drivers/dma/edma.c

   1 /*
   2  * TI EDMA DMA engine driver
   3  *
   4  * Copyright 2012 Texas Instruments
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License as
   8  * published by the Free Software Foundation version 2.
   9  *
  10  * This program is distributed "as is" WITHOUT ANY WARRANTY of any
  11  * kind, whether express or implied; without even the implied warranty
  12  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  */
  15
  16 #include <linux/dmaengine.h>
  17 #include <linux/dma-mapping.h>
  18 #include <linux/err.h>
  19 #include <linux/init.h>
  20 #include <linux/interrupt.h>
  21 #include <linux/list.h>
  22 #include <linux/module.h>
  23 #include <linux/platform_device.h>
  24 #include <linux/slab.h>
  25 #include <linux/spinlock.h>
  26
  27 #include <linux/platform_data/edma.h>
  28
  29 #include "dmaengine.h"
  30 #include "virt-dma.h"
  31
  32 /*
  33  * This will go away when the private EDMA API is folded
  34  * into this driver and the platform device(s) are
  35  * instantiated in the arch code. We can only get away
  36  * with this simplification because DA8XX may not be built
  37  * in the same kernel image with other DaVinci parts. This
  38  * avoids having to sprinkle dmaengine driver platform devices
  39  * and data throughout all the existing board files.
  40  */
  41 #ifdef CONFIG_ARCH_DAVINCI_DA8XX
  42 #define EDMA_CTLRS      2
  43 #define EDMA_CHANS      32
  44 #else
  45 #define EDMA_CTLRS      1
  46 #define EDMA_CHANS      64
  47 #endif /* CONFIG_ARCH_DAVINCI_DA8XX */
  48
  49 /*
  50  * Max of 20 segments per channel to conserve PaRAM slots
  51  * Also note that MAX_NR_SG should be atleast the no.of periods
  52  * that are required for ASoC, otherwise DMA prep calls will
  53  * fail. Today davinci-pcm is the only user of this driver and
  54  * requires atleast 17 slots, so we setup the default to 20.
  55  */
  56 #define MAX_NR_SG               20
  57 #define EDMA_MAX_SLOTS          MAX_NR_SG
  58 #define EDMA_DESCRIPTORS        16
  59
  60 struct edma_desc {
  61         struct virt_dma_desc            vdesc;
  62         struct list_head                node;
  63         int                             cyclic;
  64         int                             absync;
  65         int                             pset_nr;
  66         int                             processed;
  67         struct edmacc_param             pset[0];
  68 };
  69
  70 struct edma_cc;
  71
  72 struct edma_chan {
  73         struct virt_dma_chan            vchan;
  74         struct list_head                node;
  75         struct edma_desc                *edesc;
  76         struct edma_cc                  *ecc;
  77         int                             ch_num;
  78         bool                            alloced;
  79         int                             slot[EDMA_MAX_SLOTS];
  80         int                             missed;
  81         struct dma_slave_config         cfg;
  82 };
  83
  84 struct edma_cc {
  85         int                             ctlr;
  86         struct dma_device               dma_slave;
  87         struct edma_chan                slave_chans[EDMA_CHANS];
  88         int                             num_slave_chans;
  89         int                             dummy_slot;
  90 };
  91
  92 static inline struct edma_cc *to_edma_cc(struct dma_device *d)
  93 {
  94         return container_of(d, struct edma_cc, dma_slave);
  95 }
  96
  97 static inline struct edma_chan *to_edma_chan(struct dma_chan *c)
  98 {
  99         return container_of(c, struct edma_chan, vchan.chan);
 100 }
 101
 102 static inline struct edma_desc
 103 *to_edma_desc(struct dma_async_tx_descriptor *tx)
 104 {
 105         return container_of(tx, struct edma_desc, vdesc.tx);
 106 }
 107
 108 static void edma_desc_free(struct virt_dma_desc *vdesc)
 109 {
 110         kfree(container_of(vdesc, struct edma_desc, vdesc));
 111 }
 112
 113 /* Dispatch a queued descriptor to the controller (caller holds lock) */
 114 static void edma_execute(struct edma_chan *echan)
 115 {
 116         struct virt_dma_desc *vdesc;
 117         struct edma_desc *edesc;
 118         struct device *dev = echan->vchan.chan.device->dev;
 119         int i, j, left, nslots;
 120
 121         /* If either we processed all psets or we're still not started */
 122         if (!echan->edesc ||
 123             echan->edesc->pset_nr == echan->edesc->processed) {
 124                 /* Get next vdesc */
 125                 vdesc = vchan_next_desc(&echan->vchan);
 126                 if (!vdesc) {
 127                         echan->edesc = NULL;
 128                         return;
 129                 }
 130                 list_del(&vdesc->node);
 131                 echan->edesc = to_edma_desc(&vdesc->tx);
 132         }
 133
 134         edesc = echan->edesc;
 135
 136         /* Find out how many left */
 137         left = edesc->pset_nr - edesc->processed;
 138         nslots = min(MAX_NR_SG, left);
 139
 140         /* Write descriptor PaRAM set(s) */
 141         for (i = 0; i < nslots; i++) {
 142                 j = i + edesc->processed;
 143                 edma_write_slot(echan->slot[i], &edesc->pset[j]);
 144                 dev_dbg(echan->vchan.chan.device->dev,
 145                         "\n pset[%d]:\n"
 146                         "  chnum\t%d\n"
 147                         "  slot\t%d\n"
 148                         "  opt\t%08x\n"
 149                         "  src\t%08x\n"
 150                         "  dst\t%08x\n"
 151                         "  abcnt\t%08x\n"
 152                         "  ccnt\t%08x\n"
 153                         "  bidx\t%08x\n"
 154                         "  cidx\t%08x\n"
 155                         "  lkrld\t%08x\n",
 156                         j, echan->ch_num, echan->slot[i],
 157                         edesc->pset[j].opt,
 158                         edesc->pset[j].src,
 159                         edesc->pset[j].dst,
 160                         edesc->pset[j].a_b_cnt,
 161                         edesc->pset[j].ccnt,
 162                         edesc->pset[j].src_dst_bidx,
 163                         edesc->pset[j].src_dst_cidx,
 164                         edesc->pset[j].link_bcntrld);
 165                 /* Link to the previous slot if not the last set */
 166                 if (i != (nslots - 1))
 167                         edma_link(echan->slot[i], echan->slot[i+1]);
 168         }
 169
 170         edesc->processed += nslots;
 171
 172         /*
 173          * If this is either the last set in a set of SG-list transactions
 174          * then setup a link to the dummy slot, this results in all future
 175          * events being absorbed and that's OK because we're done
 176          */
 177         if (edesc->processed == edesc->pset_nr) {
 178                 if (edesc->cyclic)
 179                         edma_link(echan->slot[nslots-1], echan->slot[1]);
 180                 else
 181                         edma_link(echan->slot[nslots-1],
 182                                   echan->ecc->dummy_slot);
 183         }
 184
 185         if (edesc->processed <= MAX_NR_SG) {
 186                 dev_dbg(dev, "first transfer starting %d\n", echan->ch_num);
 187                 edma_start(echan->ch_num);
 188         } else {
 189                 dev_dbg(dev, "chan: %d: completed %d elements, resuming\n",
 190                         echan->ch_num, edesc->processed);
 191                 edma_resume(echan->ch_num);
 192         }
 193
 194         /*
 195          * This happens due to setup times between intermediate transfers
 196          * in long SG lists which have to be broken up into transfers of
 197          * MAX_NR_SG
 198          */
 199         if (echan->missed) {
 200                 dev_dbg(dev, "missed event in execute detected\n");
 201                 edma_clean_channel(echan->ch_num);
 202                 edma_stop(echan->ch_num);
 203                 edma_start(echan->ch_num);
 204                 edma_trigger_channel(echan->ch_num);
 205                 echan->missed = 0;
 206         }
 207 }
 208
 209 static int edma_terminate_all(struct edma_chan *echan)
 210 {
 211         unsigned long flags;
 212         LIST_HEAD(head);
 213
 214         spin_lock_irqsave(&echan->vchan.lock, flags);
 215
 216         /*
 217          * Stop DMA activity: we assume the callback will not be called
 218          * after edma_dma() returns (even if it does, it will see
 219          * echan->edesc is NULL and exit.)
 220          */
 221         if (echan->edesc) {
 222                 echan->edesc = NULL;
 223                 edma_stop(echan->ch_num);
 224         }
 225
 226         vchan_get_all_descriptors(&echan->vchan, &head);
 227         spin_unlock_irqrestore(&echan->vchan.lock, flags);
 228         vchan_dma_desc_free_list(&echan->vchan, &head);
 229
 230         return 0;
 231 }
 232
 233 static int edma_slave_config(struct edma_chan *echan,
 234         struct dma_slave_config *cfg)
 235 {
 236         if (cfg->src_addr_width == DMA_SLAVE_BUSWIDTH_8_BYTES ||
 237             cfg->dst_addr_width == DMA_SLAVE_BUSWIDTH_8_BYTES)
 238                 return -EINVAL;
 239
 240         memcpy(&echan->cfg, cfg, sizeof(echan->cfg));
 241
 242         return 0;
 243 }
 244
 245 static int edma_control(struct dma_chan *chan, enum dma_ctrl_cmd cmd,
 246                         unsigned long arg)
 247 {
 248         int ret = 0;
 249         struct dma_slave_config *config;
 250         struct edma_chan *echan = to_edma_chan(chan);
 251
 252         switch (cmd) {
 253         case DMA_TERMINATE_ALL:
 254                 edma_terminate_all(echan);
 255                 break;
 256         case DMA_SLAVE_CONFIG:
 257                 config = (struct dma_slave_config *)arg;
 258                 ret = edma_slave_config(echan, config);
 259                 break;
 260         default:
 261                 ret = -ENOSYS;
 262         }
 263
 264         return ret;
 265 }
 266
 267 /*
 268  * A PaRAM set configuration abstraction used by other modes
 269  * @chan: Channel who's PaRAM set we're configuring
 270  * @pset: PaRAM set to initialize and setup.
 271  * @src_addr: Source address of the DMA
 272  * @dst_addr: Destination address of the DMA
 273  * @burst: In units of dev_width, how much to send
 274  * @dev_width: How much is the dev_width
 275  * @dma_length: Total length of the DMA transfer
 276  * @direction: Direction of the transfer
 277  */
 278 static int edma_config_pset(struct dma_chan *chan, struct edmacc_param *pset,
 279         dma_addr_t src_addr, dma_addr_t dst_addr, u32 burst,
 280         enum dma_slave_buswidth dev_width, unsigned int dma_length,
 281         enum dma_transfer_direction direction)
 282 {
 283         struct edma_chan *echan = to_edma_chan(chan);
 284         struct device *dev = chan->device->dev;
 285         int acnt, bcnt, ccnt, cidx;
 286         int src_bidx, dst_bidx, src_cidx, dst_cidx;
 287         int absync;
 288
 289         acnt = dev_width;
 290         /*
 291          * If the maxburst is equal to the fifo width, use
 292          * A-synced transfers. This allows for large contiguous
 293          * buffer transfers using only one PaRAM set.
 294          */
 295         if (burst == 1) {
 296                 /*
 297                  * For the A-sync case, bcnt and ccnt are the remainder
 298                  * and quotient respectively of the division of:
 299                  * (dma_length / acnt) by (SZ_64K -1). This is so
 300                  * that in case bcnt over flows, we have ccnt to use.
 301                  * Note: In A-sync tranfer only, bcntrld is used, but it
 302                  * only applies for sg_dma_len(sg) >= SZ_64K.
 303                  * In this case, the best way adopted is- bccnt for the
 304                  * first frame will be the remainder below. Then for
 305                  * every successive frame, bcnt will be SZ_64K-1. This
 306                  * is assured as bcntrld = 0xffff in end of function.
 307                  */
 308                 absync = false;
 309                 ccnt = dma_length / acnt / (SZ_64K - 1);
 310                 bcnt = dma_length / acnt - ccnt * (SZ_64K - 1);
 311                 /*
 312                  * If bcnt is non-zero, we have a remainder and hence an
 313                  * extra frame to transfer, so increment ccnt.
 314                  */
 315                 if (bcnt)
 316                         ccnt++;
 317                 else
 318                         bcnt = SZ_64K - 1;
 319                 cidx = acnt;
 320         } else {
 321                 /*
 322                  * If maxburst is greater than the fifo address_width,
 323                  * use AB-synced transfers where A count is the fifo
 324                  * address_width and B count is the maxburst. In this
 325                  * case, we are limited to transfers of C count frames
 326                  * of (address_width * maxburst) where C count is limited
 327                  * to SZ_64K-1. This places an upper bound on the length
 328                  * of an SG segment that can be handled.
 329                  */
 330                 absync = true;
 331                 bcnt = burst;
 332                 ccnt = dma_length / (acnt * bcnt);
 333                 if (ccnt > (SZ_64K - 1)) {
 334                         dev_err(dev, "Exceeded max SG segment size\n");
 335                         return -EINVAL;
 336                 }
 337                 cidx = acnt * bcnt;
 338         }
 339
 340         if (direction == DMA_MEM_TO_DEV) {
 341                 src_bidx = acnt;
 342                 src_cidx = cidx;
 343                 dst_bidx = 0;
 344                 dst_cidx = 0;
 345         } else if (direction == DMA_DEV_TO_MEM)  {
 346                 src_bidx = 0;
 347                 src_cidx = 0;
 348                 dst_bidx = acnt;
 349                 dst_cidx = cidx;
 350         } else {
 351                 dev_err(dev, "%s: direction not implemented yet\n", __func__);
 352                 return -EINVAL;
 353         }
 354
 355         pset->opt = EDMA_TCC(EDMA_CHAN_SLOT(echan->ch_num));
 356         /* Configure A or AB synchronized transfers */
 357         if (absync)
 358                 pset->opt |= SYNCDIM;
 359
 360         pset->src = src_addr;
 361         pset->dst = dst_addr;
 362
 363         pset->src_dst_bidx = (dst_bidx << 16) | src_bidx;
 364         pset->src_dst_cidx = (dst_cidx << 16) | src_cidx;
 365
 366         pset->a_b_cnt = bcnt << 16 | acnt;
 367         pset->ccnt = ccnt;
 368         /*
 369          * Only time when (bcntrld) auto reload is required is for
 370          * A-sync case, and in this case, a requirement of reload value
 371          * of SZ_64K-1 only is assured. 'link' is initially set to NULL
 372          * and then later will be populated by edma_execute.
 373          */
 374         pset->link_bcntrld = 0xffffffff;
 375         return absync;
 376 }
 377
 378 static struct dma_async_tx_descriptor *edma_prep_slave_sg(
 379         struct dma_chan *chan, struct scatterlist *sgl,
 380         unsigned int sg_len, enum dma_transfer_direction direction,
 381         unsigned long tx_flags, void *context)
 382 {
 383         struct edma_chan *echan = to_edma_chan(chan);
 384         struct device *dev = chan->device->dev;
 385         struct edma_desc *edesc;
 386         dma_addr_t src_addr = 0, dst_addr = 0;
 387         enum dma_slave_buswidth dev_width;
 388         u32 burst;
 389         struct scatterlist *sg;
 390         int i, nslots, ret;
 391
 392         if (unlikely(!echan || !sgl || !sg_len))
 393                 return NULL;
 394
 395         if (direction == DMA_DEV_TO_MEM) {
 396                 src_addr = echan->cfg.src_addr;
 397                 dev_width = echan->cfg.src_addr_width;
 398                 burst = echan->cfg.src_maxburst;
 399         } else if (direction == DMA_MEM_TO_DEV) {
 400                 dst_addr = echan->cfg.dst_addr;
 401                 dev_width = echan->cfg.dst_addr_width;
 402                 burst = echan->cfg.dst_maxburst;
 403         } else {
 404                 dev_err(dev, "%s: bad direction?\n", __func__);
 405                 return NULL;
 406         }
 407
 408         if (dev_width == DMA_SLAVE_BUSWIDTH_UNDEFINED) {
 409                 dev_err(dev, "Undefined slave buswidth\n");
 410                 return NULL;
 411         }
 412
 413         edesc = kzalloc(sizeof(*edesc) + sg_len *
 414                 sizeof(edesc->pset[0]), GFP_ATOMIC);
 415         if (!edesc) {
 416                 dev_dbg(dev, "Failed to allocate a descriptor\n");
 417                 return NULL;
 418         }
 419
 420         edesc->pset_nr = sg_len;
 421
 422         /* Allocate a PaRAM slot, if needed */
 423         nslots = min_t(unsigned, MAX_NR_SG, sg_len);
 424
 425         for (i = 0; i < nslots; i++) {
 426                 if (echan->slot[i] < 0) {
 427                         echan->slot[i] =
 428                                 edma_alloc_slot(EDMA_CTLR(echan->ch_num),
 429                                                 EDMA_SLOT_ANY);
 430                         if (echan->slot[i] < 0) {
 431                                 kfree(edesc);
 432                                 dev_err(dev, "Failed to allocate slot\n");
 433                                 return NULL;
 434                         }
 435                 }
 436         }
 437
 438         /* Configure PaRAM sets for each SG */
 439         for_each_sg(sgl, sg, sg_len, i) {
 440                 /* Get address for each SG */
 441                 if (direction == DMA_DEV_TO_MEM)
 442                         dst_addr = sg_dma_address(sg);
 443                 else
 444                         src_addr = sg_dma_address(sg);
 445
 446                 ret = edma_config_pset(chan, &edesc->pset[i], src_addr,
 447                                        dst_addr, burst, dev_width,
 448                                        sg_dma_len(sg), direction);
 449                 if (ret < 0) {
 450                         kfree(edesc);
 451                         return NULL;
 452                 }
 453
 454                 edesc->absync = ret;
 455
 456                 /* If this is the last in a current SG set of transactions,
 457                    enable interrupts so that next set is processed */
 458                 if (!((i+1) % MAX_NR_SG))
 459                         edesc->pset[i].opt |= TCINTEN;
 460
 461                 /* If this is the last set, enable completion interrupt flag */
 462                 if (i == sg_len - 1)
 463                         edesc->pset[i].opt |= TCINTEN;
 464         }
 465
 466         return vchan_tx_prep(&echan->vchan, &edesc->vdesc, tx_flags);
 467 }
 468
 469 static struct dma_async_tx_descriptor *edma_prep_dma_cyclic(
 470         struct dma_chan *chan, dma_addr_t buf_addr, size_t buf_len,
 471         size_t period_len, enum dma_transfer_direction direction,
 472         unsigned long tx_flags, void *context)
 473 {
 474         struct edma_chan *echan = to_edma_chan(chan);
 475         struct device *dev = chan->device->dev;
 476         struct edma_desc *edesc;
 477         dma_addr_t src_addr, dst_addr;
 478         enum dma_slave_buswidth dev_width;
 479         u32 burst;
 480         int i, ret, nslots;
 481
 482         if (unlikely(!echan || !buf_len || !period_len))
 483                 return NULL;
 484
 485         if (direction == DMA_DEV_TO_MEM) {
 486                 src_addr = echan->cfg.src_addr;
 487                 dst_addr = buf_addr;
 488                 dev_width = echan->cfg.src_addr_width;
 489                 burst = echan->cfg.src_maxburst;
 490         } else if (direction == DMA_MEM_TO_DEV) {
 491                 src_addr = buf_addr;
 492                 dst_addr = echan->cfg.dst_addr;
 493                 dev_width = echan->cfg.dst_addr_width;
 494                 burst = echan->cfg.dst_maxburst;
 495         } else {
 496                 dev_err(dev, "%s: bad direction?\n", __func__);
 497                 return NULL;
 498         }
 499
 500         if (dev_width == DMA_SLAVE_BUSWIDTH_UNDEFINED) {
 501                 dev_err(dev, "Undefined slave buswidth\n");
 502                 return NULL;
 503         }
 504
 505         if (unlikely(buf_len % period_len)) {
 506                 dev_err(dev, "Period should be multiple of Buffer length\n");
 507                 return NULL;
 508         }
 509
 510         nslots = (buf_len / period_len) + 1;
 511
 512         /*
 513          * Cyclic DMA users such as audio cannot tolerate delays introduced
 514          * by cases where the number of periods is more than the maximum
 515          * number of SGs the EDMA driver can handle at a time. For DMA types
 516          * such as Slave SGs, such delays are tolerable and synchronized,
 517          * but the synchronization is difficult to achieve with Cyclic and
 518          * cannot be guaranteed, so we error out early.
 519          */
 520         if (nslots > MAX_NR_SG)
 521                 return NULL;
 522
 523         edesc = kzalloc(sizeof(*edesc) + nslots *
 524                 sizeof(edesc->pset[0]), GFP_ATOMIC);
 525         if (!edesc) {
 526                 dev_dbg(dev, "Failed to allocate a descriptor\n");
 527                 return NULL;
 528         }
 529
 530         edesc->cyclic = 1;
 531         edesc->pset_nr = nslots;
 532
 533         dev_dbg(dev, "%s: nslots=%d\n", __func__, nslots);
 534         dev_dbg(dev, "%s: period_len=%d\n", __func__, period_len);
 535         dev_dbg(dev, "%s: buf_len=%d\n", __func__, buf_len);
 536
 537         for (i = 0; i < nslots; i++) {
 538                 /* Allocate a PaRAM slot, if needed */
 539                 if (echan->slot[i] < 0) {
 540                         echan->slot[i] =
 541                                 edma_alloc_slot(EDMA_CTLR(echan->ch_num),
 542                                                 EDMA_SLOT_ANY);
 543                         if (echan->slot[i] < 0) {
 544                                 dev_err(dev, "Failed to allocate slot\n");
 545                                 return NULL;
 546                         }
 547                 }
 548
 549                 if (i == nslots - 1) {
 550                         memcpy(&edesc->pset[i], &edesc->pset[0],
 551                                sizeof(edesc->pset[0]));
 552                         break;
 553                 }
 554
 555                 ret = edma_config_pset(chan, &edesc->pset[i], src_addr,
 556                                        dst_addr, burst, dev_width, period_len,
 557                                        direction);
 558                 if (ret < 0)
 559                         return NULL;
 560
 561                 if (direction == DMA_DEV_TO_MEM)
 562                         dst_addr += period_len;
 563                 else
 564                         src_addr += period_len;
 565
 566                 dev_dbg(dev, "%s: Configure period %d of buf:\n", __func__, i);
 567                 dev_dbg(dev,
 568                         "\n pset[%d]:\n"
 569                         "  chnum\t%d\n"
 570                         "  slot\t%d\n"
 571                         "  opt\t%08x\n"
 572                         "  src\t%08x\n"
 573                         "  dst\t%08x\n"
 574                         "  abcnt\t%08x\n"
 575                         "  ccnt\t%08x\n"
 576                         "  bidx\t%08x\n"
 577                         "  cidx\t%08x\n"
 578                         "  lkrld\t%08x\n",
 579                         i, echan->ch_num, echan->slot[i],
 580                         edesc->pset[i].opt,
 581                         edesc->pset[i].src,
 582                         edesc->pset[i].dst,
 583                         edesc->pset[i].a_b_cnt,
 584                         edesc->pset[i].ccnt,
 585                         edesc->pset[i].src_dst_bidx,
 586                         edesc->pset[i].src_dst_cidx,
 587                         edesc->pset[i].link_bcntrld);
 588
 589                 edesc->absync = ret;
 590
 591                 /*
 592                  * Enable interrupts for every period because callback
 593                  * has to be called for every period.
 594                  */
 595                 edesc->pset[i].opt |= TCINTEN;
 596         }
 597
 598         return vchan_tx_prep(&echan->vchan, &edesc->vdesc, tx_flags);
 599 }
 600
 601 static void edma_callback(unsigned ch_num, u16 ch_status, void *data)
 602 {
 603         struct edma_chan *echan = data;
 604         struct device *dev = echan->vchan.chan.device->dev;
 605         struct edma_desc *edesc;
 606         unsigned long flags;
 607         struct edmacc_param p;
 608
 609         edesc = echan->edesc;
 610
 611         /* Pause the channel for non-cyclic */
 612         if (!edesc || (edesc && !edesc->cyclic))
 613                 edma_pause(echan->ch_num);
 614
 615         switch (ch_status) {
 616         case EDMA_DMA_COMPLETE:
 617                 spin_lock_irqsave(&echan->vchan.lock, flags);
 618
 619                 if (edesc) {
 620                         if (edesc->cyclic) {
 621                                 vchan_cyclic_callback(&edesc->vdesc);
 622                         } else if (edesc->processed == edesc->pset_nr) {
 623                                 dev_dbg(dev, "Transfer complete, stopping channel %d\n", ch_num);
 624                                 edma_stop(echan->ch_num);
 625                                 vchan_cookie_complete(&edesc->vdesc);
 626                                 edma_execute(echan);
 627                         } else {
 628                                 dev_dbg(dev, "Intermediate transfer complete on channel %d\n", ch_num);
 629                                 edma_execute(echan);
 630                         }
 631                 }
 632
 633                 spin_unlock_irqrestore(&echan->vchan.lock, flags);
 634
 635                 break;
 636         case EDMA_DMA_CC_ERROR:
 637                 spin_lock_irqsave(&echan->vchan.lock, flags);
 638
 639                 edma_read_slot(EDMA_CHAN_SLOT(echan->slot[0]), &p);
 640
 641                 /*
 642                  * Issue later based on missed flag which will be sure
 643                  * to happen as:
 644                  * (1) we finished transmitting an intermediate slot and
 645                  *     edma_execute is coming up.
 646                  * (2) or we finished current transfer and issue will
 647                  *     call edma_execute.
 648                  *
 649                  * Important note: issuing can be dangerous here and
 650                  * lead to some nasty recursion when we are in a NULL
 651                  * slot. So we avoid doing so and set the missed flag.
 652                  */
 653                 if (p.a_b_cnt == 0 && p.ccnt == 0) {
 654                         dev_dbg(dev, "Error occurred, looks like slot is null, just setting miss\n");
 655                         echan->missed = 1;
 656                 } else {
 657                         /*
 658                          * The slot is already programmed but the event got
 659                          * missed, so its safe to issue it here.
 660                          */
 661                         dev_dbg(dev, "Error occurred but slot is non-null, TRIGGERING\n");
 662                         edma_clean_channel(echan->ch_num);
 663                         edma_stop(echan->ch_num);
 664                         edma_start(echan->ch_num);
 665                         edma_trigger_channel(echan->ch_num);
 666                 }
 667
 668                 spin_unlock_irqrestore(&echan->vchan.lock, flags);
 669
 670                 break;
 671         default:
 672                 break;
 673         }
 674 }
 675
 676 /* Alloc channel resources */
 677 static int edma_alloc_chan_resources(struct dma_chan *chan)
 678 {
 679         struct edma_chan *echan = to_edma_chan(chan);
 680         struct device *dev = chan->device->dev;
 681         int ret;
 682         int a_ch_num;
 683         LIST_HEAD(descs);
 684
 685         a_ch_num = edma_alloc_channel(echan->ch_num, edma_callback,
 686                                         chan, EVENTQ_DEFAULT);
 687
 688         if (a_ch_num < 0) {
 689                 ret = -ENODEV;
 690                 goto err_no_chan;
 691         }
 692
 693         if (a_ch_num != echan->ch_num) {
 694                 dev_err(dev, "failed to allocate requested channel %u:%u\n",
 695                         EDMA_CTLR(echan->ch_num),
 696                         EDMA_CHAN_SLOT(echan->ch_num));
 697                 ret = -ENODEV;
 698                 goto err_wrong_chan;
 699         }
 700
 701         echan->alloced = true;
 702         echan->slot[0] = echan->ch_num;
 703
 704         dev_dbg(dev, "allocated channel for %u:%u\n",
 705                 EDMA_CTLR(echan->ch_num), EDMA_CHAN_SLOT(echan->ch_num));
 706
 707         return 0;
 708
 709 err_wrong_chan:
 710         edma_free_channel(a_ch_num);
 711 err_no_chan:
 712         return ret;
 713 }
 714
 715 /* Free channel resources */
 716 static void edma_free_chan_resources(struct dma_chan *chan)
 717 {
 718         struct edma_chan *echan = to_edma_chan(chan);
 719         struct device *dev = chan->device->dev;
 720         int i;
 721
 722         /* Terminate transfers */
 723         edma_stop(echan->ch_num);
 724
 725         vchan_free_chan_resources(&echan->vchan);
 726
 727         /* Free EDMA PaRAM slots */
 728         for (i = 1; i < EDMA_MAX_SLOTS; i++) {
 729                 if (echan->slot[i] >= 0) {
 730                         edma_free_slot(echan->slot[i]);
 731                         echan->slot[i] = -1;
 732                 }
 733         }
 734
 735         /* Free EDMA channel */
 736         if (echan->alloced) {
 737                 edma_free_channel(echan->ch_num);
 738                 echan->alloced = false;
 739         }
 740
 741         dev_dbg(dev, "freeing channel for %u\n", echan->ch_num);
 742 }
 743
 744 /* Send pending descriptor to hardware */
 745 static void edma_issue_pending(struct dma_chan *chan)
 746 {
 747         struct edma_chan *echan = to_edma_chan(chan);
 748         unsigned long flags;
 749
 750         spin_lock_irqsave(&echan->vchan.lock, flags);
 751         if (vchan_issue_pending(&echan->vchan) && !echan->edesc)
 752                 edma_execute(echan);
 753         spin_unlock_irqrestore(&echan->vchan.lock, flags);
 754 }
 755
 756 static size_t edma_desc_size(struct edma_desc *edesc)
 757 {
 758         int i;
 759         size_t size;
 760
 761         if (edesc->absync)
 762                 for (size = i = 0; i < edesc->pset_nr; i++)
 763                         size += (edesc->pset[i].a_b_cnt & 0xffff) *
 764                                 (edesc->pset[i].a_b_cnt >> 16) *
 765                                  edesc->pset[i].ccnt;
 766         else
 767                 size = (edesc->pset[0].a_b_cnt & 0xffff) *
 768                         (edesc->pset[0].a_b_cnt >> 16) +
 769                         (edesc->pset[0].a_b_cnt & 0xffff) *
 770                         (SZ_64K - 1) * edesc->pset[0].ccnt;
 771
 772         return size;
 773 }
 774
 775 /* Check request completion status */
 776 static enum dma_status edma_tx_status(struct dma_chan *chan,
 777                                       dma_cookie_t cookie,
 778                                       struct dma_tx_state *txstate)
 779 {
 780         struct edma_chan *echan = to_edma_chan(chan);
 781         struct virt_dma_desc *vdesc;
 782         enum dma_status ret;
 783         unsigned long flags;
 784
 785         ret = dma_cookie_status(chan, cookie, txstate);
 786         if (ret == DMA_COMPLETE || !txstate)
 787                 return ret;
 788
 789         spin_lock_irqsave(&echan->vchan.lock, flags);
 790         vdesc = vchan_find_desc(&echan->vchan, cookie);
 791         if (vdesc) {
 792                 txstate->residue = edma_desc_size(to_edma_desc(&vdesc->tx));
 793         } else if (echan->edesc && echan->edesc->vdesc.tx.cookie == cookie) {
 794                 struct edma_desc *edesc = echan->edesc;
 795                 txstate->residue = edma_desc_size(edesc);
 796         }
 797         spin_unlock_irqrestore(&echan->vchan.lock, flags);
 798
 799         return ret;
 800 }
 801
 802 static void __init edma_chan_init(struct edma_cc *ecc,
 803                                   struct dma_device *dma,
 804                                   struct edma_chan *echans)
 805 {
 806         int i, j;
 807
 808         for (i = 0; i < EDMA_CHANS; i++) {
 809                 struct edma_chan *echan = &echans[i];
 810                 echan->ch_num = EDMA_CTLR_CHAN(ecc->ctlr, i);
 811                 echan->ecc = ecc;
 812                 echan->vchan.desc_free = edma_desc_free;
 813
 814                 vchan_init(&echan->vchan, dma);
 815
 816                 INIT_LIST_HEAD(&echan->node);
 817                 for (j = 0; j < EDMA_MAX_SLOTS; j++)
 818                         echan->slot[j] = -1;
 819         }
 820 }
 821
 822 static void edma_dma_init(struct edma_cc *ecc, struct dma_device *dma,
 823                           struct device *dev)
 824 {
 825         dma->device_prep_slave_sg = edma_prep_slave_sg;
 826         dma->device_prep_dma_cyclic = edma_prep_dma_cyclic;
 827         dma->device_alloc_chan_resources = edma_alloc_chan_resources;
 828         dma->device_free_chan_resources = edma_free_chan_resources;
 829         dma->device_issue_pending = edma_issue_pending;
 830         dma->device_tx_status = edma_tx_status;
 831         dma->device_control = edma_control;
 832         dma->dev = dev;
 833
 834         INIT_LIST_HEAD(&dma->channels);
 835 }
 836
 837 static int edma_probe(struct platform_device *pdev)
 838 {
 839         struct edma_cc *ecc;
 840         int ret;
 841
 842         ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32));
 843         if (ret)
 844                 return ret;
 845
 846         ecc = devm_kzalloc(&pdev->dev, sizeof(*ecc), GFP_KERNEL);
 847         if (!ecc) {
 848                 dev_err(&pdev->dev, "Can't allocate controller\n");
 849                 return -ENOMEM;
 850         }
 851
 852         ecc->ctlr = pdev->id;
 853         ecc->dummy_slot = edma_alloc_slot(ecc->ctlr, EDMA_SLOT_ANY);
 854         if (ecc->dummy_slot < 0) {
 855                 dev_err(&pdev->dev, "Can't allocate PaRAM dummy slot\n");
 856                 return -EIO;
 857         }
 858
 859         dma_cap_zero(ecc->dma_slave.cap_mask);
 860         dma_cap_set(DMA_SLAVE, ecc->dma_slave.cap_mask);
 861
 862         edma_dma_init(ecc, &ecc->dma_slave, &pdev->dev);
 863
 864         edma_chan_init(ecc, &ecc->dma_slave, ecc->slave_chans);
 865
 866         ret = dma_async_device_register(&ecc->dma_slave);
 867         if (ret)
 868                 goto err_reg1;
 869
 870         platform_set_drvdata(pdev, ecc);
 871
 872         dev_info(&pdev->dev, "TI EDMA DMA engine driver\n");
 873
 874         return 0;
 875
 876 err_reg1:
 877         edma_free_slot(ecc->dummy_slot);
 878         return ret;
 879 }
 880
 881 static int edma_remove(struct platform_device *pdev)
 882 {
 883         struct device *dev = &pdev->dev;
 884         struct edma_cc *ecc = dev_get_drvdata(dev);
 885
 886         dma_async_device_unregister(&ecc->dma_slave);
 887         edma_free_slot(ecc->dummy_slot);
 888
 889         return 0;
 890 }
 891
 892 static struct platform_driver edma_driver = {
 893         .probe          = edma_probe,
 894         .remove         = edma_remove,
 895         .driver = {
 896                 .name = "edma-dma-engine",
 897                 .owner = THIS_MODULE,
 898         },
 899 };
 900
 901 bool edma_filter_fn(struct dma_chan *chan, void *param)
 902 {
 903         if (chan->device->dev->driver == &edma_driver.driver) {
 904                 struct edma_chan *echan = to_edma_chan(chan);
 905                 unsigned ch_req = *(unsigned *)param;
 906                 return ch_req == echan->ch_num;
 907         }
 908         return false;
 909 }
 910 EXPORT_SYMBOL(edma_filter_fn);
 911
 912 static struct platform_device *pdev0, *pdev1;
 913
 914 static const struct platform_device_info edma_dev_info0 = {
 915         .name = "edma-dma-engine",
 916         .id = 0,
 917         .dma_mask = DMA_BIT_MASK(32),
 918 };
 919
 920 static const struct platform_device_info edma_dev_info1 = {
 921         .name = "edma-dma-engine",
 922         .id = 1,
 923         .dma_mask = DMA_BIT_MASK(32),
 924 };
 925
 926 static int edma_init(void)
 927 {
 928         int ret = platform_driver_register(&edma_driver);
 929
 930         if (ret == 0) {
 931                 pdev0 = platform_device_register_full(&edma_dev_info0);
 932                 if (IS_ERR(pdev0)) {
 933                         platform_driver_unregister(&edma_driver);
 934                         ret = PTR_ERR(pdev0);
 935                         goto out;
 936                 }
 937         }
 938
 939         if (EDMA_CTLRS == 2) {
 940                 pdev1 = platform_device_register_full(&edma_dev_info1);
 941                 if (IS_ERR(pdev1)) {
 942                         platform_driver_unregister(&edma_driver);
 943                         platform_device_unregister(pdev0);
 944                         ret = PTR_ERR(pdev1);
 945                 }
 946         }
 947
 948 out:
 949         return ret;
 950 }
 951 subsys_initcall(edma_init);
 952
 953 static void __exit edma_exit(void)
 954 {
 955         platform_device_unregister(pdev0);
 956         if (pdev1)
 957                 platform_device_unregister(pdev1);
 958         platform_driver_unregister(&edma_driver);
 959 }
 960 module_exit(edma_exit);
 961
 962 MODULE_AUTHOR("Matt Porter <matt.porter@linaro.org>");
 963 MODULE_DESCRIPTION("TI EDMA DMA engine driver");
 964 MODULE_LICENSE("GPL v2");