else
307 			tmp += buf_size;
308 
309 		/* Probably should schedule every 5s to prevent soft hang. */
310 		if (unlikely((jiffies - last_sleep) > 5 * HZ)) {
311 			last_sleep = jiffies;
312 			set_current_state(TASK_INTERRUPTIBLE);
313 			schedule_timeout(1);
314 		}
315 
316 		if (unlikely(kthread_should_stop()))
317 			break;
318 	}
319 
320 	if (use_dma) {
321 		pr_debug("%s: All DMA descriptors submitted\n", current->comm);
322 		while (atomic_read(&pctx->dma_sync) != 0) {
323 			if (kthread_should_stop())
324 				break;
325 			msleep(20);
326 		}
327 	}
328 
329 	kstop = ktime_get();
330 	kdiff = ktime_sub(kstop, kstart);
331 	diff_us = ktime_to_us(kdiff);
332 
333 	pr_debug("%s: copied %llu bytes\n", current->comm, copied);
334 
335 	pr_debug("%s: lasted %llu usecs\n", current->comm, diff_us);
336 
337 	perf = div64_u64(copied, diff_us);
338 
339 	pr_debug("%s: MBytes/s: %llu\n", current->comm, perf);
340 
341 	pctx->copied = copied;
342 	pctx->diff_us = diff_us;
343 
344 	return 0;
345 }
346 
347 static bool perf_dma_filter_fn(struct dma_chan *chan, void *node)
348 {
349 	/* Is the channel required to be on the same node as the device? */
350 	if (!on_node)
351 		return true;
352 
353 	return dev_to_node(&chan->dev->device) == (int)(unsigned long)node;
354 }
355 
356 static int ntb_perf_thread(void *data)
357 {
358 	struct pthr_ctx *pctx = data;
359 	struct perf_ctx *perf = pctx->perf;
360 	struct pci_dev *pdev = perf->ntb->pdev;
361 	struct perf_mw *mw = &perf->mw;
362 	char __iomem *dst;
363 	u64 win_size, buf_size, total;
364 	void *src;
365 	int rc, node, i;
366 	struct dma_chan *dma_chan = NULL;
367 
368 	pr_debug("kthread %s starting...\n", current->comm);
369 
370 	node = on_node ? dev_to_node(&pdev->dev) : NUMA_NO_NODE;
371 
372 	if (use_dma && !pctx->dma_chan) {
373 		dma_cap_mask_t dma_mask;
374 
375 		dma_cap_zero(dma_mask);
376 		dma_cap_set(DMA_MEMCPY, dma_mask);
377 		dma_chan = dma_request_channel(dma_mask, perf_dma_filter_fn,
378 					       (void *)(unsigned long)node);
379 		if (!dma_chan) {
380 			pr_warn("%s: cannot acquire DMA channel, quitting\n",
381 				current->comm);
382 			return -ENODEV;
383 		}
384 		pctx->dma_chan = dma_chan;
385 	}
386 
387 	for (i = 0; i < MAX_SRCS; i++) {
388 		pctx->srcs[i] = kmalloc_node(MAX_TEST_SIZE, GFP_KERNEL, node);
389 		if (!pctx->srcs[i]) {
390 			rc = -ENOMEM;
391 			goto err;
392 		}
393 	}
394 
395 	win_size = mw->phys_size;
396 	buf_size = 1ULL << seg_order;
397 	total = 1ULL << run_order;
398 
399 	if (buf_size > MAX_TEST_SIZE)
400 		buf_size = MAX_TEST_SIZE;
401 
402 	dst = (char __iomem *)mw->vbase;
403 
404 	atomic_inc(&perf->tsync);
405 	while (atomic_read(&perf->tsync) != perf->perf_threads)
406 		schedule();
407 
408 	src = pctx->srcs[pctx->src_idx];
409 	pctx->src_idx = (pctx->src_idx + 1) & (MAX_SRCS - 1);
410 
411 	rc = perf_move_data(pctx, dst, src, buf_size, win_size, total);
412 
413 	atomic_dec(&perf->tsync);
414 
415 	if (rc < 0) {
416 		pr_err("%s: failed\n", current->comm);
417 		rc = -ENXIO;
418 		goto err;
419 	}
420 
421 	for (i = 0; i < MAX_SRCS; i++) {
422 		kfree(pctx->srcs[i]);
423 		pctx->srcs[i] = NULL;
424 	}
425 
426 	atomic_inc(&perf->tdone);
427 	wake_up(pctx->wq);
428 	rc = 0;
429 	goto done;
430 
431 err:
432 	for (i = 0; i < MAX_SRCS; i++) {
433 		kfree(pctx->srcs[i]);
434 		pctx->srcs[i] = NULL;
435 	}
436 
437 	if (dma_chan) {
438 		dma_release_channel(dma_chan);
439 		pctx->dma_chan = NULL;
440 	}
441 
442 done:
443 	/* Wait until we are told to stop */
444 	for (;;) {
445 		set_current_state(TASK_INTERRUPTIBLE);
446 		if (kthread_should_stop())
447 			break;
448 		schedule();
449 	}
450 	__set_current_state(TASK_RUNNING);
451 
452 	return rc;
453 }
454 
455 static void perf_free_mw(struct perf_ctx *perf)
456 {
457 	struct perf_mw *mw = &perf->mw;
458 	struct pci_dev *pdev = perf->ntb->pdev;
459 
460 	if (!mw->virt_addr)
461 		return;
462 
463 	ntb_mw_clear_trans(perf->ntb, PIDX, 0);
464 	dma_free_coherent(&pdev->dev, mw->buf_size,
465 			  mw->virt_addr, mw->dma_addr);
466 	mw->xlat_size = 0;
467 	mw->buf_size = 0;
468 	mw->virt_addr = NULL;
469 }
470 
471 static int perf_set_mw(struct perf_ctx *perf, resource_size_t size)
472 {
473 	struct perf_mw *mw = &perf->mw;
474 	size_t xlat_size, buf_size;
475 	int rc;
476 
477 	if (!size)
478 		return -EINVAL;
479 
480 	xlat_size = round_up(size, mw->xlat_align_size);
481 	buf_size = round_up(size, mw->xlat_align);
482 
483 	if (mw->xlat_size == xlat_size)
484 		return 0;
485 
486 	if (mw->buf_size)
487 		perf_free_mw(perf);
488 
489 	mw->xlat_size = xlat_size;
490 	mw->buf_size = buf_size;
491 
492 	mw->virt_addr = dma_alloc_coherent(&perf->ntb->pdev->dev, buf_size,
493 					   &mw->dma_addr, GFP_KERNEL);
494 	if (!mw->virt_addr) {
495 		mw->xlat_size = 0;
496 		mw->buf_size = 0;
497 	}
498 
499 	rc = ntb_mw_set_trans(perf->ntb, PIDX, 0, mw->dma_addr, mw->xlat_size);
500 	if (rc) {
501 		dev_err(&perf->ntb->dev, "Unable to set mw0 translation\n");
502 		perf_free_mw(perf);
503 		return -EIO;
504 	}
505 
506 	return 0;
507 }
508 
509 static void perf_link_work(struct work_struct *work)
510 {
511 	struct perf_ctx *perf =
512 		container_of(work, struct perf_ctx, link_work.work);
513 	struct ntb_dev *ndev = perf->ntb;
514 	struct pci_dev *pdev = ndev->pdev;
515 	u32 val;
516 	u64 size;
517 	int rc;
518 
519 	dev_dbg(&perf->ntb->pdev->dev, "%s called\n", __func__);
520 
521 	size = perf->mw.phys_size;
522 
523 	if (max_mw_size && size > max_mw_size)
524 		size = max_mw_size;
525 
526 	ntb_peer_spad_write(ndev, PIDX, MW_SZ_HIGH, upper_32_bits(size));
527 	ntb_peer_spad_write(ndev, PIDX, MW_SZ_LOW, lower_32_bits(size));
528 	ntb_peer_spad_write(ndev, PIDX, VERSION, PERF_VERSION);
529 
530 	/* now read what peer wrote */
531 	val = ntb_spad_read(ndev, VERSION);
532 	if (val != PERF_VERSION) {
533 		dev_dbg(&pdev->dev, "Remote version = %#x\n", val);
534 		goto out;
535 	}
536 
537 	val = ntb_spad_read(ndev, MW_SZ_HIGH);
538 	size = (u64)val << 32;
539 
540 	val = ntb_spad_read(ndev, MW_SZ_LOW);
541 	size |= val;
542 
543 	dev_dbg(&pdev->dev, "Remote MW size = %#llx\n", size);
544 
545 	rc = perf_set_mw(perf, size);
546 	if (rc)
547 		goto out1;
548 
549 	perf->link_is_up = true;
550 	wake_up(&perf->link_wq);
551 
552 	return;
553 
554 out1:
555 	perf_free_mw(perf);
556 
557 out:
558 	if (ntb_link_is_up(ndev, NULL, NULL) == 1)
559 		schedule_delayed_work(&perf->link_work,
560 				      msecs_to_jiffies(PERF_LINK_DOWN_TIMEOUT));
561 }
562 
563 static int perf_setup_mw(struct ntb_dev *ntb, struct perf_ctx *perf)
564 {
565 	struct perf_mw *mw;
566 	int rc;
567 
568 	mw = &perf->mw;
569 
570 	rc = ntb_mw_get_align(ntb, PIDX, 0, &mw->xlat_align,
571 			      &mw->xlat_align_size, NULL);
572 	if (rc)
573 		return rc;
574 
575 	rc = ntb_peer_mw_get_addr(ntb, 0, &mw->phys_addr, &mw->phys_size);
576 	if (rc)
577 		return rc;
578 
579 	perf->mw.vbase = ioremap_wc(mw->phys_addr, mw->phys_size);
580 	if (!mw->vbase)
581 		return -ENOMEM;
582 
583 	return 0;
584 }
585 
586 static ssize_t debugfs_run_read(struct file *filp, char __user *ubuf,
587 				size_t count, loff_t *offp)
588 {
589 	struct perf_ctx *perf = filp->private_data;
590 	char *buf;
591 	ssize_t ret, out_off = 0;
592 	struct pthr_ctx *pctx;
593 	int i;
594 	u64 rate;
595 
596 	if (!perf)
597 		return 0;
598 
599 	buf = kmalloc(1024, GFP_KERNEL);
600 	if (!buf)
601 		return -ENOMEM;
602 
603 	if (mutex_is_locked(&perf->run_mutex)) {
604 		out_off = scnprintf(buf, 64, "running\n");
605 		goto read_from_buf;
606 	}
607 
608 	for (i = 0; i < MAX_THREADS; i++) {
609 		pctx = &perf->pthr_ctx[i];
610 
611 		if (pctx->status == -ENODATA)
612 			break;
613 
614 		if (pctx->status) {
615 			out_off += scnprintf(buf + out_off, 1024 - out_off,
616 					    "%d: error %d\n", i,
617 					    pctx->status);
618 			continue;
619 		}
620 
621 		rate = div64_u64(pctx->copied, pctx->diff_us);
622 		out_off += scnprintf(buf + out_off, 1024 - out_off,
623 			"%d: copied %llu bytes in %llu usecs, %llu MBytes/s\n",
624 			i, pctx->copied, pctx->diff_us, rate);
625 	}
626 
627 read_from_buf:
628 	ret = simple_read_from_buffer(ubuf, count, offp, buf, out_off);
629 	kfree(buf);
630 
631 	return ret;
632 }
633 
634 static void threads_cleanup(struct perf_ctx *perf)
635 {
636 	struct pthr_ctx *pctx;
637 	int i;
638 
639 	for (i = 0; i < MAX_THREADS; i++) {
640 		pctx = &perf->pthr_ctx[i];
641 		if (pctx->thread) {
642 			pctx->status = kthread_stop(pctx->thread);
643 			pctx->thread = NULL;
644 		}
645 	}
646 }
647 
648 static void perf_clear_thread_status(struct perf_ctx *perf)
649 {
650 	int i;
651 
652 	for (i = 0; i < MAX_THREADS; i++)
653 		perf->pthr_ctx[i].status = -ENODATA;
654 }
655 
656 static ssize_t debugfs_run_write(struct file *filp, const char __user *ubuf,
657 				 size_t count, loff_t *offp)
658 {
659 	struct perf_ctx *perf = filp->private_data;
660 	int node, i;
661 	DECLARE_WAIT_QUEUE_HEAD(wq);
662 
663 	if (wait_event_interruptible(perf->link_wq, perf->link_is_up))
664 		return -ENOLINK;
665 
666 	if (perf->perf_threads == 0)
667 		return -EINVAL;
668 
669 	if (!mutex_trylock(&perf->run_mutex))
670 		return -EBUSY;
671 
672 	perf_clear_thread_status(perf);
673 
674 	if (perf->perf_threads > MAX_THREADS) {
675 		perf->perf_threads = MAX_THREADS;
676 		pr_info("Reset total threads to: %u\n", MAX_THREADS);
677 	}
678 
679 	/* no greater than 1M */
680 	if (seg_order > MAX_SEG_ORDER) {
681 		seg_order = MAX_SEG_ORDER;
682 		pr_info("Fix seg_order to %u\n", seg_order);
683 	}
684 
685 	if (run_order < seg_order) {
686 		run_order = seg_order;
687 		pr_info("Fix run_order to %u\n", run_order);
688 	}
689 
690 	node = on_node ? dev_to_node(&perf->ntb->pdev->dev)
691 		       : NUMA_NO_NODE;
692 	atomic_set(&perf->tdone, 0);
693 
694 	/* launch kernel thread */
695 	for (i = 0; i < perf->perf_threads; i++) {
696 		struct pthr_ctx *pctx;
697 
698 		pctx = &perf->pthr_ctx[i];
699 		atomic_set(&pctx->dma_sync, 0);
700 		pctx->perf = perf;
701 		pctx->wq = &wq;
702 		pctx->thread =
703 			kthread_create_on_node(ntb_perf_thread,
704 					       (void *)pctx,
705 					       node, "ntb_perf %d", i);
706 		if (IS_ERR(pctx->thread)) {
707 			pctx->thread = NULL;
708 			goto err;
709 		} else {
710 			wake_up_process(pctx->thread);
711 		}
712 	}
713 
714 	wait_event_interruptible(wq,
715 		atomic_read(&perf->tdone) == perf->perf_threads);
716 
717 	threads_cleanup(perf);
718 	mutex_unlock(&perf->run_mutex);
719 	return count;
720 
721 err:
722 	threads_cleanup(perf);
723 	mutex_unlock(&perf->run_mutex);
724 	return -ENXIO;
725 }
726 
727 static const struct file_operations ntb_perf_debugfs_run = {
728 	.owner = THIS_MODULE,
729 	.open = simple_open,
730 	.read = debugfs_run_read,
731 	.write = debugfs_run_write,
732 };
733 
734 static int perf_debugfs_setup(struct perf_ctx *perf)
735 {
736 	struct pci_dev *pdev = perf->ntb->pdev;
737 	struct dentry *debugfs_node_dir;
738 	struct dentry *debugfs_run;
739 	struct dentry *debugfs_threads;
740 	struct dentry *debugfs_seg_order;
741 	struct dentry *debugfs_run_order;
742 	struct dentry *debugfs_use_dma;
743 	struct dentry *debugfs_on_node;
744 
745 	if (!debugfs_initialized())
746 		return -ENODEV;
747 
748 	/* Assumpion: only one NTB device in the system */
749 	if (!perf_debugfs_dir) {
750 		perf_debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
751 		if (!perf_debugfs_dir)
752 			return -ENODEV;
753 	}
754 
755 	debugfs_node_dir = debugfs_create_dir(pci_name(pdev),
756 					      perf_debugfs_dir);
757 	if (!debugfs_node_dir)
758 		goto err;
759 
760 	debugfs_run = debugfs_create_file("run", S_IRUSR | S_IWUSR,
761 					  debugfs_node_dir, perf,
762 					  &ntb_perf_debugfs_run);
763 	if (!debugfs_run)
764 		goto err;
765 
766 	debugfs_threads = debugfs_create_u8("threads", S_IRUSR | S_IWUSR,
767 					    debugfs_node_dir,
768 					    &perf->perf_threads);
769 	if (!debugfs_threads)
770 		goto err;
771 
772 	debugfs_seg_order = debugfs_create_u32("seg_order", 0600,
773 					       debugfs_node_dir,
774 					       &seg_order);
775 	if (!debugfs_seg_order)
776 		goto err;
777 
778 	debugfs_run_order = debugfs_create_u32("run_order", 0600,
779 					       debugfs_node_dir,
780 					       &run_order);
781 	if (!debugfs_run_order)
782 		goto err;
783 
784 	debugfs_use_dma = debugfs_create_bool("use_dma", 0600,
785 					       debugfs_node_dir,
786 					       &use_dma);
787 	if (!debugfs_use_dma)
788 		goto err;
789 
790 	debugfs_on_node = debugfs_create_bool("on_node", 0600,
791 					      debugfs_node_dir,
792 					      &on_node);
793 	if (!debugfs_on_node)
794 		goto err;
795 
796 	return 0;
797 
798 err:
799 	debugfs_remove_recursive(perf_debugfs_dir);
800 	perf_debugfs_dir = NULL;
801 	return -ENODEV;
802 }
803 
804 static int perf_probe(struct ntb_client *client, struct ntb_dev *ntb)
805 {
806 	struct pci_dev *pdev = ntb->pdev;
807 	struct perf_ctx *perf;
808 	int node;
809 	int rc = 0;
810 
811 	if (ntb_spad_count(ntb) < MAX_SPAD) {
812 		dev_err(&ntb->dev, "Not enough scratch pad registers for %s",
813 			DRIVER_NAME);
814 		return -EIO;
815 	}
816 
817 	if (!ntb->ops->mw_set_trans) {
818 		dev_err(&ntb->dev, "Need inbound MW based NTB API\n");
819 		return -EINVAL;
820 	}
821 
822 	if (ntb_peer_port_count(ntb) != NTB_DEF_PEER_CNT)
823 		dev_warn(&ntb->dev, "Multi-port NTB devices unsupported\n");
824 
825 	node = on_node ? dev_to_node(&pdev->dev) : NUMA_NO_NODE;
826 	perf = kzalloc_node(sizeof(*perf), GFP_KERNEL, node);
827 	if (!perf) {
828 		rc = -ENOMEM;
829 		goto err_perf;
830 	}
831 
832 	perf->ntb = ntb;
833 	perf->perf_threads = 1;
834 	atomic_set(&perf->tsync, 0);
835 	mutex_init(&perf->run_mutex);
836 	spin_lock_init(&perf->db_lock);
837 	perf_setup_mw(ntb, perf);
838 	init_waitqueue_head(&perf->link_wq);
839 	INIT_DELAYED_WORK(&perf->link_work, perf_link_work);
840 
841 	rc = ntb_set_ctx(ntb, perf, &perf_ops);
842 	if (rc)
843 		goto err_ctx;
844 
845 	perf->link_is_up = false;
846 	ntb_link_enable(ntb, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
847 	ntb_link_event(ntb);
848 
849 	rc = perf_debugfs_setup(perf);
850 	if (rc)
851 		goto err_ctx;
852 
853 	perf_clear_thread_status(perf);
854 
855 	return 0;
856 
857 err_ctx:
858 	cancel_delayed_work_sync(&perf->link_work);
859 	kfree(perf);
860 err_perf:
861 	return rc;
862 }
863 
864 static void perf_remove(struct ntb_client *client, struct ntb_dev *ntb)
865 {
866 	struct perf_ctx *perf = ntb->ctx;
867 	int i;
868 
869 	dev_dbg(&perf->ntb->dev, "%s called\n", __func__);
870 
871 	mutex_lock(&perf->run_mutex);
872 
873 	cancel_delayed_work_sync(&perf->link_work);
874 
875 	ntb_clear_ctx(ntb);
876 	ntb_link_disable(ntb);
877 
878 	debugfs_remove_recursive(perf_debugfs_dir);
879 	perf_debugfs_dir = NULL;
880 
881 	if (use_dma) {
882 		for (i = 0; i < MAX_THREADS; i++) {
883 			struct pthr_ctx *pctx = &perf->pthr_ctx[i];
884 
885 			if (pctx->dma_chan)
886 				dma_release_channel(pctx->dma_chan);
887 		}
888 	}
889 
890 	kfree(perf);
891 }
892 
893 static struct ntb_client perf_client = {
894 	.ops = {
895 		.probe = perf_probe,
896 		.remove = perf_remove,
897 	},
898 };
899 module_ntb_client(perf_client);
900