drivers/dma/amba-pl08x.c

   1 /*
   2  * Copyright (c) 2006 ARM Ltd.
   3  * Copyright (c) 2010 ST-Ericsson SA
   4  *
   5  * Author: Peter Pearse <peter.pearse@arm.com>
   6  * Author: Linus Walleij <linus.walleij@stericsson.com>
   7  *
   8  * This program is free software; you can redistribute it and/or modify it
   9  * under the terms of the GNU General Public License as published by the Free
  10  * Software Foundation; either version 2 of the License, or (at your option)
  11  * any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful, but WITHOUT
  14  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  15  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  16  * more details.
  17  *
  18  * You should have received a copy of the GNU General Public License along with
  19  * this program; if not, write to the Free Software Foundation, Inc., 59
  20  * Temple Place - Suite 330, Boston, MA  02111-1307, USA.
  21  *
  22  * The full GNU General Public License is in this distribution in the
  23  * file called COPYING.
  24  *
  25  * Documentation: ARM DDI 0196G == PL080
  26  * Documentation: ARM DDI 0218E == PL081
  27  *
  28  * PL080 & PL081 both have 16 sets of DMA signals that can be routed to
  29  * any channel.
  30  *
  31  * The PL080 has 8 channels available for simultaneous use, and the PL081
  32  * has only two channels. So on these DMA controllers the number of channels
  33  * and the number of incoming DMA signals are two totally different things.
  34  * It is usually not possible to theoretically handle all physical signals,
  35  * so a multiplexing scheme with possible denial of use is necessary.
  36  *
  37  * The PL080 has a dual bus master, PL081 has a single master.
  38  *
  39  * Memory to peripheral transfer may be visualized as
  40  *      Get data from memory to DMAC
  41  *      Until no data left
  42  *              On burst request from peripheral
  43  *                      Destination burst from DMAC to peripheral
  44  *                      Clear burst request
  45  *      Raise terminal count interrupt
  46  *
  47  * For peripherals with a FIFO:
  48  * Source      burst size == half the depth of the peripheral FIFO
  49  * Destination burst size == the depth of the peripheral FIFO
  50  *
  51  * (Bursts are irrelevant for mem to mem transfers - there are no burst
  52  * signals, the DMA controller will simply facilitate its AHB master.)
  53  *
  54  * ASSUMES default (little) endianness for DMA transfers
  55  *
  56  * The PL08x has two flow control settings:
  57  *  - DMAC flow control: the transfer size defines the number of transfers
  58  *    which occur for the current LLI entry, and the DMAC raises TC at the
  59  *    end of every LLI entry.  Observed behaviour shows the DMAC listening
  60  *    to both the BREQ and SREQ signals (contrary to documented),
  61  *    transferring data if either is active.  The LBREQ and LSREQ signals
  62  *    are ignored.
  63  *
  64  *  - Peripheral flow control: the transfer size is ignored (and should be
  65  *    zero).  The data is transferred from the current LLI entry, until
  66  *    after the final transfer signalled by LBREQ or LSREQ.  The DMAC
  67  *    will then move to the next LLI entry.
  68  *
  69  * Only the former works sanely with scatter lists, so we only implement
  70  * the DMAC flow control method.  However, peripherals which use the LBREQ
  71  * and LSREQ signals (eg, MMCI) are unable to use this mode, which through
  72  * these hardware restrictions prevents them from using scatter DMA.
  73  *
  74  * Global TODO:
  75  * - Break out common code from arch/arm/mach-s3c64xx and share
  76  */
  77 #include <linux/device.h>
  78 #include <linux/init.h>
  79 #include <linux/module.h>
  80 #include <linux/interrupt.h>
  81 #include <linux/slab.h>
  82 #include <linux/dmapool.h>
  83 #include <linux/dmaengine.h>
  84 #include <linux/amba/bus.h>
  85 #include <linux/amba/pl08x.h>
  86 #include <linux/debugfs.h>
  87 #include <linux/seq_file.h>
  88
  89 #include <asm/hardware/pl080.h>
  90
  91 #define DRIVER_NAME     "pl08xdmac"
  92
  93 /**
  94  * struct vendor_data - vendor-specific config parameters
  95  * for PL08x derivatives
  96  * @channels: the number of channels available in this variant
  97  * @dualmaster: whether this version supports dual AHB masters
  98  * or not.
  99  */
 100 struct vendor_data {
 101         u8 channels;
 102         bool dualmaster;
 103 };
 104
 105 /*
 106  * PL08X private data structures
 107  * An LLI struct - see PL08x TRM.  Note that next uses bit[0] as a bus bit,
 108  * start & end do not - their bus bit info is in cctl.  Also note that these
 109  * are fixed 32-bit quantities.
 110  */
 111 struct pl08x_lli {
 112         u32 src;
 113         u32 dst;
 114         u32 lli;
 115         u32 cctl;
 116 };
 117
 118 /**
 119  * struct pl08x_driver_data - the local state holder for the PL08x
 120  * @slave: slave engine for this instance
 121  * @memcpy: memcpy engine for this instance
 122  * @base: virtual memory base (remapped) for the PL08x
 123  * @adev: the corresponding AMBA (PrimeCell) bus entry
 124  * @vd: vendor data for this PL08x variant
 125  * @pd: platform data passed in from the platform/machine
 126  * @phy_chans: array of data for the physical channels
 127  * @pool: a pool for the LLI descriptors
 128  * @pool_ctr: counter of LLIs in the pool
 129  * @lli_buses: bitmask to or in to LLI pointer selecting AHB port for LLI fetches
 130  * @mem_buses: set to indicate memory transfers on AHB2.
 131  * @lock: a spinlock for this struct
 132  */
 133 struct pl08x_driver_data {
 134         struct dma_device slave;
 135         struct dma_device memcpy;
 136         void __iomem *base;
 137         struct amba_device *adev;
 138         const struct vendor_data *vd;
 139         struct pl08x_platform_data *pd;
 140         struct pl08x_phy_chan *phy_chans;
 141         struct dma_pool *pool;
 142         int pool_ctr;
 143         u8 lli_buses;
 144         u8 mem_buses;
 145         spinlock_t lock;
 146 };
 147
 148 /*
 149  * PL08X specific defines
 150  */
 151
 152 /*
 153  * Memory boundaries: the manual for PL08x says that the controller
 154  * cannot read past a 1KiB boundary, so these defines are used to
 155  * create transfer LLIs that do not cross such boundaries.
 156  */
 157 #define PL08X_BOUNDARY_SHIFT            (10)    /* 1KB 0x400 */
 158 #define PL08X_BOUNDARY_SIZE             (1 << PL08X_BOUNDARY_SHIFT)
 159
 160 /* Minimum period between work queue runs */
 161 #define PL08X_WQ_PERIODMIN      20
 162
 163 /* Size (bytes) of each LLI buffer allocated for one transfer */
 164 # define PL08X_LLI_TSFR_SIZE    0x2000
 165
 166 /* Maximum times we call dma_pool_alloc on this pool without freeing */
 167 #define PL08X_MAX_ALLOCS        0x40
 168 #define MAX_NUM_TSFR_LLIS       (PL08X_LLI_TSFR_SIZE/sizeof(struct pl08x_lli))
 169 #define PL08X_ALIGN             8
 170
 171 static inline struct pl08x_dma_chan *to_pl08x_chan(struct dma_chan *chan)
 172 {
 173         return container_of(chan, struct pl08x_dma_chan, chan);
 174 }
 175
 176 /*
 177  * Physical channel handling
 178  */
 179
 180 /* Whether a certain channel is busy or not */
 181 static int pl08x_phy_channel_busy(struct pl08x_phy_chan *ch)
 182 {
 183         unsigned int val;
 184
 185         val = readl(ch->base + PL080_CH_CONFIG);
 186         return val & PL080_CONFIG_ACTIVE;
 187 }
 188
 189 /*
 190  * Set the initial DMA register values i.e. those for the first LLI
 191  * The next LLI pointer and the configuration interrupt bit have
 192  * been set when the LLIs were constructed.  Poke them into the hardware
 193  * and start the transfer.
 194  */
 195 static void pl08x_start_txd(struct pl08x_dma_chan *plchan,
 196         struct pl08x_txd *txd)
 197 {
 198         struct pl08x_driver_data *pl08x = plchan->host;
 199         struct pl08x_phy_chan *phychan = plchan->phychan;
 200         struct pl08x_lli *lli = &txd->llis_va[0];
 201         u32 val;
 202
 203         plchan->at = txd;
 204
 205         /* Wait for channel inactive */
 206         while (pl08x_phy_channel_busy(phychan))
 207                 cpu_relax();
 208
 209         dev_vdbg(&pl08x->adev->dev,
 210                 "WRITE channel %d: csrc=0x%08x, cdst=0x%08x, "
 211                 "clli=0x%08x, cctl=0x%08x, ccfg=0x%08x\n",
 212                 phychan->id, lli->src, lli->dst, lli->lli, lli->cctl,
 213                 txd->ccfg);
 214
 215         writel(lli->src, phychan->base + PL080_CH_SRC_ADDR);
 216         writel(lli->dst, phychan->base + PL080_CH_DST_ADDR);
 217         writel(lli->lli, phychan->base + PL080_CH_LLI);
 218         writel(lli->cctl, phychan->base + PL080_CH_CONTROL);
 219         writel(txd->ccfg, phychan->base + PL080_CH_CONFIG);
 220
 221         /* Enable the DMA channel */
 222         /* Do not access config register until channel shows as disabled */
 223         while (readl(pl08x->base + PL080_EN_CHAN) & (1 << phychan->id))
 224                 cpu_relax();
 225
 226         /* Do not access config register until channel shows as inactive */
 227         val = readl(phychan->base + PL080_CH_CONFIG);
 228         while ((val & PL080_CONFIG_ACTIVE) || (val & PL080_CONFIG_ENABLE))
 229                 val = readl(phychan->base + PL080_CH_CONFIG);
 230
 231         writel(val | PL080_CONFIG_ENABLE, phychan->base + PL080_CH_CONFIG);
 232 }
 233
 234 /*
 235  * Overall DMAC remains enabled always.
 236  *
 237  * Disabling individual channels could lose data.
 238  *
 239  * Disable the peripheral DMA after disabling the DMAC
 240  * in order to allow the DMAC FIFO to drain, and
 241  * hence allow the channel to show inactive
 242  *
 243  */
 244 static void pl08x_pause_phy_chan(struct pl08x_phy_chan *ch)
 245 {
 246         u32 val;
 247
 248         /* Set the HALT bit and wait for the FIFO to drain */
 249         val = readl(ch->base + PL080_CH_CONFIG);
 250         val |= PL080_CONFIG_HALT;
 251         writel(val, ch->base + PL080_CH_CONFIG);
 252
 253         /* Wait for channel inactive */
 254         while (pl08x_phy_channel_busy(ch))
 255                 cpu_relax();
 256 }
 257
 258 static void pl08x_resume_phy_chan(struct pl08x_phy_chan *ch)
 259 {
 260         u32 val;
 261
 262         /* Clear the HALT bit */
 263         val = readl(ch->base + PL080_CH_CONFIG);
 264         val &= ~PL080_CONFIG_HALT;
 265         writel(val, ch->base + PL080_CH_CONFIG);
 266 }
 267
 268
 269 /* Stops the channel */
 270 static void pl08x_stop_phy_chan(struct pl08x_phy_chan *ch)
 271 {
 272         u32 val;
 273
 274         pl08x_pause_phy_chan(ch);
 275
 276         /* Disable channel */
 277         val = readl(ch->base + PL080_CH_CONFIG);
 278         val &= ~PL080_CONFIG_ENABLE;
 279         val &= ~PL080_CONFIG_ERR_IRQ_MASK;
 280         val &= ~PL080_CONFIG_TC_IRQ_MASK;
 281         writel(val, ch->base + PL080_CH_CONFIG);
 282 }
 283
 284 static inline u32 get_bytes_in_cctl(u32 cctl)
 285 {
 286         /* The source width defines the number of bytes */
 287         u32 bytes = cctl & PL080_CONTROL_TRANSFER_SIZE_MASK;
 288
 289         switch (cctl >> PL080_CONTROL_SWIDTH_SHIFT) {
 290         case PL080_WIDTH_8BIT:
 291                 break;
 292         case PL080_WIDTH_16BIT:
 293                 bytes *= 2;
 294                 break;
 295         case PL080_WIDTH_32BIT:
 296                 bytes *= 4;
 297                 break;
 298         }
 299         return bytes;
 300 }
 301
 302 /* The channel should be paused when calling this */
 303 static u32 pl08x_getbytes_chan(struct pl08x_dma_chan *plchan)
 304 {
 305         struct pl08x_phy_chan *ch;
 306         struct pl08x_txd *txd;
 307         unsigned long flags;
 308         size_t bytes = 0;
 309
 310         spin_lock_irqsave(&plchan->lock, flags);
 311         ch = plchan->phychan;
 312         txd = plchan->at;
 313
 314         /*
 315          * Follow the LLIs to get the number of remaining
 316          * bytes in the currently active transaction.
 317          */
 318         if (ch && txd) {
 319                 u32 clli = readl(ch->base + PL080_CH_LLI) & ~PL080_LLI_LM_AHB2;
 320
 321                 /* First get the remaining bytes in the active transfer */
 322                 bytes = get_bytes_in_cctl(readl(ch->base + PL080_CH_CONTROL));
 323
 324                 if (clli) {
 325                         struct pl08x_lli *llis_va = txd->llis_va;
 326                         dma_addr_t llis_bus = txd->llis_bus;
 327                         int index;
 328
 329                         BUG_ON(clli < llis_bus || clli >= llis_bus +
 330                                 sizeof(struct pl08x_lli) * MAX_NUM_TSFR_LLIS);
 331
 332                         /*
 333                          * Locate the next LLI - as this is an array,
 334                          * it's simple maths to find.
 335                          */
 336                         index = (clli - llis_bus) / sizeof(struct pl08x_lli);
 337
 338                         for (; index < MAX_NUM_TSFR_LLIS; index++) {
 339                                 bytes += get_bytes_in_cctl(llis_va[index].cctl);
 340
 341                                 /*
 342                                  * A LLI pointer of 0 terminates the LLI list
 343                                  */
 344                                 if (!llis_va[index].lli)
 345                                         break;
 346                         }
 347                 }
 348         }
 349
 350         /* Sum up all queued transactions */
 351         if (!list_empty(&plchan->pend_list)) {
 352                 struct pl08x_txd *txdi;
 353                 list_for_each_entry(txdi, &plchan->pend_list, node) {
 354                         bytes += txdi->len;
 355                 }
 356         }
 357
 358         spin_unlock_irqrestore(&plchan->lock, flags);
 359
 360         return bytes;
 361 }
 362
 363 /*
 364  * Allocate a physical channel for a virtual channel
 365  */
 366 static struct pl08x_phy_chan *
 367 pl08x_get_phy_channel(struct pl08x_driver_data *pl08x,
 368                       struct pl08x_dma_chan *virt_chan)
 369 {
 370         struct pl08x_phy_chan *ch = NULL;
 371         unsigned long flags;
 372         int i;
 373
 374         /*
 375          * Try to locate a physical channel to be used for
 376          * this transfer. If all are taken return NULL and
 377          * the requester will have to cope by using some fallback
 378          * PIO mode or retrying later.
 379          */
 380         for (i = 0; i < pl08x->vd->channels; i++) {
 381                 ch = &pl08x->phy_chans[i];
 382
 383                 spin_lock_irqsave(&ch->lock, flags);
 384
 385                 if (!ch->serving) {
 386                         ch->serving = virt_chan;
 387                         ch->signal = -1;
 388                         spin_unlock_irqrestore(&ch->lock, flags);
 389                         break;
 390                 }
 391
 392                 spin_unlock_irqrestore(&ch->lock, flags);
 393         }
 394
 395         if (i == pl08x->vd->channels) {
 396                 /* No physical channel available, cope with it */
 397                 return NULL;
 398         }
 399
 400         return ch;
 401 }
 402
 403 static inline void pl08x_put_phy_channel(struct pl08x_driver_data *pl08x,
 404                                          struct pl08x_phy_chan *ch)
 405 {
 406         unsigned long flags;
 407
 408         /* Stop the channel and clear its interrupts */
 409         pl08x_stop_phy_chan(ch);
 410         writel((1 << ch->id), pl08x->base + PL080_ERR_CLEAR);
 411         writel((1 << ch->id), pl08x->base + PL080_TC_CLEAR);
 412
 413         /* Mark it as free */
 414         spin_lock_irqsave(&ch->lock, flags);
 415         ch->serving = NULL;
 416         spin_unlock_irqrestore(&ch->lock, flags);
 417 }
 418
 419 /*
 420  * LLI handling
 421  */
 422
 423 static inline unsigned int pl08x_get_bytes_for_cctl(unsigned int coded)
 424 {
 425         switch (coded) {
 426         case PL080_WIDTH_8BIT:
 427                 return 1;
 428         case PL080_WIDTH_16BIT:
 429                 return 2;
 430         case PL080_WIDTH_32BIT:
 431                 return 4;
 432         default:
 433                 break;
 434         }
 435         BUG();
 436         return 0;
 437 }
 438
 439 static inline u32 pl08x_cctl_bits(u32 cctl, u8 srcwidth, u8 dstwidth,
 440                                   size_t tsize)
 441 {
 442         u32 retbits = cctl;
 443
 444         /* Remove all src, dst and transfer size bits */
 445         retbits &= ~PL080_CONTROL_DWIDTH_MASK;
 446         retbits &= ~PL080_CONTROL_SWIDTH_MASK;
 447         retbits &= ~PL080_CONTROL_TRANSFER_SIZE_MASK;
 448
 449         /* Then set the bits according to the parameters */
 450         switch (srcwidth) {
 451         case 1:
 452                 retbits |= PL080_WIDTH_8BIT << PL080_CONTROL_SWIDTH_SHIFT;
 453                 break;
 454         case 2:
 455                 retbits |= PL080_WIDTH_16BIT << PL080_CONTROL_SWIDTH_SHIFT;
 456                 break;
 457         case 4:
 458                 retbits |= PL080_WIDTH_32BIT << PL080_CONTROL_SWIDTH_SHIFT;
 459                 break;
 460         default:
 461                 BUG();
 462                 break;
 463         }
 464
 465         switch (dstwidth) {
 466         case 1:
 467                 retbits |= PL080_WIDTH_8BIT << PL080_CONTROL_DWIDTH_SHIFT;
 468                 break;
 469         case 2:
 470                 retbits |= PL080_WIDTH_16BIT << PL080_CONTROL_DWIDTH_SHIFT;
 471                 break;
 472         case 4:
 473                 retbits |= PL080_WIDTH_32BIT << PL080_CONTROL_DWIDTH_SHIFT;
 474                 break;
 475         default:
 476                 BUG();
 477                 break;
 478         }
 479
 480         retbits |= tsize << PL080_CONTROL_TRANSFER_SIZE_SHIFT;
 481         return retbits;
 482 }
 483
 484 struct pl08x_lli_build_data {
 485         struct pl08x_txd *txd;
 486         struct pl08x_driver_data *pl08x;
 487         struct pl08x_bus_data srcbus;
 488         struct pl08x_bus_data dstbus;
 489         size_t remainder;
 490 };
 491
 492 /*
 493  * Autoselect a master bus to use for the transfer
 494  * this prefers the destination bus if both available
 495  * if fixed address on one bus the other will be chosen
 496  */
 497 static void pl08x_choose_master_bus(struct pl08x_lli_build_data *bd,
 498         struct pl08x_bus_data **mbus, struct pl08x_bus_data **sbus, u32 cctl)
 499 {
 500         if (!(cctl & PL080_CONTROL_DST_INCR)) {
 501                 *mbus = &bd->srcbus;
 502                 *sbus = &bd->dstbus;
 503         } else if (!(cctl & PL080_CONTROL_SRC_INCR)) {
 504                 *mbus = &bd->dstbus;
 505                 *sbus = &bd->srcbus;
 506         } else {
 507                 if (bd->dstbus.buswidth == 4) {
 508                         *mbus = &bd->dstbus;
 509                         *sbus = &bd->srcbus;
 510                 } else if (bd->srcbus.buswidth == 4) {
 511                         *mbus = &bd->srcbus;
 512                         *sbus = &bd->dstbus;
 513                 } else if (bd->dstbus.buswidth == 2) {
 514                         *mbus = &bd->dstbus;
 515                         *sbus = &bd->srcbus;
 516                 } else if (bd->srcbus.buswidth == 2) {
 517                         *mbus = &bd->srcbus;
 518                         *sbus = &bd->dstbus;
 519                 } else {
 520                         /* bd->srcbus.buswidth == 1 */
 521                         *mbus = &bd->dstbus;
 522                         *sbus = &bd->srcbus;
 523                 }
 524         }
 525 }
 526
 527 /*
 528  * Fills in one LLI for a certain transfer descriptor
 529  * and advance the counter
 530  */
 531 static void pl08x_fill_lli_for_desc(struct pl08x_lli_build_data *bd,
 532         int num_llis, int len, u32 cctl)
 533 {
 534         struct pl08x_lli *llis_va = bd->txd->llis_va;
 535         dma_addr_t llis_bus = bd->txd->llis_bus;
 536
 537         BUG_ON(num_llis >= MAX_NUM_TSFR_LLIS);
 538
 539         llis_va[num_llis].cctl = cctl;
 540         llis_va[num_llis].src = bd->srcbus.addr;
 541         llis_va[num_llis].dst = bd->dstbus.addr;
 542         llis_va[num_llis].lli = llis_bus + (num_llis + 1) * sizeof(struct pl08x_lli);
 543         if (bd->pl08x->lli_buses & PL08X_AHB2)
 544                 llis_va[num_llis].lli |= PL080_LLI_LM_AHB2;
 545
 546         if (cctl & PL080_CONTROL_SRC_INCR)
 547                 bd->srcbus.addr += len;
 548         if (cctl & PL080_CONTROL_DST_INCR)
 549                 bd->dstbus.addr += len;
 550
 551         BUG_ON(bd->remainder < len);
 552
 553         bd->remainder -= len;
 554 }
 555
 556 /*
 557  * Return number of bytes to fill to boundary, or len.
 558  * This calculation works for any value of addr.
 559  */
 560 static inline size_t pl08x_pre_boundary(u32 addr, size_t len)
 561 {
 562         size_t boundary_len = PL08X_BOUNDARY_SIZE -
 563                         (addr & (PL08X_BOUNDARY_SIZE - 1));
 564
 565         return min(boundary_len, len);
 566 }
 567
 568 /*
 569  * This fills in the table of LLIs for the transfer descriptor
 570  * Note that we assume we never have to change the burst sizes
 571  * Return 0 for error
 572  */
 573 static int pl08x_fill_llis_for_desc(struct pl08x_driver_data *pl08x,
 574                               struct pl08x_txd *txd)
 575 {
 576         struct pl08x_bus_data *mbus, *sbus;
 577         struct pl08x_lli_build_data bd;
 578         int num_llis = 0;
 579         u32 cctl;
 580         size_t max_bytes_per_lli;
 581         size_t total_bytes = 0;
 582         struct pl08x_lli *llis_va;
 583
 584         txd->llis_va = dma_pool_alloc(pl08x->pool, GFP_NOWAIT,
 585                                       &txd->llis_bus);
 586         if (!txd->llis_va) {
 587                 dev_err(&pl08x->adev->dev, "%s no memory for llis\n", __func__);
 588                 return 0;
 589         }
 590
 591         pl08x->pool_ctr++;
 592
 593         /* Get the default CCTL */
 594         cctl = txd->cctl;
 595
 596         bd.txd = txd;
 597         bd.pl08x = pl08x;
 598         bd.srcbus.addr = txd->src_addr;
 599         bd.dstbus.addr = txd->dst_addr;
 600
 601         /* Find maximum width of the source bus */
 602         bd.srcbus.maxwidth =
 603                 pl08x_get_bytes_for_cctl((cctl & PL080_CONTROL_SWIDTH_MASK) >>
 604                                        PL080_CONTROL_SWIDTH_SHIFT);
 605
 606         /* Find maximum width of the destination bus */
 607         bd.dstbus.maxwidth =
 608                 pl08x_get_bytes_for_cctl((cctl & PL080_CONTROL_DWIDTH_MASK) >>
 609                                        PL080_CONTROL_DWIDTH_SHIFT);
 610
 611         /* Set up the bus widths to the maximum */
 612         bd.srcbus.buswidth = bd.srcbus.maxwidth;
 613         bd.dstbus.buswidth = bd.dstbus.maxwidth;
 614         dev_vdbg(&pl08x->adev->dev,
 615                  "%s source bus is %d bytes wide, dest bus is %d bytes wide\n",
 616                  __func__, bd.srcbus.buswidth, bd.dstbus.buswidth);
 617
 618
 619         /*
 620          * Bytes transferred == tsize * MIN(buswidths), not max(buswidths)
 621          */
 622         max_bytes_per_lli = min(bd.srcbus.buswidth, bd.dstbus.buswidth) *
 623                 PL080_CONTROL_TRANSFER_SIZE_MASK;
 624         dev_vdbg(&pl08x->adev->dev,
 625                  "%s max bytes per lli = %zu\n",
 626                  __func__, max_bytes_per_lli);
 627
 628         /* We need to count this down to zero */
 629         bd.remainder = txd->len;
 630         dev_vdbg(&pl08x->adev->dev,
 631                  "%s remainder = %zu\n",
 632                  __func__, bd.remainder);
 633
 634         /*
 635          * Choose bus to align to
 636          * - prefers destination bus if both available
 637          * - if fixed address on one bus chooses other
 638          * - modifies cctl to choose an appropriate master
 639          */
 640         pl08x_choose_master_bus(&bd, &mbus, &sbus, cctl);
 641
 642         if (txd->len < mbus->buswidth) {
 643                 /*
 644                  * Less than a bus width available
 645                  * - send as single bytes
 646                  */
 647                 while (bd.remainder) {
 648                         dev_vdbg(&pl08x->adev->dev,
 649                                  "%s single byte LLIs for a transfer of "
 650                                  "less than a bus width (remain 0x%08x)\n",
 651                                  __func__, bd.remainder);
 652                         cctl = pl08x_cctl_bits(cctl, 1, 1, 1);
 653                         pl08x_fill_lli_for_desc(&bd, num_llis++, 1, cctl);
 654                         total_bytes++;
 655                 }
 656         } else {
 657                 /*
 658                  *  Make one byte LLIs until master bus is aligned
 659                  *  - slave will then be aligned also
 660                  */
 661                 while ((mbus->addr) % (mbus->buswidth)) {
 662                         dev_vdbg(&pl08x->adev->dev,
 663                                 "%s adjustment lli for less than bus width "
 664                                  "(remain 0x%08x)\n",
 665                                  __func__, bd.remainder);
 666                         cctl = pl08x_cctl_bits(cctl, 1, 1, 1);
 667                         pl08x_fill_lli_for_desc(&bd, num_llis++, 1, cctl);
 668                         total_bytes++;
 669                 }
 670
 671                 /*
 672                  *  Master now aligned
 673                  * - if slave is not then we must set its width down
 674                  */
 675                 if (sbus->addr % sbus->buswidth) {
 676                         dev_dbg(&pl08x->adev->dev,
 677                                 "%s set down bus width to one byte\n",
 678                                  __func__);
 679
 680                         sbus->buswidth = 1;
 681                 }
 682
 683                 /*
 684                  * Make largest possible LLIs until less than one bus
 685                  * width left
 686                  */
 687                 while (bd.remainder > (mbus->buswidth - 1)) {
 688                         size_t lli_len, target_len, tsize, odd_bytes;
 689
 690                         /*
 691                          * If enough left try to send max possible,
 692                          * otherwise try to send the remainder
 693                          */
 694                         target_len = min(bd.remainder, max_bytes_per_lli);
 695
 696                         /*
 697                          * Set bus lengths for incrementing buses to the
 698                          * number of bytes which fill to next memory boundary,
 699                          * limiting on the target length calculated above.
 700                          */
 701                         if (cctl & PL080_CONTROL_SRC_INCR)
 702                                 bd.srcbus.fill_bytes =
 703                                         pl08x_pre_boundary(bd.srcbus.addr,
 704                                                 target_len);
 705                         else
 706                                 bd.srcbus.fill_bytes = target_len;
 707
 708                         if (cctl & PL080_CONTROL_DST_INCR)
 709                                 bd.dstbus.fill_bytes =
 710                                         pl08x_pre_boundary(bd.dstbus.addr,
 711                                                 target_len);
 712                         else
 713                                 bd.dstbus.fill_bytes = target_len;
 714
 715                         /* Find the nearest */
 716                         lli_len = min(bd.srcbus.fill_bytes,
 717                                       bd.dstbus.fill_bytes);
 718
 719                         BUG_ON(lli_len > bd.remainder);
 720
 721                         if (lli_len <= 0) {
 722                                 dev_err(&pl08x->adev->dev,
 723                                         "%s lli_len is %zu, <= 0\n",
 724                                                 __func__, lli_len);
 725                                 return 0;
 726                         }
 727
 728                         if (lli_len == target_len) {
 729                                 /*
 730                                  * Can send what we wanted
 731                                  */
 732                                 /*
 733                                  *  Maintain alignment
 734                                  */
 735                                 lli_len = (lli_len/mbus->buswidth) *
 736                                                         mbus->buswidth;
 737                                 odd_bytes = 0;
 738                         } else {
 739                                 /*
 740                                  * So now we know how many bytes to transfer
 741                                  * to get to the nearest boundary
 742                                  * The next LLI will past the boundary
 743                                  * - however we may be working to a boundary
 744                                  *   on the slave bus
 745                                  *   We need to ensure the master stays aligned
 746                                  */
 747                                 odd_bytes = lli_len % mbus->buswidth;
 748                                 /*
 749                                  * - and that we are working in multiples
 750                                  *   of the bus widths
 751                                  */
 752                                 lli_len -= odd_bytes;
 753
 754                         }
 755
 756                         if (lli_len) {
 757                                 /*
 758                                  * Check against minimum bus alignment:
 759                                  * Calculate actual transfer size in relation
 760                                  * to bus width an get a maximum remainder of
 761                                  * the smallest bus width - 1
 762                                  */
 763                                 /* FIXME: use round_down()? */
 764                                 tsize = lli_len / min(mbus->buswidth,
 765                                                       sbus->buswidth);
 766                                 lli_len = tsize * min(mbus->buswidth,
 767                                                       sbus->buswidth);
 768
 769                                 if (target_len != lli_len) {
 770                                         dev_vdbg(&pl08x->adev->dev,
 771                                         "%s can't send what we want. Desired 0x%08zx, lli of 0x%08zx bytes in txd of 0x%08zx\n",
 772                                         __func__, target_len, lli_len, txd->len);
 773                                 }
 774
 775                                 cctl = pl08x_cctl_bits(cctl,
 776                                                        bd.srcbus.buswidth,
 777                                                        bd.dstbus.buswidth,
 778                                                        tsize);
 779
 780                                 dev_vdbg(&pl08x->adev->dev,
 781                                         "%s fill lli with single lli chunk of size 0x%08zx (remainder 0x%08zx)\n",
 782                                         __func__, lli_len, bd.remainder);
 783                                 pl08x_fill_lli_for_desc(&bd, num_llis++,
 784                                         lli_len, cctl);
 785                                 total_bytes += lli_len;
 786                         }
 787
 788
 789                         if (odd_bytes) {
 790                                 /*
 791                                  * Creep past the boundary,
 792                                  * maintaining master alignment
 793                                  */
 794                                 int j;
 795                                 for (j = 0; (j < mbus->buswidth)
 796                                                 && (bd.remainder); j++) {
 797                                         cctl = pl08x_cctl_bits(cctl, 1, 1, 1);
 798                                         dev_vdbg(&pl08x->adev->dev,
 799                                                 "%s align with boundary, single byte (remain 0x%08zx)\n",
 800                                                 __func__, bd.remainder);
 801                                         pl08x_fill_lli_for_desc(&bd,
 802                                                 num_llis++, 1, cctl);
 803                                         total_bytes++;
 804                                 }
 805                         }
 806                 }
 807
 808                 /*
 809                  * Send any odd bytes
 810                  */
 811                 while (bd.remainder) {
 812                         cctl = pl08x_cctl_bits(cctl, 1, 1, 1);
 813                         dev_vdbg(&pl08x->adev->dev,
 814                                 "%s align with boundary, single odd byte (remain %zu)\n",
 815                                 __func__, bd.remainder);
 816                         pl08x_fill_lli_for_desc(&bd, num_llis++, 1, cctl);
 817                         total_bytes++;
 818                 }
 819         }
 820         if (total_bytes != txd->len) {
 821                 dev_err(&pl08x->adev->dev,
 822                         "%s size of encoded lli:s don't match total txd, transferred 0x%08zx from size 0x%08zx\n",
 823                         __func__, total_bytes, txd->len);
 824                 return 0;
 825         }
 826
 827         if (num_llis >= MAX_NUM_TSFR_LLIS) {
 828                 dev_err(&pl08x->adev->dev,
 829                         "%s need to increase MAX_NUM_TSFR_LLIS from 0x%08x\n",
 830                         __func__, (u32) MAX_NUM_TSFR_LLIS);
 831                 return 0;
 832         }
 833
 834         llis_va = txd->llis_va;
 835         /*
 836          * The final LLI terminates the LLI.
 837          */
 838         llis_va[num_llis - 1].lli = 0;
 839         /*
 840          * The final LLI element shall also fire an interrupt
 841          */
 842         llis_va[num_llis - 1].cctl |= PL080_CONTROL_TC_IRQ_EN;
 843
 844 #ifdef VERBOSE_DEBUG
 845         {
 846                 int i;
 847
 848                 for (i = 0; i < num_llis; i++) {
 849                         dev_vdbg(&pl08x->adev->dev,
 850                                  "lli %d @%p: csrc=0x%08x, cdst=0x%08x, cctl=0x%08x, clli=0x%08x\n",
 851                                  i,
 852                                  &llis_va[i],
 853                                  llis_va[i].src,
 854                                  llis_va[i].dst,
 855                                  llis_va[i].cctl,
 856                                  llis_va[i].lli
 857                                 );
 858                 }
 859         }
 860 #endif
 861
 862         return num_llis;
 863 }
 864
 865 /* You should call this with the struct pl08x lock held */
 866 static void pl08x_free_txd(struct pl08x_driver_data *pl08x,
 867                            struct pl08x_txd *txd)
 868 {
 869         /* Free the LLI */
 870         dma_pool_free(pl08x->pool, txd->llis_va, txd->llis_bus);
 871
 872         pl08x->pool_ctr--;
 873
 874         kfree(txd);
 875 }
 876
 877 static void pl08x_free_txd_list(struct pl08x_driver_data *pl08x,
 878                                 struct pl08x_dma_chan *plchan)
 879 {
 880         struct pl08x_txd *txdi = NULL;
 881         struct pl08x_txd *next;
 882
 883         if (!list_empty(&plchan->pend_list)) {
 884                 list_for_each_entry_safe(txdi,
 885                                          next, &plchan->pend_list, node) {
 886                         list_del(&txdi->node);
 887                         pl08x_free_txd(pl08x, txdi);
 888                 }
 889
 890         }
 891 }
 892
 893 /*
 894  * The DMA ENGINE API
 895  */
 896 static int pl08x_alloc_chan_resources(struct dma_chan *chan)
 897 {
 898         return 0;
 899 }
 900
 901 static void pl08x_free_chan_resources(struct dma_chan *chan)
 902 {
 903 }
 904
 905 /*
 906  * This should be called with the channel plchan->lock held
 907  */
 908 static int prep_phy_channel(struct pl08x_dma_chan *plchan,
 909                             struct pl08x_txd *txd)
 910 {
 911         struct pl08x_driver_data *pl08x = plchan->host;
 912         struct pl08x_phy_chan *ch;
 913         int ret;
 914
 915         /* Check if we already have a channel */
 916         if (plchan->phychan)
 917                 return 0;
 918
 919         ch = pl08x_get_phy_channel(pl08x, plchan);
 920         if (!ch) {
 921                 /* No physical channel available, cope with it */
 922                 dev_dbg(&pl08x->adev->dev, "no physical channel available for xfer on %s\n", plchan->name);
 923                 return -EBUSY;
 924         }
 925
 926         /*
 927          * OK we have a physical channel: for memcpy() this is all we
 928          * need, but for slaves the physical signals may be muxed!
 929          * Can the platform allow us to use this channel?
 930          */
 931         if (plchan->slave &&
 932             ch->signal < 0 &&
 933             pl08x->pd->get_signal) {
 934                 ret = pl08x->pd->get_signal(plchan);
 935                 if (ret < 0) {
 936                         dev_dbg(&pl08x->adev->dev,
 937                                 "unable to use physical channel %d for transfer on %s due to platform restrictions\n",
 938                                 ch->id, plchan->name);
 939                         /* Release physical channel & return */
 940                         pl08x_put_phy_channel(pl08x, ch);
 941                         return -EBUSY;
 942                 }
 943                 ch->signal = ret;
 944
 945                 /* Assign the flow control signal to this channel */
 946                 if (txd->direction == DMA_TO_DEVICE)
 947                         txd->ccfg |= ch->signal << PL080_CONFIG_DST_SEL_SHIFT;
 948                 else if (txd->direction == DMA_FROM_DEVICE)
 949                         txd->ccfg |= ch->signal << PL080_CONFIG_SRC_SEL_SHIFT;
 950         }
 951
 952         dev_dbg(&pl08x->adev->dev, "allocated physical channel %d and signal %d for xfer on %s\n",
 953                  ch->id,
 954                  ch->signal,
 955                  plchan->name);
 956
 957         plchan->phychan = ch;
 958
 959         return 0;
 960 }
 961
 962 static void release_phy_channel(struct pl08x_dma_chan *plchan)
 963 {
 964         struct pl08x_driver_data *pl08x = plchan->host;
 965
 966         if ((plchan->phychan->signal >= 0) && pl08x->pd->put_signal) {
 967                 pl08x->pd->put_signal(plchan);
 968                 plchan->phychan->signal = -1;
 969         }
 970         pl08x_put_phy_channel(pl08x, plchan->phychan);
 971         plchan->phychan = NULL;
 972 }
 973
 974 static dma_cookie_t pl08x_tx_submit(struct dma_async_tx_descriptor *tx)
 975 {
 976         struct pl08x_dma_chan *plchan = to_pl08x_chan(tx->chan);
 977
 978         plchan->chan.cookie += 1;
 979         if (plchan->chan.cookie < 0)
 980                 plchan->chan.cookie = 1;
 981         tx->cookie = plchan->chan.cookie;
 982         /* This unlock follows the lock in the prep() function */
 983         spin_unlock_irqrestore(&plchan->lock, plchan->lockflags);
 984
 985         return tx->cookie;
 986 }
 987
 988 static struct dma_async_tx_descriptor *pl08x_prep_dma_interrupt(
 989                 struct dma_chan *chan, unsigned long flags)
 990 {
 991         struct dma_async_tx_descriptor *retval = NULL;
 992
 993         return retval;
 994 }
 995
 996 /*
 997  * Code accessing dma_async_is_complete() in a tight loop
 998  * may give problems - could schedule where indicated.
 999  * If slaves are relying on interrupts to signal completion this
1000  * function must not be called with interrupts disabled
1001  */
1002 static enum dma_status
1003 pl08x_dma_tx_status(struct dma_chan *chan,
1004                     dma_cookie_t cookie,
1005                     struct dma_tx_state *txstate)
1006 {
1007         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1008         dma_cookie_t last_used;
1009         dma_cookie_t last_complete;
1010         enum dma_status ret;
1011         u32 bytesleft = 0;
1012
1013         last_used = plchan->chan.cookie;
1014         last_complete = plchan->lc;
1015
1016         ret = dma_async_is_complete(cookie, last_complete, last_used);
1017         if (ret == DMA_SUCCESS) {
1018                 dma_set_tx_state(txstate, last_complete, last_used, 0);
1019                 return ret;
1020         }
1021
1022         /*
1023          * schedule(); could be inserted here
1024          */
1025
1026         /*
1027          * This cookie not complete yet
1028          */
1029         last_used = plchan->chan.cookie;
1030         last_complete = plchan->lc;
1031
1032         /* Get number of bytes left in the active transactions and queue */
1033         bytesleft = pl08x_getbytes_chan(plchan);
1034
1035         dma_set_tx_state(txstate, last_complete, last_used,
1036                          bytesleft);
1037
1038         if (plchan->state == PL08X_CHAN_PAUSED)
1039                 return DMA_PAUSED;
1040
1041         /* Whether waiting or running, we're in progress */
1042         return DMA_IN_PROGRESS;
1043 }
1044
1045 /* PrimeCell DMA extension */
1046 struct burst_table {
1047         int burstwords;
1048         u32 reg;
1049 };
1050
1051 static const struct burst_table burst_sizes[] = {
1052         {
1053                 .burstwords = 256,
1054                 .reg = (PL080_BSIZE_256 << PL080_CONTROL_SB_SIZE_SHIFT) |
1055                         (PL080_BSIZE_256 << PL080_CONTROL_DB_SIZE_SHIFT),
1056         },
1057         {
1058                 .burstwords = 128,
1059                 .reg = (PL080_BSIZE_128 << PL080_CONTROL_SB_SIZE_SHIFT) |
1060                         (PL080_BSIZE_128 << PL080_CONTROL_DB_SIZE_SHIFT),
1061         },
1062         {
1063                 .burstwords = 64,
1064                 .reg = (PL080_BSIZE_64 << PL080_CONTROL_SB_SIZE_SHIFT) |
1065                         (PL080_BSIZE_64 << PL080_CONTROL_DB_SIZE_SHIFT),
1066         },
1067         {
1068                 .burstwords = 32,
1069                 .reg = (PL080_BSIZE_32 << PL080_CONTROL_SB_SIZE_SHIFT) |
1070                         (PL080_BSIZE_32 << PL080_CONTROL_DB_SIZE_SHIFT),
1071         },
1072         {
1073                 .burstwords = 16,
1074                 .reg = (PL080_BSIZE_16 << PL080_CONTROL_SB_SIZE_SHIFT) |
1075                         (PL080_BSIZE_16 << PL080_CONTROL_DB_SIZE_SHIFT),
1076         },
1077         {
1078                 .burstwords = 8,
1079                 .reg = (PL080_BSIZE_8 << PL080_CONTROL_SB_SIZE_SHIFT) |
1080                         (PL080_BSIZE_8 << PL080_CONTROL_DB_SIZE_SHIFT),
1081         },
1082         {
1083                 .burstwords = 4,
1084                 .reg = (PL080_BSIZE_4 << PL080_CONTROL_SB_SIZE_SHIFT) |
1085                         (PL080_BSIZE_4 << PL080_CONTROL_DB_SIZE_SHIFT),
1086         },
1087         {
1088                 .burstwords = 1,
1089                 .reg = (PL080_BSIZE_1 << PL080_CONTROL_SB_SIZE_SHIFT) |
1090                         (PL080_BSIZE_1 << PL080_CONTROL_DB_SIZE_SHIFT),
1091         },
1092 };
1093
1094 static void dma_set_runtime_config(struct dma_chan *chan,
1095                                struct dma_slave_config *config)
1096 {
1097         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1098         struct pl08x_driver_data *pl08x = plchan->host;
1099         struct pl08x_channel_data *cd = plchan->cd;
1100         enum dma_slave_buswidth addr_width;
1101         u32 maxburst;
1102         u32 cctl = 0;
1103         int i;
1104
1105         /* Transfer direction */
1106         plchan->runtime_direction = config->direction;
1107         if (config->direction == DMA_TO_DEVICE) {
1108                 plchan->runtime_addr = config->dst_addr;
1109                 addr_width = config->dst_addr_width;
1110                 maxburst = config->dst_maxburst;
1111         } else if (config->direction == DMA_FROM_DEVICE) {
1112                 plchan->runtime_addr = config->src_addr;
1113                 addr_width = config->src_addr_width;
1114                 maxburst = config->src_maxburst;
1115         } else {
1116                 dev_err(&pl08x->adev->dev,
1117                         "bad runtime_config: alien transfer direction\n");
1118                 return;
1119         }
1120
1121         switch (addr_width) {
1122         case DMA_SLAVE_BUSWIDTH_1_BYTE:
1123                 cctl |= (PL080_WIDTH_8BIT << PL080_CONTROL_SWIDTH_SHIFT) |
1124                         (PL080_WIDTH_8BIT << PL080_CONTROL_DWIDTH_SHIFT);
1125                 break;
1126         case DMA_SLAVE_BUSWIDTH_2_BYTES:
1127                 cctl |= (PL080_WIDTH_16BIT << PL080_CONTROL_SWIDTH_SHIFT) |
1128                         (PL080_WIDTH_16BIT << PL080_CONTROL_DWIDTH_SHIFT);
1129                 break;
1130         case DMA_SLAVE_BUSWIDTH_4_BYTES:
1131                 cctl |= (PL080_WIDTH_32BIT << PL080_CONTROL_SWIDTH_SHIFT) |
1132                         (PL080_WIDTH_32BIT << PL080_CONTROL_DWIDTH_SHIFT);
1133                 break;
1134         default:
1135                 dev_err(&pl08x->adev->dev,
1136                         "bad runtime_config: alien address width\n");
1137                 return;
1138         }
1139
1140         /*
1141          * Now decide on a maxburst:
1142          * If this channel will only request single transfers, set this
1143          * down to ONE element.  Also select one element if no maxburst
1144          * is specified.
1145          */
1146         if (plchan->cd->single || maxburst == 0) {
1147                 cctl |= (PL080_BSIZE_1 << PL080_CONTROL_SB_SIZE_SHIFT) |
1148                         (PL080_BSIZE_1 << PL080_CONTROL_DB_SIZE_SHIFT);
1149         } else {
1150                 for (i = 0; i < ARRAY_SIZE(burst_sizes); i++)
1151                         if (burst_sizes[i].burstwords <= maxburst)
1152                                 break;
1153                 cctl |= burst_sizes[i].reg;
1154         }
1155
1156         /* Modify the default channel data to fit PrimeCell request */
1157         cd->cctl = cctl;
1158
1159         dev_dbg(&pl08x->adev->dev,
1160                 "configured channel %s (%s) for %s, data width %d, "
1161                 "maxburst %d words, LE, CCTL=0x%08x\n",
1162                 dma_chan_name(chan), plchan->name,
1163                 (config->direction == DMA_FROM_DEVICE) ? "RX" : "TX",
1164                 addr_width,
1165                 maxburst,
1166                 cctl);
1167 }
1168
1169 /*
1170  * Slave transactions callback to the slave device to allow
1171  * synchronization of slave DMA signals with the DMAC enable
1172  */
1173 static void pl08x_issue_pending(struct dma_chan *chan)
1174 {
1175         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1176         unsigned long flags;
1177
1178         spin_lock_irqsave(&plchan->lock, flags);
1179         /* Something is already active, or we're waiting for a channel... */
1180         if (plchan->at || plchan->state == PL08X_CHAN_WAITING) {
1181                 spin_unlock_irqrestore(&plchan->lock, flags);
1182                 return;
1183         }
1184
1185         /* Take the first element in the queue and execute it */
1186         if (!list_empty(&plchan->pend_list)) {
1187                 struct pl08x_txd *next;
1188
1189                 next = list_first_entry(&plchan->pend_list,
1190                                         struct pl08x_txd,
1191                                         node);
1192                 list_del(&next->node);
1193                 plchan->state = PL08X_CHAN_RUNNING;
1194
1195                 pl08x_start_txd(plchan, next);
1196         }
1197
1198         spin_unlock_irqrestore(&plchan->lock, flags);
1199 }
1200
1201 static int pl08x_prep_channel_resources(struct pl08x_dma_chan *plchan,
1202                                         struct pl08x_txd *txd)
1203 {
1204         int num_llis;
1205         struct pl08x_driver_data *pl08x = plchan->host;
1206         int ret;
1207
1208         num_llis = pl08x_fill_llis_for_desc(pl08x, txd);
1209         if (!num_llis) {
1210                 kfree(txd);
1211                 return -EINVAL;
1212         }
1213
1214         spin_lock_irqsave(&plchan->lock, plchan->lockflags);
1215
1216         list_add_tail(&txd->node, &plchan->pend_list);
1217
1218         /*
1219          * See if we already have a physical channel allocated,
1220          * else this is the time to try to get one.
1221          */
1222         ret = prep_phy_channel(plchan, txd);
1223         if (ret) {
1224                 /*
1225                  * No physical channel available, we will
1226                  * stack up the memcpy channels until there is a channel
1227                  * available to handle it whereas slave transfers may
1228                  * have been denied due to platform channel muxing restrictions
1229                  * and since there is no guarantee that this will ever be
1230                  * resolved, and since the signal must be acquired AFTER
1231                  * acquiring the physical channel, we will let them be NACK:ed
1232                  * with -EBUSY here. The drivers can alway retry the prep()
1233                  * call if they are eager on doing this using DMA.
1234                  */
1235                 if (plchan->slave) {
1236                         pl08x_free_txd_list(pl08x, plchan);
1237                         spin_unlock_irqrestore(&plchan->lock, plchan->lockflags);
1238                         return -EBUSY;
1239                 }
1240                 /* Do this memcpy whenever there is a channel ready */
1241                 plchan->state = PL08X_CHAN_WAITING;
1242                 plchan->waiting = txd;
1243         } else
1244                 /*
1245                  * Else we're all set, paused and ready to roll,
1246                  * status will switch to PL08X_CHAN_RUNNING when
1247                  * we call issue_pending(). If there is something
1248                  * running on the channel already we don't change
1249                  * its state.
1250                  */
1251                 if (plchan->state == PL08X_CHAN_IDLE)
1252                         plchan->state = PL08X_CHAN_PAUSED;
1253
1254         /*
1255          * Notice that we leave plchan->lock locked on purpose:
1256          * it will be unlocked in the subsequent tx_submit()
1257          * call. This is a consequence of the current API.
1258          */
1259
1260         return 0;
1261 }
1262
1263 /*
1264  * Given the source and destination available bus masks, select which
1265  * will be routed to each port.  We try to have source and destination
1266  * on separate ports, but always respect the allowable settings.
1267  */
1268 static u32 pl08x_select_bus(struct pl08x_driver_data *pl08x, u8 src, u8 dst)
1269 {
1270         u32 cctl = 0;
1271
1272         if (!(dst & PL08X_AHB1) || ((dst & PL08X_AHB2) && (src & PL08X_AHB1)))
1273                 cctl |= PL080_CONTROL_DST_AHB2;
1274         if (!(src & PL08X_AHB1) || ((src & PL08X_AHB2) && !(dst & PL08X_AHB2)))
1275                 cctl |= PL080_CONTROL_SRC_AHB2;
1276
1277         return cctl;
1278 }
1279
1280 static struct pl08x_txd *pl08x_get_txd(struct pl08x_dma_chan *plchan,
1281         unsigned long flags)
1282 {
1283         struct pl08x_txd *txd = kzalloc(sizeof(struct pl08x_txd), GFP_NOWAIT);
1284
1285         if (txd) {
1286                 dma_async_tx_descriptor_init(&txd->tx, &plchan->chan);
1287                 txd->tx.flags = flags;
1288                 txd->tx.tx_submit = pl08x_tx_submit;
1289                 INIT_LIST_HEAD(&txd->node);
1290
1291                 /* Always enable error and terminal interrupts */
1292                 txd->ccfg = PL080_CONFIG_ERR_IRQ_MASK |
1293                             PL080_CONFIG_TC_IRQ_MASK;
1294         }
1295         return txd;
1296 }
1297
1298 /*
1299  * Initialize a descriptor to be used by memcpy submit
1300  */
1301 static struct dma_async_tx_descriptor *pl08x_prep_dma_memcpy(
1302                 struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
1303                 size_t len, unsigned long flags)
1304 {
1305         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1306         struct pl08x_driver_data *pl08x = plchan->host;
1307         struct pl08x_txd *txd;
1308         int ret;
1309
1310         txd = pl08x_get_txd(plchan, flags);
1311         if (!txd) {
1312                 dev_err(&pl08x->adev->dev,
1313                         "%s no memory for descriptor\n", __func__);
1314                 return NULL;
1315         }
1316
1317         txd->direction = DMA_NONE;
1318         txd->src_addr = src;
1319         txd->dst_addr = dest;
1320         txd->len = len;
1321
1322         /* Set platform data for m2m */
1323         txd->ccfg |= PL080_FLOW_MEM2MEM << PL080_CONFIG_FLOW_CONTROL_SHIFT;
1324         txd->cctl = pl08x->pd->memcpy_channel.cctl &
1325                         ~(PL080_CONTROL_DST_AHB2 | PL080_CONTROL_SRC_AHB2);
1326
1327         /* Both to be incremented or the code will break */
1328         txd->cctl |= PL080_CONTROL_SRC_INCR | PL080_CONTROL_DST_INCR;
1329
1330         if (pl08x->vd->dualmaster)
1331                 txd->cctl |= pl08x_select_bus(pl08x,
1332                                         pl08x->mem_buses, pl08x->mem_buses);
1333
1334         ret = pl08x_prep_channel_resources(plchan, txd);
1335         if (ret)
1336                 return NULL;
1337         /*
1338          * NB: the channel lock is held at this point so tx_submit()
1339          * must be called in direct succession.
1340          */
1341
1342         return &txd->tx;
1343 }
1344
1345 static struct dma_async_tx_descriptor *pl08x_prep_slave_sg(
1346                 struct dma_chan *chan, struct scatterlist *sgl,
1347                 unsigned int sg_len, enum dma_data_direction direction,
1348                 unsigned long flags)
1349 {
1350         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1351         struct pl08x_driver_data *pl08x = plchan->host;
1352         struct pl08x_txd *txd;
1353         u8 src_buses, dst_buses;
1354         int ret;
1355
1356         /*
1357          * Current implementation ASSUMES only one sg
1358          */
1359         if (sg_len != 1) {
1360                 dev_err(&pl08x->adev->dev, "%s prepared too long sglist\n",
1361                         __func__);
1362                 BUG();
1363         }
1364
1365         dev_dbg(&pl08x->adev->dev, "%s prepare transaction of %d bytes from %s\n",
1366                 __func__, sgl->length, plchan->name);
1367
1368         txd = pl08x_get_txd(plchan, flags);
1369         if (!txd) {
1370                 dev_err(&pl08x->adev->dev, "%s no txd\n", __func__);
1371                 return NULL;
1372         }
1373
1374         if (direction != plchan->runtime_direction)
1375                 dev_err(&pl08x->adev->dev, "%s DMA setup does not match "
1376                         "the direction configured for the PrimeCell\n",
1377                         __func__);
1378
1379         /*
1380          * Set up addresses, the PrimeCell configured address
1381          * will take precedence since this may configure the
1382          * channel target address dynamically at runtime.
1383          */
1384         txd->direction = direction;
1385         txd->len = sgl->length;
1386
1387         txd->cctl = plchan->cd->cctl &
1388                         ~(PL080_CONTROL_SRC_AHB2 | PL080_CONTROL_DST_AHB2 |
1389                           PL080_CONTROL_SRC_INCR | PL080_CONTROL_DST_INCR |
1390                           PL080_CONTROL_PROT_MASK);
1391
1392         /* Access the cell in privileged mode, non-bufferable, non-cacheable */
1393         txd->cctl |= PL080_CONTROL_PROT_SYS;
1394
1395         if (direction == DMA_TO_DEVICE) {
1396                 txd->ccfg |= PL080_FLOW_MEM2PER << PL080_CONFIG_FLOW_CONTROL_SHIFT;
1397                 txd->cctl |= PL080_CONTROL_SRC_INCR;
1398                 txd->src_addr = sgl->dma_address;
1399                 if (plchan->runtime_addr)
1400                         txd->dst_addr = plchan->runtime_addr;
1401                 else
1402                         txd->dst_addr = plchan->cd->addr;
1403                 src_buses = pl08x->mem_buses;
1404                 dst_buses = plchan->cd->periph_buses;
1405         } else if (direction == DMA_FROM_DEVICE) {
1406                 txd->ccfg |= PL080_FLOW_PER2MEM << PL080_CONFIG_FLOW_CONTROL_SHIFT;
1407                 txd->cctl |= PL080_CONTROL_DST_INCR;
1408                 if (plchan->runtime_addr)
1409                         txd->src_addr = plchan->runtime_addr;
1410                 else
1411                         txd->src_addr = plchan->cd->addr;
1412                 txd->dst_addr = sgl->dma_address;
1413                 src_buses = plchan->cd->periph_buses;
1414                 dst_buses = pl08x->mem_buses;
1415         } else {
1416                 dev_err(&pl08x->adev->dev,
1417                         "%s direction unsupported\n", __func__);
1418                 return NULL;
1419         }
1420
1421         txd->cctl |= pl08x_select_bus(pl08x, src_buses, dst_buses);
1422
1423         ret = pl08x_prep_channel_resources(plchan, txd);
1424         if (ret)
1425                 return NULL;
1426         /*
1427          * NB: the channel lock is held at this point so tx_submit()
1428          * must be called in direct succession.
1429          */
1430
1431         return &txd->tx;
1432 }
1433
1434 static int pl08x_control(struct dma_chan *chan, enum dma_ctrl_cmd cmd,
1435                          unsigned long arg)
1436 {
1437         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1438         struct pl08x_driver_data *pl08x = plchan->host;
1439         unsigned long flags;
1440         int ret = 0;
1441
1442         /* Controls applicable to inactive channels */
1443         if (cmd == DMA_SLAVE_CONFIG) {
1444                 dma_set_runtime_config(chan,
1445                                        (struct dma_slave_config *)
1446                                        arg);
1447                 return 0;
1448         }
1449
1450         /*
1451          * Anything succeeds on channels with no physical allocation and
1452          * no queued transfers.
1453          */
1454         spin_lock_irqsave(&plchan->lock, flags);
1455         if (!plchan->phychan && !plchan->at) {
1456                 spin_unlock_irqrestore(&plchan->lock, flags);
1457                 return 0;
1458         }
1459
1460         switch (cmd) {
1461         case DMA_TERMINATE_ALL:
1462                 plchan->state = PL08X_CHAN_IDLE;
1463
1464                 if (plchan->phychan) {
1465                         pl08x_stop_phy_chan(plchan->phychan);
1466
1467                         /*
1468                          * Mark physical channel as free and free any slave
1469                          * signal
1470                          */
1471                         release_phy_channel(plchan);
1472                 }
1473                 /* Dequeue jobs and free LLIs */
1474                 if (plchan->at) {
1475                         pl08x_free_txd(pl08x, plchan->at);
1476                         plchan->at = NULL;
1477                 }
1478                 /* Dequeue jobs not yet fired as well */
1479                 pl08x_free_txd_list(pl08x, plchan);
1480                 break;
1481         case DMA_PAUSE:
1482                 pl08x_pause_phy_chan(plchan->phychan);
1483                 plchan->state = PL08X_CHAN_PAUSED;
1484                 break;
1485         case DMA_RESUME:
1486                 pl08x_resume_phy_chan(plchan->phychan);
1487                 plchan->state = PL08X_CHAN_RUNNING;
1488                 break;
1489         default:
1490                 /* Unknown command */
1491                 ret = -ENXIO;
1492                 break;
1493         }
1494
1495         spin_unlock_irqrestore(&plchan->lock, flags);
1496
1497         return ret;
1498 }
1499
1500 bool pl08x_filter_id(struct dma_chan *chan, void *chan_id)
1501 {
1502         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1503         char *name = chan_id;
1504
1505         /* Check that the channel is not taken! */
1506         if (!strcmp(plchan->name, name))
1507                 return true;
1508
1509         return false;
1510 }
1511
1512 /*
1513  * Just check that the device is there and active
1514  * TODO: turn this bit on/off depending on the number of
1515  * physical channels actually used, if it is zero... well
1516  * shut it off. That will save some power. Cut the clock
1517  * at the same time.
1518  */
1519 static void pl08x_ensure_on(struct pl08x_driver_data *pl08x)
1520 {
1521         u32 val;
1522
1523         val = readl(pl08x->base + PL080_CONFIG);
1524         val &= ~(PL080_CONFIG_M2_BE | PL080_CONFIG_M1_BE | PL080_CONFIG_ENABLE);
1525         /* We implicitly clear bit 1 and that means little-endian mode */
1526         val |= PL080_CONFIG_ENABLE;
1527         writel(val, pl08x->base + PL080_CONFIG);
1528 }
1529
1530 static void pl08x_unmap_buffers(struct pl08x_txd *txd)
1531 {
1532         struct device *dev = txd->tx.chan->device->dev;
1533
1534         if (!(txd->tx.flags & DMA_COMPL_SKIP_SRC_UNMAP)) {
1535                 if (txd->tx.flags & DMA_COMPL_SRC_UNMAP_SINGLE)
1536                         dma_unmap_single(dev, txd->src_addr, txd->len,
1537                                 DMA_TO_DEVICE);
1538                 else
1539                         dma_unmap_page(dev, txd->src_addr, txd->len,
1540                                 DMA_TO_DEVICE);
1541         }
1542         if (!(txd->tx.flags & DMA_COMPL_SKIP_DEST_UNMAP)) {
1543                 if (txd->tx.flags & DMA_COMPL_DEST_UNMAP_SINGLE)
1544                         dma_unmap_single(dev, txd->dst_addr, txd->len,
1545                                 DMA_FROM_DEVICE);
1546                 else
1547                         dma_unmap_page(dev, txd->dst_addr, txd->len,
1548                                 DMA_FROM_DEVICE);
1549         }
1550 }
1551
1552 static void pl08x_tasklet(unsigned long data)
1553 {
1554         struct pl08x_dma_chan *plchan = (struct pl08x_dma_chan *) data;
1555         struct pl08x_driver_data *pl08x = plchan->host;
1556         struct pl08x_txd *txd;
1557         unsigned long flags;
1558
1559         spin_lock_irqsave(&plchan->lock, flags);
1560
1561         txd = plchan->at;
1562         plchan->at = NULL;
1563
1564         if (txd) {
1565                 /*
1566                  * Update last completed
1567                  */
1568                 plchan->lc = txd->tx.cookie;
1569         }
1570         /*
1571          * If a new descriptor is queued, set it up
1572          * plchan->at is NULL here
1573          */
1574         if (!list_empty(&plchan->pend_list)) {
1575                 struct pl08x_txd *next;
1576
1577                 next = list_first_entry(&plchan->pend_list,
1578                                         struct pl08x_txd,
1579                                         node);
1580                 list_del(&next->node);
1581
1582                 pl08x_start_txd(plchan, next);
1583         } else {
1584                 struct pl08x_dma_chan *waiting = NULL;
1585
1586                 /*
1587                  * No more jobs, so free up the physical channel
1588                  * Free any allocated signal on slave transfers too
1589                  */
1590                 release_phy_channel(plchan);
1591                 plchan->state = PL08X_CHAN_IDLE;
1592
1593                 /*
1594                  * And NOW before anyone else can grab that free:d
1595                  * up physical channel, see if there is some memcpy
1596                  * pending that seriously needs to start because of
1597                  * being stacked up while we were choking the
1598                  * physical channels with data.
1599                  */
1600                 list_for_each_entry(waiting, &pl08x->memcpy.channels,
1601                                     chan.device_node) {
1602                   if (waiting->state == PL08X_CHAN_WAITING &&
1603                             waiting->waiting != NULL) {
1604                                 int ret;
1605
1606                                 /* This should REALLY not fail now */
1607                                 ret = prep_phy_channel(waiting,
1608                                                        waiting->waiting);
1609                                 BUG_ON(ret);
1610                                 waiting->state = PL08X_CHAN_RUNNING;
1611                                 waiting->waiting = NULL;
1612                                 pl08x_issue_pending(&waiting->chan);
1613                                 break;
1614                         }
1615                 }
1616         }
1617
1618         spin_unlock_irqrestore(&plchan->lock, flags);
1619
1620         if (txd) {
1621                 dma_async_tx_callback callback = txd->tx.callback;
1622                 void *callback_param = txd->tx.callback_param;
1623
1624                 /* Don't try to unmap buffers on slave channels */
1625                 if (!plchan->slave)
1626                         pl08x_unmap_buffers(txd);
1627
1628                 /* Free the descriptor */
1629                 spin_lock_irqsave(&plchan->lock, flags);
1630                 pl08x_free_txd(pl08x, txd);
1631                 spin_unlock_irqrestore(&plchan->lock, flags);
1632
1633                 /* Callback to signal completion */
1634                 if (callback)
1635                         callback(callback_param);
1636         }
1637 }
1638
1639 static irqreturn_t pl08x_irq(int irq, void *dev)
1640 {
1641         struct pl08x_driver_data *pl08x = dev;
1642         u32 mask = 0;
1643         u32 val;
1644         int i;
1645
1646         val = readl(pl08x->base + PL080_ERR_STATUS);
1647         if (val) {
1648                 /*
1649                  * An error interrupt (on one or more channels)
1650                  */
1651                 dev_err(&pl08x->adev->dev,
1652                         "%s error interrupt, register value 0x%08x\n",
1653                                 __func__, val);
1654                 /*
1655                  * Simply clear ALL PL08X error interrupts,
1656                  * regardless of channel and cause
1657                  * FIXME: should be 0x00000003 on PL081 really.
1658                  */
1659                 writel(0x000000FF, pl08x->base + PL080_ERR_CLEAR);
1660         }
1661         val = readl(pl08x->base + PL080_INT_STATUS);
1662         for (i = 0; i < pl08x->vd->channels; i++) {
1663                 if ((1 << i) & val) {
1664                         /* Locate physical channel */
1665                         struct pl08x_phy_chan *phychan = &pl08x->phy_chans[i];
1666                         struct pl08x_dma_chan *plchan = phychan->serving;
1667
1668                         /* Schedule tasklet on this channel */
1669                         tasklet_schedule(&plchan->tasklet);
1670
1671                         mask |= (1 << i);
1672                 }
1673         }
1674         /*
1675          * Clear only the terminal interrupts on channels we processed
1676          */
1677         writel(mask, pl08x->base + PL080_TC_CLEAR);
1678
1679         return mask ? IRQ_HANDLED : IRQ_NONE;
1680 }
1681
1682 /*
1683  * Initialise the DMAC memcpy/slave channels.
1684  * Make a local wrapper to hold required data
1685  */
1686 static int pl08x_dma_init_virtual_channels(struct pl08x_driver_data *pl08x,
1687                                            struct dma_device *dmadev,
1688                                            unsigned int channels,
1689                                            bool slave)
1690 {
1691         struct pl08x_dma_chan *chan;
1692         int i;
1693
1694         INIT_LIST_HEAD(&dmadev->channels);
1695         /*
1696          * Register as many many memcpy as we have physical channels,
1697          * we won't always be able to use all but the code will have
1698          * to cope with that situation.
1699          */
1700         for (i = 0; i < channels; i++) {
1701                 chan = kzalloc(sizeof(struct pl08x_dma_chan), GFP_KERNEL);
1702                 if (!chan) {
1703                         dev_err(&pl08x->adev->dev,
1704                                 "%s no memory for channel\n", __func__);
1705                         return -ENOMEM;
1706                 }
1707
1708                 chan->host = pl08x;
1709                 chan->state = PL08X_CHAN_IDLE;
1710
1711                 if (slave) {
1712                         chan->slave = true;
1713                         chan->name = pl08x->pd->slave_channels[i].bus_id;
1714                         chan->cd = &pl08x->pd->slave_channels[i];
1715                 } else {
1716                         chan->cd = &pl08x->pd->memcpy_channel;
1717                         chan->name = kasprintf(GFP_KERNEL, "memcpy%d", i);
1718                         if (!chan->name) {
1719                                 kfree(chan);
1720                                 return -ENOMEM;
1721                         }
1722                 }
1723                 if (chan->cd->circular_buffer) {
1724                         dev_err(&pl08x->adev->dev,
1725                                 "channel %s: circular buffers not supported\n",
1726                                 chan->name);
1727                         kfree(chan);
1728                         continue;
1729                 }
1730                 dev_info(&pl08x->adev->dev,
1731                          "initialize virtual channel \"%s\"\n",
1732                          chan->name);
1733
1734                 chan->chan.device = dmadev;
1735                 chan->chan.cookie = 0;
1736                 chan->lc = 0;
1737
1738                 spin_lock_init(&chan->lock);
1739                 INIT_LIST_HEAD(&chan->pend_list);
1740                 tasklet_init(&chan->tasklet, pl08x_tasklet,
1741                              (unsigned long) chan);
1742
1743                 list_add_tail(&chan->chan.device_node, &dmadev->channels);
1744         }
1745         dev_info(&pl08x->adev->dev, "initialized %d virtual %s channels\n",
1746                  i, slave ? "slave" : "memcpy");
1747         return i;
1748 }
1749
1750 static void pl08x_free_virtual_channels(struct dma_device *dmadev)
1751 {
1752         struct pl08x_dma_chan *chan = NULL;
1753         struct pl08x_dma_chan *next;
1754
1755         list_for_each_entry_safe(chan,
1756                                  next, &dmadev->channels, chan.device_node) {
1757                 list_del(&chan->chan.device_node);
1758                 kfree(chan);
1759         }
1760 }
1761
1762 #ifdef CONFIG_DEBUG_FS
1763 static const char *pl08x_state_str(enum pl08x_dma_chan_state state)
1764 {
1765         switch (state) {
1766         case PL08X_CHAN_IDLE:
1767                 return "idle";
1768         case PL08X_CHAN_RUNNING:
1769                 return "running";
1770         case PL08X_CHAN_PAUSED:
1771                 return "paused";
1772         case PL08X_CHAN_WAITING:
1773                 return "waiting";
1774         default:
1775                 break;
1776         }
1777         return "UNKNOWN STATE";
1778 }
1779
1780 static int pl08x_debugfs_show(struct seq_file *s, void *data)
1781 {
1782         struct pl08x_driver_data *pl08x = s->private;
1783         struct pl08x_dma_chan *chan;
1784         struct pl08x_phy_chan *ch;
1785         unsigned long flags;
1786         int i;
1787
1788         seq_printf(s, "PL08x physical channels:\n");
1789         seq_printf(s, "CHANNEL:\tUSER:\n");
1790         seq_printf(s, "--------\t-----\n");
1791         for (i = 0; i < pl08x->vd->channels; i++) {
1792                 struct pl08x_dma_chan *virt_chan;
1793
1794                 ch = &pl08x->phy_chans[i];
1795
1796                 spin_lock_irqsave(&ch->lock, flags);
1797                 virt_chan = ch->serving;
1798
1799                 seq_printf(s, "%d\t\t%s\n",
1800                            ch->id, virt_chan ? virt_chan->name : "(none)");
1801
1802                 spin_unlock_irqrestore(&ch->lock, flags);
1803         }
1804
1805         seq_printf(s, "\nPL08x virtual memcpy channels:\n");
1806         seq_printf(s, "CHANNEL:\tSTATE:\n");
1807         seq_printf(s, "--------\t------\n");
1808         list_for_each_entry(chan, &pl08x->memcpy.channels, chan.device_node) {
1809                 seq_printf(s, "%s\t\t%s\n", chan->name,
1810                            pl08x_state_str(chan->state));
1811         }
1812
1813         seq_printf(s, "\nPL08x virtual slave channels:\n");
1814         seq_printf(s, "CHANNEL:\tSTATE:\n");
1815         seq_printf(s, "--------\t------\n");
1816         list_for_each_entry(chan, &pl08x->slave.channels, chan.device_node) {
1817                 seq_printf(s, "%s\t\t%s\n", chan->name,
1818                            pl08x_state_str(chan->state));
1819         }
1820
1821         return 0;
1822 }
1823
1824 static int pl08x_debugfs_open(struct inode *inode, struct file *file)
1825 {
1826         return single_open(file, pl08x_debugfs_show, inode->i_private);
1827 }
1828
1829 static const struct file_operations pl08x_debugfs_operations = {
1830         .open           = pl08x_debugfs_open,
1831         .read           = seq_read,
1832         .llseek         = seq_lseek,
1833         .release        = single_release,
1834 };
1835
1836 static void init_pl08x_debugfs(struct pl08x_driver_data *pl08x)
1837 {
1838         /* Expose a simple debugfs interface to view all clocks */
1839         (void) debugfs_create_file(dev_name(&pl08x->adev->dev), S_IFREG | S_IRUGO,
1840                                    NULL, pl08x,
1841                                    &pl08x_debugfs_operations);
1842 }
1843
1844 #else
1845 static inline void init_pl08x_debugfs(struct pl08x_driver_data *pl08x)
1846 {
1847 }
1848 #endif
1849
1850 static int pl08x_probe(struct amba_device *adev, struct amba_id *id)
1851 {
1852         struct pl08x_driver_data *pl08x;
1853         const struct vendor_data *vd = id->data;
1854         int ret = 0;
1855         int i;
1856
1857         ret = amba_request_regions(adev, NULL);
1858         if (ret)
1859                 return ret;
1860
1861         /* Create the driver state holder */
1862         pl08x = kzalloc(sizeof(struct pl08x_driver_data), GFP_KERNEL);
1863         if (!pl08x) {
1864                 ret = -ENOMEM;
1865                 goto out_no_pl08x;
1866         }
1867
1868         /* Initialize memcpy engine */
1869         dma_cap_set(DMA_MEMCPY, pl08x->memcpy.cap_mask);
1870         pl08x->memcpy.dev = &adev->dev;
1871         pl08x->memcpy.device_alloc_chan_resources = pl08x_alloc_chan_resources;
1872         pl08x->memcpy.device_free_chan_resources = pl08x_free_chan_resources;
1873         pl08x->memcpy.device_prep_dma_memcpy = pl08x_prep_dma_memcpy;
1874         pl08x->memcpy.device_prep_dma_interrupt = pl08x_prep_dma_interrupt;
1875         pl08x->memcpy.device_tx_status = pl08x_dma_tx_status;
1876         pl08x->memcpy.device_issue_pending = pl08x_issue_pending;
1877         pl08x->memcpy.device_control = pl08x_control;
1878
1879         /* Initialize slave engine */
1880         dma_cap_set(DMA_SLAVE, pl08x->slave.cap_mask);
1881         pl08x->slave.dev = &adev->dev;
1882         pl08x->slave.device_alloc_chan_resources = pl08x_alloc_chan_resources;
1883         pl08x->slave.device_free_chan_resources = pl08x_free_chan_resources;
1884         pl08x->slave.device_prep_dma_interrupt = pl08x_prep_dma_interrupt;
1885         pl08x->slave.device_tx_status = pl08x_dma_tx_status;
1886         pl08x->slave.device_issue_pending = pl08x_issue_pending;
1887         pl08x->slave.device_prep_slave_sg = pl08x_prep_slave_sg;
1888         pl08x->slave.device_control = pl08x_control;
1889
1890         /* Get the platform data */
1891         pl08x->pd = dev_get_platdata(&adev->dev);
1892         if (!pl08x->pd) {
1893                 dev_err(&adev->dev, "no platform data supplied\n");
1894                 goto out_no_platdata;
1895         }
1896
1897         /* Assign useful pointers to the driver state */
1898         pl08x->adev = adev;
1899         pl08x->vd = vd;
1900
1901         /* By default, AHB1 only.  If dualmaster, from platform */
1902         pl08x->lli_buses = PL08X_AHB1;
1903         pl08x->mem_buses = PL08X_AHB1;
1904         if (pl08x->vd->dualmaster) {
1905                 pl08x->lli_buses = pl08x->pd->lli_buses;
1906                 pl08x->mem_buses = pl08x->pd->mem_buses;
1907         }
1908
1909         /* A DMA memory pool for LLIs, align on 1-byte boundary */
1910         pl08x->pool = dma_pool_create(DRIVER_NAME, &pl08x->adev->dev,
1911                         PL08X_LLI_TSFR_SIZE, PL08X_ALIGN, 0);
1912         if (!pl08x->pool) {
1913                 ret = -ENOMEM;
1914                 goto out_no_lli_pool;
1915         }
1916
1917         spin_lock_init(&pl08x->lock);
1918
1919         pl08x->base = ioremap(adev->res.start, resource_size(&adev->res));
1920         if (!pl08x->base) {
1921                 ret = -ENOMEM;
1922                 goto out_no_ioremap;
1923         }
1924
1925         /* Turn on the PL08x */
1926         pl08x_ensure_on(pl08x);
1927
1928         /*
1929          * Attach the interrupt handler
1930          */
1931         writel(0x000000FF, pl08x->base + PL080_ERR_CLEAR);
1932         writel(0x000000FF, pl08x->base + PL080_TC_CLEAR);
1933
1934         ret = request_irq(adev->irq[0], pl08x_irq, IRQF_DISABLED,
1935                           DRIVER_NAME, pl08x);
1936         if (ret) {
1937                 dev_err(&adev->dev, "%s failed to request interrupt %d\n",
1938                         __func__, adev->irq[0]);
1939                 goto out_no_irq;
1940         }
1941
1942         /* Initialize physical channels */
1943         pl08x->phy_chans = kmalloc((vd->channels * sizeof(struct pl08x_phy_chan)),
1944                         GFP_KERNEL);
1945         if (!pl08x->phy_chans) {
1946                 dev_err(&adev->dev, "%s failed to allocate "
1947                         "physical channel holders\n",
1948                         __func__);
1949                 goto out_no_phychans;
1950         }
1951
1952         for (i = 0; i < vd->channels; i++) {
1953                 struct pl08x_phy_chan *ch = &pl08x->phy_chans[i];
1954
1955                 ch->id = i;
1956                 ch->base = pl08x->base + PL080_Cx_BASE(i);
1957                 spin_lock_init(&ch->lock);
1958                 ch->serving = NULL;
1959                 ch->signal = -1;
1960                 dev_info(&adev->dev,
1961                          "physical channel %d is %s\n", i,
1962                          pl08x_phy_channel_busy(ch) ? "BUSY" : "FREE");
1963         }
1964
1965         /* Register as many memcpy channels as there are physical channels */
1966         ret = pl08x_dma_init_virtual_channels(pl08x, &pl08x->memcpy,
1967                                               pl08x->vd->channels, false);
1968         if (ret <= 0) {
1969                 dev_warn(&pl08x->adev->dev,
1970                          "%s failed to enumerate memcpy channels - %d\n",
1971                          __func__, ret);
1972                 goto out_no_memcpy;
1973         }
1974         pl08x->memcpy.chancnt = ret;
1975
1976         /* Register slave channels */
1977         ret = pl08x_dma_init_virtual_channels(pl08x, &pl08x->slave,
1978                                               pl08x->pd->num_slave_channels,
1979                                               true);
1980         if (ret <= 0) {
1981                 dev_warn(&pl08x->adev->dev,
1982                         "%s failed to enumerate slave channels - %d\n",
1983                                 __func__, ret);
1984                 goto out_no_slave;
1985         }
1986         pl08x->slave.chancnt = ret;
1987
1988         ret = dma_async_device_register(&pl08x->memcpy);
1989         if (ret) {
1990                 dev_warn(&pl08x->adev->dev,
1991                         "%s failed to register memcpy as an async device - %d\n",
1992                         __func__, ret);
1993                 goto out_no_memcpy_reg;
1994         }
1995
1996         ret = dma_async_device_register(&pl08x->slave);
1997         if (ret) {
1998                 dev_warn(&pl08x->adev->dev,
1999                         "%s failed to register slave as an async device - %d\n",
2000                         __func__, ret);
2001                 goto out_no_slave_reg;
2002         }
2003
2004         amba_set_drvdata(adev, pl08x);
2005         init_pl08x_debugfs(pl08x);
2006         dev_info(&pl08x->adev->dev, "DMA: PL%03x rev%u at 0x%08llx irq %d\n",
2007                  amba_part(adev), amba_rev(adev),
2008                  (unsigned long long)adev->res.start, adev->irq[0]);
2009         return 0;
2010
2011 out_no_slave_reg:
2012         dma_async_device_unregister(&pl08x->memcpy);
2013 out_no_memcpy_reg:
2014         pl08x_free_virtual_channels(&pl08x->slave);
2015 out_no_slave:
2016         pl08x_free_virtual_channels(&pl08x->memcpy);
2017 out_no_memcpy:
2018         kfree(pl08x->phy_chans);
2019 out_no_phychans:
2020         free_irq(adev->irq[0], pl08x);
2021 out_no_irq:
2022         iounmap(pl08x->base);
2023 out_no_ioremap:
2024         dma_pool_destroy(pl08x->pool);
2025 out_no_lli_pool:
2026 out_no_platdata:
2027         kfree(pl08x);
2028 out_no_pl08x:
2029         amba_release_regions(adev);
2030         return ret;
2031 }
2032
2033 /* PL080 has 8 channels and the PL080 have just 2 */
2034 static struct vendor_data vendor_pl080 = {
2035         .channels = 8,
2036         .dualmaster = true,
2037 };
2038
2039 static struct vendor_data vendor_pl081 = {
2040         .channels = 2,
2041         .dualmaster = false,
2042 };
2043
2044 static struct amba_id pl08x_ids[] = {
2045         /* PL080 */
2046         {
2047                 .id     = 0x00041080,
2048                 .mask   = 0x000fffff,
2049                 .data   = &vendor_pl080,
2050         },
2051         /* PL081 */
2052         {
2053                 .id     = 0x00041081,
2054                 .mask   = 0x000fffff,
2055                 .data   = &vendor_pl081,
2056         },
2057         /* Nomadik 8815 PL080 variant */
2058         {
2059                 .id     = 0x00280880,
2060                 .mask   = 0x00ffffff,
2061                 .data   = &vendor_pl080,
2062         },
2063         { 0, 0 },
2064 };
2065
2066 static struct amba_driver pl08x_amba_driver = {
2067         .drv.name       = DRIVER_NAME,
2068         .id_table       = pl08x_ids,
2069         .probe          = pl08x_probe,
2070 };
2071
2072 static int __init pl08x_init(void)
2073 {
2074         int retval;
2075         retval = amba_driver_register(&pl08x_amba_driver);
2076         if (retval)
2077                 printk(KERN_WARNING DRIVER_NAME
2078                        "failed to register as an AMBA device (%d)\n",
2079                        retval);
2080         return retval;
2081 }
2082 subsys_initcall(pl08x_init);