drivers/dma/dmatest.c

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