Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Apr 14-17, 2025
Register
Loading...
v5.9
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * DMA Engine test module
   4 *
   5 * Copyright (C) 2007 Atmel Corporation
   6 * Copyright (C) 2013 Intel Corporation
   7 */
   8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
 
  10#include <linux/delay.h>
  11#include <linux/dma-mapping.h>
  12#include <linux/dmaengine.h>
  13#include <linux/freezer.h>
  14#include <linux/init.h>
  15#include <linux/kthread.h>
  16#include <linux/sched/task.h>
  17#include <linux/module.h>
  18#include <linux/moduleparam.h>
  19#include <linux/random.h>
  20#include <linux/slab.h>
  21#include <linux/wait.h>
  22
  23static unsigned int test_buf_size = 16384;
  24module_param(test_buf_size, uint, S_IRUGO | S_IWUSR);
  25MODULE_PARM_DESC(test_buf_size, "Size of the memcpy test buffer");
  26
  27static char test_device[32];
  28module_param_string(device, test_device, sizeof(test_device),
  29		S_IRUGO | S_IWUSR);
  30MODULE_PARM_DESC(device, "Bus ID of the DMA Engine to test (default: any)");
  31
  32static unsigned int threads_per_chan = 1;
  33module_param(threads_per_chan, uint, S_IRUGO | S_IWUSR);
  34MODULE_PARM_DESC(threads_per_chan,
  35		"Number of threads to start per channel (default: 1)");
  36
  37static unsigned int max_channels;
  38module_param(max_channels, uint, S_IRUGO | S_IWUSR);
  39MODULE_PARM_DESC(max_channels,
  40		"Maximum number of channels to use (default: all)");
  41
  42static unsigned int iterations;
  43module_param(iterations, uint, S_IRUGO | S_IWUSR);
  44MODULE_PARM_DESC(iterations,
  45		"Iterations before stopping test (default: infinite)");
  46
  47static unsigned int dmatest;
  48module_param(dmatest, uint, S_IRUGO | S_IWUSR);
  49MODULE_PARM_DESC(dmatest,
  50		"dmatest 0-memcpy 1-memset (default: 0)");
  51
  52static unsigned int xor_sources = 3;
  53module_param(xor_sources, uint, S_IRUGO | S_IWUSR);
  54MODULE_PARM_DESC(xor_sources,
  55		"Number of xor source buffers (default: 3)");
  56
  57static unsigned int pq_sources = 3;
  58module_param(pq_sources, uint, S_IRUGO | S_IWUSR);
  59MODULE_PARM_DESC(pq_sources,
  60		"Number of p+q source buffers (default: 3)");
  61
  62static int timeout = 3000;
  63module_param(timeout, int, S_IRUGO | S_IWUSR);
  64MODULE_PARM_DESC(timeout, "Transfer Timeout in msec (default: 3000), "
  65		 "Pass -1 for infinite timeout");
  66
  67static bool noverify;
  68module_param(noverify, bool, S_IRUGO | S_IWUSR);
  69MODULE_PARM_DESC(noverify, "Disable data verification (default: verify)");
  70
  71static bool norandom;
  72module_param(norandom, bool, 0644);
  73MODULE_PARM_DESC(norandom, "Disable random offset setup (default: random)");
  74
  75static bool verbose;
  76module_param(verbose, bool, S_IRUGO | S_IWUSR);
  77MODULE_PARM_DESC(verbose, "Enable \"success\" result messages (default: off)");
  78
  79static int alignment = -1;
  80module_param(alignment, int, 0644);
  81MODULE_PARM_DESC(alignment, "Custom data address alignment taken as 2^(alignment) (default: not used (-1))");
  82
  83static unsigned int transfer_size;
  84module_param(transfer_size, uint, 0644);
  85MODULE_PARM_DESC(transfer_size, "Optional custom transfer size in bytes (default: not used (0))");
  86
  87static bool polled;
  88module_param(polled, bool, S_IRUGO | S_IWUSR);
  89MODULE_PARM_DESC(polled, "Use polling for completion instead of interrupts");
  90
  91/**
  92 * struct dmatest_params - test parameters.
  93 * @buf_size:		size of the memcpy test buffer
  94 * @channel:		bus ID of the channel to test
  95 * @device:		bus ID of the DMA Engine to test
  96 * @threads_per_chan:	number of threads to start per channel
  97 * @max_channels:	maximum number of channels to use
  98 * @iterations:		iterations before stopping test
  99 * @xor_sources:	number of xor source buffers
 100 * @pq_sources:		number of p+q source buffers
 101 * @timeout:		transfer timeout in msec, -1 for infinite timeout
 102 * @noverify:		disable data verification
 103 * @norandom:		disable random offset setup
 104 * @alignment:		custom data address alignment taken as 2^alignment
 105 * @transfer_size:	custom transfer size in bytes
 106 * @polled:		use polling for completion instead of interrupts
 107 */
 108struct dmatest_params {
 109	unsigned int	buf_size;
 110	char		channel[20];
 111	char		device[32];
 112	unsigned int	threads_per_chan;
 113	unsigned int	max_channels;
 114	unsigned int	iterations;
 115	unsigned int	xor_sources;
 116	unsigned int	pq_sources;
 117	int		timeout;
 118	bool		noverify;
 119	bool		norandom;
 120	int		alignment;
 121	unsigned int	transfer_size;
 122	bool		polled;
 123};
 124
 125/**
 126 * struct dmatest_info - test information.
 127 * @params:		test parameters
 128 * @channels:		channels under test
 129 * @nr_channels:	number of channels under test
 130 * @lock:		access protection to the fields of this structure
 131 * @did_init:		module has been initialized completely
 132 * @last_error:		test has faced configuration issues
 133 */
 134static struct dmatest_info {
 135	/* Test parameters */
 136	struct dmatest_params	params;
 137
 138	/* Internal state */
 139	struct list_head	channels;
 140	unsigned int		nr_channels;
 141	int			last_error;
 142	struct mutex		lock;
 143	bool			did_init;
 144} test_info = {
 145	.channels = LIST_HEAD_INIT(test_info.channels),
 146	.lock = __MUTEX_INITIALIZER(test_info.lock),
 147};
 148
 149static int dmatest_run_set(const char *val, const struct kernel_param *kp);
 150static int dmatest_run_get(char *val, const struct kernel_param *kp);
 151static const struct kernel_param_ops run_ops = {
 152	.set = dmatest_run_set,
 153	.get = dmatest_run_get,
 154};
 155static bool dmatest_run;
 156module_param_cb(run, &run_ops, &dmatest_run, S_IRUGO | S_IWUSR);
 157MODULE_PARM_DESC(run, "Run the test (default: false)");
 158
 159static int dmatest_chan_set(const char *val, const struct kernel_param *kp);
 160static int dmatest_chan_get(char *val, const struct kernel_param *kp);
 161static const struct kernel_param_ops multi_chan_ops = {
 162	.set = dmatest_chan_set,
 163	.get = dmatest_chan_get,
 164};
 165
 166static char test_channel[20];
 167static struct kparam_string newchan_kps = {
 168	.string = test_channel,
 169	.maxlen = 20,
 170};
 171module_param_cb(channel, &multi_chan_ops, &newchan_kps, 0644);
 172MODULE_PARM_DESC(channel, "Bus ID of the channel to test (default: any)");
 173
 174static int dmatest_test_list_get(char *val, const struct kernel_param *kp);
 175static const struct kernel_param_ops test_list_ops = {
 176	.get = dmatest_test_list_get,
 177};
 178module_param_cb(test_list, &test_list_ops, NULL, 0444);
 179MODULE_PARM_DESC(test_list, "Print current test list");
 180
 181/* Maximum amount of mismatched bytes in buffer to print */
 182#define MAX_ERROR_COUNT		32
 183
 184/*
 185 * Initialization patterns. All bytes in the source buffer has bit 7
 186 * set, all bytes in the destination buffer has bit 7 cleared.
 187 *
 188 * Bit 6 is set for all bytes which are to be copied by the DMA
 189 * engine. Bit 5 is set for all bytes which are to be overwritten by
 190 * the DMA engine.
 191 *
 192 * The remaining bits are the inverse of a counter which increments by
 193 * one for each byte address.
 194 */
 195#define PATTERN_SRC		0x80
 196#define PATTERN_DST		0x00
 197#define PATTERN_COPY		0x40
 198#define PATTERN_OVERWRITE	0x20
 199#define PATTERN_COUNT_MASK	0x1f
 200#define PATTERN_MEMSET_IDX	0x01
 201
 202/* Fixed point arithmetic ops */
 203#define FIXPT_SHIFT		8
 204#define FIXPNT_MASK		0xFF
 205#define FIXPT_TO_INT(a)	((a) >> FIXPT_SHIFT)
 206#define INT_TO_FIXPT(a)	((a) << FIXPT_SHIFT)
 207#define FIXPT_GET_FRAC(a)	((((a) & FIXPNT_MASK) * 100) >> FIXPT_SHIFT)
 208
 209/* poor man's completion - we want to use wait_event_freezable() on it */
 210struct dmatest_done {
 211	bool			done;
 212	wait_queue_head_t	*wait;
 213};
 214
 215struct dmatest_data {
 216	u8		**raw;
 217	u8		**aligned;
 218	unsigned int	cnt;
 219	unsigned int	off;
 220};
 221
 222struct dmatest_thread {
 223	struct list_head	node;
 224	struct dmatest_info	*info;
 225	struct task_struct	*task;
 226	struct dma_chan		*chan;
 227	struct dmatest_data	src;
 228	struct dmatest_data	dst;
 229	enum dma_transaction_type type;
 230	wait_queue_head_t done_wait;
 231	struct dmatest_done test_done;
 232	bool			done;
 233	bool			pending;
 234};
 235
 236struct dmatest_chan {
 237	struct list_head	node;
 238	struct dma_chan		*chan;
 239	struct list_head	threads;
 240};
 241
 242static DECLARE_WAIT_QUEUE_HEAD(thread_wait);
 243static bool wait;
 244
 245static bool is_threaded_test_run(struct dmatest_info *info)
 246{
 247	struct dmatest_chan *dtc;
 248
 249	list_for_each_entry(dtc, &info->channels, node) {
 250		struct dmatest_thread *thread;
 251
 252		list_for_each_entry(thread, &dtc->threads, node) {
 253			if (!thread->done && !thread->pending)
 254				return true;
 255		}
 256	}
 257
 258	return false;
 259}
 260
 261static bool is_threaded_test_pending(struct dmatest_info *info)
 262{
 263	struct dmatest_chan *dtc;
 264
 265	list_for_each_entry(dtc, &info->channels, node) {
 266		struct dmatest_thread *thread;
 267
 268		list_for_each_entry(thread, &dtc->threads, node) {
 269			if (thread->pending)
 270				return true;
 271		}
 272	}
 273
 274	return false;
 275}
 276
 277static int dmatest_wait_get(char *val, const struct kernel_param *kp)
 278{
 279	struct dmatest_info *info = &test_info;
 280	struct dmatest_params *params = &info->params;
 281
 282	if (params->iterations)
 283		wait_event(thread_wait, !is_threaded_test_run(info));
 284	wait = true;
 285	return param_get_bool(val, kp);
 286}
 287
 288static const struct kernel_param_ops wait_ops = {
 289	.get = dmatest_wait_get,
 290	.set = param_set_bool,
 291};
 292module_param_cb(wait, &wait_ops, &wait, S_IRUGO);
 293MODULE_PARM_DESC(wait, "Wait for tests to complete (default: false)");
 294
 295static bool dmatest_match_channel(struct dmatest_params *params,
 296		struct dma_chan *chan)
 297{
 298	if (params->channel[0] == '\0')
 299		return true;
 300	return strcmp(dma_chan_name(chan), params->channel) == 0;
 301}
 302
 303static bool dmatest_match_device(struct dmatest_params *params,
 304		struct dma_device *device)
 305{
 306	if (params->device[0] == '\0')
 307		return true;
 308	return strcmp(dev_name(device->dev), params->device) == 0;
 309}
 310
 311static unsigned long dmatest_random(void)
 312{
 313	unsigned long buf;
 314
 315	prandom_bytes(&buf, sizeof(buf));
 316	return buf;
 317}
 318
 319static inline u8 gen_inv_idx(u8 index, bool is_memset)
 320{
 321	u8 val = is_memset ? PATTERN_MEMSET_IDX : index;
 322
 323	return ~val & PATTERN_COUNT_MASK;
 324}
 325
 326static inline u8 gen_src_value(u8 index, bool is_memset)
 327{
 328	return PATTERN_SRC | gen_inv_idx(index, is_memset);
 329}
 330
 331static inline u8 gen_dst_value(u8 index, bool is_memset)
 332{
 333	return PATTERN_DST | gen_inv_idx(index, is_memset);
 334}
 335
 336static void dmatest_init_srcs(u8 **bufs, unsigned int start, unsigned int len,
 337		unsigned int buf_size, bool is_memset)
 338{
 339	unsigned int i;
 340	u8 *buf;
 341
 342	for (; (buf = *bufs); bufs++) {
 343		for (i = 0; i < start; i++)
 344			buf[i] = gen_src_value(i, is_memset);
 345		for ( ; i < start + len; i++)
 346			buf[i] = gen_src_value(i, is_memset) | PATTERN_COPY;
 347		for ( ; i < buf_size; i++)
 348			buf[i] = gen_src_value(i, is_memset);
 349		buf++;
 350	}
 351}
 352
 353static void dmatest_init_dsts(u8 **bufs, unsigned int start, unsigned int len,
 354		unsigned int buf_size, bool is_memset)
 355{
 356	unsigned int i;
 357	u8 *buf;
 358
 359	for (; (buf = *bufs); bufs++) {
 360		for (i = 0; i < start; i++)
 361			buf[i] = gen_dst_value(i, is_memset);
 362		for ( ; i < start + len; i++)
 363			buf[i] = gen_dst_value(i, is_memset) |
 364						PATTERN_OVERWRITE;
 365		for ( ; i < buf_size; i++)
 366			buf[i] = gen_dst_value(i, is_memset);
 367	}
 368}
 369
 370static void dmatest_mismatch(u8 actual, u8 pattern, unsigned int index,
 371		unsigned int counter, bool is_srcbuf, bool is_memset)
 372{
 373	u8		diff = actual ^ pattern;
 374	u8		expected = pattern | gen_inv_idx(counter, is_memset);
 375	const char	*thread_name = current->comm;
 376
 377	if (is_srcbuf)
 378		pr_warn("%s: srcbuf[0x%x] overwritten! Expected %02x, got %02x\n",
 379			thread_name, index, expected, actual);
 380	else if ((pattern & PATTERN_COPY)
 381			&& (diff & (PATTERN_COPY | PATTERN_OVERWRITE)))
 382		pr_warn("%s: dstbuf[0x%x] not copied! Expected %02x, got %02x\n",
 383			thread_name, index, expected, actual);
 384	else if (diff & PATTERN_SRC)
 385		pr_warn("%s: dstbuf[0x%x] was copied! Expected %02x, got %02x\n",
 386			thread_name, index, expected, actual);
 387	else
 388		pr_warn("%s: dstbuf[0x%x] mismatch! Expected %02x, got %02x\n",
 389			thread_name, index, expected, actual);
 390}
 391
 392static unsigned int dmatest_verify(u8 **bufs, unsigned int start,
 393		unsigned int end, unsigned int counter, u8 pattern,
 394		bool is_srcbuf, bool is_memset)
 395{
 396	unsigned int i;
 397	unsigned int error_count = 0;
 398	u8 actual;
 399	u8 expected;
 400	u8 *buf;
 401	unsigned int counter_orig = counter;
 402
 403	for (; (buf = *bufs); bufs++) {
 404		counter = counter_orig;
 405		for (i = start; i < end; i++) {
 406			actual = buf[i];
 407			expected = pattern | gen_inv_idx(counter, is_memset);
 408			if (actual != expected) {
 409				if (error_count < MAX_ERROR_COUNT)
 410					dmatest_mismatch(actual, pattern, i,
 411							 counter, is_srcbuf,
 412							 is_memset);
 413				error_count++;
 414			}
 415			counter++;
 416		}
 417	}
 418
 419	if (error_count > MAX_ERROR_COUNT)
 420		pr_warn("%s: %u errors suppressed\n",
 421			current->comm, error_count - MAX_ERROR_COUNT);
 422
 423	return error_count;
 424}
 425
 426
 427static void dmatest_callback(void *arg)
 428{
 429	struct dmatest_done *done = arg;
 430	struct dmatest_thread *thread =
 431		container_of(done, struct dmatest_thread, test_done);
 432	if (!thread->done) {
 433		done->done = true;
 434		wake_up_all(done->wait);
 435	} else {
 436		/*
 437		 * If thread->done, it means that this callback occurred
 438		 * after the parent thread has cleaned up. This can
 439		 * happen in the case that driver doesn't implement
 440		 * the terminate_all() functionality and a dma operation
 441		 * did not occur within the timeout period
 442		 */
 443		WARN(1, "dmatest: Kernel memory may be corrupted!!\n");
 444	}
 445}
 446
 447static unsigned int min_odd(unsigned int x, unsigned int y)
 448{
 449	unsigned int val = min(x, y);
 450
 451	return val % 2 ? val : val - 1;
 452}
 453
 454static void result(const char *err, unsigned int n, unsigned int src_off,
 455		   unsigned int dst_off, unsigned int len, unsigned long data)
 456{
 457	pr_info("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%lu)\n",
 458		current->comm, n, err, src_off, dst_off, len, data);
 
 
 
 
 
 459}
 460
 461static void dbg_result(const char *err, unsigned int n, unsigned int src_off,
 462		       unsigned int dst_off, unsigned int len,
 463		       unsigned long data)
 464{
 465	pr_debug("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%lu)\n",
 466		 current->comm, n, err, src_off, dst_off, len, data);
 467}
 468
 469#define verbose_result(err, n, src_off, dst_off, len, data) ({	\
 470	if (verbose)						\
 471		result(err, n, src_off, dst_off, len, data);	\
 472	else							\
 473		dbg_result(err, n, src_off, dst_off, len, data);\
 474})
 475
 476static unsigned long long dmatest_persec(s64 runtime, unsigned int val)
 477{
 478	unsigned long long per_sec = 1000000;
 479
 480	if (runtime <= 0)
 481		return 0;
 482
 483	/* drop precision until runtime is 32-bits */
 484	while (runtime > UINT_MAX) {
 485		runtime >>= 1;
 486		per_sec <<= 1;
 487	}
 488
 489	per_sec *= val;
 490	per_sec = INT_TO_FIXPT(per_sec);
 491	do_div(per_sec, runtime);
 492
 493	return per_sec;
 494}
 495
 496static unsigned long long dmatest_KBs(s64 runtime, unsigned long long len)
 497{
 498	return FIXPT_TO_INT(dmatest_persec(runtime, len >> 10));
 499}
 500
 501static void __dmatest_free_test_data(struct dmatest_data *d, unsigned int cnt)
 502{
 503	unsigned int i;
 504
 505	for (i = 0; i < cnt; i++)
 506		kfree(d->raw[i]);
 507
 508	kfree(d->aligned);
 509	kfree(d->raw);
 510}
 511
 512static void dmatest_free_test_data(struct dmatest_data *d)
 513{
 514	__dmatest_free_test_data(d, d->cnt);
 515}
 516
 517static int dmatest_alloc_test_data(struct dmatest_data *d,
 518		unsigned int buf_size, u8 align)
 519{
 520	unsigned int i = 0;
 521
 522	d->raw = kcalloc(d->cnt + 1, sizeof(u8 *), GFP_KERNEL);
 523	if (!d->raw)
 524		return -ENOMEM;
 525
 526	d->aligned = kcalloc(d->cnt + 1, sizeof(u8 *), GFP_KERNEL);
 527	if (!d->aligned)
 528		goto err;
 529
 530	for (i = 0; i < d->cnt; i++) {
 531		d->raw[i] = kmalloc(buf_size + align, GFP_KERNEL);
 532		if (!d->raw[i])
 533			goto err;
 534
 535		/* align to alignment restriction */
 536		if (align)
 537			d->aligned[i] = PTR_ALIGN(d->raw[i], align);
 538		else
 539			d->aligned[i] = d->raw[i];
 540	}
 541
 542	return 0;
 543err:
 544	__dmatest_free_test_data(d, i);
 545	return -ENOMEM;
 546}
 547
 548/*
 549 * This function repeatedly tests DMA transfers of various lengths and
 550 * offsets for a given operation type until it is told to exit by
 551 * kthread_stop(). There may be multiple threads running this function
 552 * in parallel for a single channel, and there may be multiple channels
 553 * being tested in parallel.
 554 *
 555 * Before each test, the source and destination buffer is initialized
 556 * with a known pattern. This pattern is different depending on
 557 * whether it's in an area which is supposed to be copied or
 558 * overwritten, and different in the source and destination buffers.
 559 * So if the DMA engine doesn't copy exactly what we tell it to copy,
 560 * we'll notice.
 561 */
 562static int dmatest_func(void *data)
 563{
 564	struct dmatest_thread	*thread = data;
 565	struct dmatest_done	*done = &thread->test_done;
 566	struct dmatest_info	*info;
 567	struct dmatest_params	*params;
 568	struct dma_chan		*chan;
 569	struct dma_device	*dev;
 
 570	unsigned int		error_count;
 571	unsigned int		failed_tests = 0;
 572	unsigned int		total_tests = 0;
 573	dma_cookie_t		cookie;
 574	enum dma_status		status;
 575	enum dma_ctrl_flags 	flags;
 576	u8			*pq_coefs = NULL;
 577	int			ret;
 578	unsigned int 		buf_size;
 579	struct dmatest_data	*src;
 580	struct dmatest_data	*dst;
 581	int			i;
 582	ktime_t			ktime, start, diff;
 583	ktime_t			filltime = 0;
 584	ktime_t			comparetime = 0;
 585	s64			runtime = 0;
 586	unsigned long long	total_len = 0;
 587	unsigned long long	iops = 0;
 588	u8			align = 0;
 589	bool			is_memset = false;
 590	dma_addr_t		*srcs;
 591	dma_addr_t		*dma_pq;
 592
 593	set_freezable();
 594
 595	ret = -ENOMEM;
 596
 597	smp_rmb();
 598	thread->pending = false;
 599	info = thread->info;
 600	params = &info->params;
 601	chan = thread->chan;
 602	dev = chan->device;
 
 
 603	src = &thread->src;
 604	dst = &thread->dst;
 605	if (thread->type == DMA_MEMCPY) {
 606		align = params->alignment < 0 ? dev->copy_align :
 607						params->alignment;
 608		src->cnt = dst->cnt = 1;
 609	} else if (thread->type == DMA_MEMSET) {
 610		align = params->alignment < 0 ? dev->fill_align :
 611						params->alignment;
 612		src->cnt = dst->cnt = 1;
 613		is_memset = true;
 614	} else if (thread->type == DMA_XOR) {
 615		/* force odd to ensure dst = src */
 616		src->cnt = min_odd(params->xor_sources | 1, dev->max_xor);
 617		dst->cnt = 1;
 618		align = params->alignment < 0 ? dev->xor_align :
 619						params->alignment;
 620	} else if (thread->type == DMA_PQ) {
 621		/* force odd to ensure dst = src */
 622		src->cnt = min_odd(params->pq_sources | 1, dma_maxpq(dev, 0));
 623		dst->cnt = 2;
 624		align = params->alignment < 0 ? dev->pq_align :
 625						params->alignment;
 626
 627		pq_coefs = kmalloc(params->pq_sources + 1, GFP_KERNEL);
 628		if (!pq_coefs)
 629			goto err_thread_type;
 630
 631		for (i = 0; i < src->cnt; i++)
 632			pq_coefs[i] = 1;
 633	} else
 634		goto err_thread_type;
 635
 636	/* Check if buffer count fits into map count variable (u8) */
 637	if ((src->cnt + dst->cnt) >= 255) {
 638		pr_err("too many buffers (%d of 255 supported)\n",
 639		       src->cnt + dst->cnt);
 640		goto err_free_coefs;
 641	}
 642
 643	buf_size = params->buf_size;
 644	if (1 << align > buf_size) {
 645		pr_err("%u-byte buffer too small for %d-byte alignment\n",
 646		       buf_size, 1 << align);
 647		goto err_free_coefs;
 648	}
 649
 650	if (dmatest_alloc_test_data(src, buf_size, align) < 0)
 651		goto err_free_coefs;
 652
 653	if (dmatest_alloc_test_data(dst, buf_size, align) < 0)
 654		goto err_src;
 655
 656	set_user_nice(current, 10);
 657
 658	srcs = kcalloc(src->cnt, sizeof(dma_addr_t), GFP_KERNEL);
 659	if (!srcs)
 660		goto err_dst;
 661
 662	dma_pq = kcalloc(dst->cnt, sizeof(dma_addr_t), GFP_KERNEL);
 663	if (!dma_pq)
 664		goto err_srcs_array;
 665
 666	/*
 667	 * src and dst buffers are freed by ourselves below
 668	 */
 669	if (params->polled)
 670		flags = DMA_CTRL_ACK;
 671	else
 672		flags = DMA_CTRL_ACK | DMA_PREP_INTERRUPT;
 673
 674	ktime = ktime_get();
 675	while (!(kthread_should_stop() ||
 676	       (params->iterations && total_tests >= params->iterations))) {
 677		struct dma_async_tx_descriptor *tx = NULL;
 678		struct dmaengine_unmap_data *um;
 679		dma_addr_t *dsts;
 680		unsigned int len;
 681
 682		total_tests++;
 683
 684		if (params->transfer_size) {
 685			if (params->transfer_size >= buf_size) {
 686				pr_err("%u-byte transfer size must be lower than %u-buffer size\n",
 687				       params->transfer_size, buf_size);
 688				break;
 689			}
 690			len = params->transfer_size;
 691		} else if (params->norandom) {
 692			len = buf_size;
 693		} else {
 694			len = dmatest_random() % buf_size + 1;
 695		}
 696
 697		/* Do not alter transfer size explicitly defined by user */
 698		if (!params->transfer_size) {
 699			len = (len >> align) << align;
 700			if (!len)
 701				len = 1 << align;
 702		}
 703		total_len += len;
 704
 705		if (params->norandom) {
 706			src->off = 0;
 707			dst->off = 0;
 708		} else {
 709			src->off = dmatest_random() % (buf_size - len + 1);
 710			dst->off = dmatest_random() % (buf_size - len + 1);
 711
 712			src->off = (src->off >> align) << align;
 713			dst->off = (dst->off >> align) << align;
 714		}
 715
 716		if (!params->noverify) {
 717			start = ktime_get();
 718			dmatest_init_srcs(src->aligned, src->off, len,
 719					  buf_size, is_memset);
 720			dmatest_init_dsts(dst->aligned, dst->off, len,
 721					  buf_size, is_memset);
 722
 723			diff = ktime_sub(ktime_get(), start);
 724			filltime = ktime_add(filltime, diff);
 725		}
 726
 727		um = dmaengine_get_unmap_data(dev->dev, src->cnt + dst->cnt,
 728					      GFP_KERNEL);
 729		if (!um) {
 730			failed_tests++;
 731			result("unmap data NULL", total_tests,
 732			       src->off, dst->off, len, ret);
 733			continue;
 734		}
 735
 736		um->len = buf_size;
 737		for (i = 0; i < src->cnt; i++) {
 738			void *buf = src->aligned[i];
 739			struct page *pg = virt_to_page(buf);
 740			unsigned long pg_off = offset_in_page(buf);
 741
 742			um->addr[i] = dma_map_page(dev->dev, pg, pg_off,
 743						   um->len, DMA_TO_DEVICE);
 744			srcs[i] = um->addr[i] + src->off;
 745			ret = dma_mapping_error(dev->dev, um->addr[i]);
 746			if (ret) {
 747				result("src mapping error", total_tests,
 748				       src->off, dst->off, len, ret);
 749				goto error_unmap_continue;
 750			}
 751			um->to_cnt++;
 752		}
 753		/* map with DMA_BIDIRECTIONAL to force writeback/invalidate */
 754		dsts = &um->addr[src->cnt];
 755		for (i = 0; i < dst->cnt; i++) {
 756			void *buf = dst->aligned[i];
 757			struct page *pg = virt_to_page(buf);
 758			unsigned long pg_off = offset_in_page(buf);
 759
 760			dsts[i] = dma_map_page(dev->dev, pg, pg_off, um->len,
 761					       DMA_BIDIRECTIONAL);
 762			ret = dma_mapping_error(dev->dev, dsts[i]);
 763			if (ret) {
 764				result("dst mapping error", total_tests,
 765				       src->off, dst->off, len, ret);
 766				goto error_unmap_continue;
 767			}
 768			um->bidi_cnt++;
 769		}
 770
 771		if (thread->type == DMA_MEMCPY)
 772			tx = dev->device_prep_dma_memcpy(chan,
 773							 dsts[0] + dst->off,
 774							 srcs[0], len, flags);
 775		else if (thread->type == DMA_MEMSET)
 776			tx = dev->device_prep_dma_memset(chan,
 777						dsts[0] + dst->off,
 778						*(src->aligned[0] + src->off),
 779						len, flags);
 780		else if (thread->type == DMA_XOR)
 781			tx = dev->device_prep_dma_xor(chan,
 782						      dsts[0] + dst->off,
 783						      srcs, src->cnt,
 784						      len, flags);
 785		else if (thread->type == DMA_PQ) {
 786			for (i = 0; i < dst->cnt; i++)
 787				dma_pq[i] = dsts[i] + dst->off;
 788			tx = dev->device_prep_dma_pq(chan, dma_pq, srcs,
 789						     src->cnt, pq_coefs,
 790						     len, flags);
 791		}
 792
 793		if (!tx) {
 794			result("prep error", total_tests, src->off,
 795			       dst->off, len, ret);
 796			msleep(100);
 797			goto error_unmap_continue;
 798		}
 799
 800		done->done = false;
 801		if (!params->polled) {
 802			tx->callback = dmatest_callback;
 803			tx->callback_param = done;
 804		}
 805		cookie = tx->tx_submit(tx);
 806
 807		if (dma_submit_error(cookie)) {
 808			result("submit error", total_tests, src->off,
 809			       dst->off, len, ret);
 810			msleep(100);
 811			goto error_unmap_continue;
 812		}
 813
 814		if (params->polled) {
 815			status = dma_sync_wait(chan, cookie);
 816			dmaengine_terminate_sync(chan);
 817			if (status == DMA_COMPLETE)
 818				done->done = true;
 819		} else {
 820			dma_async_issue_pending(chan);
 821
 822			wait_event_freezable_timeout(thread->done_wait,
 823					done->done,
 824					msecs_to_jiffies(params->timeout));
 825
 826			status = dma_async_is_tx_complete(chan, cookie, NULL,
 827							  NULL);
 828		}
 829
 830		if (!done->done) {
 831			result("test timed out", total_tests, src->off, dst->off,
 832			       len, 0);
 833			goto error_unmap_continue;
 834		} else if (status != DMA_COMPLETE &&
 835			   !(dma_has_cap(DMA_COMPLETION_NO_ORDER,
 836					 dev->cap_mask) &&
 837			     status == DMA_OUT_OF_ORDER)) {
 838			result(status == DMA_ERROR ?
 839			       "completion error status" :
 840			       "completion busy status", total_tests, src->off,
 841			       dst->off, len, ret);
 842			goto error_unmap_continue;
 843		}
 844
 845		dmaengine_unmap_put(um);
 846
 847		if (params->noverify) {
 848			verbose_result("test passed", total_tests, src->off,
 849				       dst->off, len, 0);
 850			continue;
 851		}
 852
 853		start = ktime_get();
 854		pr_debug("%s: verifying source buffer...\n", current->comm);
 855		error_count = dmatest_verify(src->aligned, 0, src->off,
 856				0, PATTERN_SRC, true, is_memset);
 857		error_count += dmatest_verify(src->aligned, src->off,
 858				src->off + len, src->off,
 859				PATTERN_SRC | PATTERN_COPY, true, is_memset);
 860		error_count += dmatest_verify(src->aligned, src->off + len,
 861				buf_size, src->off + len,
 862				PATTERN_SRC, true, is_memset);
 863
 864		pr_debug("%s: verifying dest buffer...\n", current->comm);
 865		error_count += dmatest_verify(dst->aligned, 0, dst->off,
 866				0, PATTERN_DST, false, is_memset);
 867
 868		error_count += dmatest_verify(dst->aligned, dst->off,
 869				dst->off + len, src->off,
 870				PATTERN_SRC | PATTERN_COPY, false, is_memset);
 871
 872		error_count += dmatest_verify(dst->aligned, dst->off + len,
 873				buf_size, dst->off + len,
 874				PATTERN_DST, false, is_memset);
 875
 876		diff = ktime_sub(ktime_get(), start);
 877		comparetime = ktime_add(comparetime, diff);
 878
 879		if (error_count) {
 880			result("data error", total_tests, src->off, dst->off,
 881			       len, error_count);
 882			failed_tests++;
 883		} else {
 884			verbose_result("test passed", total_tests, src->off,
 885				       dst->off, len, 0);
 886		}
 887
 888		continue;
 889
 890error_unmap_continue:
 891		dmaengine_unmap_put(um);
 892		failed_tests++;
 893	}
 894	ktime = ktime_sub(ktime_get(), ktime);
 895	ktime = ktime_sub(ktime, comparetime);
 896	ktime = ktime_sub(ktime, filltime);
 897	runtime = ktime_to_us(ktime);
 898
 899	ret = 0;
 900	kfree(dma_pq);
 901err_srcs_array:
 902	kfree(srcs);
 903err_dst:
 904	dmatest_free_test_data(dst);
 905err_src:
 906	dmatest_free_test_data(src);
 907err_free_coefs:
 908	kfree(pq_coefs);
 909err_thread_type:
 910	iops = dmatest_persec(runtime, total_tests);
 911	pr_info("%s: summary %u tests, %u failures %llu.%02llu iops %llu KB/s (%d)\n",
 912		current->comm, total_tests, failed_tests,
 913		FIXPT_TO_INT(iops), FIXPT_GET_FRAC(iops),
 914		dmatest_KBs(runtime, total_len), ret);
 915
 916	/* terminate all transfers on specified channels */
 917	if (ret || failed_tests)
 918		dmaengine_terminate_sync(chan);
 919
 920	thread->done = true;
 921	wake_up(&thread_wait);
 922
 923	return ret;
 924}
 925
 926static void dmatest_cleanup_channel(struct dmatest_chan *dtc)
 927{
 928	struct dmatest_thread	*thread;
 929	struct dmatest_thread	*_thread;
 930	int			ret;
 931
 932	list_for_each_entry_safe(thread, _thread, &dtc->threads, node) {
 933		ret = kthread_stop(thread->task);
 934		pr_debug("thread %s exited with status %d\n",
 935			 thread->task->comm, ret);
 936		list_del(&thread->node);
 937		put_task_struct(thread->task);
 938		kfree(thread);
 939	}
 940
 941	/* terminate all transfers on specified channels */
 942	dmaengine_terminate_sync(dtc->chan);
 943
 944	kfree(dtc);
 945}
 946
 947static int dmatest_add_threads(struct dmatest_info *info,
 948		struct dmatest_chan *dtc, enum dma_transaction_type type)
 949{
 950	struct dmatest_params *params = &info->params;
 951	struct dmatest_thread *thread;
 952	struct dma_chan *chan = dtc->chan;
 953	char *op;
 954	unsigned int i;
 955
 956	if (type == DMA_MEMCPY)
 957		op = "copy";
 958	else if (type == DMA_MEMSET)
 959		op = "set";
 960	else if (type == DMA_XOR)
 961		op = "xor";
 962	else if (type == DMA_PQ)
 963		op = "pq";
 964	else
 965		return -EINVAL;
 966
 967	for (i = 0; i < params->threads_per_chan; i++) {
 968		thread = kzalloc(sizeof(struct dmatest_thread), GFP_KERNEL);
 969		if (!thread) {
 970			pr_warn("No memory for %s-%s%u\n",
 971				dma_chan_name(chan), op, i);
 972			break;
 973		}
 974		thread->info = info;
 975		thread->chan = dtc->chan;
 976		thread->type = type;
 977		thread->test_done.wait = &thread->done_wait;
 978		init_waitqueue_head(&thread->done_wait);
 979		smp_wmb();
 980		thread->task = kthread_create(dmatest_func, thread, "%s-%s%u",
 981				dma_chan_name(chan), op, i);
 982		if (IS_ERR(thread->task)) {
 983			pr_warn("Failed to create thread %s-%s%u\n",
 984				dma_chan_name(chan), op, i);
 985			kfree(thread);
 986			break;
 987		}
 988
 989		/* srcbuf and dstbuf are allocated by the thread itself */
 990		get_task_struct(thread->task);
 991		list_add_tail(&thread->node, &dtc->threads);
 992		thread->pending = true;
 993	}
 994
 995	return i;
 996}
 997
 998static int dmatest_add_channel(struct dmatest_info *info,
 999		struct dma_chan *chan)
1000{
1001	struct dmatest_chan	*dtc;
1002	struct dma_device	*dma_dev = chan->device;
1003	unsigned int		thread_count = 0;
1004	int cnt;
1005
1006	dtc = kmalloc(sizeof(struct dmatest_chan), GFP_KERNEL);
1007	if (!dtc) {
1008		pr_warn("No memory for %s\n", dma_chan_name(chan));
1009		return -ENOMEM;
1010	}
1011
1012	dtc->chan = chan;
1013	INIT_LIST_HEAD(&dtc->threads);
1014
1015	if (dma_has_cap(DMA_COMPLETION_NO_ORDER, dma_dev->cap_mask) &&
1016	    info->params.polled) {
1017		info->params.polled = false;
1018		pr_warn("DMA_COMPLETION_NO_ORDER, polled disabled\n");
1019	}
1020
1021	if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask)) {
1022		if (dmatest == 0) {
1023			cnt = dmatest_add_threads(info, dtc, DMA_MEMCPY);
1024			thread_count += cnt > 0 ? cnt : 0;
1025		}
1026	}
1027
1028	if (dma_has_cap(DMA_MEMSET, dma_dev->cap_mask)) {
1029		if (dmatest == 1) {
1030			cnt = dmatest_add_threads(info, dtc, DMA_MEMSET);
1031			thread_count += cnt > 0 ? cnt : 0;
1032		}
1033	}
1034
1035	if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1036		cnt = dmatest_add_threads(info, dtc, DMA_XOR);
1037		thread_count += cnt > 0 ? cnt : 0;
1038	}
1039	if (dma_has_cap(DMA_PQ, dma_dev->cap_mask)) {
1040		cnt = dmatest_add_threads(info, dtc, DMA_PQ);
1041		thread_count += cnt > 0 ? cnt : 0;
1042	}
1043
1044	pr_info("Added %u threads using %s\n",
1045		thread_count, dma_chan_name(chan));
1046
1047	list_add_tail(&dtc->node, &info->channels);
1048	info->nr_channels++;
1049
1050	return 0;
1051}
1052
1053static bool filter(struct dma_chan *chan, void *param)
1054{
1055	struct dmatest_params *params = param;
1056
1057	if (!dmatest_match_channel(params, chan) ||
1058	    !dmatest_match_device(params, chan->device))
1059		return false;
1060	else
1061		return true;
1062}
1063
1064static void request_channels(struct dmatest_info *info,
1065			     enum dma_transaction_type type)
1066{
1067	dma_cap_mask_t mask;
1068
1069	dma_cap_zero(mask);
1070	dma_cap_set(type, mask);
1071	for (;;) {
1072		struct dmatest_params *params = &info->params;
1073		struct dma_chan *chan;
1074
1075		chan = dma_request_channel(mask, filter, params);
1076		if (chan) {
1077			if (dmatest_add_channel(info, chan)) {
1078				dma_release_channel(chan);
1079				break; /* add_channel failed, punt */
1080			}
1081		} else
1082			break; /* no more channels available */
1083		if (params->max_channels &&
1084		    info->nr_channels >= params->max_channels)
1085			break; /* we have all we need */
1086	}
1087}
1088
1089static void add_threaded_test(struct dmatest_info *info)
1090{
1091	struct dmatest_params *params = &info->params;
1092
1093	/* Copy test parameters */
1094	params->buf_size = test_buf_size;
1095	strlcpy(params->channel, strim(test_channel), sizeof(params->channel));
1096	strlcpy(params->device, strim(test_device), sizeof(params->device));
1097	params->threads_per_chan = threads_per_chan;
1098	params->max_channels = max_channels;
1099	params->iterations = iterations;
1100	params->xor_sources = xor_sources;
1101	params->pq_sources = pq_sources;
1102	params->timeout = timeout;
1103	params->noverify = noverify;
1104	params->norandom = norandom;
1105	params->alignment = alignment;
1106	params->transfer_size = transfer_size;
1107	params->polled = polled;
1108
1109	request_channels(info, DMA_MEMCPY);
1110	request_channels(info, DMA_MEMSET);
1111	request_channels(info, DMA_XOR);
1112	request_channels(info, DMA_PQ);
1113}
1114
1115static void run_pending_tests(struct dmatest_info *info)
1116{
1117	struct dmatest_chan *dtc;
1118	unsigned int thread_count = 0;
1119
1120	list_for_each_entry(dtc, &info->channels, node) {
1121		struct dmatest_thread *thread;
1122
1123		thread_count = 0;
1124		list_for_each_entry(thread, &dtc->threads, node) {
1125			wake_up_process(thread->task);
1126			thread_count++;
1127		}
1128		pr_info("Started %u threads using %s\n",
1129			thread_count, dma_chan_name(dtc->chan));
1130	}
1131}
1132
1133static void stop_threaded_test(struct dmatest_info *info)
1134{
1135	struct dmatest_chan *dtc, *_dtc;
1136	struct dma_chan *chan;
1137
1138	list_for_each_entry_safe(dtc, _dtc, &info->channels, node) {
1139		list_del(&dtc->node);
1140		chan = dtc->chan;
1141		dmatest_cleanup_channel(dtc);
1142		pr_debug("dropped channel %s\n", dma_chan_name(chan));
1143		dma_release_channel(chan);
1144	}
1145
1146	info->nr_channels = 0;
1147}
1148
1149static void start_threaded_tests(struct dmatest_info *info)
1150{
1151	/* we might be called early to set run=, defer running until all
1152	 * parameters have been evaluated
1153	 */
1154	if (!info->did_init)
1155		return;
1156
1157	run_pending_tests(info);
1158}
1159
1160static int dmatest_run_get(char *val, const struct kernel_param *kp)
1161{
1162	struct dmatest_info *info = &test_info;
1163
1164	mutex_lock(&info->lock);
1165	if (is_threaded_test_run(info)) {
1166		dmatest_run = true;
1167	} else {
1168		if (!is_threaded_test_pending(info))
1169			stop_threaded_test(info);
1170		dmatest_run = false;
1171	}
1172	mutex_unlock(&info->lock);
1173
1174	return param_get_bool(val, kp);
1175}
1176
1177static int dmatest_run_set(const char *val, const struct kernel_param *kp)
1178{
1179	struct dmatest_info *info = &test_info;
1180	int ret;
1181
1182	mutex_lock(&info->lock);
1183	ret = param_set_bool(val, kp);
1184	if (ret) {
1185		mutex_unlock(&info->lock);
1186		return ret;
1187	} else if (dmatest_run) {
1188		if (!is_threaded_test_pending(info)) {
1189			/*
1190			 * We have nothing to run. This can be due to:
1191			 */
1192			ret = info->last_error;
1193			if (ret) {
1194				/* 1) Misconfiguration */
1195				pr_err("Channel misconfigured, can't continue\n");
1196				mutex_unlock(&info->lock);
1197				return ret;
1198			} else {
1199				/* 2) We rely on defaults */
1200				pr_info("No channels configured, continue with any\n");
1201				if (!is_threaded_test_run(info))
1202					stop_threaded_test(info);
1203				add_threaded_test(info);
1204			}
1205		}
1206		start_threaded_tests(info);
1207	} else {
1208		stop_threaded_test(info);
1209	}
1210
1211	mutex_unlock(&info->lock);
1212
1213	return ret;
1214}
1215
1216static int dmatest_chan_set(const char *val, const struct kernel_param *kp)
1217{
1218	struct dmatest_info *info = &test_info;
1219	struct dmatest_chan *dtc;
1220	char chan_reset_val[20];
1221	int ret;
1222
1223	mutex_lock(&info->lock);
1224	ret = param_set_copystring(val, kp);
1225	if (ret) {
1226		mutex_unlock(&info->lock);
1227		return ret;
1228	}
1229	/*Clear any previously run threads */
1230	if (!is_threaded_test_run(info) && !is_threaded_test_pending(info))
1231		stop_threaded_test(info);
1232	/* Reject channels that are already registered */
1233	if (is_threaded_test_pending(info)) {
1234		list_for_each_entry(dtc, &info->channels, node) {
1235			if (strcmp(dma_chan_name(dtc->chan),
1236				   strim(test_channel)) == 0) {
1237				dtc = list_last_entry(&info->channels,
1238						      struct dmatest_chan,
1239						      node);
1240				strlcpy(chan_reset_val,
1241					dma_chan_name(dtc->chan),
1242					sizeof(chan_reset_val));
1243				ret = -EBUSY;
1244				goto add_chan_err;
1245			}
1246		}
1247	}
1248
1249	add_threaded_test(info);
1250
1251	/* Check if channel was added successfully */
1252	dtc = list_last_entry(&info->channels, struct dmatest_chan, node);
1253
1254	if (dtc->chan) {
1255		/*
1256		 * if new channel was not successfully added, revert the
1257		 * "test_channel" string to the name of the last successfully
1258		 * added channel. exception for when users issues empty string
1259		 * to channel parameter.
1260		 */
 
1261		if ((strcmp(dma_chan_name(dtc->chan), strim(test_channel)) != 0)
1262		    && (strcmp("", strim(test_channel)) != 0)) {
1263			ret = -EINVAL;
1264			strlcpy(chan_reset_val, dma_chan_name(dtc->chan),
1265				sizeof(chan_reset_val));
1266			goto add_chan_err;
1267		}
1268
1269	} else {
1270		/* Clear test_channel if no channels were added successfully */
1271		strlcpy(chan_reset_val, "", sizeof(chan_reset_val));
1272		ret = -EBUSY;
1273		goto add_chan_err;
1274	}
1275
1276	info->last_error = ret;
1277	mutex_unlock(&info->lock);
1278
1279	return ret;
1280
1281add_chan_err:
1282	param_set_copystring(chan_reset_val, kp);
1283	info->last_error = ret;
1284	mutex_unlock(&info->lock);
1285
1286	return ret;
1287}
1288
1289static int dmatest_chan_get(char *val, const struct kernel_param *kp)
1290{
1291	struct dmatest_info *info = &test_info;
1292
1293	mutex_lock(&info->lock);
1294	if (!is_threaded_test_run(info) && !is_threaded_test_pending(info)) {
1295		stop_threaded_test(info);
1296		strlcpy(test_channel, "", sizeof(test_channel));
1297	}
1298	mutex_unlock(&info->lock);
1299
1300	return param_get_string(val, kp);
1301}
1302
1303static int dmatest_test_list_get(char *val, const struct kernel_param *kp)
1304{
1305	struct dmatest_info *info = &test_info;
1306	struct dmatest_chan *dtc;
1307	unsigned int thread_count = 0;
1308
1309	list_for_each_entry(dtc, &info->channels, node) {
1310		struct dmatest_thread *thread;
1311
1312		thread_count = 0;
1313		list_for_each_entry(thread, &dtc->threads, node) {
1314			thread_count++;
1315		}
1316		pr_info("%u threads using %s\n",
1317			thread_count, dma_chan_name(dtc->chan));
1318	}
1319
1320	return 0;
1321}
1322
1323static int __init dmatest_init(void)
1324{
1325	struct dmatest_info *info = &test_info;
1326	struct dmatest_params *params = &info->params;
1327
1328	if (dmatest_run) {
1329		mutex_lock(&info->lock);
1330		add_threaded_test(info);
1331		run_pending_tests(info);
1332		mutex_unlock(&info->lock);
1333	}
1334
1335	if (params->iterations && wait)
1336		wait_event(thread_wait, !is_threaded_test_run(info));
1337
1338	/* module parameters are stable, inittime tests are started,
1339	 * let userspace take over 'run' control
1340	 */
1341	info->did_init = true;
1342
1343	return 0;
1344}
1345/* when compiled-in wait for drivers to load first */
1346late_initcall(dmatest_init);
1347
1348static void __exit dmatest_exit(void)
1349{
1350	struct dmatest_info *info = &test_info;
1351
1352	mutex_lock(&info->lock);
1353	stop_threaded_test(info);
1354	mutex_unlock(&info->lock);
1355}
1356module_exit(dmatest_exit);
1357
1358MODULE_AUTHOR("Haavard Skinnemoen (Atmel)");
1359MODULE_LICENSE("GPL v2");
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * DMA Engine test module
   4 *
   5 * Copyright (C) 2007 Atmel Corporation
   6 * Copyright (C) 2013 Intel Corporation
   7 */
   8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10#include <linux/err.h>
  11#include <linux/delay.h>
  12#include <linux/dma-mapping.h>
  13#include <linux/dmaengine.h>
  14#include <linux/freezer.h>
  15#include <linux/init.h>
  16#include <linux/kthread.h>
  17#include <linux/sched/task.h>
  18#include <linux/module.h>
  19#include <linux/moduleparam.h>
  20#include <linux/random.h>
  21#include <linux/slab.h>
  22#include <linux/wait.h>
  23
  24static unsigned int test_buf_size = 16384;
  25module_param(test_buf_size, uint, 0644);
  26MODULE_PARM_DESC(test_buf_size, "Size of the memcpy test buffer");
  27
  28static char test_device[32];
  29module_param_string(device, test_device, sizeof(test_device), 0644);
 
  30MODULE_PARM_DESC(device, "Bus ID of the DMA Engine to test (default: any)");
  31
  32static unsigned int threads_per_chan = 1;
  33module_param(threads_per_chan, uint, 0644);
  34MODULE_PARM_DESC(threads_per_chan,
  35		"Number of threads to start per channel (default: 1)");
  36
  37static unsigned int max_channels;
  38module_param(max_channels, uint, 0644);
  39MODULE_PARM_DESC(max_channels,
  40		"Maximum number of channels to use (default: all)");
  41
  42static unsigned int iterations;
  43module_param(iterations, uint, 0644);
  44MODULE_PARM_DESC(iterations,
  45		"Iterations before stopping test (default: infinite)");
  46
  47static unsigned int dmatest;
  48module_param(dmatest, uint, 0644);
  49MODULE_PARM_DESC(dmatest,
  50		"dmatest 0-memcpy 1-memset (default: 0)");
  51
  52static unsigned int xor_sources = 3;
  53module_param(xor_sources, uint, 0644);
  54MODULE_PARM_DESC(xor_sources,
  55		"Number of xor source buffers (default: 3)");
  56
  57static unsigned int pq_sources = 3;
  58module_param(pq_sources, uint, 0644);
  59MODULE_PARM_DESC(pq_sources,
  60		"Number of p+q source buffers (default: 3)");
  61
  62static int timeout = 3000;
  63module_param(timeout, int, 0644);
  64MODULE_PARM_DESC(timeout, "Transfer Timeout in msec (default: 3000), "
  65		 "Pass -1 for infinite timeout");
  66
  67static bool noverify;
  68module_param(noverify, bool, 0644);
  69MODULE_PARM_DESC(noverify, "Disable data verification (default: verify)");
  70
  71static bool norandom;
  72module_param(norandom, bool, 0644);
  73MODULE_PARM_DESC(norandom, "Disable random offset setup (default: random)");
  74
  75static bool verbose;
  76module_param(verbose, bool, 0644);
  77MODULE_PARM_DESC(verbose, "Enable \"success\" result messages (default: off)");
  78
  79static int alignment = -1;
  80module_param(alignment, int, 0644);
  81MODULE_PARM_DESC(alignment, "Custom data address alignment taken as 2^(alignment) (default: not used (-1))");
  82
  83static unsigned int transfer_size;
  84module_param(transfer_size, uint, 0644);
  85MODULE_PARM_DESC(transfer_size, "Optional custom transfer size in bytes (default: not used (0))");
  86
  87static bool polled;
  88module_param(polled, bool, 0644);
  89MODULE_PARM_DESC(polled, "Use polling for completion instead of interrupts");
  90
  91/**
  92 * struct dmatest_params - test parameters.
  93 * @buf_size:		size of the memcpy test buffer
  94 * @channel:		bus ID of the channel to test
  95 * @device:		bus ID of the DMA Engine to test
  96 * @threads_per_chan:	number of threads to start per channel
  97 * @max_channels:	maximum number of channels to use
  98 * @iterations:		iterations before stopping test
  99 * @xor_sources:	number of xor source buffers
 100 * @pq_sources:		number of p+q source buffers
 101 * @timeout:		transfer timeout in msec, -1 for infinite timeout
 102 * @noverify:		disable data verification
 103 * @norandom:		disable random offset setup
 104 * @alignment:		custom data address alignment taken as 2^alignment
 105 * @transfer_size:	custom transfer size in bytes
 106 * @polled:		use polling for completion instead of interrupts
 107 */
 108struct dmatest_params {
 109	unsigned int	buf_size;
 110	char		channel[20];
 111	char		device[32];
 112	unsigned int	threads_per_chan;
 113	unsigned int	max_channels;
 114	unsigned int	iterations;
 115	unsigned int	xor_sources;
 116	unsigned int	pq_sources;
 117	int		timeout;
 118	bool		noverify;
 119	bool		norandom;
 120	int		alignment;
 121	unsigned int	transfer_size;
 122	bool		polled;
 123};
 124
 125/**
 126 * struct dmatest_info - test information.
 127 * @params:		test parameters
 128 * @channels:		channels under test
 129 * @nr_channels:	number of channels under test
 130 * @lock:		access protection to the fields of this structure
 131 * @did_init:		module has been initialized completely
 132 * @last_error:		test has faced configuration issues
 133 */
 134static struct dmatest_info {
 135	/* Test parameters */
 136	struct dmatest_params	params;
 137
 138	/* Internal state */
 139	struct list_head	channels;
 140	unsigned int		nr_channels;
 141	int			last_error;
 142	struct mutex		lock;
 143	bool			did_init;
 144} test_info = {
 145	.channels = LIST_HEAD_INIT(test_info.channels),
 146	.lock = __MUTEX_INITIALIZER(test_info.lock),
 147};
 148
 149static int dmatest_run_set(const char *val, const struct kernel_param *kp);
 150static int dmatest_run_get(char *val, const struct kernel_param *kp);
 151static const struct kernel_param_ops run_ops = {
 152	.set = dmatest_run_set,
 153	.get = dmatest_run_get,
 154};
 155static bool dmatest_run;
 156module_param_cb(run, &run_ops, &dmatest_run, 0644);
 157MODULE_PARM_DESC(run, "Run the test (default: false)");
 158
 159static int dmatest_chan_set(const char *val, const struct kernel_param *kp);
 160static int dmatest_chan_get(char *val, const struct kernel_param *kp);
 161static const struct kernel_param_ops multi_chan_ops = {
 162	.set = dmatest_chan_set,
 163	.get = dmatest_chan_get,
 164};
 165
 166static char test_channel[20];
 167static struct kparam_string newchan_kps = {
 168	.string = test_channel,
 169	.maxlen = 20,
 170};
 171module_param_cb(channel, &multi_chan_ops, &newchan_kps, 0644);
 172MODULE_PARM_DESC(channel, "Bus ID of the channel to test (default: any)");
 173
 174static int dmatest_test_list_get(char *val, const struct kernel_param *kp);
 175static const struct kernel_param_ops test_list_ops = {
 176	.get = dmatest_test_list_get,
 177};
 178module_param_cb(test_list, &test_list_ops, NULL, 0444);
 179MODULE_PARM_DESC(test_list, "Print current test list");
 180
 181/* Maximum amount of mismatched bytes in buffer to print */
 182#define MAX_ERROR_COUNT		32
 183
 184/*
 185 * Initialization patterns. All bytes in the source buffer has bit 7
 186 * set, all bytes in the destination buffer has bit 7 cleared.
 187 *
 188 * Bit 6 is set for all bytes which are to be copied by the DMA
 189 * engine. Bit 5 is set for all bytes which are to be overwritten by
 190 * the DMA engine.
 191 *
 192 * The remaining bits are the inverse of a counter which increments by
 193 * one for each byte address.
 194 */
 195#define PATTERN_SRC		0x80
 196#define PATTERN_DST		0x00
 197#define PATTERN_COPY		0x40
 198#define PATTERN_OVERWRITE	0x20
 199#define PATTERN_COUNT_MASK	0x1f
 200#define PATTERN_MEMSET_IDX	0x01
 201
 202/* Fixed point arithmetic ops */
 203#define FIXPT_SHIFT		8
 204#define FIXPNT_MASK		0xFF
 205#define FIXPT_TO_INT(a)	((a) >> FIXPT_SHIFT)
 206#define INT_TO_FIXPT(a)	((a) << FIXPT_SHIFT)
 207#define FIXPT_GET_FRAC(a)	((((a) & FIXPNT_MASK) * 100) >> FIXPT_SHIFT)
 208
 209/* poor man's completion - we want to use wait_event_freezable() on it */
 210struct dmatest_done {
 211	bool			done;
 212	wait_queue_head_t	*wait;
 213};
 214
 215struct dmatest_data {
 216	u8		**raw;
 217	u8		**aligned;
 218	unsigned int	cnt;
 219	unsigned int	off;
 220};
 221
 222struct dmatest_thread {
 223	struct list_head	node;
 224	struct dmatest_info	*info;
 225	struct task_struct	*task;
 226	struct dma_chan		*chan;
 227	struct dmatest_data	src;
 228	struct dmatest_data	dst;
 229	enum dma_transaction_type type;
 230	wait_queue_head_t done_wait;
 231	struct dmatest_done test_done;
 232	bool			done;
 233	bool			pending;
 234};
 235
 236struct dmatest_chan {
 237	struct list_head	node;
 238	struct dma_chan		*chan;
 239	struct list_head	threads;
 240};
 241
 242static DECLARE_WAIT_QUEUE_HEAD(thread_wait);
 243static bool wait;
 244
 245static bool is_threaded_test_run(struct dmatest_info *info)
 246{
 247	struct dmatest_chan *dtc;
 248
 249	list_for_each_entry(dtc, &info->channels, node) {
 250		struct dmatest_thread *thread;
 251
 252		list_for_each_entry(thread, &dtc->threads, node) {
 253			if (!thread->done && !thread->pending)
 254				return true;
 255		}
 256	}
 257
 258	return false;
 259}
 260
 261static bool is_threaded_test_pending(struct dmatest_info *info)
 262{
 263	struct dmatest_chan *dtc;
 264
 265	list_for_each_entry(dtc, &info->channels, node) {
 266		struct dmatest_thread *thread;
 267
 268		list_for_each_entry(thread, &dtc->threads, node) {
 269			if (thread->pending)
 270				return true;
 271		}
 272	}
 273
 274	return false;
 275}
 276
 277static int dmatest_wait_get(char *val, const struct kernel_param *kp)
 278{
 279	struct dmatest_info *info = &test_info;
 280	struct dmatest_params *params = &info->params;
 281
 282	if (params->iterations)
 283		wait_event(thread_wait, !is_threaded_test_run(info));
 284	wait = true;
 285	return param_get_bool(val, kp);
 286}
 287
 288static const struct kernel_param_ops wait_ops = {
 289	.get = dmatest_wait_get,
 290	.set = param_set_bool,
 291};
 292module_param_cb(wait, &wait_ops, &wait, 0444);
 293MODULE_PARM_DESC(wait, "Wait for tests to complete (default: false)");
 294
 295static bool dmatest_match_channel(struct dmatest_params *params,
 296		struct dma_chan *chan)
 297{
 298	if (params->channel[0] == '\0')
 299		return true;
 300	return strcmp(dma_chan_name(chan), params->channel) == 0;
 301}
 302
 303static bool dmatest_match_device(struct dmatest_params *params,
 304		struct dma_device *device)
 305{
 306	if (params->device[0] == '\0')
 307		return true;
 308	return strcmp(dev_name(device->dev), params->device) == 0;
 309}
 310
 311static unsigned long dmatest_random(void)
 312{
 313	unsigned long buf;
 314
 315	get_random_bytes(&buf, sizeof(buf));
 316	return buf;
 317}
 318
 319static inline u8 gen_inv_idx(u8 index, bool is_memset)
 320{
 321	u8 val = is_memset ? PATTERN_MEMSET_IDX : index;
 322
 323	return ~val & PATTERN_COUNT_MASK;
 324}
 325
 326static inline u8 gen_src_value(u8 index, bool is_memset)
 327{
 328	return PATTERN_SRC | gen_inv_idx(index, is_memset);
 329}
 330
 331static inline u8 gen_dst_value(u8 index, bool is_memset)
 332{
 333	return PATTERN_DST | gen_inv_idx(index, is_memset);
 334}
 335
 336static void dmatest_init_srcs(u8 **bufs, unsigned int start, unsigned int len,
 337		unsigned int buf_size, bool is_memset)
 338{
 339	unsigned int i;
 340	u8 *buf;
 341
 342	for (; (buf = *bufs); bufs++) {
 343		for (i = 0; i < start; i++)
 344			buf[i] = gen_src_value(i, is_memset);
 345		for ( ; i < start + len; i++)
 346			buf[i] = gen_src_value(i, is_memset) | PATTERN_COPY;
 347		for ( ; i < buf_size; i++)
 348			buf[i] = gen_src_value(i, is_memset);
 349		buf++;
 350	}
 351}
 352
 353static void dmatest_init_dsts(u8 **bufs, unsigned int start, unsigned int len,
 354		unsigned int buf_size, bool is_memset)
 355{
 356	unsigned int i;
 357	u8 *buf;
 358
 359	for (; (buf = *bufs); bufs++) {
 360		for (i = 0; i < start; i++)
 361			buf[i] = gen_dst_value(i, is_memset);
 362		for ( ; i < start + len; i++)
 363			buf[i] = gen_dst_value(i, is_memset) |
 364						PATTERN_OVERWRITE;
 365		for ( ; i < buf_size; i++)
 366			buf[i] = gen_dst_value(i, is_memset);
 367	}
 368}
 369
 370static void dmatest_mismatch(u8 actual, u8 pattern, unsigned int index,
 371		unsigned int counter, bool is_srcbuf, bool is_memset)
 372{
 373	u8		diff = actual ^ pattern;
 374	u8		expected = pattern | gen_inv_idx(counter, is_memset);
 375	const char	*thread_name = current->comm;
 376
 377	if (is_srcbuf)
 378		pr_warn("%s: srcbuf[0x%x] overwritten! Expected %02x, got %02x\n",
 379			thread_name, index, expected, actual);
 380	else if ((pattern & PATTERN_COPY)
 381			&& (diff & (PATTERN_COPY | PATTERN_OVERWRITE)))
 382		pr_warn("%s: dstbuf[0x%x] not copied! Expected %02x, got %02x\n",
 383			thread_name, index, expected, actual);
 384	else if (diff & PATTERN_SRC)
 385		pr_warn("%s: dstbuf[0x%x] was copied! Expected %02x, got %02x\n",
 386			thread_name, index, expected, actual);
 387	else
 388		pr_warn("%s: dstbuf[0x%x] mismatch! Expected %02x, got %02x\n",
 389			thread_name, index, expected, actual);
 390}
 391
 392static unsigned int dmatest_verify(u8 **bufs, unsigned int start,
 393		unsigned int end, unsigned int counter, u8 pattern,
 394		bool is_srcbuf, bool is_memset)
 395{
 396	unsigned int i;
 397	unsigned int error_count = 0;
 398	u8 actual;
 399	u8 expected;
 400	u8 *buf;
 401	unsigned int counter_orig = counter;
 402
 403	for (; (buf = *bufs); bufs++) {
 404		counter = counter_orig;
 405		for (i = start; i < end; i++) {
 406			actual = buf[i];
 407			expected = pattern | gen_inv_idx(counter, is_memset);
 408			if (actual != expected) {
 409				if (error_count < MAX_ERROR_COUNT)
 410					dmatest_mismatch(actual, pattern, i,
 411							 counter, is_srcbuf,
 412							 is_memset);
 413				error_count++;
 414			}
 415			counter++;
 416		}
 417	}
 418
 419	if (error_count > MAX_ERROR_COUNT)
 420		pr_warn("%s: %u errors suppressed\n",
 421			current->comm, error_count - MAX_ERROR_COUNT);
 422
 423	return error_count;
 424}
 425
 426
 427static void dmatest_callback(void *arg)
 428{
 429	struct dmatest_done *done = arg;
 430	struct dmatest_thread *thread =
 431		container_of(done, struct dmatest_thread, test_done);
 432	if (!thread->done) {
 433		done->done = true;
 434		wake_up_all(done->wait);
 435	} else {
 436		/*
 437		 * If thread->done, it means that this callback occurred
 438		 * after the parent thread has cleaned up. This can
 439		 * happen in the case that driver doesn't implement
 440		 * the terminate_all() functionality and a dma operation
 441		 * did not occur within the timeout period
 442		 */
 443		WARN(1, "dmatest: Kernel memory may be corrupted!!\n");
 444	}
 445}
 446
 447static unsigned int min_odd(unsigned int x, unsigned int y)
 448{
 449	unsigned int val = min(x, y);
 450
 451	return val % 2 ? val : val - 1;
 452}
 453
 454static void result(const char *err, unsigned int n, unsigned int src_off,
 455		   unsigned int dst_off, unsigned int len, unsigned long data)
 456{
 457	if (IS_ERR_VALUE(data)) {
 458		pr_info("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%ld)\n",
 459			current->comm, n, err, src_off, dst_off, len, data);
 460	} else {
 461		pr_info("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%lu)\n",
 462			current->comm, n, err, src_off, dst_off, len, data);
 463	}
 464}
 465
 466static void dbg_result(const char *err, unsigned int n, unsigned int src_off,
 467		       unsigned int dst_off, unsigned int len,
 468		       unsigned long data)
 469{
 470	pr_debug("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%lu)\n",
 471		 current->comm, n, err, src_off, dst_off, len, data);
 472}
 473
 474#define verbose_result(err, n, src_off, dst_off, len, data) ({	\
 475	if (verbose)						\
 476		result(err, n, src_off, dst_off, len, data);	\
 477	else							\
 478		dbg_result(err, n, src_off, dst_off, len, data);\
 479})
 480
 481static unsigned long long dmatest_persec(s64 runtime, unsigned int val)
 482{
 483	unsigned long long per_sec = 1000000;
 484
 485	if (runtime <= 0)
 486		return 0;
 487
 488	/* drop precision until runtime is 32-bits */
 489	while (runtime > UINT_MAX) {
 490		runtime >>= 1;
 491		per_sec <<= 1;
 492	}
 493
 494	per_sec *= val;
 495	per_sec = INT_TO_FIXPT(per_sec);
 496	do_div(per_sec, runtime);
 497
 498	return per_sec;
 499}
 500
 501static unsigned long long dmatest_KBs(s64 runtime, unsigned long long len)
 502{
 503	return FIXPT_TO_INT(dmatest_persec(runtime, len >> 10));
 504}
 505
 506static void __dmatest_free_test_data(struct dmatest_data *d, unsigned int cnt)
 507{
 508	unsigned int i;
 509
 510	for (i = 0; i < cnt; i++)
 511		kfree(d->raw[i]);
 512
 513	kfree(d->aligned);
 514	kfree(d->raw);
 515}
 516
 517static void dmatest_free_test_data(struct dmatest_data *d)
 518{
 519	__dmatest_free_test_data(d, d->cnt);
 520}
 521
 522static int dmatest_alloc_test_data(struct dmatest_data *d,
 523		unsigned int buf_size, u8 align)
 524{
 525	unsigned int i = 0;
 526
 527	d->raw = kcalloc(d->cnt + 1, sizeof(u8 *), GFP_KERNEL);
 528	if (!d->raw)
 529		return -ENOMEM;
 530
 531	d->aligned = kcalloc(d->cnt + 1, sizeof(u8 *), GFP_KERNEL);
 532	if (!d->aligned)
 533		goto err;
 534
 535	for (i = 0; i < d->cnt; i++) {
 536		d->raw[i] = kmalloc(buf_size + align, GFP_KERNEL);
 537		if (!d->raw[i])
 538			goto err;
 539
 540		/* align to alignment restriction */
 541		if (align)
 542			d->aligned[i] = PTR_ALIGN(d->raw[i], align);
 543		else
 544			d->aligned[i] = d->raw[i];
 545	}
 546
 547	return 0;
 548err:
 549	__dmatest_free_test_data(d, i);
 550	return -ENOMEM;
 551}
 552
 553/*
 554 * This function repeatedly tests DMA transfers of various lengths and
 555 * offsets for a given operation type until it is told to exit by
 556 * kthread_stop(). There may be multiple threads running this function
 557 * in parallel for a single channel, and there may be multiple channels
 558 * being tested in parallel.
 559 *
 560 * Before each test, the source and destination buffer is initialized
 561 * with a known pattern. This pattern is different depending on
 562 * whether it's in an area which is supposed to be copied or
 563 * overwritten, and different in the source and destination buffers.
 564 * So if the DMA engine doesn't copy exactly what we tell it to copy,
 565 * we'll notice.
 566 */
 567static int dmatest_func(void *data)
 568{
 569	struct dmatest_thread	*thread = data;
 570	struct dmatest_done	*done = &thread->test_done;
 571	struct dmatest_info	*info;
 572	struct dmatest_params	*params;
 573	struct dma_chan		*chan;
 574	struct dma_device	*dev;
 575	struct device		*dma_dev;
 576	unsigned int		error_count;
 577	unsigned int		failed_tests = 0;
 578	unsigned int		total_tests = 0;
 579	dma_cookie_t		cookie;
 580	enum dma_status		status;
 581	enum dma_ctrl_flags	flags;
 582	u8			*pq_coefs = NULL;
 583	int			ret;
 584	unsigned int		buf_size;
 585	struct dmatest_data	*src;
 586	struct dmatest_data	*dst;
 587	int			i;
 588	ktime_t			ktime, start, diff;
 589	ktime_t			filltime = 0;
 590	ktime_t			comparetime = 0;
 591	s64			runtime = 0;
 592	unsigned long long	total_len = 0;
 593	unsigned long long	iops = 0;
 594	u8			align = 0;
 595	bool			is_memset = false;
 596	dma_addr_t		*srcs;
 597	dma_addr_t		*dma_pq;
 598
 599	set_freezable();
 600
 601	ret = -ENOMEM;
 602
 603	smp_rmb();
 604	thread->pending = false;
 605	info = thread->info;
 606	params = &info->params;
 607	chan = thread->chan;
 608	dev = chan->device;
 609	dma_dev = dmaengine_get_dma_device(chan);
 610
 611	src = &thread->src;
 612	dst = &thread->dst;
 613	if (thread->type == DMA_MEMCPY) {
 614		align = params->alignment < 0 ? dev->copy_align :
 615						params->alignment;
 616		src->cnt = dst->cnt = 1;
 617	} else if (thread->type == DMA_MEMSET) {
 618		align = params->alignment < 0 ? dev->fill_align :
 619						params->alignment;
 620		src->cnt = dst->cnt = 1;
 621		is_memset = true;
 622	} else if (thread->type == DMA_XOR) {
 623		/* force odd to ensure dst = src */
 624		src->cnt = min_odd(params->xor_sources | 1, dev->max_xor);
 625		dst->cnt = 1;
 626		align = params->alignment < 0 ? dev->xor_align :
 627						params->alignment;
 628	} else if (thread->type == DMA_PQ) {
 629		/* force odd to ensure dst = src */
 630		src->cnt = min_odd(params->pq_sources | 1, dma_maxpq(dev, 0));
 631		dst->cnt = 2;
 632		align = params->alignment < 0 ? dev->pq_align :
 633						params->alignment;
 634
 635		pq_coefs = kmalloc(params->pq_sources + 1, GFP_KERNEL);
 636		if (!pq_coefs)
 637			goto err_thread_type;
 638
 639		for (i = 0; i < src->cnt; i++)
 640			pq_coefs[i] = 1;
 641	} else
 642		goto err_thread_type;
 643
 644	/* Check if buffer count fits into map count variable (u8) */
 645	if ((src->cnt + dst->cnt) >= 255) {
 646		pr_err("too many buffers (%d of 255 supported)\n",
 647		       src->cnt + dst->cnt);
 648		goto err_free_coefs;
 649	}
 650
 651	buf_size = params->buf_size;
 652	if (1 << align > buf_size) {
 653		pr_err("%u-byte buffer too small for %d-byte alignment\n",
 654		       buf_size, 1 << align);
 655		goto err_free_coefs;
 656	}
 657
 658	if (dmatest_alloc_test_data(src, buf_size, align) < 0)
 659		goto err_free_coefs;
 660
 661	if (dmatest_alloc_test_data(dst, buf_size, align) < 0)
 662		goto err_src;
 663
 664	set_user_nice(current, 10);
 665
 666	srcs = kcalloc(src->cnt, sizeof(dma_addr_t), GFP_KERNEL);
 667	if (!srcs)
 668		goto err_dst;
 669
 670	dma_pq = kcalloc(dst->cnt, sizeof(dma_addr_t), GFP_KERNEL);
 671	if (!dma_pq)
 672		goto err_srcs_array;
 673
 674	/*
 675	 * src and dst buffers are freed by ourselves below
 676	 */
 677	if (params->polled)
 678		flags = DMA_CTRL_ACK;
 679	else
 680		flags = DMA_CTRL_ACK | DMA_PREP_INTERRUPT;
 681
 682	ktime = ktime_get();
 683	while (!(kthread_should_stop() ||
 684	       (params->iterations && total_tests >= params->iterations))) {
 685		struct dma_async_tx_descriptor *tx = NULL;
 686		struct dmaengine_unmap_data *um;
 687		dma_addr_t *dsts;
 688		unsigned int len;
 689
 690		total_tests++;
 691
 692		if (params->transfer_size) {
 693			if (params->transfer_size >= buf_size) {
 694				pr_err("%u-byte transfer size must be lower than %u-buffer size\n",
 695				       params->transfer_size, buf_size);
 696				break;
 697			}
 698			len = params->transfer_size;
 699		} else if (params->norandom) {
 700			len = buf_size;
 701		} else {
 702			len = dmatest_random() % buf_size + 1;
 703		}
 704
 705		/* Do not alter transfer size explicitly defined by user */
 706		if (!params->transfer_size) {
 707			len = (len >> align) << align;
 708			if (!len)
 709				len = 1 << align;
 710		}
 711		total_len += len;
 712
 713		if (params->norandom) {
 714			src->off = 0;
 715			dst->off = 0;
 716		} else {
 717			src->off = dmatest_random() % (buf_size - len + 1);
 718			dst->off = dmatest_random() % (buf_size - len + 1);
 719
 720			src->off = (src->off >> align) << align;
 721			dst->off = (dst->off >> align) << align;
 722		}
 723
 724		if (!params->noverify) {
 725			start = ktime_get();
 726			dmatest_init_srcs(src->aligned, src->off, len,
 727					  buf_size, is_memset);
 728			dmatest_init_dsts(dst->aligned, dst->off, len,
 729					  buf_size, is_memset);
 730
 731			diff = ktime_sub(ktime_get(), start);
 732			filltime = ktime_add(filltime, diff);
 733		}
 734
 735		um = dmaengine_get_unmap_data(dma_dev, src->cnt + dst->cnt,
 736					      GFP_KERNEL);
 737		if (!um) {
 738			failed_tests++;
 739			result("unmap data NULL", total_tests,
 740			       src->off, dst->off, len, ret);
 741			continue;
 742		}
 743
 744		um->len = buf_size;
 745		for (i = 0; i < src->cnt; i++) {
 746			void *buf = src->aligned[i];
 747			struct page *pg = virt_to_page(buf);
 748			unsigned long pg_off = offset_in_page(buf);
 749
 750			um->addr[i] = dma_map_page(dma_dev, pg, pg_off,
 751						   um->len, DMA_TO_DEVICE);
 752			srcs[i] = um->addr[i] + src->off;
 753			ret = dma_mapping_error(dma_dev, um->addr[i]);
 754			if (ret) {
 755				result("src mapping error", total_tests,
 756				       src->off, dst->off, len, ret);
 757				goto error_unmap_continue;
 758			}
 759			um->to_cnt++;
 760		}
 761		/* map with DMA_BIDIRECTIONAL to force writeback/invalidate */
 762		dsts = &um->addr[src->cnt];
 763		for (i = 0; i < dst->cnt; i++) {
 764			void *buf = dst->aligned[i];
 765			struct page *pg = virt_to_page(buf);
 766			unsigned long pg_off = offset_in_page(buf);
 767
 768			dsts[i] = dma_map_page(dma_dev, pg, pg_off, um->len,
 769					       DMA_BIDIRECTIONAL);
 770			ret = dma_mapping_error(dma_dev, dsts[i]);
 771			if (ret) {
 772				result("dst mapping error", total_tests,
 773				       src->off, dst->off, len, ret);
 774				goto error_unmap_continue;
 775			}
 776			um->bidi_cnt++;
 777		}
 778
 779		if (thread->type == DMA_MEMCPY)
 780			tx = dev->device_prep_dma_memcpy(chan,
 781							 dsts[0] + dst->off,
 782							 srcs[0], len, flags);
 783		else if (thread->type == DMA_MEMSET)
 784			tx = dev->device_prep_dma_memset(chan,
 785						dsts[0] + dst->off,
 786						*(src->aligned[0] + src->off),
 787						len, flags);
 788		else if (thread->type == DMA_XOR)
 789			tx = dev->device_prep_dma_xor(chan,
 790						      dsts[0] + dst->off,
 791						      srcs, src->cnt,
 792						      len, flags);
 793		else if (thread->type == DMA_PQ) {
 794			for (i = 0; i < dst->cnt; i++)
 795				dma_pq[i] = dsts[i] + dst->off;
 796			tx = dev->device_prep_dma_pq(chan, dma_pq, srcs,
 797						     src->cnt, pq_coefs,
 798						     len, flags);
 799		}
 800
 801		if (!tx) {
 802			result("prep error", total_tests, src->off,
 803			       dst->off, len, ret);
 804			msleep(100);
 805			goto error_unmap_continue;
 806		}
 807
 808		done->done = false;
 809		if (!params->polled) {
 810			tx->callback = dmatest_callback;
 811			tx->callback_param = done;
 812		}
 813		cookie = tx->tx_submit(tx);
 814
 815		if (dma_submit_error(cookie)) {
 816			result("submit error", total_tests, src->off,
 817			       dst->off, len, ret);
 818			msleep(100);
 819			goto error_unmap_continue;
 820		}
 821
 822		if (params->polled) {
 823			status = dma_sync_wait(chan, cookie);
 824			dmaengine_terminate_sync(chan);
 825			if (status == DMA_COMPLETE)
 826				done->done = true;
 827		} else {
 828			dma_async_issue_pending(chan);
 829
 830			wait_event_freezable_timeout(thread->done_wait,
 831					done->done,
 832					msecs_to_jiffies(params->timeout));
 833
 834			status = dma_async_is_tx_complete(chan, cookie, NULL,
 835							  NULL);
 836		}
 837
 838		if (!done->done) {
 839			result("test timed out", total_tests, src->off, dst->off,
 840			       len, 0);
 841			goto error_unmap_continue;
 842		} else if (status != DMA_COMPLETE &&
 843			   !(dma_has_cap(DMA_COMPLETION_NO_ORDER,
 844					 dev->cap_mask) &&
 845			     status == DMA_OUT_OF_ORDER)) {
 846			result(status == DMA_ERROR ?
 847			       "completion error status" :
 848			       "completion busy status", total_tests, src->off,
 849			       dst->off, len, ret);
 850			goto error_unmap_continue;
 851		}
 852
 853		dmaengine_unmap_put(um);
 854
 855		if (params->noverify) {
 856			verbose_result("test passed", total_tests, src->off,
 857				       dst->off, len, 0);
 858			continue;
 859		}
 860
 861		start = ktime_get();
 862		pr_debug("%s: verifying source buffer...\n", current->comm);
 863		error_count = dmatest_verify(src->aligned, 0, src->off,
 864				0, PATTERN_SRC, true, is_memset);
 865		error_count += dmatest_verify(src->aligned, src->off,
 866				src->off + len, src->off,
 867				PATTERN_SRC | PATTERN_COPY, true, is_memset);
 868		error_count += dmatest_verify(src->aligned, src->off + len,
 869				buf_size, src->off + len,
 870				PATTERN_SRC, true, is_memset);
 871
 872		pr_debug("%s: verifying dest buffer...\n", current->comm);
 873		error_count += dmatest_verify(dst->aligned, 0, dst->off,
 874				0, PATTERN_DST, false, is_memset);
 875
 876		error_count += dmatest_verify(dst->aligned, dst->off,
 877				dst->off + len, src->off,
 878				PATTERN_SRC | PATTERN_COPY, false, is_memset);
 879
 880		error_count += dmatest_verify(dst->aligned, dst->off + len,
 881				buf_size, dst->off + len,
 882				PATTERN_DST, false, is_memset);
 883
 884		diff = ktime_sub(ktime_get(), start);
 885		comparetime = ktime_add(comparetime, diff);
 886
 887		if (error_count) {
 888			result("data error", total_tests, src->off, dst->off,
 889			       len, error_count);
 890			failed_tests++;
 891		} else {
 892			verbose_result("test passed", total_tests, src->off,
 893				       dst->off, len, 0);
 894		}
 895
 896		continue;
 897
 898error_unmap_continue:
 899		dmaengine_unmap_put(um);
 900		failed_tests++;
 901	}
 902	ktime = ktime_sub(ktime_get(), ktime);
 903	ktime = ktime_sub(ktime, comparetime);
 904	ktime = ktime_sub(ktime, filltime);
 905	runtime = ktime_to_us(ktime);
 906
 907	ret = 0;
 908	kfree(dma_pq);
 909err_srcs_array:
 910	kfree(srcs);
 911err_dst:
 912	dmatest_free_test_data(dst);
 913err_src:
 914	dmatest_free_test_data(src);
 915err_free_coefs:
 916	kfree(pq_coefs);
 917err_thread_type:
 918	iops = dmatest_persec(runtime, total_tests);
 919	pr_info("%s: summary %u tests, %u failures %llu.%02llu iops %llu KB/s (%d)\n",
 920		current->comm, total_tests, failed_tests,
 921		FIXPT_TO_INT(iops), FIXPT_GET_FRAC(iops),
 922		dmatest_KBs(runtime, total_len), ret);
 923
 924	/* terminate all transfers on specified channels */
 925	if (ret || failed_tests)
 926		dmaengine_terminate_sync(chan);
 927
 928	thread->done = true;
 929	wake_up(&thread_wait);
 930
 931	return ret;
 932}
 933
 934static void dmatest_cleanup_channel(struct dmatest_chan *dtc)
 935{
 936	struct dmatest_thread	*thread;
 937	struct dmatest_thread	*_thread;
 938	int			ret;
 939
 940	list_for_each_entry_safe(thread, _thread, &dtc->threads, node) {
 941		ret = kthread_stop(thread->task);
 942		pr_debug("thread %s exited with status %d\n",
 943			 thread->task->comm, ret);
 944		list_del(&thread->node);
 945		put_task_struct(thread->task);
 946		kfree(thread);
 947	}
 948
 949	/* terminate all transfers on specified channels */
 950	dmaengine_terminate_sync(dtc->chan);
 951
 952	kfree(dtc);
 953}
 954
 955static int dmatest_add_threads(struct dmatest_info *info,
 956		struct dmatest_chan *dtc, enum dma_transaction_type type)
 957{
 958	struct dmatest_params *params = &info->params;
 959	struct dmatest_thread *thread;
 960	struct dma_chan *chan = dtc->chan;
 961	char *op;
 962	unsigned int i;
 963
 964	if (type == DMA_MEMCPY)
 965		op = "copy";
 966	else if (type == DMA_MEMSET)
 967		op = "set";
 968	else if (type == DMA_XOR)
 969		op = "xor";
 970	else if (type == DMA_PQ)
 971		op = "pq";
 972	else
 973		return -EINVAL;
 974
 975	for (i = 0; i < params->threads_per_chan; i++) {
 976		thread = kzalloc(sizeof(struct dmatest_thread), GFP_KERNEL);
 977		if (!thread) {
 978			pr_warn("No memory for %s-%s%u\n",
 979				dma_chan_name(chan), op, i);
 980			break;
 981		}
 982		thread->info = info;
 983		thread->chan = dtc->chan;
 984		thread->type = type;
 985		thread->test_done.wait = &thread->done_wait;
 986		init_waitqueue_head(&thread->done_wait);
 987		smp_wmb();
 988		thread->task = kthread_create(dmatest_func, thread, "%s-%s%u",
 989				dma_chan_name(chan), op, i);
 990		if (IS_ERR(thread->task)) {
 991			pr_warn("Failed to create thread %s-%s%u\n",
 992				dma_chan_name(chan), op, i);
 993			kfree(thread);
 994			break;
 995		}
 996
 997		/* srcbuf and dstbuf are allocated by the thread itself */
 998		get_task_struct(thread->task);
 999		list_add_tail(&thread->node, &dtc->threads);
1000		thread->pending = true;
1001	}
1002
1003	return i;
1004}
1005
1006static int dmatest_add_channel(struct dmatest_info *info,
1007		struct dma_chan *chan)
1008{
1009	struct dmatest_chan	*dtc;
1010	struct dma_device	*dma_dev = chan->device;
1011	unsigned int		thread_count = 0;
1012	int cnt;
1013
1014	dtc = kmalloc(sizeof(struct dmatest_chan), GFP_KERNEL);
1015	if (!dtc) {
1016		pr_warn("No memory for %s\n", dma_chan_name(chan));
1017		return -ENOMEM;
1018	}
1019
1020	dtc->chan = chan;
1021	INIT_LIST_HEAD(&dtc->threads);
1022
1023	if (dma_has_cap(DMA_COMPLETION_NO_ORDER, dma_dev->cap_mask) &&
1024	    info->params.polled) {
1025		info->params.polled = false;
1026		pr_warn("DMA_COMPLETION_NO_ORDER, polled disabled\n");
1027	}
1028
1029	if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask)) {
1030		if (dmatest == 0) {
1031			cnt = dmatest_add_threads(info, dtc, DMA_MEMCPY);
1032			thread_count += cnt > 0 ? cnt : 0;
1033		}
1034	}
1035
1036	if (dma_has_cap(DMA_MEMSET, dma_dev->cap_mask)) {
1037		if (dmatest == 1) {
1038			cnt = dmatest_add_threads(info, dtc, DMA_MEMSET);
1039			thread_count += cnt > 0 ? cnt : 0;
1040		}
1041	}
1042
1043	if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1044		cnt = dmatest_add_threads(info, dtc, DMA_XOR);
1045		thread_count += cnt > 0 ? cnt : 0;
1046	}
1047	if (dma_has_cap(DMA_PQ, dma_dev->cap_mask)) {
1048		cnt = dmatest_add_threads(info, dtc, DMA_PQ);
1049		thread_count += cnt > 0 ? cnt : 0;
1050	}
1051
1052	pr_info("Added %u threads using %s\n",
1053		thread_count, dma_chan_name(chan));
1054
1055	list_add_tail(&dtc->node, &info->channels);
1056	info->nr_channels++;
1057
1058	return 0;
1059}
1060
1061static bool filter(struct dma_chan *chan, void *param)
1062{
1063	return dmatest_match_channel(param, chan) && dmatest_match_device(param, chan->device);
 
 
 
 
 
 
1064}
1065
1066static void request_channels(struct dmatest_info *info,
1067			     enum dma_transaction_type type)
1068{
1069	dma_cap_mask_t mask;
1070
1071	dma_cap_zero(mask);
1072	dma_cap_set(type, mask);
1073	for (;;) {
1074		struct dmatest_params *params = &info->params;
1075		struct dma_chan *chan;
1076
1077		chan = dma_request_channel(mask, filter, params);
1078		if (chan) {
1079			if (dmatest_add_channel(info, chan)) {
1080				dma_release_channel(chan);
1081				break; /* add_channel failed, punt */
1082			}
1083		} else
1084			break; /* no more channels available */
1085		if (params->max_channels &&
1086		    info->nr_channels >= params->max_channels)
1087			break; /* we have all we need */
1088	}
1089}
1090
1091static void add_threaded_test(struct dmatest_info *info)
1092{
1093	struct dmatest_params *params = &info->params;
1094
1095	/* Copy test parameters */
1096	params->buf_size = test_buf_size;
1097	strscpy(params->channel, strim(test_channel), sizeof(params->channel));
1098	strscpy(params->device, strim(test_device), sizeof(params->device));
1099	params->threads_per_chan = threads_per_chan;
1100	params->max_channels = max_channels;
1101	params->iterations = iterations;
1102	params->xor_sources = xor_sources;
1103	params->pq_sources = pq_sources;
1104	params->timeout = timeout;
1105	params->noverify = noverify;
1106	params->norandom = norandom;
1107	params->alignment = alignment;
1108	params->transfer_size = transfer_size;
1109	params->polled = polled;
1110
1111	request_channels(info, DMA_MEMCPY);
1112	request_channels(info, DMA_MEMSET);
1113	request_channels(info, DMA_XOR);
1114	request_channels(info, DMA_PQ);
1115}
1116
1117static void run_pending_tests(struct dmatest_info *info)
1118{
1119	struct dmatest_chan *dtc;
1120	unsigned int thread_count = 0;
1121
1122	list_for_each_entry(dtc, &info->channels, node) {
1123		struct dmatest_thread *thread;
1124
1125		thread_count = 0;
1126		list_for_each_entry(thread, &dtc->threads, node) {
1127			wake_up_process(thread->task);
1128			thread_count++;
1129		}
1130		pr_info("Started %u threads using %s\n",
1131			thread_count, dma_chan_name(dtc->chan));
1132	}
1133}
1134
1135static void stop_threaded_test(struct dmatest_info *info)
1136{
1137	struct dmatest_chan *dtc, *_dtc;
1138	struct dma_chan *chan;
1139
1140	list_for_each_entry_safe(dtc, _dtc, &info->channels, node) {
1141		list_del(&dtc->node);
1142		chan = dtc->chan;
1143		dmatest_cleanup_channel(dtc);
1144		pr_debug("dropped channel %s\n", dma_chan_name(chan));
1145		dma_release_channel(chan);
1146	}
1147
1148	info->nr_channels = 0;
1149}
1150
1151static void start_threaded_tests(struct dmatest_info *info)
1152{
1153	/* we might be called early to set run=, defer running until all
1154	 * parameters have been evaluated
1155	 */
1156	if (!info->did_init)
1157		return;
1158
1159	run_pending_tests(info);
1160}
1161
1162static int dmatest_run_get(char *val, const struct kernel_param *kp)
1163{
1164	struct dmatest_info *info = &test_info;
1165
1166	mutex_lock(&info->lock);
1167	if (is_threaded_test_run(info)) {
1168		dmatest_run = true;
1169	} else {
1170		if (!is_threaded_test_pending(info))
1171			stop_threaded_test(info);
1172		dmatest_run = false;
1173	}
1174	mutex_unlock(&info->lock);
1175
1176	return param_get_bool(val, kp);
1177}
1178
1179static int dmatest_run_set(const char *val, const struct kernel_param *kp)
1180{
1181	struct dmatest_info *info = &test_info;
1182	int ret;
1183
1184	mutex_lock(&info->lock);
1185	ret = param_set_bool(val, kp);
1186	if (ret) {
1187		mutex_unlock(&info->lock);
1188		return ret;
1189	} else if (dmatest_run) {
1190		if (!is_threaded_test_pending(info)) {
1191			/*
1192			 * We have nothing to run. This can be due to:
1193			 */
1194			ret = info->last_error;
1195			if (ret) {
1196				/* 1) Misconfiguration */
1197				pr_err("Channel misconfigured, can't continue\n");
1198				mutex_unlock(&info->lock);
1199				return ret;
1200			} else {
1201				/* 2) We rely on defaults */
1202				pr_info("No channels configured, continue with any\n");
1203				if (!is_threaded_test_run(info))
1204					stop_threaded_test(info);
1205				add_threaded_test(info);
1206			}
1207		}
1208		start_threaded_tests(info);
1209	} else {
1210		stop_threaded_test(info);
1211	}
1212
1213	mutex_unlock(&info->lock);
1214
1215	return ret;
1216}
1217
1218static int dmatest_chan_set(const char *val, const struct kernel_param *kp)
1219{
1220	struct dmatest_info *info = &test_info;
1221	struct dmatest_chan *dtc;
1222	char chan_reset_val[20];
1223	int ret;
1224
1225	mutex_lock(&info->lock);
1226	ret = param_set_copystring(val, kp);
1227	if (ret) {
1228		mutex_unlock(&info->lock);
1229		return ret;
1230	}
1231	/*Clear any previously run threads */
1232	if (!is_threaded_test_run(info) && !is_threaded_test_pending(info))
1233		stop_threaded_test(info);
1234	/* Reject channels that are already registered */
1235	if (is_threaded_test_pending(info)) {
1236		list_for_each_entry(dtc, &info->channels, node) {
1237			if (strcmp(dma_chan_name(dtc->chan),
1238				   strim(test_channel)) == 0) {
1239				dtc = list_last_entry(&info->channels,
1240						      struct dmatest_chan,
1241						      node);
1242				strscpy(chan_reset_val,
1243					dma_chan_name(dtc->chan),
1244					sizeof(chan_reset_val));
1245				ret = -EBUSY;
1246				goto add_chan_err;
1247			}
1248		}
1249	}
1250
1251	add_threaded_test(info);
1252
1253	/* Check if channel was added successfully */
1254	if (!list_empty(&info->channels)) {
 
 
1255		/*
1256		 * if new channel was not successfully added, revert the
1257		 * "test_channel" string to the name of the last successfully
1258		 * added channel. exception for when users issues empty string
1259		 * to channel parameter.
1260		 */
1261		dtc = list_last_entry(&info->channels, struct dmatest_chan, node);
1262		if ((strcmp(dma_chan_name(dtc->chan), strim(test_channel)) != 0)
1263		    && (strcmp("", strim(test_channel)) != 0)) {
1264			ret = -EINVAL;
1265			strscpy(chan_reset_val, dma_chan_name(dtc->chan),
1266				sizeof(chan_reset_val));
1267			goto add_chan_err;
1268		}
1269
1270	} else {
1271		/* Clear test_channel if no channels were added successfully */
1272		strscpy(chan_reset_val, "", sizeof(chan_reset_val));
1273		ret = -EBUSY;
1274		goto add_chan_err;
1275	}
1276
1277	info->last_error = ret;
1278	mutex_unlock(&info->lock);
1279
1280	return ret;
1281
1282add_chan_err:
1283	param_set_copystring(chan_reset_val, kp);
1284	info->last_error = ret;
1285	mutex_unlock(&info->lock);
1286
1287	return ret;
1288}
1289
1290static int dmatest_chan_get(char *val, const struct kernel_param *kp)
1291{
1292	struct dmatest_info *info = &test_info;
1293
1294	mutex_lock(&info->lock);
1295	if (!is_threaded_test_run(info) && !is_threaded_test_pending(info)) {
1296		stop_threaded_test(info);
1297		strscpy(test_channel, "", sizeof(test_channel));
1298	}
1299	mutex_unlock(&info->lock);
1300
1301	return param_get_string(val, kp);
1302}
1303
1304static int dmatest_test_list_get(char *val, const struct kernel_param *kp)
1305{
1306	struct dmatest_info *info = &test_info;
1307	struct dmatest_chan *dtc;
1308	unsigned int thread_count = 0;
1309
1310	list_for_each_entry(dtc, &info->channels, node) {
1311		struct dmatest_thread *thread;
1312
1313		thread_count = 0;
1314		list_for_each_entry(thread, &dtc->threads, node) {
1315			thread_count++;
1316		}
1317		pr_info("%u threads using %s\n",
1318			thread_count, dma_chan_name(dtc->chan));
1319	}
1320
1321	return 0;
1322}
1323
1324static int __init dmatest_init(void)
1325{
1326	struct dmatest_info *info = &test_info;
1327	struct dmatest_params *params = &info->params;
1328
1329	if (dmatest_run) {
1330		mutex_lock(&info->lock);
1331		add_threaded_test(info);
1332		run_pending_tests(info);
1333		mutex_unlock(&info->lock);
1334	}
1335
1336	if (params->iterations && wait)
1337		wait_event(thread_wait, !is_threaded_test_run(info));
1338
1339	/* module parameters are stable, inittime tests are started,
1340	 * let userspace take over 'run' control
1341	 */
1342	info->did_init = true;
1343
1344	return 0;
1345}
1346/* when compiled-in wait for drivers to load first */
1347late_initcall(dmatest_init);
1348
1349static void __exit dmatest_exit(void)
1350{
1351	struct dmatest_info *info = &test_info;
1352
1353	mutex_lock(&info->lock);
1354	stop_threaded_test(info);
1355	mutex_unlock(&info->lock);
1356}
1357module_exit(dmatest_exit);
1358
1359MODULE_AUTHOR("Haavard Skinnemoen (Atmel)");
1360MODULE_LICENSE("GPL v2");