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->desc_list)) {
 352                 struct pl08x_txd *txdi;
 353                 list_for_each_entry(txdi, &plchan->desc_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 /*
 485  * Autoselect a master bus to use for the transfer
 486  * this prefers the destination bus if both available
 487  * if fixed address on one bus the other will be chosen
 488  */
 489 static void pl08x_choose_master_bus(struct pl08x_bus_data *src_bus,
 490         struct pl08x_bus_data *dst_bus, struct pl08x_bus_data **mbus,
 491         struct pl08x_bus_data **sbus, u32 cctl)
 492 {
 493         if (!(cctl & PL080_CONTROL_DST_INCR)) {
 494                 *mbus = src_bus;
 495                 *sbus = dst_bus;
 496         } else if (!(cctl & PL080_CONTROL_SRC_INCR)) {
 497                 *mbus = dst_bus;
 498                 *sbus = src_bus;
 499         } else {
 500                 if (dst_bus->buswidth == 4) {
 501                         *mbus = dst_bus;
 502                         *sbus = src_bus;
 503                 } else if (src_bus->buswidth == 4) {
 504                         *mbus = src_bus;
 505                         *sbus = dst_bus;
 506                 } else if (dst_bus->buswidth == 2) {
 507                         *mbus = dst_bus;
 508                         *sbus = src_bus;
 509                 } else if (src_bus->buswidth == 2) {
 510                         *mbus = src_bus;
 511                         *sbus = dst_bus;
 512                 } else {
 513                         /* src_bus->buswidth == 1 */
 514                         *mbus = dst_bus;
 515                         *sbus = src_bus;
 516                 }
 517         }
 518 }
 519
 520 /*
 521  * Fills in one LLI for a certain transfer descriptor
 522  * and advance the counter
 523  */
 524 static void pl08x_fill_lli_for_desc(struct pl08x_driver_data *pl08x,
 525         struct pl08x_txd *txd, int num_llis, int len, u32 cctl, u32 *remainder)
 526 {
 527         struct pl08x_lli *llis_va = txd->llis_va;
 528         dma_addr_t llis_bus = txd->llis_bus;
 529
 530         BUG_ON(num_llis >= MAX_NUM_TSFR_LLIS);
 531
 532         llis_va[num_llis].cctl = cctl;
 533         llis_va[num_llis].src = txd->srcbus.addr;
 534         llis_va[num_llis].dst = txd->dstbus.addr;
 535         llis_va[num_llis].lli = llis_bus + (num_llis + 1) * sizeof(struct pl08x_lli);
 536         if (pl08x->lli_buses & PL08X_AHB2)
 537                 llis_va[num_llis].lli |= PL080_LLI_LM_AHB2;
 538
 539         if (cctl & PL080_CONTROL_SRC_INCR)
 540                 txd->srcbus.addr += len;
 541         if (cctl & PL080_CONTROL_DST_INCR)
 542                 txd->dstbus.addr += len;
 543
 544         BUG_ON(*remainder < len);
 545
 546         *remainder -= len;
 547 }
 548
 549 /*
 550  * Return number of bytes to fill to boundary, or len.
 551  * This calculation works for any value of addr.
 552  */
 553 static inline size_t pl08x_pre_boundary(u32 addr, size_t len)
 554 {
 555         size_t boundary_len = PL08X_BOUNDARY_SIZE -
 556                         (addr & (PL08X_BOUNDARY_SIZE - 1));
 557
 558         return min(boundary_len, len);
 559 }
 560
 561 /*
 562  * This fills in the table of LLIs for the transfer descriptor
 563  * Note that we assume we never have to change the burst sizes
 564  * Return 0 for error
 565  */
 566 static int pl08x_fill_llis_for_desc(struct pl08x_driver_data *pl08x,
 567                               struct pl08x_txd *txd)
 568 {
 569         struct pl08x_bus_data *mbus, *sbus;
 570         size_t remainder;
 571         int num_llis = 0;
 572         u32 cctl;
 573         size_t max_bytes_per_lli;
 574         size_t total_bytes = 0;
 575         struct pl08x_lli *llis_va;
 576
 577         txd->llis_va = dma_pool_alloc(pl08x->pool, GFP_NOWAIT,
 578                                       &txd->llis_bus);
 579         if (!txd->llis_va) {
 580                 dev_err(&pl08x->adev->dev, "%s no memory for llis\n", __func__);
 581                 return 0;
 582         }
 583
 584         pl08x->pool_ctr++;
 585
 586         /* Get the default CCTL */
 587         cctl = txd->cctl;
 588
 589         /* Find maximum width of the source bus */
 590         txd->srcbus.maxwidth =
 591                 pl08x_get_bytes_for_cctl((cctl & PL080_CONTROL_SWIDTH_MASK) >>
 592                                        PL080_CONTROL_SWIDTH_SHIFT);
 593
 594         /* Find maximum width of the destination bus */
 595         txd->dstbus.maxwidth =
 596                 pl08x_get_bytes_for_cctl((cctl & PL080_CONTROL_DWIDTH_MASK) >>
 597                                        PL080_CONTROL_DWIDTH_SHIFT);
 598
 599         /* Set up the bus widths to the maximum */
 600         txd->srcbus.buswidth = txd->srcbus.maxwidth;
 601         txd->dstbus.buswidth = txd->dstbus.maxwidth;
 602         dev_vdbg(&pl08x->adev->dev,
 603                  "%s source bus is %d bytes wide, dest bus is %d bytes wide\n",
 604                  __func__, txd->srcbus.buswidth, txd->dstbus.buswidth);
 605
 606
 607         /*
 608          * Bytes transferred == tsize * MIN(buswidths), not max(buswidths)
 609          */
 610         max_bytes_per_lli = min(txd->srcbus.buswidth, txd->dstbus.buswidth) *
 611                 PL080_CONTROL_TRANSFER_SIZE_MASK;
 612         dev_vdbg(&pl08x->adev->dev,
 613                  "%s max bytes per lli = %zu\n",
 614                  __func__, max_bytes_per_lli);
 615
 616         /* We need to count this down to zero */
 617         remainder = txd->len;
 618         dev_vdbg(&pl08x->adev->dev,
 619                  "%s remainder = %zu\n",
 620                  __func__, remainder);
 621
 622         /*
 623          * Choose bus to align to
 624          * - prefers destination bus if both available
 625          * - if fixed address on one bus chooses other
 626          * - modifies cctl to choose an appropriate master
 627          */
 628         pl08x_choose_master_bus(&txd->srcbus, &txd->dstbus,
 629                                 &mbus, &sbus, cctl);
 630
 631         if (txd->len < mbus->buswidth) {
 632                 /*
 633                  * Less than a bus width available
 634                  * - send as single bytes
 635                  */
 636                 while (remainder) {
 637                         dev_vdbg(&pl08x->adev->dev,
 638                                  "%s single byte LLIs for a transfer of "
 639                                  "less than a bus width (remain 0x%08x)\n",
 640                                  __func__, remainder);
 641                         cctl = pl08x_cctl_bits(cctl, 1, 1, 1);
 642                         pl08x_fill_lli_for_desc(pl08x, txd, num_llis++, 1,
 643                                         cctl, &remainder);
 644                         total_bytes++;
 645                 }
 646         } else {
 647                 /*
 648                  *  Make one byte LLIs until master bus is aligned
 649                  *  - slave will then be aligned also
 650                  */
 651                 while ((mbus->addr) % (mbus->buswidth)) {
 652                         dev_vdbg(&pl08x->adev->dev,
 653                                 "%s adjustment lli for less than bus width "
 654                                  "(remain 0x%08x)\n",
 655                                  __func__, remainder);
 656                         cctl = pl08x_cctl_bits(cctl, 1, 1, 1);
 657                         pl08x_fill_lli_for_desc(pl08x, txd, num_llis++, 1,
 658                                         cctl, &remainder);
 659                         total_bytes++;
 660                 }
 661
 662                 /*
 663                  *  Master now aligned
 664                  * - if slave is not then we must set its width down
 665                  */
 666                 if (sbus->addr % sbus->buswidth) {
 667                         dev_dbg(&pl08x->adev->dev,
 668                                 "%s set down bus width to one byte\n",
 669                                  __func__);
 670
 671                         sbus->buswidth = 1;
 672                 }
 673
 674                 /*
 675                  * Make largest possible LLIs until less than one bus
 676                  * width left
 677                  */
 678                 while (remainder > (mbus->buswidth - 1)) {
 679                         size_t lli_len, target_len, tsize, odd_bytes;
 680
 681                         /*
 682                          * If enough left try to send max possible,
 683                          * otherwise try to send the remainder
 684                          */
 685                         target_len = min(remainder, max_bytes_per_lli);
 686
 687                         /*
 688                          * Set bus lengths for incrementing buses to the
 689                          * number of bytes which fill to next memory boundary,
 690                          * limiting on the target length calculated above.
 691                          */
 692                         if (cctl & PL080_CONTROL_SRC_INCR)
 693                                 txd->srcbus.fill_bytes =
 694                                         pl08x_pre_boundary(txd->srcbus.addr,
 695                                                 target_len);
 696                         else
 697                                 txd->srcbus.fill_bytes = target_len;
 698
 699                         if (cctl & PL080_CONTROL_DST_INCR)
 700                                 txd->dstbus.fill_bytes =
 701                                         pl08x_pre_boundary(txd->dstbus.addr,
 702                                                 target_len);
 703                         else
 704                                 txd->dstbus.fill_bytes = target_len;
 705
 706                         /* Find the nearest */
 707                         lli_len = min(txd->srcbus.fill_bytes,
 708                                 txd->dstbus.fill_bytes);
 709
 710                         BUG_ON(lli_len > remainder);
 711
 712                         if (lli_len <= 0) {
 713                                 dev_err(&pl08x->adev->dev,
 714                                         "%s lli_len is %zu, <= 0\n",
 715                                                 __func__, lli_len);
 716                                 return 0;
 717                         }
 718
 719                         if (lli_len == target_len) {
 720                                 /*
 721                                  * Can send what we wanted
 722                                  */
 723                                 /*
 724                                  *  Maintain alignment
 725                                  */
 726                                 lli_len = (lli_len/mbus->buswidth) *
 727                                                         mbus->buswidth;
 728                                 odd_bytes = 0;
 729                         } else {
 730                                 /*
 731                                  * So now we know how many bytes to transfer
 732                                  * to get to the nearest boundary
 733                                  * The next LLI will past the boundary
 734                                  * - however we may be working to a boundary
 735                                  *   on the slave bus
 736                                  *   We need to ensure the master stays aligned
 737                                  */
 738                                 odd_bytes = lli_len % mbus->buswidth;
 739                                 /*
 740                                  * - and that we are working in multiples
 741                                  *   of the bus widths
 742                                  */
 743                                 lli_len -= odd_bytes;
 744
 745                         }
 746
 747                         if (lli_len) {
 748                                 /*
 749                                  * Check against minimum bus alignment:
 750                                  * Calculate actual transfer size in relation
 751                                  * to bus width an get a maximum remainder of
 752                                  * the smallest bus width - 1
 753                                  */
 754                                 /* FIXME: use round_down()? */
 755                                 tsize = lli_len / min(mbus->buswidth,
 756                                                       sbus->buswidth);
 757                                 lli_len = tsize * min(mbus->buswidth,
 758                                                       sbus->buswidth);
 759
 760                                 if (target_len != lli_len) {
 761                                         dev_vdbg(&pl08x->adev->dev,
 762                                         "%s can't send what we want. Desired 0x%08zx, lli of 0x%08zx bytes in txd of 0x%08zx\n",
 763                                         __func__, target_len, lli_len, txd->len);
 764                                 }
 765
 766                                 cctl = pl08x_cctl_bits(cctl,
 767                                                        txd->srcbus.buswidth,
 768                                                        txd->dstbus.buswidth,
 769                                                        tsize);
 770
 771                                 dev_vdbg(&pl08x->adev->dev,
 772                                         "%s fill lli with single lli chunk of size 0x%08zx (remainder 0x%08zx)\n",
 773                                         __func__, lli_len, remainder);
 774                                 pl08x_fill_lli_for_desc(pl08x, txd, num_llis++,
 775                                                 lli_len, cctl, &remainder);
 776                                 total_bytes += lli_len;
 777                         }
 778
 779
 780                         if (odd_bytes) {
 781                                 /*
 782                                  * Creep past the boundary,
 783                                  * maintaining master alignment
 784                                  */
 785                                 int j;
 786                                 for (j = 0; (j < mbus->buswidth)
 787                                                 && (remainder); j++) {
 788                                         cctl = pl08x_cctl_bits(cctl, 1, 1, 1);
 789                                         dev_vdbg(&pl08x->adev->dev,
 790                                                 "%s align with boundary, single byte (remain 0x%08zx)\n",
 791                                                 __func__, remainder);
 792                                         pl08x_fill_lli_for_desc(pl08x, txd,
 793                                                         num_llis++, 1, cctl,
 794                                                         &remainder);
 795                                         total_bytes++;
 796                                 }
 797                         }
 798                 }
 799
 800                 /*
 801                  * Send any odd bytes
 802                  */
 803                 while (remainder) {
 804                         cctl = pl08x_cctl_bits(cctl, 1, 1, 1);
 805                         dev_vdbg(&pl08x->adev->dev,
 806                                 "%s align with boundary, single odd byte (remain %zu)\n",
 807                                 __func__, remainder);
 808                         pl08x_fill_lli_for_desc(pl08x, txd, num_llis++, 1,
 809                                         cctl, &remainder);
 810                         total_bytes++;
 811                 }
 812         }
 813         if (total_bytes != txd->len) {
 814                 dev_err(&pl08x->adev->dev,
 815                         "%s size of encoded lli:s don't match total txd, transferred 0x%08zx from size 0x%08zx\n",
 816                         __func__, total_bytes, txd->len);
 817                 return 0;
 818         }
 819
 820         if (num_llis >= MAX_NUM_TSFR_LLIS) {
 821                 dev_err(&pl08x->adev->dev,
 822                         "%s need to increase MAX_NUM_TSFR_LLIS from 0x%08x\n",
 823                         __func__, (u32) MAX_NUM_TSFR_LLIS);
 824                 return 0;
 825         }
 826
 827         llis_va = txd->llis_va;
 828         /*
 829          * The final LLI terminates the LLI.
 830          */
 831         llis_va[num_llis - 1].lli = 0;
 832         /*
 833          * The final LLI element shall also fire an interrupt
 834          */
 835         llis_va[num_llis - 1].cctl |= PL080_CONTROL_TC_IRQ_EN;
 836
 837 #ifdef VERBOSE_DEBUG
 838         {
 839                 int i;
 840
 841                 for (i = 0; i < num_llis; i++) {
 842                         dev_vdbg(&pl08x->adev->dev,
 843                                  "lli %d @%p: csrc=0x%08x, cdst=0x%08x, cctl=0x%08x, clli=0x%08x\n",
 844                                  i,
 845                                  &llis_va[i],
 846                                  llis_va[i].src,
 847                                  llis_va[i].dst,
 848                                  llis_va[i].cctl,
 849                                  llis_va[i].lli
 850                                 );
 851                 }
 852         }
 853 #endif
 854
 855         return num_llis;
 856 }
 857
 858 /* You should call this with the struct pl08x lock held */
 859 static void pl08x_free_txd(struct pl08x_driver_data *pl08x,
 860                            struct pl08x_txd *txd)
 861 {
 862         /* Free the LLI */
 863         dma_pool_free(pl08x->pool, txd->llis_va, txd->llis_bus);
 864
 865         pl08x->pool_ctr--;
 866
 867         kfree(txd);
 868 }
 869
 870 static void pl08x_free_txd_list(struct pl08x_driver_data *pl08x,
 871                                 struct pl08x_dma_chan *plchan)
 872 {
 873         struct pl08x_txd *txdi = NULL;
 874         struct pl08x_txd *next;
 875
 876         if (!list_empty(&plchan->desc_list)) {
 877                 list_for_each_entry_safe(txdi,
 878                                          next, &plchan->desc_list, node) {
 879                         list_del(&txdi->node);
 880                         pl08x_free_txd(pl08x, txdi);
 881                 }
 882
 883         }
 884 }
 885
 886 /*
 887  * The DMA ENGINE API
 888  */
 889 static int pl08x_alloc_chan_resources(struct dma_chan *chan)
 890 {
 891         return 0;
 892 }
 893
 894 static void pl08x_free_chan_resources(struct dma_chan *chan)
 895 {
 896 }
 897
 898 /*
 899  * This should be called with the channel plchan->lock held
 900  */
 901 static int prep_phy_channel(struct pl08x_dma_chan *plchan,
 902                             struct pl08x_txd *txd)
 903 {
 904         struct pl08x_driver_data *pl08x = plchan->host;
 905         struct pl08x_phy_chan *ch;
 906         int ret;
 907
 908         /* Check if we already have a channel */
 909         if (plchan->phychan)
 910                 return 0;
 911
 912         ch = pl08x_get_phy_channel(pl08x, plchan);
 913         if (!ch) {
 914                 /* No physical channel available, cope with it */
 915                 dev_dbg(&pl08x->adev->dev, "no physical channel available for xfer on %s\n", plchan->name);
 916                 return -EBUSY;
 917         }
 918
 919         /*
 920          * OK we have a physical channel: for memcpy() this is all we
 921          * need, but for slaves the physical signals may be muxed!
 922          * Can the platform allow us to use this channel?
 923          */
 924         if (plchan->slave &&
 925             ch->signal < 0 &&
 926             pl08x->pd->get_signal) {
 927                 ret = pl08x->pd->get_signal(plchan);
 928                 if (ret < 0) {
 929                         dev_dbg(&pl08x->adev->dev,
 930                                 "unable to use physical channel %d for transfer on %s due to platform restrictions\n",
 931                                 ch->id, plchan->name);
 932                         /* Release physical channel & return */
 933                         pl08x_put_phy_channel(pl08x, ch);
 934                         return -EBUSY;
 935                 }
 936                 ch->signal = ret;
 937
 938                 /* Assign the flow control signal to this channel */
 939                 if (txd->direction == DMA_TO_DEVICE)
 940                         txd->ccfg |= ch->signal << PL080_CONFIG_DST_SEL_SHIFT;
 941                 else if (txd->direction == DMA_FROM_DEVICE)
 942                         txd->ccfg |= ch->signal << PL080_CONFIG_SRC_SEL_SHIFT;
 943         }
 944
 945         dev_dbg(&pl08x->adev->dev, "allocated physical channel %d and signal %d for xfer on %s\n",
 946                  ch->id,
 947                  ch->signal,
 948                  plchan->name);
 949
 950         plchan->phychan = ch;
 951
 952         return 0;
 953 }
 954
 955 static void release_phy_channel(struct pl08x_dma_chan *plchan)
 956 {
 957         struct pl08x_driver_data *pl08x = plchan->host;
 958
 959         if ((plchan->phychan->signal >= 0) && pl08x->pd->put_signal) {
 960                 pl08x->pd->put_signal(plchan);
 961                 plchan->phychan->signal = -1;
 962         }
 963         pl08x_put_phy_channel(pl08x, plchan->phychan);
 964         plchan->phychan = NULL;
 965 }
 966
 967 static dma_cookie_t pl08x_tx_submit(struct dma_async_tx_descriptor *tx)
 968 {
 969         struct pl08x_dma_chan *plchan = to_pl08x_chan(tx->chan);
 970
 971         plchan->chan.cookie += 1;
 972         if (plchan->chan.cookie < 0)
 973                 plchan->chan.cookie = 1;
 974         tx->cookie = plchan->chan.cookie;
 975         /* This unlock follows the lock in the prep() function */
 976         spin_unlock_irqrestore(&plchan->lock, plchan->lockflags);
 977
 978         return tx->cookie;
 979 }
 980
 981 static struct dma_async_tx_descriptor *pl08x_prep_dma_interrupt(
 982                 struct dma_chan *chan, unsigned long flags)
 983 {
 984         struct dma_async_tx_descriptor *retval = NULL;
 985
 986         return retval;
 987 }
 988
 989 /*
 990  * Code accessing dma_async_is_complete() in a tight loop
 991  * may give problems - could schedule where indicated.
 992  * If slaves are relying on interrupts to signal completion this
 993  * function must not be called with interrupts disabled
 994  */
 995 static enum dma_status
 996 pl08x_dma_tx_status(struct dma_chan *chan,
 997                     dma_cookie_t cookie,
 998                     struct dma_tx_state *txstate)
 999 {
1000         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1001         dma_cookie_t last_used;
1002         dma_cookie_t last_complete;
1003         enum dma_status ret;
1004         u32 bytesleft = 0;
1005
1006         last_used = plchan->chan.cookie;
1007         last_complete = plchan->lc;
1008
1009         ret = dma_async_is_complete(cookie, last_complete, last_used);
1010         if (ret == DMA_SUCCESS) {
1011                 dma_set_tx_state(txstate, last_complete, last_used, 0);
1012                 return ret;
1013         }
1014
1015         /*
1016          * schedule(); could be inserted here
1017          */
1018
1019         /*
1020          * This cookie not complete yet
1021          */
1022         last_used = plchan->chan.cookie;
1023         last_complete = plchan->lc;
1024
1025         /* Get number of bytes left in the active transactions and queue */
1026         bytesleft = pl08x_getbytes_chan(plchan);
1027
1028         dma_set_tx_state(txstate, last_complete, last_used,
1029                          bytesleft);
1030
1031         if (plchan->state == PL08X_CHAN_PAUSED)
1032                 return DMA_PAUSED;
1033
1034         /* Whether waiting or running, we're in progress */
1035         return DMA_IN_PROGRESS;
1036 }
1037
1038 /* PrimeCell DMA extension */
1039 struct burst_table {
1040         int burstwords;
1041         u32 reg;
1042 };
1043
1044 static const struct burst_table burst_sizes[] = {
1045         {
1046                 .burstwords = 256,
1047                 .reg = (PL080_BSIZE_256 << PL080_CONTROL_SB_SIZE_SHIFT) |
1048                         (PL080_BSIZE_256 << PL080_CONTROL_DB_SIZE_SHIFT),
1049         },
1050         {
1051                 .burstwords = 128,
1052                 .reg = (PL080_BSIZE_128 << PL080_CONTROL_SB_SIZE_SHIFT) |
1053                         (PL080_BSIZE_128 << PL080_CONTROL_DB_SIZE_SHIFT),
1054         },
1055         {
1056                 .burstwords = 64,
1057                 .reg = (PL080_BSIZE_64 << PL080_CONTROL_SB_SIZE_SHIFT) |
1058                         (PL080_BSIZE_64 << PL080_CONTROL_DB_SIZE_SHIFT),
1059         },
1060         {
1061                 .burstwords = 32,
1062                 .reg = (PL080_BSIZE_32 << PL080_CONTROL_SB_SIZE_SHIFT) |
1063                         (PL080_BSIZE_32 << PL080_CONTROL_DB_SIZE_SHIFT),
1064         },
1065         {
1066                 .burstwords = 16,
1067                 .reg = (PL080_BSIZE_16 << PL080_CONTROL_SB_SIZE_SHIFT) |
1068                         (PL080_BSIZE_16 << PL080_CONTROL_DB_SIZE_SHIFT),
1069         },
1070         {
1071                 .burstwords = 8,
1072                 .reg = (PL080_BSIZE_8 << PL080_CONTROL_SB_SIZE_SHIFT) |
1073                         (PL080_BSIZE_8 << PL080_CONTROL_DB_SIZE_SHIFT),
1074         },
1075         {
1076                 .burstwords = 4,
1077                 .reg = (PL080_BSIZE_4 << PL080_CONTROL_SB_SIZE_SHIFT) |
1078                         (PL080_BSIZE_4 << PL080_CONTROL_DB_SIZE_SHIFT),
1079         },
1080         {
1081                 .burstwords = 1,
1082                 .reg = (PL080_BSIZE_1 << PL080_CONTROL_SB_SIZE_SHIFT) |
1083                         (PL080_BSIZE_1 << PL080_CONTROL_DB_SIZE_SHIFT),
1084         },
1085 };
1086
1087 static void dma_set_runtime_config(struct dma_chan *chan,
1088                                struct dma_slave_config *config)
1089 {
1090         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1091         struct pl08x_driver_data *pl08x = plchan->host;
1092         struct pl08x_channel_data *cd = plchan->cd;
1093         enum dma_slave_buswidth addr_width;
1094         u32 maxburst;
1095         u32 cctl = 0;
1096         int i;
1097
1098         /* Transfer direction */
1099         plchan->runtime_direction = config->direction;
1100         if (config->direction == DMA_TO_DEVICE) {
1101                 plchan->runtime_addr = config->dst_addr;
1102                 addr_width = config->dst_addr_width;
1103                 maxburst = config->dst_maxburst;
1104         } else if (config->direction == DMA_FROM_DEVICE) {
1105                 plchan->runtime_addr = config->src_addr;
1106                 addr_width = config->src_addr_width;
1107                 maxburst = config->src_maxburst;
1108         } else {
1109                 dev_err(&pl08x->adev->dev,
1110                         "bad runtime_config: alien transfer direction\n");
1111                 return;
1112         }
1113
1114         switch (addr_width) {
1115         case DMA_SLAVE_BUSWIDTH_1_BYTE:
1116                 cctl |= (PL080_WIDTH_8BIT << PL080_CONTROL_SWIDTH_SHIFT) |
1117                         (PL080_WIDTH_8BIT << PL080_CONTROL_DWIDTH_SHIFT);
1118                 break;
1119         case DMA_SLAVE_BUSWIDTH_2_BYTES:
1120                 cctl |= (PL080_WIDTH_16BIT << PL080_CONTROL_SWIDTH_SHIFT) |
1121                         (PL080_WIDTH_16BIT << PL080_CONTROL_DWIDTH_SHIFT);
1122                 break;
1123         case DMA_SLAVE_BUSWIDTH_4_BYTES:
1124                 cctl |= (PL080_WIDTH_32BIT << PL080_CONTROL_SWIDTH_SHIFT) |
1125                         (PL080_WIDTH_32BIT << PL080_CONTROL_DWIDTH_SHIFT);
1126                 break;
1127         default:
1128                 dev_err(&pl08x->adev->dev,
1129                         "bad runtime_config: alien address width\n");
1130                 return;
1131         }
1132
1133         /*
1134          * Now decide on a maxburst:
1135          * If this channel will only request single transfers, set this
1136          * down to ONE element.  Also select one element if no maxburst
1137          * is specified.
1138          */
1139         if (plchan->cd->single || maxburst == 0) {
1140                 cctl |= (PL080_BSIZE_1 << PL080_CONTROL_SB_SIZE_SHIFT) |
1141                         (PL080_BSIZE_1 << PL080_CONTROL_DB_SIZE_SHIFT);
1142         } else {
1143                 for (i = 0; i < ARRAY_SIZE(burst_sizes); i++)
1144                         if (burst_sizes[i].burstwords <= maxburst)
1145                                 break;
1146                 cctl |= burst_sizes[i].reg;
1147         }
1148
1149         /* Modify the default channel data to fit PrimeCell request */
1150         cd->cctl = cctl;
1151
1152         dev_dbg(&pl08x->adev->dev,
1153                 "configured channel %s (%s) for %s, data width %d, "
1154                 "maxburst %d words, LE, CCTL=0x%08x\n",
1155                 dma_chan_name(chan), plchan->name,
1156                 (config->direction == DMA_FROM_DEVICE) ? "RX" : "TX",
1157                 addr_width,
1158                 maxburst,
1159                 cctl);
1160 }
1161
1162 /*
1163  * Slave transactions callback to the slave device to allow
1164  * synchronization of slave DMA signals with the DMAC enable
1165  */
1166 static void pl08x_issue_pending(struct dma_chan *chan)
1167 {
1168         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1169         unsigned long flags;
1170
1171         spin_lock_irqsave(&plchan->lock, flags);
1172         /* Something is already active, or we're waiting for a channel... */
1173         if (plchan->at || plchan->state == PL08X_CHAN_WAITING) {
1174                 spin_unlock_irqrestore(&plchan->lock, flags);
1175                 return;
1176         }
1177
1178         /* Take the first element in the queue and execute it */
1179         if (!list_empty(&plchan->desc_list)) {
1180                 struct pl08x_txd *next;
1181
1182                 next = list_first_entry(&plchan->desc_list,
1183                                         struct pl08x_txd,
1184                                         node);
1185                 list_del(&next->node);
1186                 plchan->state = PL08X_CHAN_RUNNING;
1187
1188                 pl08x_start_txd(plchan, next);
1189         }
1190
1191         spin_unlock_irqrestore(&plchan->lock, flags);
1192 }
1193
1194 static int pl08x_prep_channel_resources(struct pl08x_dma_chan *plchan,
1195                                         struct pl08x_txd *txd)
1196 {
1197         int num_llis;
1198         struct pl08x_driver_data *pl08x = plchan->host;
1199         int ret;
1200
1201         num_llis = pl08x_fill_llis_for_desc(pl08x, txd);
1202         if (!num_llis) {
1203                 kfree(txd);
1204                 return -EINVAL;
1205         }
1206
1207         spin_lock_irqsave(&plchan->lock, plchan->lockflags);
1208
1209         list_add_tail(&txd->node, &plchan->desc_list);
1210
1211         /*
1212          * See if we already have a physical channel allocated,
1213          * else this is the time to try to get one.
1214          */
1215         ret = prep_phy_channel(plchan, txd);
1216         if (ret) {
1217                 /*
1218                  * No physical channel available, we will
1219                  * stack up the memcpy channels until there is a channel
1220                  * available to handle it whereas slave transfers may
1221                  * have been denied due to platform channel muxing restrictions
1222                  * and since there is no guarantee that this will ever be
1223                  * resolved, and since the signal must be acquired AFTER
1224                  * acquiring the physical channel, we will let them be NACK:ed
1225                  * with -EBUSY here. The drivers can alway retry the prep()
1226                  * call if they are eager on doing this using DMA.
1227                  */
1228                 if (plchan->slave) {
1229                         pl08x_free_txd_list(pl08x, plchan);
1230                         spin_unlock_irqrestore(&plchan->lock, plchan->lockflags);
1231                         return -EBUSY;
1232                 }
1233                 /* Do this memcpy whenever there is a channel ready */
1234                 plchan->state = PL08X_CHAN_WAITING;
1235                 plchan->waiting = txd;
1236         } else
1237                 /*
1238                  * Else we're all set, paused and ready to roll,
1239                  * status will switch to PL08X_CHAN_RUNNING when
1240                  * we call issue_pending(). If there is something
1241                  * running on the channel already we don't change
1242                  * its state.
1243                  */
1244                 if (plchan->state == PL08X_CHAN_IDLE)
1245                         plchan->state = PL08X_CHAN_PAUSED;
1246
1247         /*
1248          * Notice that we leave plchan->lock locked on purpose:
1249          * it will be unlocked in the subsequent tx_submit()
1250          * call. This is a consequence of the current API.
1251          */
1252
1253         return 0;
1254 }
1255
1256 /*
1257  * Given the source and destination available bus masks, select which
1258  * will be routed to each port.  We try to have source and destination
1259  * on separate ports, but always respect the allowable settings.
1260  */
1261 static u32 pl08x_select_bus(struct pl08x_driver_data *pl08x, u8 src, u8 dst)
1262 {
1263         u32 cctl = 0;
1264
1265         if (!(dst & PL08X_AHB1) || ((dst & PL08X_AHB2) && (src & PL08X_AHB1)))
1266                 cctl |= PL080_CONTROL_DST_AHB2;
1267         if (!(src & PL08X_AHB1) || ((src & PL08X_AHB2) && !(dst & PL08X_AHB2)))
1268                 cctl |= PL080_CONTROL_SRC_AHB2;
1269
1270         return cctl;
1271 }
1272
1273 static struct pl08x_txd *pl08x_get_txd(struct pl08x_dma_chan *plchan)
1274 {
1275         struct pl08x_txd *txd = kzalloc(sizeof(struct pl08x_txd), GFP_NOWAIT);
1276
1277         if (txd) {
1278                 dma_async_tx_descriptor_init(&txd->tx, &plchan->chan);
1279                 txd->tx.tx_submit = pl08x_tx_submit;
1280                 INIT_LIST_HEAD(&txd->node);
1281
1282                 /* Always enable error and terminal interrupts */
1283                 txd->ccfg = PL080_CONFIG_ERR_IRQ_MASK |
1284                             PL080_CONFIG_TC_IRQ_MASK;
1285         }
1286         return txd;
1287 }
1288
1289 /*
1290  * Initialize a descriptor to be used by memcpy submit
1291  */
1292 static struct dma_async_tx_descriptor *pl08x_prep_dma_memcpy(
1293                 struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
1294                 size_t len, unsigned long flags)
1295 {
1296         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1297         struct pl08x_driver_data *pl08x = plchan->host;
1298         struct pl08x_txd *txd;
1299         int ret;
1300
1301         txd = pl08x_get_txd(plchan);
1302         if (!txd) {
1303                 dev_err(&pl08x->adev->dev,
1304                         "%s no memory for descriptor\n", __func__);
1305                 return NULL;
1306         }
1307
1308         txd->direction = DMA_NONE;
1309         txd->srcbus.addr = src;
1310         txd->dstbus.addr = dest;
1311         txd->len = len;
1312
1313         /* Set platform data for m2m */
1314         txd->ccfg |= PL080_FLOW_MEM2MEM << PL080_CONFIG_FLOW_CONTROL_SHIFT;
1315         txd->cctl = pl08x->pd->memcpy_channel.cctl &
1316                         ~(PL080_CONTROL_DST_AHB2 | PL080_CONTROL_SRC_AHB2);
1317
1318         /* Both to be incremented or the code will break */
1319         txd->cctl |= PL080_CONTROL_SRC_INCR | PL080_CONTROL_DST_INCR;
1320
1321         if (pl08x->vd->dualmaster)
1322                 txd->cctl |= pl08x_select_bus(pl08x,
1323                                         pl08x->mem_buses, pl08x->mem_buses);
1324
1325         ret = pl08x_prep_channel_resources(plchan, txd);
1326         if (ret)
1327                 return NULL;
1328         /*
1329          * NB: the channel lock is held at this point so tx_submit()
1330          * must be called in direct succession.
1331          */
1332
1333         return &txd->tx;
1334 }
1335
1336 static struct dma_async_tx_descriptor *pl08x_prep_slave_sg(
1337                 struct dma_chan *chan, struct scatterlist *sgl,
1338                 unsigned int sg_len, enum dma_data_direction direction,
1339                 unsigned long flags)
1340 {
1341         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1342         struct pl08x_driver_data *pl08x = plchan->host;
1343         struct pl08x_txd *txd;
1344         u8 src_buses, dst_buses;
1345         int ret;
1346
1347         /*
1348          * Current implementation ASSUMES only one sg
1349          */
1350         if (sg_len != 1) {
1351                 dev_err(&pl08x->adev->dev, "%s prepared too long sglist\n",
1352                         __func__);
1353                 BUG();
1354         }
1355
1356         dev_dbg(&pl08x->adev->dev, "%s prepare transaction of %d bytes from %s\n",
1357                 __func__, sgl->length, plchan->name);
1358
1359         txd = pl08x_get_txd(plchan);
1360         if (!txd) {
1361                 dev_err(&pl08x->adev->dev, "%s no txd\n", __func__);
1362                 return NULL;
1363         }
1364
1365         if (direction != plchan->runtime_direction)
1366                 dev_err(&pl08x->adev->dev, "%s DMA setup does not match "
1367                         "the direction configured for the PrimeCell\n",
1368                         __func__);
1369
1370         /*
1371          * Set up addresses, the PrimeCell configured address
1372          * will take precedence since this may configure the
1373          * channel target address dynamically at runtime.
1374          */
1375         txd->direction = direction;
1376         txd->len = sgl->length;
1377
1378         txd->cctl = plchan->cd->cctl &
1379                         ~(PL080_CONTROL_SRC_AHB2 | PL080_CONTROL_DST_AHB2 |
1380                           PL080_CONTROL_SRC_INCR | PL080_CONTROL_DST_INCR |
1381                           PL080_CONTROL_PROT_MASK);
1382
1383         /* Access the cell in privileged mode, non-bufferable, non-cacheable */
1384         txd->cctl |= PL080_CONTROL_PROT_SYS;
1385
1386         if (direction == DMA_TO_DEVICE) {
1387                 txd->ccfg |= PL080_FLOW_MEM2PER << PL080_CONFIG_FLOW_CONTROL_SHIFT;
1388                 txd->cctl |= PL080_CONTROL_SRC_INCR;
1389                 txd->srcbus.addr = sgl->dma_address;
1390                 if (plchan->runtime_addr)
1391                         txd->dstbus.addr = plchan->runtime_addr;
1392                 else
1393                         txd->dstbus.addr = plchan->cd->addr;
1394                 src_buses = pl08x->mem_buses;
1395                 dst_buses = plchan->cd->periph_buses;
1396         } else if (direction == DMA_FROM_DEVICE) {
1397                 txd->ccfg |= PL080_FLOW_PER2MEM << PL080_CONFIG_FLOW_CONTROL_SHIFT;
1398                 txd->cctl |= PL080_CONTROL_DST_INCR;
1399                 if (plchan->runtime_addr)
1400                         txd->srcbus.addr = plchan->runtime_addr;
1401                 else
1402                         txd->srcbus.addr = plchan->cd->addr;
1403                 txd->dstbus.addr = sgl->dma_address;
1404                 src_buses = plchan->cd->periph_buses;
1405                 dst_buses = pl08x->mem_buses;
1406         } else {
1407                 dev_err(&pl08x->adev->dev,
1408                         "%s direction unsupported\n", __func__);
1409                 return NULL;
1410         }
1411
1412         txd->cctl |= pl08x_select_bus(pl08x, src_buses, dst_buses);
1413
1414         ret = pl08x_prep_channel_resources(plchan, txd);
1415         if (ret)
1416                 return NULL;
1417         /*
1418          * NB: the channel lock is held at this point so tx_submit()
1419          * must be called in direct succession.
1420          */
1421
1422         return &txd->tx;
1423 }
1424
1425 static int pl08x_control(struct dma_chan *chan, enum dma_ctrl_cmd cmd,
1426                          unsigned long arg)
1427 {
1428         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1429         struct pl08x_driver_data *pl08x = plchan->host;
1430         unsigned long flags;
1431         int ret = 0;
1432
1433         /* Controls applicable to inactive channels */
1434         if (cmd == DMA_SLAVE_CONFIG) {
1435                 dma_set_runtime_config(chan,
1436                                        (struct dma_slave_config *)
1437                                        arg);
1438                 return 0;
1439         }
1440
1441         /*
1442          * Anything succeeds on channels with no physical allocation and
1443          * no queued transfers.
1444          */
1445         spin_lock_irqsave(&plchan->lock, flags);
1446         if (!plchan->phychan && !plchan->at) {
1447                 spin_unlock_irqrestore(&plchan->lock, flags);
1448                 return 0;
1449         }
1450
1451         switch (cmd) {
1452         case DMA_TERMINATE_ALL:
1453                 plchan->state = PL08X_CHAN_IDLE;
1454
1455                 if (plchan->phychan) {
1456                         pl08x_stop_phy_chan(plchan->phychan);
1457
1458                         /*
1459                          * Mark physical channel as free and free any slave
1460                          * signal
1461                          */
1462                         release_phy_channel(plchan);
1463                 }
1464                 /* Dequeue jobs and free LLIs */
1465                 if (plchan->at) {
1466                         pl08x_free_txd(pl08x, plchan->at);
1467                         plchan->at = NULL;
1468                 }
1469                 /* Dequeue jobs not yet fired as well */
1470                 pl08x_free_txd_list(pl08x, plchan);
1471                 break;
1472         case DMA_PAUSE:
1473                 pl08x_pause_phy_chan(plchan->phychan);
1474                 plchan->state = PL08X_CHAN_PAUSED;
1475                 break;
1476         case DMA_RESUME:
1477                 pl08x_resume_phy_chan(plchan->phychan);
1478                 plchan->state = PL08X_CHAN_RUNNING;
1479                 break;
1480         default:
1481                 /* Unknown command */
1482                 ret = -ENXIO;
1483                 break;
1484         }
1485
1486         spin_unlock_irqrestore(&plchan->lock, flags);
1487
1488         return ret;
1489 }
1490
1491 bool pl08x_filter_id(struct dma_chan *chan, void *chan_id)
1492 {
1493         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1494         char *name = chan_id;
1495
1496         /* Check that the channel is not taken! */
1497         if (!strcmp(plchan->name, name))
1498                 return true;
1499
1500         return false;
1501 }
1502
1503 /*
1504  * Just check that the device is there and active
1505  * TODO: turn this bit on/off depending on the number of
1506  * physical channels actually used, if it is zero... well
1507  * shut it off. That will save some power. Cut the clock
1508  * at the same time.
1509  */
1510 static void pl08x_ensure_on(struct pl08x_driver_data *pl08x)
1511 {
1512         u32 val;
1513
1514         val = readl(pl08x->base + PL080_CONFIG);
1515         val &= ~(PL080_CONFIG_M2_BE | PL080_CONFIG_M1_BE | PL080_CONFIG_ENABLE);
1516         /* We implicitly clear bit 1 and that means little-endian mode */
1517         val |= PL080_CONFIG_ENABLE;
1518         writel(val, pl08x->base + PL080_CONFIG);
1519 }
1520
1521 static void pl08x_tasklet(unsigned long data)
1522 {
1523         struct pl08x_dma_chan *plchan = (struct pl08x_dma_chan *) data;
1524         struct pl08x_driver_data *pl08x = plchan->host;
1525         struct pl08x_txd *txd;
1526         dma_async_tx_callback callback = NULL;
1527         void *callback_param = NULL;
1528         unsigned long flags;
1529
1530         spin_lock_irqsave(&plchan->lock, flags);
1531
1532         txd = plchan->at;
1533         plchan->at = NULL;
1534
1535         if (txd) {
1536                 callback = txd->tx.callback;
1537                 callback_param = txd->tx.callback_param;
1538
1539                 /*
1540                  * Update last completed
1541                  */
1542                 plchan->lc = txd->tx.cookie;
1543
1544                 /*
1545                  * Free the descriptor
1546                  */
1547                 pl08x_free_txd(pl08x, txd);
1548         }
1549         /*
1550          * If a new descriptor is queued, set it up
1551          * plchan->at is NULL here
1552          */
1553         if (!list_empty(&plchan->desc_list)) {
1554                 struct pl08x_txd *next;
1555
1556                 next = list_first_entry(&plchan->desc_list,
1557                                         struct pl08x_txd,
1558                                         node);
1559                 list_del(&next->node);
1560
1561                 pl08x_start_txd(plchan, next);
1562         } else {
1563                 struct pl08x_dma_chan *waiting = NULL;
1564
1565                 /*
1566                  * No more jobs, so free up the physical channel
1567                  * Free any allocated signal on slave transfers too
1568                  */
1569                 release_phy_channel(plchan);
1570                 plchan->state = PL08X_CHAN_IDLE;
1571
1572                 /*
1573                  * And NOW before anyone else can grab that free:d
1574                  * up physical channel, see if there is some memcpy
1575                  * pending that seriously needs to start because of
1576                  * being stacked up while we were choking the
1577                  * physical channels with data.
1578                  */
1579                 list_for_each_entry(waiting, &pl08x->memcpy.channels,
1580                                     chan.device_node) {
1581                   if (waiting->state == PL08X_CHAN_WAITING &&
1582                             waiting->waiting != NULL) {
1583                                 int ret;
1584
1585                                 /* This should REALLY not fail now */
1586                                 ret = prep_phy_channel(waiting,
1587                                                        waiting->waiting);
1588                                 BUG_ON(ret);
1589                                 waiting->state = PL08X_CHAN_RUNNING;
1590                                 waiting->waiting = NULL;
1591                                 pl08x_issue_pending(&waiting->chan);
1592                                 break;
1593                         }
1594                 }
1595         }
1596
1597         spin_unlock_irqrestore(&plchan->lock, flags);
1598
1599         /* Callback to signal completion */
1600         if (callback)
1601                 callback(callback_param);
1602 }
1603
1604 static irqreturn_t pl08x_irq(int irq, void *dev)
1605 {
1606         struct pl08x_driver_data *pl08x = dev;
1607         u32 mask = 0;
1608         u32 val;
1609         int i;
1610
1611         val = readl(pl08x->base + PL080_ERR_STATUS);
1612         if (val) {
1613                 /*
1614                  * An error interrupt (on one or more channels)
1615                  */
1616                 dev_err(&pl08x->adev->dev,
1617                         "%s error interrupt, register value 0x%08x\n",
1618                                 __func__, val);
1619                 /*
1620                  * Simply clear ALL PL08X error interrupts,
1621                  * regardless of channel and cause
1622                  * FIXME: should be 0x00000003 on PL081 really.
1623                  */
1624                 writel(0x000000FF, pl08x->base + PL080_ERR_CLEAR);
1625         }
1626         val = readl(pl08x->base + PL080_INT_STATUS);
1627         for (i = 0; i < pl08x->vd->channels; i++) {
1628                 if ((1 << i) & val) {
1629                         /* Locate physical channel */
1630                         struct pl08x_phy_chan *phychan = &pl08x->phy_chans[i];
1631                         struct pl08x_dma_chan *plchan = phychan->serving;
1632
1633                         /* Schedule tasklet on this channel */
1634                         tasklet_schedule(&plchan->tasklet);
1635
1636                         mask |= (1 << i);
1637                 }
1638         }
1639         /*
1640          * Clear only the terminal interrupts on channels we processed
1641          */
1642         writel(mask, pl08x->base + PL080_TC_CLEAR);
1643
1644         return mask ? IRQ_HANDLED : IRQ_NONE;
1645 }
1646
1647 /*
1648  * Initialise the DMAC memcpy/slave channels.
1649  * Make a local wrapper to hold required data
1650  */
1651 static int pl08x_dma_init_virtual_channels(struct pl08x_driver_data *pl08x,
1652                                            struct dma_device *dmadev,
1653                                            unsigned int channels,
1654                                            bool slave)
1655 {
1656         struct pl08x_dma_chan *chan;
1657         int i;
1658
1659         INIT_LIST_HEAD(&dmadev->channels);
1660         /*
1661          * Register as many many memcpy as we have physical channels,
1662          * we won't always be able to use all but the code will have
1663          * to cope with that situation.
1664          */
1665         for (i = 0; i < channels; i++) {
1666                 chan = kzalloc(sizeof(struct pl08x_dma_chan), GFP_KERNEL);
1667                 if (!chan) {
1668                         dev_err(&pl08x->adev->dev,
1669                                 "%s no memory for channel\n", __func__);
1670                         return -ENOMEM;
1671                 }
1672
1673                 chan->host = pl08x;
1674                 chan->state = PL08X_CHAN_IDLE;
1675
1676                 if (slave) {
1677                         chan->slave = true;
1678                         chan->name = pl08x->pd->slave_channels[i].bus_id;
1679                         chan->cd = &pl08x->pd->slave_channels[i];
1680                 } else {
1681                         chan->cd = &pl08x->pd->memcpy_channel;
1682                         chan->name = kasprintf(GFP_KERNEL, "memcpy%d", i);
1683                         if (!chan->name) {
1684                                 kfree(chan);
1685                                 return -ENOMEM;
1686                         }
1687                 }
1688                 if (chan->cd->circular_buffer) {
1689                         dev_err(&pl08x->adev->dev,
1690                                 "channel %s: circular buffers not supported\n",
1691                                 chan->name);
1692                         kfree(chan);
1693                         continue;
1694                 }
1695                 dev_info(&pl08x->adev->dev,
1696                          "initialize virtual channel \"%s\"\n",
1697                          chan->name);
1698
1699                 chan->chan.device = dmadev;
1700                 chan->chan.cookie = 0;
1701                 chan->lc = 0;
1702
1703                 spin_lock_init(&chan->lock);
1704                 INIT_LIST_HEAD(&chan->desc_list);
1705                 tasklet_init(&chan->tasklet, pl08x_tasklet,
1706                              (unsigned long) chan);
1707
1708                 list_add_tail(&chan->chan.device_node, &dmadev->channels);
1709         }
1710         dev_info(&pl08x->adev->dev, "initialized %d virtual %s channels\n",
1711                  i, slave ? "slave" : "memcpy");
1712         return i;
1713 }
1714
1715 static void pl08x_free_virtual_channels(struct dma_device *dmadev)
1716 {
1717         struct pl08x_dma_chan *chan = NULL;
1718         struct pl08x_dma_chan *next;
1719
1720         list_for_each_entry_safe(chan,
1721                                  next, &dmadev->channels, chan.device_node) {
1722                 list_del(&chan->chan.device_node);
1723                 kfree(chan);
1724         }
1725 }
1726
1727 #ifdef CONFIG_DEBUG_FS
1728 static const char *pl08x_state_str(enum pl08x_dma_chan_state state)
1729 {
1730         switch (state) {
1731         case PL08X_CHAN_IDLE:
1732                 return "idle";
1733         case PL08X_CHAN_RUNNING:
1734                 return "running";
1735         case PL08X_CHAN_PAUSED:
1736                 return "paused";
1737         case PL08X_CHAN_WAITING:
1738                 return "waiting";
1739         default:
1740                 break;
1741         }
1742         return "UNKNOWN STATE";
1743 }
1744
1745 static int pl08x_debugfs_show(struct seq_file *s, void *data)
1746 {
1747         struct pl08x_driver_data *pl08x = s->private;
1748         struct pl08x_dma_chan *chan;
1749         struct pl08x_phy_chan *ch;
1750         unsigned long flags;
1751         int i;
1752
1753         seq_printf(s, "PL08x physical channels:\n");
1754         seq_printf(s, "CHANNEL:\tUSER:\n");
1755         seq_printf(s, "--------\t-----\n");
1756         for (i = 0; i < pl08x->vd->channels; i++) {
1757                 struct pl08x_dma_chan *virt_chan;
1758
1759                 ch = &pl08x->phy_chans[i];
1760
1761                 spin_lock_irqsave(&ch->lock, flags);
1762                 virt_chan = ch->serving;
1763
1764                 seq_printf(s, "%d\t\t%s\n",
1765                            ch->id, virt_chan ? virt_chan->name : "(none)");
1766
1767                 spin_unlock_irqrestore(&ch->lock, flags);
1768         }
1769
1770         seq_printf(s, "\nPL08x virtual memcpy channels:\n");
1771         seq_printf(s, "CHANNEL:\tSTATE:\n");
1772         seq_printf(s, "--------\t------\n");
1773         list_for_each_entry(chan, &pl08x->memcpy.channels, chan.device_node) {
1774                 seq_printf(s, "%s\t\t%s\n", chan->name,
1775                            pl08x_state_str(chan->state));
1776         }
1777
1778         seq_printf(s, "\nPL08x virtual slave channels:\n");
1779         seq_printf(s, "CHANNEL:\tSTATE:\n");
1780         seq_printf(s, "--------\t------\n");
1781         list_for_each_entry(chan, &pl08x->slave.channels, chan.device_node) {
1782                 seq_printf(s, "%s\t\t%s\n", chan->name,
1783                            pl08x_state_str(chan->state));
1784         }
1785
1786         return 0;
1787 }
1788
1789 static int pl08x_debugfs_open(struct inode *inode, struct file *file)
1790 {
1791         return single_open(file, pl08x_debugfs_show, inode->i_private);
1792 }
1793
1794 static const struct file_operations pl08x_debugfs_operations = {
1795         .open           = pl08x_debugfs_open,
1796         .read           = seq_read,
1797         .llseek         = seq_lseek,
1798         .release        = single_release,
1799 };
1800
1801 static void init_pl08x_debugfs(struct pl08x_driver_data *pl08x)
1802 {
1803         /* Expose a simple debugfs interface to view all clocks */
1804         (void) debugfs_create_file(dev_name(&pl08x->adev->dev), S_IFREG | S_IRUGO,
1805                                    NULL, pl08x,
1806                                    &pl08x_debugfs_operations);
1807 }
1808
1809 #else
1810 static inline void init_pl08x_debugfs(struct pl08x_driver_data *pl08x)
1811 {
1812 }
1813 #endif
1814
1815 static int pl08x_probe(struct amba_device *adev, struct amba_id *id)
1816 {
1817         struct pl08x_driver_data *pl08x;
1818         const struct vendor_data *vd = id->data;
1819         int ret = 0;
1820         int i;
1821
1822         ret = amba_request_regions(adev, NULL);
1823         if (ret)
1824                 return ret;
1825
1826         /* Create the driver state holder */
1827         pl08x = kzalloc(sizeof(struct pl08x_driver_data), GFP_KERNEL);
1828         if (!pl08x) {
1829                 ret = -ENOMEM;
1830                 goto out_no_pl08x;
1831         }
1832
1833         /* Initialize memcpy engine */
1834         dma_cap_set(DMA_MEMCPY, pl08x->memcpy.cap_mask);
1835         pl08x->memcpy.dev = &adev->dev;
1836         pl08x->memcpy.device_alloc_chan_resources = pl08x_alloc_chan_resources;
1837         pl08x->memcpy.device_free_chan_resources = pl08x_free_chan_resources;
1838         pl08x->memcpy.device_prep_dma_memcpy = pl08x_prep_dma_memcpy;
1839         pl08x->memcpy.device_prep_dma_interrupt = pl08x_prep_dma_interrupt;
1840         pl08x->memcpy.device_tx_status = pl08x_dma_tx_status;
1841         pl08x->memcpy.device_issue_pending = pl08x_issue_pending;
1842         pl08x->memcpy.device_control = pl08x_control;
1843
1844         /* Initialize slave engine */
1845         dma_cap_set(DMA_SLAVE, pl08x->slave.cap_mask);
1846         pl08x->slave.dev = &adev->dev;
1847         pl08x->slave.device_alloc_chan_resources = pl08x_alloc_chan_resources;
1848         pl08x->slave.device_free_chan_resources = pl08x_free_chan_resources;
1849         pl08x->slave.device_prep_dma_interrupt = pl08x_prep_dma_interrupt;
1850         pl08x->slave.device_tx_status = pl08x_dma_tx_status;
1851         pl08x->slave.device_issue_pending = pl08x_issue_pending;
1852         pl08x->slave.device_prep_slave_sg = pl08x_prep_slave_sg;
1853         pl08x->slave.device_control = pl08x_control;
1854
1855         /* Get the platform data */
1856         pl08x->pd = dev_get_platdata(&adev->dev);
1857         if (!pl08x->pd) {
1858                 dev_err(&adev->dev, "no platform data supplied\n");
1859                 goto out_no_platdata;
1860         }
1861
1862         /* Assign useful pointers to the driver state */
1863         pl08x->adev = adev;
1864         pl08x->vd = vd;
1865
1866         /* By default, AHB1 only.  If dualmaster, from platform */
1867         pl08x->lli_buses = PL08X_AHB1;
1868         pl08x->mem_buses = PL08X_AHB1;
1869         if (pl08x->vd->dualmaster) {
1870                 pl08x->lli_buses = pl08x->pd->lli_buses;
1871                 pl08x->mem_buses = pl08x->pd->mem_buses;
1872         }
1873
1874         /* A DMA memory pool for LLIs, align on 1-byte boundary */
1875         pl08x->pool = dma_pool_create(DRIVER_NAME, &pl08x->adev->dev,
1876                         PL08X_LLI_TSFR_SIZE, PL08X_ALIGN, 0);
1877         if (!pl08x->pool) {
1878                 ret = -ENOMEM;
1879                 goto out_no_lli_pool;
1880         }
1881
1882         spin_lock_init(&pl08x->lock);
1883
1884         pl08x->base = ioremap(adev->res.start, resource_size(&adev->res));
1885         if (!pl08x->base) {
1886                 ret = -ENOMEM;
1887                 goto out_no_ioremap;
1888         }
1889
1890         /* Turn on the PL08x */
1891         pl08x_ensure_on(pl08x);
1892
1893         /*
1894          * Attach the interrupt handler
1895          */
1896         writel(0x000000FF, pl08x->base + PL080_ERR_CLEAR);
1897         writel(0x000000FF, pl08x->base + PL080_TC_CLEAR);
1898
1899         ret = request_irq(adev->irq[0], pl08x_irq, IRQF_DISABLED,
1900                           DRIVER_NAME, pl08x);
1901         if (ret) {
1902                 dev_err(&adev->dev, "%s failed to request interrupt %d\n",
1903                         __func__, adev->irq[0]);
1904                 goto out_no_irq;
1905         }
1906
1907         /* Initialize physical channels */
1908         pl08x->phy_chans = kmalloc((vd->channels * sizeof(struct pl08x_phy_chan)),
1909                         GFP_KERNEL);
1910         if (!pl08x->phy_chans) {
1911                 dev_err(&adev->dev, "%s failed to allocate "
1912                         "physical channel holders\n",
1913                         __func__);
1914                 goto out_no_phychans;
1915         }
1916
1917         for (i = 0; i < vd->channels; i++) {
1918                 struct pl08x_phy_chan *ch = &pl08x->phy_chans[i];
1919
1920                 ch->id = i;
1921                 ch->base = pl08x->base + PL080_Cx_BASE(i);
1922                 spin_lock_init(&ch->lock);
1923                 ch->serving = NULL;
1924                 ch->signal = -1;
1925                 dev_info(&adev->dev,
1926                          "physical channel %d is %s\n", i,
1927                          pl08x_phy_channel_busy(ch) ? "BUSY" : "FREE");
1928         }
1929
1930         /* Register as many memcpy channels as there are physical channels */
1931         ret = pl08x_dma_init_virtual_channels(pl08x, &pl08x->memcpy,
1932                                               pl08x->vd->channels, false);
1933         if (ret <= 0) {
1934                 dev_warn(&pl08x->adev->dev,
1935                          "%s failed to enumerate memcpy channels - %d\n",
1936                          __func__, ret);
1937                 goto out_no_memcpy;
1938         }
1939         pl08x->memcpy.chancnt = ret;
1940
1941         /* Register slave channels */
1942         ret = pl08x_dma_init_virtual_channels(pl08x, &pl08x->slave,
1943                                               pl08x->pd->num_slave_channels,
1944                                               true);
1945         if (ret <= 0) {
1946                 dev_warn(&pl08x->adev->dev,
1947                         "%s failed to enumerate slave channels - %d\n",
1948                                 __func__, ret);
1949                 goto out_no_slave;
1950         }
1951         pl08x->slave.chancnt = ret;
1952
1953         ret = dma_async_device_register(&pl08x->memcpy);
1954         if (ret) {
1955                 dev_warn(&pl08x->adev->dev,
1956                         "%s failed to register memcpy as an async device - %d\n",
1957                         __func__, ret);
1958                 goto out_no_memcpy_reg;
1959         }
1960
1961         ret = dma_async_device_register(&pl08x->slave);
1962         if (ret) {
1963                 dev_warn(&pl08x->adev->dev,
1964                         "%s failed to register slave as an async device - %d\n",
1965                         __func__, ret);
1966                 goto out_no_slave_reg;
1967         }
1968
1969         amba_set_drvdata(adev, pl08x);
1970         init_pl08x_debugfs(pl08x);
1971         dev_info(&pl08x->adev->dev, "DMA: PL%03x rev%u at 0x%08llx irq %d\n",
1972                  amba_part(adev), amba_rev(adev),
1973                  (unsigned long long)adev->res.start, adev->irq[0]);
1974         return 0;
1975
1976 out_no_slave_reg:
1977         dma_async_device_unregister(&pl08x->memcpy);
1978 out_no_memcpy_reg:
1979         pl08x_free_virtual_channels(&pl08x->slave);
1980 out_no_slave:
1981         pl08x_free_virtual_channels(&pl08x->memcpy);
1982 out_no_memcpy:
1983         kfree(pl08x->phy_chans);
1984 out_no_phychans:
1985         free_irq(adev->irq[0], pl08x);
1986 out_no_irq:
1987         iounmap(pl08x->base);
1988 out_no_ioremap:
1989         dma_pool_destroy(pl08x->pool);
1990 out_no_lli_pool:
1991 out_no_platdata:
1992         kfree(pl08x);
1993 out_no_pl08x:
1994         amba_release_regions(adev);
1995         return ret;
1996 }
1997
1998 /* PL080 has 8 channels and the PL080 have just 2 */
1999 static struct vendor_data vendor_pl080 = {
2000         .channels = 8,
2001         .dualmaster = true,
2002 };
2003
2004 static struct vendor_data vendor_pl081 = {
2005         .channels = 2,
2006         .dualmaster = false,
2007 };
2008
2009 static struct amba_id pl08x_ids[] = {
2010         /* PL080 */
2011         {
2012                 .id     = 0x00041080,
2013                 .mask   = 0x000fffff,
2014                 .data   = &vendor_pl080,
2015         },
2016         /* PL081 */
2017         {
2018                 .id     = 0x00041081,
2019                 .mask   = 0x000fffff,
2020                 .data   = &vendor_pl081,
2021         },
2022         /* Nomadik 8815 PL080 variant */
2023         {
2024                 .id     = 0x00280880,
2025                 .mask   = 0x00ffffff,
2026                 .data   = &vendor_pl080,
2027         },
2028         { 0, 0 },
2029 };
2030
2031 static struct amba_driver pl08x_amba_driver = {
2032         .drv.name       = DRIVER_NAME,
2033         .id_table       = pl08x_ids,
2034         .probe          = pl08x_probe,
2035 };
2036
2037 static int __init pl08x_init(void)
2038 {
2039         int retval;
2040         retval = amba_driver_register(&pl08x_amba_driver);
2041         if (retval)
2042                 printk(KERN_WARNING DRIVER_NAME
2043                        "failed to register as an AMBA device (%d)\n",
2044                        retval);
2045         return retval;
2046 }
2047 subsys_initcall(pl08x_init);