Merge remote-tracking branch 'origin/master' into perf/core
[linux-2.6-microblaze.git] / drivers / dma / dmatest.c
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
24 static unsigned int test_buf_size = 16384;
25 module_param(test_buf_size, uint, S_IRUGO | S_IWUSR);
26 MODULE_PARM_DESC(test_buf_size, "Size of the memcpy test buffer");
27
28 static char test_device[32];
29 module_param_string(device, test_device, sizeof(test_device),
30                 S_IRUGO | S_IWUSR);
31 MODULE_PARM_DESC(device, "Bus ID of the DMA Engine to test (default: any)");
32
33 static unsigned int threads_per_chan = 1;
34 module_param(threads_per_chan, uint, S_IRUGO | S_IWUSR);
35 MODULE_PARM_DESC(threads_per_chan,
36                 "Number of threads to start per channel (default: 1)");
37
38 static unsigned int max_channels;
39 module_param(max_channels, uint, S_IRUGO | S_IWUSR);
40 MODULE_PARM_DESC(max_channels,
41                 "Maximum number of channels to use (default: all)");
42
43 static unsigned int iterations;
44 module_param(iterations, uint, S_IRUGO | S_IWUSR);
45 MODULE_PARM_DESC(iterations,
46                 "Iterations before stopping test (default: infinite)");
47
48 static unsigned int dmatest;
49 module_param(dmatest, uint, S_IRUGO | S_IWUSR);
50 MODULE_PARM_DESC(dmatest,
51                 "dmatest 0-memcpy 1-memset (default: 0)");
52
53 static unsigned int xor_sources = 3;
54 module_param(xor_sources, uint, S_IRUGO | S_IWUSR);
55 MODULE_PARM_DESC(xor_sources,
56                 "Number of xor source buffers (default: 3)");
57
58 static unsigned int pq_sources = 3;
59 module_param(pq_sources, uint, S_IRUGO | S_IWUSR);
60 MODULE_PARM_DESC(pq_sources,
61                 "Number of p+q source buffers (default: 3)");
62
63 static int timeout = 3000;
64 module_param(timeout, int, S_IRUGO | S_IWUSR);
65 MODULE_PARM_DESC(timeout, "Transfer Timeout in msec (default: 3000), "
66                  "Pass -1 for infinite timeout");
67
68 static bool noverify;
69 module_param(noverify, bool, S_IRUGO | S_IWUSR);
70 MODULE_PARM_DESC(noverify, "Disable data verification (default: verify)");
71
72 static bool norandom;
73 module_param(norandom, bool, 0644);
74 MODULE_PARM_DESC(norandom, "Disable random offset setup (default: random)");
75
76 static bool verbose;
77 module_param(verbose, bool, S_IRUGO | S_IWUSR);
78 MODULE_PARM_DESC(verbose, "Enable \"success\" result messages (default: off)");
79
80 static int alignment = -1;
81 module_param(alignment, int, 0644);
82 MODULE_PARM_DESC(alignment, "Custom data address alignment taken as 2^(alignment) (default: not used (-1))");
83
84 static unsigned int transfer_size;
85 module_param(transfer_size, uint, 0644);
86 MODULE_PARM_DESC(transfer_size, "Optional custom transfer size in bytes (default: not used (0))");
87
88 static bool polled;
89 module_param(polled, bool, S_IRUGO | S_IWUSR);
90 MODULE_PARM_DESC(polled, "Use polling for completion instead of interrupts");
91
92 /**
93  * struct dmatest_params - test parameters.
94  * @buf_size:           size of the memcpy test buffer
95  * @channel:            bus ID of the channel to test
96  * @device:             bus ID of the DMA Engine to test
97  * @threads_per_chan:   number of threads to start per channel
98  * @max_channels:       maximum number of channels to use
99  * @iterations:         iterations before stopping test
100  * @xor_sources:        number of xor source buffers
101  * @pq_sources:         number of p+q source buffers
102  * @timeout:            transfer timeout in msec, -1 for infinite timeout
103  * @noverify:           disable data verification
104  * @norandom:           disable random offset setup
105  * @alignment:          custom data address alignment taken as 2^alignment
106  * @transfer_size:      custom transfer size in bytes
107  * @polled:             use polling for completion instead of interrupts
108  */
109 struct dmatest_params {
110         unsigned int    buf_size;
111         char            channel[20];
112         char            device[32];
113         unsigned int    threads_per_chan;
114         unsigned int    max_channels;
115         unsigned int    iterations;
116         unsigned int    xor_sources;
117         unsigned int    pq_sources;
118         int             timeout;
119         bool            noverify;
120         bool            norandom;
121         int             alignment;
122         unsigned int    transfer_size;
123         bool            polled;
124 };
125
126 /**
127  * struct dmatest_info - test information.
128  * @params:             test parameters
129  * @channels:           channels under test
130  * @nr_channels:        number of channels under test
131  * @lock:               access protection to the fields of this structure
132  * @did_init:           module has been initialized completely
133  * @last_error:         test has faced configuration issues
134  */
135 static struct dmatest_info {
136         /* Test parameters */
137         struct dmatest_params   params;
138
139         /* Internal state */
140         struct list_head        channels;
141         unsigned int            nr_channels;
142         int                     last_error;
143         struct mutex            lock;
144         bool                    did_init;
145 } test_info = {
146         .channels = LIST_HEAD_INIT(test_info.channels),
147         .lock = __MUTEX_INITIALIZER(test_info.lock),
148 };
149
150 static int dmatest_run_set(const char *val, const struct kernel_param *kp);
151 static int dmatest_run_get(char *val, const struct kernel_param *kp);
152 static const struct kernel_param_ops run_ops = {
153         .set = dmatest_run_set,
154         .get = dmatest_run_get,
155 };
156 static bool dmatest_run;
157 module_param_cb(run, &run_ops, &dmatest_run, S_IRUGO | S_IWUSR);
158 MODULE_PARM_DESC(run, "Run the test (default: false)");
159
160 static int dmatest_chan_set(const char *val, const struct kernel_param *kp);
161 static int dmatest_chan_get(char *val, const struct kernel_param *kp);
162 static const struct kernel_param_ops multi_chan_ops = {
163         .set = dmatest_chan_set,
164         .get = dmatest_chan_get,
165 };
166
167 static char test_channel[20];
168 static struct kparam_string newchan_kps = {
169         .string = test_channel,
170         .maxlen = 20,
171 };
172 module_param_cb(channel, &multi_chan_ops, &newchan_kps, 0644);
173 MODULE_PARM_DESC(channel, "Bus ID of the channel to test (default: any)");
174
175 static int dmatest_test_list_get(char *val, const struct kernel_param *kp);
176 static const struct kernel_param_ops test_list_ops = {
177         .get = dmatest_test_list_get,
178 };
179 module_param_cb(test_list, &test_list_ops, NULL, 0444);
180 MODULE_PARM_DESC(test_list, "Print current test list");
181
182 /* Maximum amount of mismatched bytes in buffer to print */
183 #define MAX_ERROR_COUNT         32
184
185 /*
186  * Initialization patterns. All bytes in the source buffer has bit 7
187  * set, all bytes in the destination buffer has bit 7 cleared.
188  *
189  * Bit 6 is set for all bytes which are to be copied by the DMA
190  * engine. Bit 5 is set for all bytes which are to be overwritten by
191  * the DMA engine.
192  *
193  * The remaining bits are the inverse of a counter which increments by
194  * one for each byte address.
195  */
196 #define PATTERN_SRC             0x80
197 #define PATTERN_DST             0x00
198 #define PATTERN_COPY            0x40
199 #define PATTERN_OVERWRITE       0x20
200 #define PATTERN_COUNT_MASK      0x1f
201 #define PATTERN_MEMSET_IDX      0x01
202
203 /* Fixed point arithmetic ops */
204 #define FIXPT_SHIFT             8
205 #define FIXPNT_MASK             0xFF
206 #define FIXPT_TO_INT(a) ((a) >> FIXPT_SHIFT)
207 #define INT_TO_FIXPT(a) ((a) << FIXPT_SHIFT)
208 #define FIXPT_GET_FRAC(a)       ((((a) & FIXPNT_MASK) * 100) >> FIXPT_SHIFT)
209
210 /* poor man's completion - we want to use wait_event_freezable() on it */
211 struct dmatest_done {
212         bool                    done;
213         wait_queue_head_t       *wait;
214 };
215
216 struct dmatest_data {
217         u8              **raw;
218         u8              **aligned;
219         unsigned int    cnt;
220         unsigned int    off;
221 };
222
223 struct dmatest_thread {
224         struct list_head        node;
225         struct dmatest_info     *info;
226         struct task_struct      *task;
227         struct dma_chan         *chan;
228         struct dmatest_data     src;
229         struct dmatest_data     dst;
230         enum dma_transaction_type type;
231         wait_queue_head_t done_wait;
232         struct dmatest_done test_done;
233         bool                    done;
234         bool                    pending;
235 };
236
237 struct dmatest_chan {
238         struct list_head        node;
239         struct dma_chan         *chan;
240         struct list_head        threads;
241 };
242
243 static DECLARE_WAIT_QUEUE_HEAD(thread_wait);
244 static bool wait;
245
246 static bool is_threaded_test_run(struct dmatest_info *info)
247 {
248         struct dmatest_chan *dtc;
249
250         list_for_each_entry(dtc, &info->channels, node) {
251                 struct dmatest_thread *thread;
252
253                 list_for_each_entry(thread, &dtc->threads, node) {
254                         if (!thread->done && !thread->pending)
255                                 return true;
256                 }
257         }
258
259         return false;
260 }
261
262 static bool is_threaded_test_pending(struct dmatest_info *info)
263 {
264         struct dmatest_chan *dtc;
265
266         list_for_each_entry(dtc, &info->channels, node) {
267                 struct dmatest_thread *thread;
268
269                 list_for_each_entry(thread, &dtc->threads, node) {
270                         if (thread->pending)
271                                 return true;
272                 }
273         }
274
275         return false;
276 }
277
278 static int dmatest_wait_get(char *val, const struct kernel_param *kp)
279 {
280         struct dmatest_info *info = &test_info;
281         struct dmatest_params *params = &info->params;
282
283         if (params->iterations)
284                 wait_event(thread_wait, !is_threaded_test_run(info));
285         wait = true;
286         return param_get_bool(val, kp);
287 }
288
289 static const struct kernel_param_ops wait_ops = {
290         .get = dmatest_wait_get,
291         .set = param_set_bool,
292 };
293 module_param_cb(wait, &wait_ops, &wait, S_IRUGO);
294 MODULE_PARM_DESC(wait, "Wait for tests to complete (default: false)");
295
296 static bool dmatest_match_channel(struct dmatest_params *params,
297                 struct dma_chan *chan)
298 {
299         if (params->channel[0] == '\0')
300                 return true;
301         return strcmp(dma_chan_name(chan), params->channel) == 0;
302 }
303
304 static bool dmatest_match_device(struct dmatest_params *params,
305                 struct dma_device *device)
306 {
307         if (params->device[0] == '\0')
308                 return true;
309         return strcmp(dev_name(device->dev), params->device) == 0;
310 }
311
312 static unsigned long dmatest_random(void)
313 {
314         unsigned long buf;
315
316         prandom_bytes(&buf, sizeof(buf));
317         return buf;
318 }
319
320 static inline u8 gen_inv_idx(u8 index, bool is_memset)
321 {
322         u8 val = is_memset ? PATTERN_MEMSET_IDX : index;
323
324         return ~val & PATTERN_COUNT_MASK;
325 }
326
327 static inline u8 gen_src_value(u8 index, bool is_memset)
328 {
329         return PATTERN_SRC | gen_inv_idx(index, is_memset);
330 }
331
332 static inline u8 gen_dst_value(u8 index, bool is_memset)
333 {
334         return PATTERN_DST | gen_inv_idx(index, is_memset);
335 }
336
337 static void dmatest_init_srcs(u8 **bufs, unsigned int start, unsigned int len,
338                 unsigned int buf_size, bool is_memset)
339 {
340         unsigned int i;
341         u8 *buf;
342
343         for (; (buf = *bufs); bufs++) {
344                 for (i = 0; i < start; i++)
345                         buf[i] = gen_src_value(i, is_memset);
346                 for ( ; i < start + len; i++)
347                         buf[i] = gen_src_value(i, is_memset) | PATTERN_COPY;
348                 for ( ; i < buf_size; i++)
349                         buf[i] = gen_src_value(i, is_memset);
350                 buf++;
351         }
352 }
353
354 static void dmatest_init_dsts(u8 **bufs, unsigned int start, unsigned int len,
355                 unsigned int buf_size, bool is_memset)
356 {
357         unsigned int i;
358         u8 *buf;
359
360         for (; (buf = *bufs); bufs++) {
361                 for (i = 0; i < start; i++)
362                         buf[i] = gen_dst_value(i, is_memset);
363                 for ( ; i < start + len; i++)
364                         buf[i] = gen_dst_value(i, is_memset) |
365                                                 PATTERN_OVERWRITE;
366                 for ( ; i < buf_size; i++)
367                         buf[i] = gen_dst_value(i, is_memset);
368         }
369 }
370
371 static void dmatest_mismatch(u8 actual, u8 pattern, unsigned int index,
372                 unsigned int counter, bool is_srcbuf, bool is_memset)
373 {
374         u8              diff = actual ^ pattern;
375         u8              expected = pattern | gen_inv_idx(counter, is_memset);
376         const char      *thread_name = current->comm;
377
378         if (is_srcbuf)
379                 pr_warn("%s: srcbuf[0x%x] overwritten! Expected %02x, got %02x\n",
380                         thread_name, index, expected, actual);
381         else if ((pattern & PATTERN_COPY)
382                         && (diff & (PATTERN_COPY | PATTERN_OVERWRITE)))
383                 pr_warn("%s: dstbuf[0x%x] not copied! Expected %02x, got %02x\n",
384                         thread_name, index, expected, actual);
385         else if (diff & PATTERN_SRC)
386                 pr_warn("%s: dstbuf[0x%x] was copied! Expected %02x, got %02x\n",
387                         thread_name, index, expected, actual);
388         else
389                 pr_warn("%s: dstbuf[0x%x] mismatch! Expected %02x, got %02x\n",
390                         thread_name, index, expected, actual);
391 }
392
393 static unsigned int dmatest_verify(u8 **bufs, unsigned int start,
394                 unsigned int end, unsigned int counter, u8 pattern,
395                 bool is_srcbuf, bool is_memset)
396 {
397         unsigned int i;
398         unsigned int error_count = 0;
399         u8 actual;
400         u8 expected;
401         u8 *buf;
402         unsigned int counter_orig = counter;
403
404         for (; (buf = *bufs); bufs++) {
405                 counter = counter_orig;
406                 for (i = start; i < end; i++) {
407                         actual = buf[i];
408                         expected = pattern | gen_inv_idx(counter, is_memset);
409                         if (actual != expected) {
410                                 if (error_count < MAX_ERROR_COUNT)
411                                         dmatest_mismatch(actual, pattern, i,
412                                                          counter, is_srcbuf,
413                                                          is_memset);
414                                 error_count++;
415                         }
416                         counter++;
417                 }
418         }
419
420         if (error_count > MAX_ERROR_COUNT)
421                 pr_warn("%s: %u errors suppressed\n",
422                         current->comm, error_count - MAX_ERROR_COUNT);
423
424         return error_count;
425 }
426
427
428 static void dmatest_callback(void *arg)
429 {
430         struct dmatest_done *done = arg;
431         struct dmatest_thread *thread =
432                 container_of(done, struct dmatest_thread, test_done);
433         if (!thread->done) {
434                 done->done = true;
435                 wake_up_all(done->wait);
436         } else {
437                 /*
438                  * If thread->done, it means that this callback occurred
439                  * after the parent thread has cleaned up. This can
440                  * happen in the case that driver doesn't implement
441                  * the terminate_all() functionality and a dma operation
442                  * did not occur within the timeout period
443                  */
444                 WARN(1, "dmatest: Kernel memory may be corrupted!!\n");
445         }
446 }
447
448 static unsigned int min_odd(unsigned int x, unsigned int y)
449 {
450         unsigned int val = min(x, y);
451
452         return val % 2 ? val : val - 1;
453 }
454
455 static void result(const char *err, unsigned int n, unsigned int src_off,
456                    unsigned int dst_off, unsigned int len, unsigned long data)
457 {
458         if (IS_ERR_VALUE(data)) {
459                 pr_info("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%ld)\n",
460                         current->comm, n, err, src_off, dst_off, len, data);
461         } else {
462                 pr_info("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%lu)\n",
463                         current->comm, n, err, src_off, dst_off, len, data);
464         }
465 }
466
467 static void dbg_result(const char *err, unsigned int n, unsigned int src_off,
468                        unsigned int dst_off, unsigned int len,
469                        unsigned long data)
470 {
471         pr_debug("%s: result #%u: '%s' with src_off=0x%x dst_off=0x%x len=0x%x (%lu)\n",
472                  current->comm, n, err, src_off, dst_off, len, data);
473 }
474
475 #define verbose_result(err, n, src_off, dst_off, len, data) ({  \
476         if (verbose)                                            \
477                 result(err, n, src_off, dst_off, len, data);    \
478         else                                                    \
479                 dbg_result(err, n, src_off, dst_off, len, data);\
480 })
481
482 static unsigned long long dmatest_persec(s64 runtime, unsigned int val)
483 {
484         unsigned long long per_sec = 1000000;
485
486         if (runtime <= 0)
487                 return 0;
488
489         /* drop precision until runtime is 32-bits */
490         while (runtime > UINT_MAX) {
491                 runtime >>= 1;
492                 per_sec <<= 1;
493         }
494
495         per_sec *= val;
496         per_sec = INT_TO_FIXPT(per_sec);
497         do_div(per_sec, runtime);
498
499         return per_sec;
500 }
501
502 static unsigned long long dmatest_KBs(s64 runtime, unsigned long long len)
503 {
504         return FIXPT_TO_INT(dmatest_persec(runtime, len >> 10));
505 }
506
507 static void __dmatest_free_test_data(struct dmatest_data *d, unsigned int cnt)
508 {
509         unsigned int i;
510
511         for (i = 0; i < cnt; i++)
512                 kfree(d->raw[i]);
513
514         kfree(d->aligned);
515         kfree(d->raw);
516 }
517
518 static void dmatest_free_test_data(struct dmatest_data *d)
519 {
520         __dmatest_free_test_data(d, d->cnt);
521 }
522
523 static int dmatest_alloc_test_data(struct dmatest_data *d,
524                 unsigned int buf_size, u8 align)
525 {
526         unsigned int i = 0;
527
528         d->raw = kcalloc(d->cnt + 1, sizeof(u8 *), GFP_KERNEL);
529         if (!d->raw)
530                 return -ENOMEM;
531
532         d->aligned = kcalloc(d->cnt + 1, sizeof(u8 *), GFP_KERNEL);
533         if (!d->aligned)
534                 goto err;
535
536         for (i = 0; i < d->cnt; i++) {
537                 d->raw[i] = kmalloc(buf_size + align, GFP_KERNEL);
538                 if (!d->raw[i])
539                         goto err;
540
541                 /* align to alignment restriction */
542                 if (align)
543                         d->aligned[i] = PTR_ALIGN(d->raw[i], align);
544                 else
545                         d->aligned[i] = d->raw[i];
546         }
547
548         return 0;
549 err:
550         __dmatest_free_test_data(d, i);
551         return -ENOMEM;
552 }
553
554 /*
555  * This function repeatedly tests DMA transfers of various lengths and
556  * offsets for a given operation type until it is told to exit by
557  * kthread_stop(). There may be multiple threads running this function
558  * in parallel for a single channel, and there may be multiple channels
559  * being tested in parallel.
560  *
561  * Before each test, the source and destination buffer is initialized
562  * with a known pattern. This pattern is different depending on
563  * whether it's in an area which is supposed to be copied or
564  * overwritten, and different in the source and destination buffers.
565  * So if the DMA engine doesn't copy exactly what we tell it to copy,
566  * we'll notice.
567  */
568 static int dmatest_func(void *data)
569 {
570         struct dmatest_thread   *thread = data;
571         struct dmatest_done     *done = &thread->test_done;
572         struct dmatest_info     *info;
573         struct dmatest_params   *params;
574         struct dma_chan         *chan;
575         struct dma_device       *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         src = &thread->src;
610         dst = &thread->dst;
611         if (thread->type == DMA_MEMCPY) {
612                 align = params->alignment < 0 ? dev->copy_align :
613                                                 params->alignment;
614                 src->cnt = dst->cnt = 1;
615         } else if (thread->type == DMA_MEMSET) {
616                 align = params->alignment < 0 ? dev->fill_align :
617                                                 params->alignment;
618                 src->cnt = dst->cnt = 1;
619                 is_memset = true;
620         } else if (thread->type == DMA_XOR) {
621                 /* force odd to ensure dst = src */
622                 src->cnt = min_odd(params->xor_sources | 1, dev->max_xor);
623                 dst->cnt = 1;
624                 align = params->alignment < 0 ? dev->xor_align :
625                                                 params->alignment;
626         } else if (thread->type == DMA_PQ) {
627                 /* force odd to ensure dst = src */
628                 src->cnt = min_odd(params->pq_sources | 1, dma_maxpq(dev, 0));
629                 dst->cnt = 2;
630                 align = params->alignment < 0 ? dev->pq_align :
631                                                 params->alignment;
632
633                 pq_coefs = kmalloc(params->pq_sources + 1, GFP_KERNEL);
634                 if (!pq_coefs)
635                         goto err_thread_type;
636
637                 for (i = 0; i < src->cnt; i++)
638                         pq_coefs[i] = 1;
639         } else
640                 goto err_thread_type;
641
642         /* Check if buffer count fits into map count variable (u8) */
643         if ((src->cnt + dst->cnt) >= 255) {
644                 pr_err("too many buffers (%d of 255 supported)\n",
645                        src->cnt + dst->cnt);
646                 goto err_free_coefs;
647         }
648
649         buf_size = params->buf_size;
650         if (1 << align > buf_size) {
651                 pr_err("%u-byte buffer too small for %d-byte alignment\n",
652                        buf_size, 1 << align);
653                 goto err_free_coefs;
654         }
655
656         if (dmatest_alloc_test_data(src, buf_size, align) < 0)
657                 goto err_free_coefs;
658
659         if (dmatest_alloc_test_data(dst, buf_size, align) < 0)
660                 goto err_src;
661
662         set_user_nice(current, 10);
663
664         srcs = kcalloc(src->cnt, sizeof(dma_addr_t), GFP_KERNEL);
665         if (!srcs)
666                 goto err_dst;
667
668         dma_pq = kcalloc(dst->cnt, sizeof(dma_addr_t), GFP_KERNEL);
669         if (!dma_pq)
670                 goto err_srcs_array;
671
672         /*
673          * src and dst buffers are freed by ourselves below
674          */
675         if (params->polled)
676                 flags = DMA_CTRL_ACK;
677         else
678                 flags = DMA_CTRL_ACK | DMA_PREP_INTERRUPT;
679
680         ktime = ktime_get();
681         while (!(kthread_should_stop() ||
682                (params->iterations && total_tests >= params->iterations))) {
683                 struct dma_async_tx_descriptor *tx = NULL;
684                 struct dmaengine_unmap_data *um;
685                 dma_addr_t *dsts;
686                 unsigned int len;
687
688                 total_tests++;
689
690                 if (params->transfer_size) {
691                         if (params->transfer_size >= buf_size) {
692                                 pr_err("%u-byte transfer size must be lower than %u-buffer size\n",
693                                        params->transfer_size, buf_size);
694                                 break;
695                         }
696                         len = params->transfer_size;
697                 } else if (params->norandom) {
698                         len = buf_size;
699                 } else {
700                         len = dmatest_random() % buf_size + 1;
701                 }
702
703                 /* Do not alter transfer size explicitly defined by user */
704                 if (!params->transfer_size) {
705                         len = (len >> align) << align;
706                         if (!len)
707                                 len = 1 << align;
708                 }
709                 total_len += len;
710
711                 if (params->norandom) {
712                         src->off = 0;
713                         dst->off = 0;
714                 } else {
715                         src->off = dmatest_random() % (buf_size - len + 1);
716                         dst->off = dmatest_random() % (buf_size - len + 1);
717
718                         src->off = (src->off >> align) << align;
719                         dst->off = (dst->off >> align) << align;
720                 }
721
722                 if (!params->noverify) {
723                         start = ktime_get();
724                         dmatest_init_srcs(src->aligned, src->off, len,
725                                           buf_size, is_memset);
726                         dmatest_init_dsts(dst->aligned, dst->off, len,
727                                           buf_size, is_memset);
728
729                         diff = ktime_sub(ktime_get(), start);
730                         filltime = ktime_add(filltime, diff);
731                 }
732
733                 um = dmaengine_get_unmap_data(dev->dev, src->cnt + dst->cnt,
734                                               GFP_KERNEL);
735                 if (!um) {
736                         failed_tests++;
737                         result("unmap data NULL", total_tests,
738                                src->off, dst->off, len, ret);
739                         continue;
740                 }
741
742                 um->len = buf_size;
743                 for (i = 0; i < src->cnt; i++) {
744                         void *buf = src->aligned[i];
745                         struct page *pg = virt_to_page(buf);
746                         unsigned long pg_off = offset_in_page(buf);
747
748                         um->addr[i] = dma_map_page(dev->dev, pg, pg_off,
749                                                    um->len, DMA_TO_DEVICE);
750                         srcs[i] = um->addr[i] + src->off;
751                         ret = dma_mapping_error(dev->dev, um->addr[i]);
752                         if (ret) {
753                                 result("src mapping error", total_tests,
754                                        src->off, dst->off, len, ret);
755                                 goto error_unmap_continue;
756                         }
757                         um->to_cnt++;
758                 }
759                 /* map with DMA_BIDIRECTIONAL to force writeback/invalidate */
760                 dsts = &um->addr[src->cnt];
761                 for (i = 0; i < dst->cnt; i++) {
762                         void *buf = dst->aligned[i];
763                         struct page *pg = virt_to_page(buf);
764                         unsigned long pg_off = offset_in_page(buf);
765
766                         dsts[i] = dma_map_page(dev->dev, pg, pg_off, um->len,
767                                                DMA_BIDIRECTIONAL);
768                         ret = dma_mapping_error(dev->dev, dsts[i]);
769                         if (ret) {
770                                 result("dst mapping error", total_tests,
771                                        src->off, dst->off, len, ret);
772                                 goto error_unmap_continue;
773                         }
774                         um->bidi_cnt++;
775                 }
776
777                 if (thread->type == DMA_MEMCPY)
778                         tx = dev->device_prep_dma_memcpy(chan,
779                                                          dsts[0] + dst->off,
780                                                          srcs[0], len, flags);
781                 else if (thread->type == DMA_MEMSET)
782                         tx = dev->device_prep_dma_memset(chan,
783                                                 dsts[0] + dst->off,
784                                                 *(src->aligned[0] + src->off),
785                                                 len, flags);
786                 else if (thread->type == DMA_XOR)
787                         tx = dev->device_prep_dma_xor(chan,
788                                                       dsts[0] + dst->off,
789                                                       srcs, src->cnt,
790                                                       len, flags);
791                 else if (thread->type == DMA_PQ) {
792                         for (i = 0; i < dst->cnt; i++)
793                                 dma_pq[i] = dsts[i] + dst->off;
794                         tx = dev->device_prep_dma_pq(chan, dma_pq, srcs,
795                                                      src->cnt, pq_coefs,
796                                                      len, flags);
797                 }
798
799                 if (!tx) {
800                         result("prep error", total_tests, src->off,
801                                dst->off, len, ret);
802                         msleep(100);
803                         goto error_unmap_continue;
804                 }
805
806                 done->done = false;
807                 if (!params->polled) {
808                         tx->callback = dmatest_callback;
809                         tx->callback_param = done;
810                 }
811                 cookie = tx->tx_submit(tx);
812
813                 if (dma_submit_error(cookie)) {
814                         result("submit error", total_tests, src->off,
815                                dst->off, len, ret);
816                         msleep(100);
817                         goto error_unmap_continue;
818                 }
819
820                 if (params->polled) {
821                         status = dma_sync_wait(chan, cookie);
822                         dmaengine_terminate_sync(chan);
823                         if (status == DMA_COMPLETE)
824                                 done->done = true;
825                 } else {
826                         dma_async_issue_pending(chan);
827
828                         wait_event_freezable_timeout(thread->done_wait,
829                                         done->done,
830                                         msecs_to_jiffies(params->timeout));
831
832                         status = dma_async_is_tx_complete(chan, cookie, NULL,
833                                                           NULL);
834                 }
835
836                 if (!done->done) {
837                         result("test timed out", total_tests, src->off, dst->off,
838                                len, 0);
839                         goto error_unmap_continue;
840                 } else if (status != DMA_COMPLETE &&
841                            !(dma_has_cap(DMA_COMPLETION_NO_ORDER,
842                                          dev->cap_mask) &&
843                              status == DMA_OUT_OF_ORDER)) {
844                         result(status == DMA_ERROR ?
845                                "completion error status" :
846                                "completion busy status", total_tests, src->off,
847                                dst->off, len, ret);
848                         goto error_unmap_continue;
849                 }
850
851                 dmaengine_unmap_put(um);
852
853                 if (params->noverify) {
854                         verbose_result("test passed", total_tests, src->off,
855                                        dst->off, len, 0);
856                         continue;
857                 }
858
859                 start = ktime_get();
860                 pr_debug("%s: verifying source buffer...\n", current->comm);
861                 error_count = dmatest_verify(src->aligned, 0, src->off,
862                                 0, PATTERN_SRC, true, is_memset);
863                 error_count += dmatest_verify(src->aligned, src->off,
864                                 src->off + len, src->off,
865                                 PATTERN_SRC | PATTERN_COPY, true, is_memset);
866                 error_count += dmatest_verify(src->aligned, src->off + len,
867                                 buf_size, src->off + len,
868                                 PATTERN_SRC, true, is_memset);
869
870                 pr_debug("%s: verifying dest buffer...\n", current->comm);
871                 error_count += dmatest_verify(dst->aligned, 0, dst->off,
872                                 0, PATTERN_DST, false, is_memset);
873
874                 error_count += dmatest_verify(dst->aligned, dst->off,
875                                 dst->off + len, src->off,
876                                 PATTERN_SRC | PATTERN_COPY, false, is_memset);
877
878                 error_count += dmatest_verify(dst->aligned, dst->off + len,
879                                 buf_size, dst->off + len,
880                                 PATTERN_DST, false, is_memset);
881
882                 diff = ktime_sub(ktime_get(), start);
883                 comparetime = ktime_add(comparetime, diff);
884
885                 if (error_count) {
886                         result("data error", total_tests, src->off, dst->off,
887                                len, error_count);
888                         failed_tests++;
889                 } else {
890                         verbose_result("test passed", total_tests, src->off,
891                                        dst->off, len, 0);
892                 }
893
894                 continue;
895
896 error_unmap_continue:
897                 dmaengine_unmap_put(um);
898                 failed_tests++;
899         }
900         ktime = ktime_sub(ktime_get(), ktime);
901         ktime = ktime_sub(ktime, comparetime);
902         ktime = ktime_sub(ktime, filltime);
903         runtime = ktime_to_us(ktime);
904
905         ret = 0;
906         kfree(dma_pq);
907 err_srcs_array:
908         kfree(srcs);
909 err_dst:
910         dmatest_free_test_data(dst);
911 err_src:
912         dmatest_free_test_data(src);
913 err_free_coefs:
914         kfree(pq_coefs);
915 err_thread_type:
916         iops = dmatest_persec(runtime, total_tests);
917         pr_info("%s: summary %u tests, %u failures %llu.%02llu iops %llu KB/s (%d)\n",
918                 current->comm, total_tests, failed_tests,
919                 FIXPT_TO_INT(iops), FIXPT_GET_FRAC(iops),
920                 dmatest_KBs(runtime, total_len), ret);
921
922         /* terminate all transfers on specified channels */
923         if (ret || failed_tests)
924                 dmaengine_terminate_sync(chan);
925
926         thread->done = true;
927         wake_up(&thread_wait);
928
929         return ret;
930 }
931
932 static void dmatest_cleanup_channel(struct dmatest_chan *dtc)
933 {
934         struct dmatest_thread   *thread;
935         struct dmatest_thread   *_thread;
936         int                     ret;
937
938         list_for_each_entry_safe(thread, _thread, &dtc->threads, node) {
939                 ret = kthread_stop(thread->task);
940                 pr_debug("thread %s exited with status %d\n",
941                          thread->task->comm, ret);
942                 list_del(&thread->node);
943                 put_task_struct(thread->task);
944                 kfree(thread);
945         }
946
947         /* terminate all transfers on specified channels */
948         dmaengine_terminate_sync(dtc->chan);
949
950         kfree(dtc);
951 }
952
953 static int dmatest_add_threads(struct dmatest_info *info,
954                 struct dmatest_chan *dtc, enum dma_transaction_type type)
955 {
956         struct dmatest_params *params = &info->params;
957         struct dmatest_thread *thread;
958         struct dma_chan *chan = dtc->chan;
959         char *op;
960         unsigned int i;
961
962         if (type == DMA_MEMCPY)
963                 op = "copy";
964         else if (type == DMA_MEMSET)
965                 op = "set";
966         else if (type == DMA_XOR)
967                 op = "xor";
968         else if (type == DMA_PQ)
969                 op = "pq";
970         else
971                 return -EINVAL;
972
973         for (i = 0; i < params->threads_per_chan; i++) {
974                 thread = kzalloc(sizeof(struct dmatest_thread), GFP_KERNEL);
975                 if (!thread) {
976                         pr_warn("No memory for %s-%s%u\n",
977                                 dma_chan_name(chan), op, i);
978                         break;
979                 }
980                 thread->info = info;
981                 thread->chan = dtc->chan;
982                 thread->type = type;
983                 thread->test_done.wait = &thread->done_wait;
984                 init_waitqueue_head(&thread->done_wait);
985                 smp_wmb();
986                 thread->task = kthread_create(dmatest_func, thread, "%s-%s%u",
987                                 dma_chan_name(chan), op, i);
988                 if (IS_ERR(thread->task)) {
989                         pr_warn("Failed to create thread %s-%s%u\n",
990                                 dma_chan_name(chan), op, i);
991                         kfree(thread);
992                         break;
993                 }
994
995                 /* srcbuf and dstbuf are allocated by the thread itself */
996                 get_task_struct(thread->task);
997                 list_add_tail(&thread->node, &dtc->threads);
998                 thread->pending = true;
999         }
1000
1001         return i;
1002 }
1003
1004 static int dmatest_add_channel(struct dmatest_info *info,
1005                 struct dma_chan *chan)
1006 {
1007         struct dmatest_chan     *dtc;
1008         struct dma_device       *dma_dev = chan->device;
1009         unsigned int            thread_count = 0;
1010         int cnt;
1011
1012         dtc = kmalloc(sizeof(struct dmatest_chan), GFP_KERNEL);
1013         if (!dtc) {
1014                 pr_warn("No memory for %s\n", dma_chan_name(chan));
1015                 return -ENOMEM;
1016         }
1017
1018         dtc->chan = chan;
1019         INIT_LIST_HEAD(&dtc->threads);
1020
1021         if (dma_has_cap(DMA_COMPLETION_NO_ORDER, dma_dev->cap_mask) &&
1022             info->params.polled) {
1023                 info->params.polled = false;
1024                 pr_warn("DMA_COMPLETION_NO_ORDER, polled disabled\n");
1025         }
1026
1027         if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask)) {
1028                 if (dmatest == 0) {
1029                         cnt = dmatest_add_threads(info, dtc, DMA_MEMCPY);
1030                         thread_count += cnt > 0 ? cnt : 0;
1031                 }
1032         }
1033
1034         if (dma_has_cap(DMA_MEMSET, dma_dev->cap_mask)) {
1035                 if (dmatest == 1) {
1036                         cnt = dmatest_add_threads(info, dtc, DMA_MEMSET);
1037                         thread_count += cnt > 0 ? cnt : 0;
1038                 }
1039         }
1040
1041         if (dma_has_cap(DMA_XOR, dma_dev->cap_mask)) {
1042                 cnt = dmatest_add_threads(info, dtc, DMA_XOR);
1043                 thread_count += cnt > 0 ? cnt : 0;
1044         }
1045         if (dma_has_cap(DMA_PQ, dma_dev->cap_mask)) {
1046                 cnt = dmatest_add_threads(info, dtc, DMA_PQ);
1047                 thread_count += cnt > 0 ? cnt : 0;
1048         }
1049
1050         pr_info("Added %u threads using %s\n",
1051                 thread_count, dma_chan_name(chan));
1052
1053         list_add_tail(&dtc->node, &info->channels);
1054         info->nr_channels++;
1055
1056         return 0;
1057 }
1058
1059 static bool filter(struct dma_chan *chan, void *param)
1060 {
1061         return dmatest_match_channel(param, chan) && dmatest_match_device(param, chan->device);
1062 }
1063
1064 static 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
1089 static 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
1115 static 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
1133 static 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
1149 static 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
1160 static 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
1177 static 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
1216 static 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         if (!list_empty(&info->channels)) {
1253                 /*
1254                  * if new channel was not successfully added, revert the
1255                  * "test_channel" string to the name of the last successfully
1256                  * added channel. exception for when users issues empty string
1257                  * to channel parameter.
1258                  */
1259                 dtc = list_last_entry(&info->channels, struct dmatest_chan, node);
1260                 if ((strcmp(dma_chan_name(dtc->chan), strim(test_channel)) != 0)
1261                     && (strcmp("", strim(test_channel)) != 0)) {
1262                         ret = -EINVAL;
1263                         strlcpy(chan_reset_val, dma_chan_name(dtc->chan),
1264                                 sizeof(chan_reset_val));
1265                         goto add_chan_err;
1266                 }
1267
1268         } else {
1269                 /* Clear test_channel if no channels were added successfully */
1270                 strlcpy(chan_reset_val, "", sizeof(chan_reset_val));
1271                 ret = -EBUSY;
1272                 goto add_chan_err;
1273         }
1274
1275         info->last_error = ret;
1276         mutex_unlock(&info->lock);
1277
1278         return ret;
1279
1280 add_chan_err:
1281         param_set_copystring(chan_reset_val, kp);
1282         info->last_error = ret;
1283         mutex_unlock(&info->lock);
1284
1285         return ret;
1286 }
1287
1288 static int dmatest_chan_get(char *val, const struct kernel_param *kp)
1289 {
1290         struct dmatest_info *info = &test_info;
1291
1292         mutex_lock(&info->lock);
1293         if (!is_threaded_test_run(info) && !is_threaded_test_pending(info)) {
1294                 stop_threaded_test(info);
1295                 strlcpy(test_channel, "", sizeof(test_channel));
1296         }
1297         mutex_unlock(&info->lock);
1298
1299         return param_get_string(val, kp);
1300 }
1301
1302 static int dmatest_test_list_get(char *val, const struct kernel_param *kp)
1303 {
1304         struct dmatest_info *info = &test_info;
1305         struct dmatest_chan *dtc;
1306         unsigned int thread_count = 0;
1307
1308         list_for_each_entry(dtc, &info->channels, node) {
1309                 struct dmatest_thread *thread;
1310
1311                 thread_count = 0;
1312                 list_for_each_entry(thread, &dtc->threads, node) {
1313                         thread_count++;
1314                 }
1315                 pr_info("%u threads using %s\n",
1316                         thread_count, dma_chan_name(dtc->chan));
1317         }
1318
1319         return 0;
1320 }
1321
1322 static int __init dmatest_init(void)
1323 {
1324         struct dmatest_info *info = &test_info;
1325         struct dmatest_params *params = &info->params;
1326
1327         if (dmatest_run) {
1328                 mutex_lock(&info->lock);
1329                 add_threaded_test(info);
1330                 run_pending_tests(info);
1331                 mutex_unlock(&info->lock);
1332         }
1333
1334         if (params->iterations && wait)
1335                 wait_event(thread_wait, !is_threaded_test_run(info));
1336
1337         /* module parameters are stable, inittime tests are started,
1338          * let userspace take over 'run' control
1339          */
1340         info->did_init = true;
1341
1342         return 0;
1343 }
1344 /* when compiled-in wait for drivers to load first */
1345 late_initcall(dmatest_init);
1346
1347 static void __exit dmatest_exit(void)
1348 {
1349         struct dmatest_info *info = &test_info;
1350
1351         mutex_lock(&info->lock);
1352         stop_threaded_test(info);
1353         mutex_unlock(&info->lock);
1354 }
1355 module_exit(dmatest_exit);
1356
1357 MODULE_AUTHOR("Haavard Skinnemoen (Atmel)");
1358 MODULE_LICENSE("GPL v2");