[go: nahoru, domu]

1/*
2 * DMA Engine test module
3 *
4 * Copyright (C) 2007 Atmel Corporation
5 * Copyright (C) 2013 Intel Corporation
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
10 */
11#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12
13#include <linux/delay.h>
14#include <linux/dma-mapping.h>
15#include <linux/dmaengine.h>
16#include <linux/freezer.h>
17#include <linux/init.h>
18#include <linux/kthread.h>
19#include <linux/module.h>
20#include <linux/moduleparam.h>
21#include <linux/random.h>
22#include <linux/slab.h>
23#include <linux/wait.h>
24
25static unsigned int test_buf_size = 16384;
26module_param(test_buf_size, uint, S_IRUGO | S_IWUSR);
27MODULE_PARM_DESC(test_buf_size, "Size of the memcpy test buffer");
28
29static char test_channel[20];
30module_param_string(channel, test_channel, sizeof(test_channel),
31		S_IRUGO | S_IWUSR);
32MODULE_PARM_DESC(channel, "Bus ID of the channel to test (default: any)");
33
34static char test_device[32];
35module_param_string(device, test_device, sizeof(test_device),
36		S_IRUGO | S_IWUSR);
37MODULE_PARM_DESC(device, "Bus ID of the DMA Engine to test (default: any)");
38
39static unsigned int threads_per_chan = 1;
40module_param(threads_per_chan, uint, S_IRUGO | S_IWUSR);
41MODULE_PARM_DESC(threads_per_chan,
42		"Number of threads to start per channel (default: 1)");
43
44static unsigned int max_channels;
45module_param(max_channels, uint, S_IRUGO | S_IWUSR);
46MODULE_PARM_DESC(max_channels,
47		"Maximum number of channels to use (default: all)");
48
49static unsigned int iterations;
50module_param(iterations, uint, S_IRUGO | S_IWUSR);
51MODULE_PARM_DESC(iterations,
52		"Iterations before stopping test (default: infinite)");
53
54static unsigned int xor_sources = 3;
55module_param(xor_sources, uint, S_IRUGO | S_IWUSR);
56MODULE_PARM_DESC(xor_sources,
57		"Number of xor source buffers (default: 3)");
58
59static unsigned int pq_sources = 3;
60module_param(pq_sources, uint, S_IRUGO | S_IWUSR);
61MODULE_PARM_DESC(pq_sources,
62		"Number of p+q source buffers (default: 3)");
63
64static int timeout = 3000;
65module_param(timeout, uint, S_IRUGO | S_IWUSR);
66MODULE_PARM_DESC(timeout, "Transfer Timeout in msec (default: 3000), "
67		 "Pass -1 for infinite timeout");
68
69static bool noverify;
70module_param(noverify, bool, S_IRUGO | S_IWUSR);
71MODULE_PARM_DESC(noverify, "Disable random data setup and verification");
72
73static bool verbose;
74module_param(verbose, bool, S_IRUGO | S_IWUSR);
75MODULE_PARM_DESC(verbose, "Enable \"success\" result messages (default: off)");
76
77/**
78 * struct dmatest_params - test parameters.
79 * @buf_size:		size of the memcpy test buffer
80 * @channel:		bus ID of the channel to test
81 * @device:		bus ID of the DMA Engine to test
82 * @threads_per_chan:	number of threads to start per channel
83 * @max_channels:	maximum number of channels to use
84 * @iterations:		iterations before stopping test
85 * @xor_sources:	number of xor source buffers
86 * @pq_sources:		number of p+q source buffers
87 * @timeout:		transfer timeout in msec, -1 for infinite timeout
88 */
89struct dmatest_params {
90	unsigned int	buf_size;
91	char		channel[20];
92	char		device[32];
93	unsigned int	threads_per_chan;
94	unsigned int	max_channels;
95	unsigned int	iterations;
96	unsigned int	xor_sources;
97	unsigned int	pq_sources;
98	int		timeout;
99	bool		noverify;
100};
101
102/**
103 * struct dmatest_info - test information.
104 * @params:		test parameters
105 * @lock:		access protection to the fields of this structure
106 */
107static struct dmatest_info {
108	/* Test parameters */
109	struct dmatest_params	params;
110
111	/* Internal state */
112	struct list_head	channels;
113	unsigned int		nr_channels;
114	struct mutex		lock;
115	bool			did_init;
116} test_info = {
117	.channels = LIST_HEAD_INIT(test_info.channels),
118	.lock = __MUTEX_INITIALIZER(test_info.lock),
119};
120
121static int dmatest_run_set(const char *val, const struct kernel_param *kp);
122static int dmatest_run_get(char *val, const struct kernel_param *kp);
123static struct kernel_param_ops run_ops = {
124	.set = dmatest_run_set,
125	.get = dmatest_run_get,
126};
127static bool dmatest_run;
128module_param_cb(run, &run_ops, &dmatest_run, S_IRUGO | S_IWUSR);
129MODULE_PARM_DESC(run, "Run the test (default: false)");
130
131/* Maximum amount of mismatched bytes in buffer to print */
132#define MAX_ERROR_COUNT		32
133
134/*
135 * Initialization patterns. All bytes in the source buffer has bit 7
136 * set, all bytes in the destination buffer has bit 7 cleared.
137 *
138 * Bit 6 is set for all bytes which are to be copied by the DMA
139 * engine. Bit 5 is set for all bytes which are to be overwritten by
140 * the DMA engine.
141 *
142 * The remaining bits are the inverse of a counter which increments by
143 * one for each byte address.
144 */
145#define PATTERN_SRC		0x80
146#define PATTERN_DST		0x00
147#define PATTERN_COPY		0x40
148#define PATTERN_OVERWRITE	0x20
149#define PATTERN_COUNT_MASK	0x1f
150
151struct dmatest_thread {
152	struct list_head	node;
153	struct dmatest_info	*info;
154	struct task_struct	*task;
155	struct dma_chan		*chan;
156	u8			**srcs;
157	u8			**dsts;
158	enum dma_transaction_type type;
159	bool			done;
160};
161
162struct dmatest_chan {
163	struct list_head	node;
164	struct dma_chan		*chan;
165	struct list_head	threads;
166};
167
168static DECLARE_WAIT_QUEUE_HEAD(thread_wait);
169static bool wait;
170
171static bool is_threaded_test_run(struct dmatest_info *info)
172{
173	struct dmatest_chan *dtc;
174
175	list_for_each_entry(dtc, &info->channels, node) {
176		struct dmatest_thread *thread;
177
178		list_for_each_entry(thread, &dtc->threads, node) {
179			if (!thread->done)
180				return true;
181		}
182	}
183
184	return false;
185}
186
187static int dmatest_wait_get(char *val, const struct kernel_param *kp)
188{
189	struct dmatest_info *info = &test_info;
190	struct dmatest_params *params = &info->params;
191
192	if (params->iterations)
193		wait_event(thread_wait, !is_threaded_test_run(info));
194	wait = true;
195	return param_get_bool(val, kp);
196}
197
198static struct kernel_param_ops wait_ops = {
199	.get = dmatest_wait_get,
200	.set = param_set_bool,
201};
202module_param_cb(wait, &wait_ops, &wait, S_IRUGO);
203MODULE_PARM_DESC(wait, "Wait for tests to complete (default: false)");
204
205static bool dmatest_match_channel(struct dmatest_params *params,
206		struct dma_chan *chan)
207{
208	if (params->channel[0] == '\0')
209		return true;
210	return strcmp(dma_chan_name(chan), params->channel) == 0;
211}
212
213static bool dmatest_match_device(struct dmatest_params *params,
214		struct dma_device *device)
215{
216	if (params->device[0] == '\0')
217		return true;
218	return strcmp(dev_name(device->dev), params->device) == 0;
219}
220
221static unsigned long dmatest_random(void)
222{
223	unsigned long buf;
224
225	prandom_bytes(&buf, sizeof(buf));
226	return buf;
227}
228
229static void dmatest_init_srcs(u8 **bufs, unsigned int start, unsigned int len,
230		unsigned int buf_size)
231{
232	unsigned int i;
233	u8 *buf;
234
235	for (; (buf = *bufs); bufs++) {
236		for (i = 0; i < start; i++)
237			buf[i] = PATTERN_SRC | (~i & PATTERN_COUNT_MASK);
238		for ( ; i < start + len; i++)
239			buf[i] = PATTERN_SRC | PATTERN_COPY
240				| (~i & PATTERN_COUNT_MASK);
241		for ( ; i < buf_size; i++)
242			buf[i] = PATTERN_SRC | (~i & PATTERN_COUNT_MASK);
243		buf++;
244	}
245}
246
247static void dmatest_init_dsts(u8 **bufs, unsigned int start, unsigned int len,
248		unsigned int buf_size)
249{
250	unsigned int i;
251	u8 *buf;
252
253	for (; (buf = *bufs); bufs++) {
254		for (i = 0; i < start; i++)
255			buf[i] = PATTERN_DST | (~i & PATTERN_COUNT_MASK);
256		for ( ; i < start + len; i++)
257			buf[i] = PATTERN_DST | PATTERN_OVERWRITE
258				| (~i & PATTERN_COUNT_MASK);
259		for ( ; i < buf_size; i++)
260			buf[i] = PATTERN_DST | (~i & PATTERN_COUNT_MASK);
261	}
262}
263
264static void dmatest_mismatch(u8 actual, u8 pattern, unsigned int index,
265		unsigned int counter, bool is_srcbuf)
266{
267	u8		diff = actual ^ pattern;
268	u8		expected = pattern | (~counter & PATTERN_COUNT_MASK);
269	const char	*thread_name = current->comm;
270
271	if (is_srcbuf)
272		pr_warn("%s: srcbuf[0x%x] overwritten! Expected %02x, got %02x\n",
273			thread_name, index, expected, actual);
274	else if ((pattern & PATTERN_COPY)
275			&& (diff & (PATTERN_COPY | PATTERN_OVERWRITE)))
276		pr_warn("%s: dstbuf[0x%x] not copied! Expected %02x, got %02x\n",
277			thread_name, index, expected, actual);
278	else if (diff & PATTERN_SRC)
279		pr_warn("%s: dstbuf[0x%x] was copied! Expected %02x, got %02x\n",
280			thread_name, index, expected, actual);
281	else
282		pr_warn("%s: dstbuf[0x%x] mismatch! Expected %02x, got %02x\n",
283			thread_name, index, expected, actual);
284}
285
286static unsigned int dmatest_verify(u8 **bufs, unsigned int start,
287		unsigned int end, unsigned int counter, u8 pattern,
288		bool is_srcbuf)
289{
290	unsigned int i;
291	unsigned int error_count = 0;
292	u8 actual;
293	u8 expected;
294	u8 *buf;
295	unsigned int counter_orig = counter;
296
297	for (; (buf = *bufs); bufs++) {
298		counter = counter_orig;
299		for (i = start; i < end; i++) {
300			actual = buf[i];
301			expected = pattern | (~counter & PATTERN_COUNT_MASK);
302			if (actual != expected) {
303				if (error_count < MAX_ERROR_COUNT)
304					dmatest_mismatch(actual, pattern, i,
305							 counter, is_srcbuf);
306				error_count++;
307			}
308			counter++;
309		}
310	}
311
312	if (error_count > MAX_ERROR_COUNT)
313		pr_warn("%s: %u errors suppressed\n",
314			current->comm, error_count - MAX_ERROR_COUNT);
315
316	return error_count;
317}
318
319/* poor man's completion - we want to use wait_event_freezable() on it */
320struct dmatest_done {
321	bool			done;
322	wait_queue_head_t	*wait;
323};
324
325static void dmatest_callback(void *arg)
326{
327	struct dmatest_done *done = arg;
328
329	done->done = true;
330	wake_up_all(done->wait);
331}
332
333static unsigned int min_odd(unsigned int x, unsigned int y)
334{
335	unsigned int val = min(x, y);
336
337	return val % 2 ? val : val - 1;
338}
339
340static void result(const char *err, unsigned int n, unsigned int src_off,
341		   unsigned int dst_off, unsigned int len, unsigned long data)
342{
343	pr_info("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%lu)\n",
344		current->comm, n, err, src_off, dst_off, len, data);
345}
346
347static void dbg_result(const char *err, unsigned int n, unsigned int src_off,
348		       unsigned int dst_off, unsigned int len,
349		       unsigned long data)
350{
351	pr_debug("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%lu)\n",
352		   current->comm, n, err, src_off, dst_off, len, data);
353}
354
355#define verbose_result(err, n, src_off, dst_off, len, data) ({ \
356	if (verbose) \
357		result(err, n, src_off, dst_off, len, data); \
358	else \
359		dbg_result(err, n, src_off, dst_off, len, data); \
360})
361
362static unsigned long long dmatest_persec(s64 runtime, unsigned int val)
363{
364	unsigned long long per_sec = 1000000;
365
366	if (runtime <= 0)
367		return 0;
368
369	/* drop precision until runtime is 32-bits */
370	while (runtime > UINT_MAX) {
371		runtime >>= 1;
372		per_sec <<= 1;
373	}
374
375	per_sec *= val;
376	do_div(per_sec, runtime);
377	return per_sec;
378}
379
380static unsigned long long dmatest_KBs(s64 runtime, unsigned long long len)
381{
382	return dmatest_persec(runtime, len >> 10);
383}
384
385/*
386 * This function repeatedly tests DMA transfers of various lengths and
387 * offsets for a given operation type until it is told to exit by
388 * kthread_stop(). There may be multiple threads running this function
389 * in parallel for a single channel, and there may be multiple channels
390 * being tested in parallel.
391 *
392 * Before each test, the source and destination buffer is initialized
393 * with a known pattern. This pattern is different depending on
394 * whether it's in an area which is supposed to be copied or
395 * overwritten, and different in the source and destination buffers.
396 * So if the DMA engine doesn't copy exactly what we tell it to copy,
397 * we'll notice.
398 */
399static int dmatest_func(void *data)
400{
401	DECLARE_WAIT_QUEUE_HEAD_ONSTACK(done_wait);
402	struct dmatest_thread	*thread = data;
403	struct dmatest_done	done = { .wait = &done_wait };
404	struct dmatest_info	*info;
405	struct dmatest_params	*params;
406	struct dma_chan		*chan;
407	struct dma_device	*dev;
408	unsigned int		src_off, dst_off, len;
409	unsigned int		error_count;
410	unsigned int		failed_tests = 0;
411	unsigned int		total_tests = 0;
412	dma_cookie_t		cookie;
413	enum dma_status		status;
414	enum dma_ctrl_flags 	flags;
415	u8			*pq_coefs = NULL;
416	int			ret;
417	int			src_cnt;
418	int			dst_cnt;
419	int			i;
420	ktime_t			ktime;
421	s64			runtime = 0;
422	unsigned long long	total_len = 0;
423
424	set_freezable();
425
426	ret = -ENOMEM;
427
428	smp_rmb();
429	info = thread->info;
430	params = &info->params;
431	chan = thread->chan;
432	dev = chan->device;
433	if (thread->type == DMA_MEMCPY)
434		src_cnt = dst_cnt = 1;
435	else if (thread->type == DMA_XOR) {
436		/* force odd to ensure dst = src */
437		src_cnt = min_odd(params->xor_sources | 1, dev->max_xor);
438		dst_cnt = 1;
439	} else if (thread->type == DMA_PQ) {
440		/* force odd to ensure dst = src */
441		src_cnt = min_odd(params->pq_sources | 1, dma_maxpq(dev, 0));
442		dst_cnt = 2;
443
444		pq_coefs = kmalloc(params->pq_sources+1, GFP_KERNEL);
445		if (!pq_coefs)
446			goto err_thread_type;
447
448		for (i = 0; i < src_cnt; i++)
449			pq_coefs[i] = 1;
450	} else
451		goto err_thread_type;
452
453	thread->srcs = kcalloc(src_cnt+1, sizeof(u8 *), GFP_KERNEL);
454	if (!thread->srcs)
455		goto err_srcs;
456	for (i = 0; i < src_cnt; i++) {
457		thread->srcs[i] = kmalloc(params->buf_size, GFP_KERNEL);
458		if (!thread->srcs[i])
459			goto err_srcbuf;
460	}
461	thread->srcs[i] = NULL;
462
463	thread->dsts = kcalloc(dst_cnt+1, sizeof(u8 *), GFP_KERNEL);
464	if (!thread->dsts)
465		goto err_dsts;
466	for (i = 0; i < dst_cnt; i++) {
467		thread->dsts[i] = kmalloc(params->buf_size, GFP_KERNEL);
468		if (!thread->dsts[i])
469			goto err_dstbuf;
470	}
471	thread->dsts[i] = NULL;
472
473	set_user_nice(current, 10);
474
475	/*
476	 * src and dst buffers are freed by ourselves below
477	 */
478	flags = DMA_CTRL_ACK | DMA_PREP_INTERRUPT;
479
480	ktime = ktime_get();
481	while (!kthread_should_stop()
482	       && !(params->iterations && total_tests >= params->iterations)) {
483		struct dma_async_tx_descriptor *tx = NULL;
484		struct dmaengine_unmap_data *um;
485		dma_addr_t srcs[src_cnt];
486		dma_addr_t *dsts;
487		u8 align = 0;
488
489		total_tests++;
490
491		/* honor alignment restrictions */
492		if (thread->type == DMA_MEMCPY)
493			align = dev->copy_align;
494		else if (thread->type == DMA_XOR)
495			align = dev->xor_align;
496		else if (thread->type == DMA_PQ)
497			align = dev->pq_align;
498
499		if (1 << align > params->buf_size) {
500			pr_err("%u-byte buffer too small for %d-byte alignment\n",
501			       params->buf_size, 1 << align);
502			break;
503		}
504
505		if (params->noverify) {
506			len = params->buf_size;
507			src_off = 0;
508			dst_off = 0;
509		} else {
510			len = dmatest_random() % params->buf_size + 1;
511			len = (len >> align) << align;
512			if (!len)
513				len = 1 << align;
514			src_off = dmatest_random() % (params->buf_size - len + 1);
515			dst_off = dmatest_random() % (params->buf_size - len + 1);
516
517			src_off = (src_off >> align) << align;
518			dst_off = (dst_off >> align) << align;
519
520			dmatest_init_srcs(thread->srcs, src_off, len,
521					  params->buf_size);
522			dmatest_init_dsts(thread->dsts, dst_off, len,
523					  params->buf_size);
524		}
525
526		len = (len >> align) << align;
527		if (!len)
528			len = 1 << align;
529		total_len += len;
530
531		um = dmaengine_get_unmap_data(dev->dev, src_cnt+dst_cnt,
532					      GFP_KERNEL);
533		if (!um) {
534			failed_tests++;
535			result("unmap data NULL", total_tests,
536			       src_off, dst_off, len, ret);
537			continue;
538		}
539
540		um->len = params->buf_size;
541		for (i = 0; i < src_cnt; i++) {
542			void *buf = thread->srcs[i];
543			struct page *pg = virt_to_page(buf);
544			unsigned pg_off = (unsigned long) buf & ~PAGE_MASK;
545
546			um->addr[i] = dma_map_page(dev->dev, pg, pg_off,
547						   um->len, DMA_TO_DEVICE);
548			srcs[i] = um->addr[i] + src_off;
549			ret = dma_mapping_error(dev->dev, um->addr[i]);
550			if (ret) {
551				dmaengine_unmap_put(um);
552				result("src mapping error", total_tests,
553				       src_off, dst_off, len, ret);
554				failed_tests++;
555				continue;
556			}
557			um->to_cnt++;
558		}
559		/* map with DMA_BIDIRECTIONAL to force writeback/invalidate */
560		dsts = &um->addr[src_cnt];
561		for (i = 0; i < dst_cnt; i++) {
562			void *buf = thread->dsts[i];
563			struct page *pg = virt_to_page(buf);
564			unsigned pg_off = (unsigned long) buf & ~PAGE_MASK;
565
566			dsts[i] = dma_map_page(dev->dev, pg, pg_off, um->len,
567					       DMA_BIDIRECTIONAL);
568			ret = dma_mapping_error(dev->dev, dsts[i]);
569			if (ret) {
570				dmaengine_unmap_put(um);
571				result("dst mapping error", total_tests,
572				       src_off, dst_off, len, ret);
573				failed_tests++;
574				continue;
575			}
576			um->bidi_cnt++;
577		}
578
579		if (thread->type == DMA_MEMCPY)
580			tx = dev->device_prep_dma_memcpy(chan,
581							 dsts[0] + dst_off,
582							 srcs[0], len, flags);
583		else if (thread->type == DMA_XOR)
584			tx = dev->device_prep_dma_xor(chan,
585						      dsts[0] + dst_off,
586						      srcs, src_cnt,
587						      len, flags);
588		else if (thread->type == DMA_PQ) {
589			dma_addr_t dma_pq[dst_cnt];
590
591			for (i = 0; i < dst_cnt; i++)
592				dma_pq[i] = dsts[i] + dst_off;
593			tx = dev->device_prep_dma_pq(chan, dma_pq, srcs,
594						     src_cnt, pq_coefs,
595						     len, flags);
596		}
597
598		if (!tx) {
599			dmaengine_unmap_put(um);
600			result("prep error", total_tests, src_off,
601			       dst_off, len, ret);
602			msleep(100);
603			failed_tests++;
604			continue;
605		}
606
607		done.done = false;
608		tx->callback = dmatest_callback;
609		tx->callback_param = &done;
610		cookie = tx->tx_submit(tx);
611
612		if (dma_submit_error(cookie)) {
613			dmaengine_unmap_put(um);
614			result("submit error", total_tests, src_off,
615			       dst_off, len, ret);
616			msleep(100);
617			failed_tests++;
618			continue;
619		}
620		dma_async_issue_pending(chan);
621
622		wait_event_freezable_timeout(done_wait, done.done,
623					     msecs_to_jiffies(params->timeout));
624
625		status = dma_async_is_tx_complete(chan, cookie, NULL, NULL);
626
627		if (!done.done) {
628			/*
629			 * We're leaving the timed out dma operation with
630			 * dangling pointer to done_wait.  To make this
631			 * correct, we'll need to allocate wait_done for
632			 * each test iteration and perform "who's gonna
633			 * free it this time?" dancing.  For now, just
634			 * leave it dangling.
635			 */
636			dmaengine_unmap_put(um);
637			result("test timed out", total_tests, src_off, dst_off,
638			       len, 0);
639			failed_tests++;
640			continue;
641		} else if (status != DMA_COMPLETE) {
642			dmaengine_unmap_put(um);
643			result(status == DMA_ERROR ?
644			       "completion error status" :
645			       "completion busy status", total_tests, src_off,
646			       dst_off, len, ret);
647			failed_tests++;
648			continue;
649		}
650
651		dmaengine_unmap_put(um);
652
653		if (params->noverify) {
654			verbose_result("test passed", total_tests, src_off,
655				       dst_off, len, 0);
656			continue;
657		}
658
659		pr_debug("%s: verifying source buffer...\n", current->comm);
660		error_count = dmatest_verify(thread->srcs, 0, src_off,
661				0, PATTERN_SRC, true);
662		error_count += dmatest_verify(thread->srcs, src_off,
663				src_off + len, src_off,
664				PATTERN_SRC | PATTERN_COPY, true);
665		error_count += dmatest_verify(thread->srcs, src_off + len,
666				params->buf_size, src_off + len,
667				PATTERN_SRC, true);
668
669		pr_debug("%s: verifying dest buffer...\n", current->comm);
670		error_count += dmatest_verify(thread->dsts, 0, dst_off,
671				0, PATTERN_DST, false);
672		error_count += dmatest_verify(thread->dsts, dst_off,
673				dst_off + len, src_off,
674				PATTERN_SRC | PATTERN_COPY, false);
675		error_count += dmatest_verify(thread->dsts, dst_off + len,
676				params->buf_size, dst_off + len,
677				PATTERN_DST, false);
678
679		if (error_count) {
680			result("data error", total_tests, src_off, dst_off,
681			       len, error_count);
682			failed_tests++;
683		} else {
684			verbose_result("test passed", total_tests, src_off,
685				       dst_off, len, 0);
686		}
687	}
688	runtime = ktime_us_delta(ktime_get(), ktime);
689
690	ret = 0;
691err_dstbuf:
692	for (i = 0; thread->dsts[i]; i++)
693		kfree(thread->dsts[i]);
694	kfree(thread->dsts);
695err_dsts:
696err_srcbuf:
697	for (i = 0; thread->srcs[i]; i++)
698		kfree(thread->srcs[i]);
699	kfree(thread->srcs);
700err_srcs:
701	kfree(pq_coefs);
702err_thread_type:
703	pr_info("%s: summary %u tests, %u failures %llu iops %llu KB/s (%d)\n",
704		current->comm, total_tests, failed_tests,
705		dmatest_persec(runtime, total_tests),
706		dmatest_KBs(runtime, total_len), ret);
707
708	/* terminate all transfers on specified channels */
709	if (ret)
710		dmaengine_terminate_all(chan);
711
712	thread->done = true;
713	wake_up(&thread_wait);
714
715	return ret;
716}
717
718static void dmatest_cleanup_channel(struct dmatest_chan *dtc)
719{
720	struct dmatest_thread	*thread;
721	struct dmatest_thread	*_thread;
722	int			ret;
723
724	list_for_each_entry_safe(thread, _thread, &dtc->threads, node) {
725		ret = kthread_stop(thread->task);
726		pr_debug("thread %s exited with status %d\n",
727			 thread->task->comm, ret);
728		list_del(&thread->node);
729		put_task_struct(thread->task);
730		kfree(thread);
731	}
732
733	/* terminate all transfers on specified channels */
734	dmaengine_terminate_all(dtc->chan);
735
736	kfree(dtc);
737}
738
739static int dmatest_add_threads(struct dmatest_info *info,
740		struct dmatest_chan *dtc, enum dma_transaction_type type)
741{
742	struct dmatest_params *params = &info->params;
743	struct dmatest_thread *thread;
744	struct dma_chan *chan = dtc->chan;
745	char *op;
746	unsigned int i;
747
748	if (type == DMA_MEMCPY)
749		op = "copy";
750	else if (type == DMA_XOR)
751		op = "xor";
752	else if (type == DMA_PQ)
753		op = "pq";
754	else
755		return -EINVAL;
756
757	for (i = 0; i < params->threads_per_chan; i++) {
758		thread = kzalloc(sizeof(struct dmatest_thread), GFP_KERNEL);
759		if (!thread) {
760			pr_warn("No memory for %s-%s%u\n",
761				dma_chan_name(chan), op, i);
762			break;
763		}
764		thread->info = info;
765		thread->chan = dtc->chan;
766		thread->type = type;
767		smp_wmb();
768		thread->task = kthread_create(dmatest_func, thread, "%s-%s%u",
769				dma_chan_name(chan), op, i);
770		if (IS_ERR(thread->task)) {
771			pr_warn("Failed to create thread %s-%s%u\n",
772				dma_chan_name(chan), op, i);
773			kfree(thread);
774			break;
775		}
776
777		/* srcbuf and dstbuf are allocated by the thread itself */
778		get_task_struct(thread->task);
779		list_add_tail(&thread->node, &dtc->threads);
780		wake_up_process(thread->task);
781	}
782
783	return i;
784}
785
786static int dmatest_add_channel(struct dmatest_info *info,
787		struct dma_chan *chan)
788{
789	struct dmatest_chan	*dtc;
790	struct dma_device	*dma_dev = chan->device;
791	unsigned int		thread_count = 0;
792	int cnt;
793
794	dtc = kmalloc(sizeof(struct dmatest_chan), GFP_KERNEL);
795	if (!dtc) {
796		pr_warn("No memory for %s\n", dma_chan_name(chan));
797		return -ENOMEM;
798	}
799
800	dtc->chan = chan;
801	INIT_LIST_HEAD(&dtc->threads);
802
803	if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask)) {
804		cnt = dmatest_add_threads(info, dtc, DMA_MEMCPY);
805		thread_count += cnt > 0 ? cnt : 0;
806	}
807	if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
808		cnt = dmatest_add_threads(info, dtc, DMA_XOR);
809		thread_count += cnt > 0 ? cnt : 0;
810	}
811	if (dma_has_cap(DMA_PQ, dma_dev->cap_mask)) {
812		cnt = dmatest_add_threads(info, dtc, DMA_PQ);
813		thread_count += cnt > 0 ? cnt : 0;
814	}
815
816	pr_info("Started %u threads using %s\n",
817		thread_count, dma_chan_name(chan));
818
819	list_add_tail(&dtc->node, &info->channels);
820	info->nr_channels++;
821
822	return 0;
823}
824
825static bool filter(struct dma_chan *chan, void *param)
826{
827	struct dmatest_params *params = param;
828
829	if (!dmatest_match_channel(params, chan) ||
830	    !dmatest_match_device(params, chan->device))
831		return false;
832	else
833		return true;
834}
835
836static void request_channels(struct dmatest_info *info,
837			     enum dma_transaction_type type)
838{
839	dma_cap_mask_t mask;
840
841	dma_cap_zero(mask);
842	dma_cap_set(type, mask);
843	for (;;) {
844		struct dmatest_params *params = &info->params;
845		struct dma_chan *chan;
846
847		chan = dma_request_channel(mask, filter, params);
848		if (chan) {
849			if (dmatest_add_channel(info, chan)) {
850				dma_release_channel(chan);
851				break; /* add_channel failed, punt */
852			}
853		} else
854			break; /* no more channels available */
855		if (params->max_channels &&
856		    info->nr_channels >= params->max_channels)
857			break; /* we have all we need */
858	}
859}
860
861static void run_threaded_test(struct dmatest_info *info)
862{
863	struct dmatest_params *params = &info->params;
864
865	/* Copy test parameters */
866	params->buf_size = test_buf_size;
867	strlcpy(params->channel, strim(test_channel), sizeof(params->channel));
868	strlcpy(params->device, strim(test_device), sizeof(params->device));
869	params->threads_per_chan = threads_per_chan;
870	params->max_channels = max_channels;
871	params->iterations = iterations;
872	params->xor_sources = xor_sources;
873	params->pq_sources = pq_sources;
874	params->timeout = timeout;
875	params->noverify = noverify;
876
877	request_channels(info, DMA_MEMCPY);
878	request_channels(info, DMA_XOR);
879	request_channels(info, DMA_PQ);
880}
881
882static void stop_threaded_test(struct dmatest_info *info)
883{
884	struct dmatest_chan *dtc, *_dtc;
885	struct dma_chan *chan;
886
887	list_for_each_entry_safe(dtc, _dtc, &info->channels, node) {
888		list_del(&dtc->node);
889		chan = dtc->chan;
890		dmatest_cleanup_channel(dtc);
891		pr_debug("dropped channel %s\n", dma_chan_name(chan));
892		dma_release_channel(chan);
893	}
894
895	info->nr_channels = 0;
896}
897
898static void restart_threaded_test(struct dmatest_info *info, bool run)
899{
900	/* we might be called early to set run=, defer running until all
901	 * parameters have been evaluated
902	 */
903	if (!info->did_init)
904		return;
905
906	/* Stop any running test first */
907	stop_threaded_test(info);
908
909	/* Run test with new parameters */
910	run_threaded_test(info);
911}
912
913static int dmatest_run_get(char *val, const struct kernel_param *kp)
914{
915	struct dmatest_info *info = &test_info;
916
917	mutex_lock(&info->lock);
918	if (is_threaded_test_run(info)) {
919		dmatest_run = true;
920	} else {
921		stop_threaded_test(info);
922		dmatest_run = false;
923	}
924	mutex_unlock(&info->lock);
925
926	return param_get_bool(val, kp);
927}
928
929static int dmatest_run_set(const char *val, const struct kernel_param *kp)
930{
931	struct dmatest_info *info = &test_info;
932	int ret;
933
934	mutex_lock(&info->lock);
935	ret = param_set_bool(val, kp);
936	if (ret) {
937		mutex_unlock(&info->lock);
938		return ret;
939	}
940
941	if (is_threaded_test_run(info))
942		ret = -EBUSY;
943	else if (dmatest_run)
944		restart_threaded_test(info, dmatest_run);
945
946	mutex_unlock(&info->lock);
947
948	return ret;
949}
950
951static int __init dmatest_init(void)
952{
953	struct dmatest_info *info = &test_info;
954	struct dmatest_params *params = &info->params;
955
956	if (dmatest_run) {
957		mutex_lock(&info->lock);
958		run_threaded_test(info);
959		mutex_unlock(&info->lock);
960	}
961
962	if (params->iterations && wait)
963		wait_event(thread_wait, !is_threaded_test_run(info));
964
965	/* module parameters are stable, inittime tests are started,
966	 * let userspace take over 'run' control
967	 */
968	info->did_init = true;
969
970	return 0;
971}
972/* when compiled-in wait for drivers to load first */
973late_initcall(dmatest_init);
974
975static void __exit dmatest_exit(void)
976{
977	struct dmatest_info *info = &test_info;
978
979	mutex_lock(&info->lock);
980	stop_threaded_test(info);
981	mutex_unlock(&info->lock);
982}
983module_exit(dmatest_exit);
984
985MODULE_AUTHOR("Haavard Skinnemoen (Atmel)");
986MODULE_LICENSE("GPL v2");
987