drivers/dma/amba-pl08x.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright (c) 2006 ARM Ltd.
   4  * Copyright (c) 2010 ST-Ericsson SA
   5  * Copyirght (c) 2017 Linaro Ltd.
   6  *
   7  * Author: Peter Pearse <peter.pearse@arm.com>
   8  * Author: Linus Walleij <linus.walleij@linaro.org>
   9  *
  10  * Documentation: ARM DDI 0196G == PL080
  11  * Documentation: ARM DDI 0218E == PL081
  12  * Documentation: S3C6410 User's Manual == PL080S
  13  *
  14  * PL080 & PL081 both have 16 sets of DMA signals that can be routed to any
  15  * channel.
  16  *
  17  * The PL080 has 8 channels available for simultaneous use, and the PL081
  18  * has only two channels. So on these DMA controllers the number of channels
  19  * and the number of incoming DMA signals are two totally different things.
  20  * It is usually not possible to theoretically handle all physical signals,
  21  * so a multiplexing scheme with possible denial of use is necessary.
  22  *
  23  * The PL080 has a dual bus master, PL081 has a single master.
  24  *
  25  * PL080S is a version modified by Samsung and used in S3C64xx SoCs.
  26  * It differs in following aspects:
  27  * - CH_CONFIG register at different offset,
  28  * - separate CH_CONTROL2 register for transfer size,
  29  * - bigger maximum transfer size,
  30  * - 8-word aligned LLI, instead of 4-word, due to extra CCTL2 word,
  31  * - no support for peripheral flow control.
  32  *
  33  * Memory to peripheral transfer may be visualized as
  34  *      Get data from memory to DMAC
  35  *      Until no data left
  36  *              On burst request from peripheral
  37  *                      Destination burst from DMAC to peripheral
  38  *                      Clear burst request
  39  *      Raise terminal count interrupt
  40  *
  41  * For peripherals with a FIFO:
  42  * Source      burst size == half the depth of the peripheral FIFO
  43  * Destination burst size == the depth of the peripheral FIFO
  44  *
  45  * (Bursts are irrelevant for mem to mem transfers - there are no burst
  46  * signals, the DMA controller will simply facilitate its AHB master.)
  47  *
  48  * ASSUMES default (little) endianness for DMA transfers
  49  *
  50  * The PL08x has two flow control settings:
  51  *  - DMAC flow control: the transfer size defines the number of transfers
  52  *    which occur for the current LLI entry, and the DMAC raises TC at the
  53  *    end of every LLI entry.  Observed behaviour shows the DMAC listening
  54  *    to both the BREQ and SREQ signals (contrary to documented),
  55  *    transferring data if either is active.  The LBREQ and LSREQ signals
  56  *    are ignored.
  57  *
  58  *  - Peripheral flow control: the transfer size is ignored (and should be
  59  *    zero).  The data is transferred from the current LLI entry, until
  60  *    after the final transfer signalled by LBREQ or LSREQ.  The DMAC
  61  *    will then move to the next LLI entry. Unsupported by PL080S.
  62  */
  63 #include <linux/amba/bus.h>
  64 #include <linux/amba/pl08x.h>
  65 #include <linux/debugfs.h>
  66 #include <linux/delay.h>
  67 #include <linux/device.h>
  68 #include <linux/dmaengine.h>
  69 #include <linux/dmapool.h>
  70 #include <linux/dma-mapping.h>
  71 #include <linux/export.h>
  72 #include <linux/init.h>
  73 #include <linux/interrupt.h>
  74 #include <linux/module.h>
  75 #include <linux/of.h>
  76 #include <linux/of_dma.h>
  77 #include <linux/pm_runtime.h>
  78 #include <linux/seq_file.h>
  79 #include <linux/slab.h>
  80 #include <linux/amba/pl080.h>
  81
  82 #include "dmaengine.h"
  83 #include "virt-dma.h"
  84
  85 #define DRIVER_NAME     "pl08xdmac"
  86
  87 #define PL80X_DMA_BUSWIDTHS \
  88         BIT(DMA_SLAVE_BUSWIDTH_UNDEFINED) | \
  89         BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) | \
  90         BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) | \
  91         BIT(DMA_SLAVE_BUSWIDTH_4_BYTES)
  92
  93 static struct amba_driver pl08x_amba_driver;
  94 struct pl08x_driver_data;
  95
  96 /**
  97  * struct vendor_data - vendor-specific config parameters for PL08x derivatives
  98  * @config_offset: offset to the configuration register
  99  * @channels: the number of channels available in this variant
 100  * @signals: the number of request signals available from the hardware
 101  * @dualmaster: whether this version supports dual AHB masters or not.
 102  * @nomadik: whether this variant is a ST Microelectronics Nomadik, where the
 103  *      channels have Nomadik security extension bits that need to be checked
 104  *      for permission before use and some registers are missing
 105  * @pl080s: whether this variant is a Samsung PL080S, which has separate
 106  *      register and LLI word for transfer size.
 107  * @ftdmac020: whether this variant is a Faraday Technology FTDMAC020
 108  * @max_transfer_size: the maximum single element transfer size for this
 109  *      PL08x variant.
 110  */
 111 struct vendor_data {
 112         u8 config_offset;
 113         u8 channels;
 114         u8 signals;
 115         bool dualmaster;
 116         bool nomadik;
 117         bool pl080s;
 118         bool ftdmac020;
 119         u32 max_transfer_size;
 120 };
 121
 122 /**
 123  * struct pl08x_bus_data - information of source or destination
 124  * busses for a transfer
 125  * @addr: current address
 126  * @maxwidth: the maximum width of a transfer on this bus
 127  * @buswidth: the width of this bus in bytes: 1, 2 or 4
 128  */
 129 struct pl08x_bus_data {
 130         dma_addr_t addr;
 131         u8 maxwidth;
 132         u8 buswidth;
 133 };
 134
 135 #define IS_BUS_ALIGNED(bus) IS_ALIGNED((bus)->addr, (bus)->buswidth)
 136
 137 /**
 138  * struct pl08x_phy_chan - holder for the physical channels
 139  * @id: physical index to this channel
 140  * @base: memory base address for this physical channel
 141  * @reg_config: configuration address for this physical channel
 142  * @reg_control: control address for this physical channel
 143  * @reg_src: transfer source address register
 144  * @reg_dst: transfer destination address register
 145  * @reg_lli: transfer LLI address register
 146  * @reg_busy: if the variant has a special per-channel busy register,
 147  * this contains a pointer to it
 148  * @lock: a lock to use when altering an instance of this struct
 149  * @serving: the virtual channel currently being served by this physical
 150  * channel
 151  * @locked: channel unavailable for the system, e.g. dedicated to secure
 152  * world
 153  * @ftdmac020: channel is on a FTDMAC020
 154  * @pl080s: channel is on a PL08s
 155  */
 156 struct pl08x_phy_chan {
 157         unsigned int id;
 158         void __iomem *base;
 159         void __iomem *reg_config;
 160         void __iomem *reg_control;
 161         void __iomem *reg_src;
 162         void __iomem *reg_dst;
 163         void __iomem *reg_lli;
 164         void __iomem *reg_busy;
 165         spinlock_t lock;
 166         struct pl08x_dma_chan *serving;
 167         bool locked;
 168         bool ftdmac020;
 169         bool pl080s;
 170 };
 171
 172 /**
 173  * struct pl08x_sg - structure containing data per sg
 174  * @src_addr: src address of sg
 175  * @dst_addr: dst address of sg
 176  * @len: transfer len in bytes
 177  * @node: node for txd's dsg_list
 178  */
 179 struct pl08x_sg {
 180         dma_addr_t src_addr;
 181         dma_addr_t dst_addr;
 182         size_t len;
 183         struct list_head node;
 184 };
 185
 186 /**
 187  * struct pl08x_txd - wrapper for struct dma_async_tx_descriptor
 188  * @vd: virtual DMA descriptor
 189  * @dsg_list: list of children sg's
 190  * @llis_bus: DMA memory address (physical) start for the LLIs
 191  * @llis_va: virtual memory address start for the LLIs
 192  * @cctl: control reg values for current txd
 193  * @ccfg: config reg values for current txd
 194  * @done: this marks completed descriptors, which should not have their
 195  *   mux released.
 196  * @cyclic: indicate cyclic transfers
 197  */
 198 struct pl08x_txd {
 199         struct virt_dma_desc vd;
 200         struct list_head dsg_list;
 201         dma_addr_t llis_bus;
 202         u32 *llis_va;
 203         /* Default cctl value for LLIs */
 204         u32 cctl;
 205         /*
 206          * Settings to be put into the physical channel when we
 207          * trigger this txd.  Other registers are in llis_va[0].
 208          */
 209         u32 ccfg;
 210         bool done;
 211         bool cyclic;
 212 };
 213
 214 /**
 215  * enum pl08x_dma_chan_state - holds the PL08x specific virtual channel
 216  * states
 217  * @PL08X_CHAN_IDLE: the channel is idle
 218  * @PL08X_CHAN_RUNNING: the channel has allocated a physical transport
 219  * channel and is running a transfer on it
 220  * @PL08X_CHAN_PAUSED: the channel has allocated a physical transport
 221  * channel, but the transfer is currently paused
 222  * @PL08X_CHAN_WAITING: the channel is waiting for a physical transport
 223  * channel to become available (only pertains to memcpy channels)
 224  */
 225 enum pl08x_dma_chan_state {
 226         PL08X_CHAN_IDLE,
 227         PL08X_CHAN_RUNNING,
 228         PL08X_CHAN_PAUSED,
 229         PL08X_CHAN_WAITING,
 230 };
 231
 232 /**
 233  * struct pl08x_dma_chan - this structure wraps a DMA ENGINE channel
 234  * @vc: wrappped virtual channel
 235  * @phychan: the physical channel utilized by this channel, if there is one
 236  * @name: name of channel
 237  * @cd: channel platform data
 238  * @cfg: slave configuration
 239  * @at: active transaction on this channel
 240  * @host: a pointer to the host (internal use)
 241  * @state: whether the channel is idle, paused, running etc
 242  * @slave: whether this channel is a device (slave) or for memcpy
 243  * @signal: the physical DMA request signal which this channel is using
 244  * @mux_use: count of descriptors using this DMA request signal setting
 245  * @waiting_at: time in jiffies when this channel moved to waiting state
 246  */
 247 struct pl08x_dma_chan {
 248         struct virt_dma_chan vc;
 249         struct pl08x_phy_chan *phychan;
 250         const char *name;
 251         struct pl08x_channel_data *cd;
 252         struct dma_slave_config cfg;
 253         struct pl08x_txd *at;
 254         struct pl08x_driver_data *host;
 255         enum pl08x_dma_chan_state state;
 256         bool slave;
 257         int signal;
 258         unsigned mux_use;
 259         unsigned long waiting_at;
 260 };
 261
 262 /**
 263  * struct pl08x_driver_data - the local state holder for the PL08x
 264  * @slave: optional slave engine for this instance
 265  * @memcpy: memcpy engine for this instance
 266  * @has_slave: the PL08x has a slave engine (routed signals)
 267  * @base: virtual memory base (remapped) for the PL08x
 268  * @adev: the corresponding AMBA (PrimeCell) bus entry
 269  * @vd: vendor data for this PL08x variant
 270  * @pd: platform data passed in from the platform/machine
 271  * @phy_chans: array of data for the physical channels
 272  * @pool: a pool for the LLI descriptors
 273  * @lli_buses: bitmask to or in to LLI pointer selecting AHB port for LLI
 274  * fetches
 275  * @mem_buses: set to indicate memory transfers on AHB2.
 276  * @lli_words: how many words are used in each LLI item for this variant
 277  */
 278 struct pl08x_driver_data {
 279         struct dma_device slave;
 280         struct dma_device memcpy;
 281         bool has_slave;
 282         void __iomem *base;
 283         struct amba_device *adev;
 284         const struct vendor_data *vd;
 285         struct pl08x_platform_data *pd;
 286         struct pl08x_phy_chan *phy_chans;
 287         struct dma_pool *pool;
 288         u8 lli_buses;
 289         u8 mem_buses;
 290         u8 lli_words;
 291 };
 292
 293 /*
 294  * PL08X specific defines
 295  */
 296
 297 /* The order of words in an LLI. */
 298 #define PL080_LLI_SRC           0
 299 #define PL080_LLI_DST           1
 300 #define PL080_LLI_LLI           2
 301 #define PL080_LLI_CCTL          3
 302 #define PL080S_LLI_CCTL2        4
 303
 304 /* Total words in an LLI. */
 305 #define PL080_LLI_WORDS         4
 306 #define PL080S_LLI_WORDS        8
 307
 308 /*
 309  * Number of LLIs in each LLI buffer allocated for one transfer
 310  * (maximum times we call dma_pool_alloc on this pool without freeing)
 311  */
 312 #define MAX_NUM_TSFR_LLIS       512
 313 #define PL08X_ALIGN             8
 314
 315 static inline struct pl08x_dma_chan *to_pl08x_chan(struct dma_chan *chan)
 316 {
 317         return container_of(chan, struct pl08x_dma_chan, vc.chan);
 318 }
 319
 320 static inline struct pl08x_txd *to_pl08x_txd(struct dma_async_tx_descriptor *tx)
 321 {
 322         return container_of(tx, struct pl08x_txd, vd.tx);
 323 }
 324
 325 /*
 326  * Mux handling.
 327  *
 328  * This gives us the DMA request input to the PL08x primecell which the
 329  * peripheral described by the channel data will be routed to, possibly
 330  * via a board/SoC specific external MUX.  One important point to note
 331  * here is that this does not depend on the physical channel.
 332  */
 333 static int pl08x_request_mux(struct pl08x_dma_chan *plchan)
 334 {
 335         const struct pl08x_platform_data *pd = plchan->host->pd;
 336         int ret;
 337
 338         if (plchan->mux_use++ == 0 && pd->get_xfer_signal) {
 339                 ret = pd->get_xfer_signal(plchan->cd);
 340                 if (ret < 0) {
 341                         plchan->mux_use = 0;
 342                         return ret;
 343                 }
 344
 345                 plchan->signal = ret;
 346         }
 347         return 0;
 348 }
 349
 350 static void pl08x_release_mux(struct pl08x_dma_chan *plchan)
 351 {
 352         const struct pl08x_platform_data *pd = plchan->host->pd;
 353
 354         if (plchan->signal >= 0) {
 355                 WARN_ON(plchan->mux_use == 0);
 356
 357                 if (--plchan->mux_use == 0 && pd->put_xfer_signal) {
 358                         pd->put_xfer_signal(plchan->cd, plchan->signal);
 359                         plchan->signal = -1;
 360                 }
 361         }
 362 }
 363
 364 /*
 365  * Physical channel handling
 366  */
 367
 368 /* Whether a certain channel is busy or not */
 369 static int pl08x_phy_channel_busy(struct pl08x_phy_chan *ch)
 370 {
 371         unsigned int val;
 372
 373         /* If we have a special busy register, take a shortcut */
 374         if (ch->reg_busy) {
 375                 val = readl(ch->reg_busy);
 376                 return !!(val & BIT(ch->id));
 377         }
 378         val = readl(ch->reg_config);
 379         return val & PL080_CONFIG_ACTIVE;
 380 }
 381
 382 /*
 383  * pl08x_write_lli() - Write an LLI into the DMA controller.
 384  *
 385  * The PL08x derivatives support linked lists, but the first item of the
 386  * list containing the source, destination, control word and next LLI is
 387  * ignored. Instead the driver has to write those values directly into the
 388  * SRC, DST, LLI and control registers. On FTDMAC020 also the SIZE
 389  * register need to be set up for the first transfer.
 390  */
 391 static void pl08x_write_lli(struct pl08x_driver_data *pl08x,
 392                 struct pl08x_phy_chan *phychan, const u32 *lli, u32 ccfg)
 393 {
 394         if (pl08x->vd->pl080s)
 395                 dev_vdbg(&pl08x->adev->dev,
 396                         "WRITE channel %d: csrc=0x%08x, cdst=0x%08x, "
 397                         "clli=0x%08x, cctl=0x%08x, cctl2=0x%08x, ccfg=0x%08x\n",
 398                         phychan->id, lli[PL080_LLI_SRC], lli[PL080_LLI_DST],
 399                         lli[PL080_LLI_LLI], lli[PL080_LLI_CCTL],
 400                         lli[PL080S_LLI_CCTL2], ccfg);
 401         else
 402                 dev_vdbg(&pl08x->adev->dev,
 403                         "WRITE channel %d: csrc=0x%08x, cdst=0x%08x, "
 404                         "clli=0x%08x, cctl=0x%08x, ccfg=0x%08x\n",
 405                         phychan->id, lli[PL080_LLI_SRC], lli[PL080_LLI_DST],
 406                         lli[PL080_LLI_LLI], lli[PL080_LLI_CCTL], ccfg);
 407
 408         writel_relaxed(lli[PL080_LLI_SRC], phychan->reg_src);
 409         writel_relaxed(lli[PL080_LLI_DST], phychan->reg_dst);
 410         writel_relaxed(lli[PL080_LLI_LLI], phychan->reg_lli);
 411
 412         /*
 413          * The FTMAC020 has a different layout in the CCTL word of the LLI
 414          * and the CCTL register which is split in CSR and SIZE registers.
 415          * Convert the LLI item CCTL into the proper values to write into
 416          * the CSR and SIZE registers.
 417          */
 418         if (phychan->ftdmac020) {
 419                 u32 llictl = lli[PL080_LLI_CCTL];
 420                 u32 val = 0;
 421
 422                 /* Write the transfer size (12 bits) to the size register */
 423                 writel_relaxed(llictl & FTDMAC020_LLI_TRANSFER_SIZE_MASK,
 424                                phychan->base + FTDMAC020_CH_SIZE);
 425                 /*
 426                  * Then write the control bits 28..16 to the control register
 427                  * by shuffleing the bits around to where they are in the
 428                  * main register. The mapping is as follows:
 429                  * Bit 28: TC_MSK - mask on all except last LLI
 430                  * Bit 27..25: SRC_WIDTH
 431                  * Bit 24..22: DST_WIDTH
 432                  * Bit 21..20: SRCAD_CTRL
 433                  * Bit 19..17: DSTAD_CTRL
 434                  * Bit 17: SRC_SEL
 435                  * Bit 16: DST_SEL
 436                  */
 437                 if (llictl & FTDMAC020_LLI_TC_MSK)
 438                         val |= FTDMAC020_CH_CSR_TC_MSK;
 439                 val |= ((llictl  & FTDMAC020_LLI_SRC_WIDTH_MSK) >>
 440                         (FTDMAC020_LLI_SRC_WIDTH_SHIFT -
 441                          FTDMAC020_CH_CSR_SRC_WIDTH_SHIFT));
 442                 val |= ((llictl  & FTDMAC020_LLI_DST_WIDTH_MSK) >>
 443                         (FTDMAC020_LLI_DST_WIDTH_SHIFT -
 444                          FTDMAC020_CH_CSR_DST_WIDTH_SHIFT));
 445                 val |= ((llictl  & FTDMAC020_LLI_SRCAD_CTL_MSK) >>
 446                         (FTDMAC020_LLI_SRCAD_CTL_SHIFT -
 447                          FTDMAC020_CH_CSR_SRCAD_CTL_SHIFT));
 448                 val |= ((llictl  & FTDMAC020_LLI_DSTAD_CTL_MSK) >>
 449                         (FTDMAC020_LLI_DSTAD_CTL_SHIFT -
 450                          FTDMAC020_CH_CSR_DSTAD_CTL_SHIFT));
 451                 if (llictl & FTDMAC020_LLI_SRC_SEL)
 452                         val |= FTDMAC020_CH_CSR_SRC_SEL;
 453                 if (llictl & FTDMAC020_LLI_DST_SEL)
 454                         val |= FTDMAC020_CH_CSR_DST_SEL;
 455
 456                 /*
 457                  * Set up the bits that exist in the CSR but are not
 458                  * part the LLI, i.e. only gets written to the control
 459                  * register right here.
 460                  *
 461                  * FIXME: do not just handle memcpy, also handle slave DMA.
 462                  */
 463                 switch (pl08x->pd->memcpy_burst_size) {
 464                 default:
 465                 case PL08X_BURST_SZ_1:
 466                         val |= PL080_BSIZE_1 <<
 467                                 FTDMAC020_CH_CSR_SRC_SIZE_SHIFT;
 468                         break;
 469                 case PL08X_BURST_SZ_4:
 470                         val |= PL080_BSIZE_4 <<
 471                                 FTDMAC020_CH_CSR_SRC_SIZE_SHIFT;
 472                         break;
 473                 case PL08X_BURST_SZ_8:
 474                         val |= PL080_BSIZE_8 <<
 475                                 FTDMAC020_CH_CSR_SRC_SIZE_SHIFT;
 476                         break;
 477                 case PL08X_BURST_SZ_16:
 478                         val |= PL080_BSIZE_16 <<
 479                                 FTDMAC020_CH_CSR_SRC_SIZE_SHIFT;
 480                         break;
 481                 case PL08X_BURST_SZ_32:
 482                         val |= PL080_BSIZE_32 <<
 483                                 FTDMAC020_CH_CSR_SRC_SIZE_SHIFT;
 484                         break;
 485                 case PL08X_BURST_SZ_64:
 486                         val |= PL080_BSIZE_64 <<
 487                                 FTDMAC020_CH_CSR_SRC_SIZE_SHIFT;
 488                         break;
 489                 case PL08X_BURST_SZ_128:
 490                         val |= PL080_BSIZE_128 <<
 491                                 FTDMAC020_CH_CSR_SRC_SIZE_SHIFT;
 492                         break;
 493                 case PL08X_BURST_SZ_256:
 494                         val |= PL080_BSIZE_256 <<
 495                                 FTDMAC020_CH_CSR_SRC_SIZE_SHIFT;
 496                         break;
 497                 }
 498
 499                 /* Protection flags */
 500                 if (pl08x->pd->memcpy_prot_buff)
 501                         val |= FTDMAC020_CH_CSR_PROT2;
 502                 if (pl08x->pd->memcpy_prot_cache)
 503                         val |= FTDMAC020_CH_CSR_PROT3;
 504                 /* We are the kernel, so we are in privileged mode */
 505                 val |= FTDMAC020_CH_CSR_PROT1;
 506
 507                 writel_relaxed(val, phychan->reg_control);
 508         } else {
 509                 /* Bits are just identical */
 510                 writel_relaxed(lli[PL080_LLI_CCTL], phychan->reg_control);
 511         }
 512
 513         /* Second control word on the PL080s */
 514         if (pl08x->vd->pl080s)
 515                 writel_relaxed(lli[PL080S_LLI_CCTL2],
 516                                 phychan->base + PL080S_CH_CONTROL2);
 517
 518         writel(ccfg, phychan->reg_config);
 519 }
 520
 521 /*
 522  * Set the initial DMA register values i.e. those for the first LLI
 523  * The next LLI pointer and the configuration interrupt bit have
 524  * been set when the LLIs were constructed.  Poke them into the hardware
 525  * and start the transfer.
 526  */
 527 static void pl08x_start_next_txd(struct pl08x_dma_chan *plchan)
 528 {
 529         struct pl08x_driver_data *pl08x = plchan->host;
 530         struct pl08x_phy_chan *phychan = plchan->phychan;
 531         struct virt_dma_desc *vd = vchan_next_desc(&plchan->vc);
 532         struct pl08x_txd *txd = to_pl08x_txd(&vd->tx);
 533         u32 val;
 534
 535         list_del(&txd->vd.node);
 536
 537         plchan->at = txd;
 538
 539         /* Wait for channel inactive */
 540         while (pl08x_phy_channel_busy(phychan))
 541                 cpu_relax();
 542
 543         pl08x_write_lli(pl08x, phychan, &txd->llis_va[0], txd->ccfg);
 544
 545         /* Enable the DMA channel */
 546         /* Do not access config register until channel shows as disabled */
 547         while (readl(pl08x->base + PL080_EN_CHAN) & BIT(phychan->id))
 548                 cpu_relax();
 549
 550         /* Do not access config register until channel shows as inactive */
 551         if (phychan->ftdmac020) {
 552                 val = readl(phychan->reg_config);
 553                 while (val & FTDMAC020_CH_CFG_BUSY)
 554                         val = readl(phychan->reg_config);
 555
 556                 val = readl(phychan->reg_control);
 557                 while (val & FTDMAC020_CH_CSR_EN)
 558                         val = readl(phychan->reg_control);
 559
 560                 writel(val | FTDMAC020_CH_CSR_EN,
 561                        phychan->reg_control);
 562         } else {
 563                 val = readl(phychan->reg_config);
 564                 while ((val & PL080_CONFIG_ACTIVE) ||
 565                        (val & PL080_CONFIG_ENABLE))
 566                         val = readl(phychan->reg_config);
 567
 568                 writel(val | PL080_CONFIG_ENABLE, phychan->reg_config);
 569         }
 570 }
 571
 572 /*
 573  * Pause the channel by setting the HALT bit.
 574  *
 575  * For M->P transfers, pause the DMAC first and then stop the peripheral -
 576  * the FIFO can only drain if the peripheral is still requesting data.
 577  * (note: this can still timeout if the DMAC FIFO never drains of data.)
 578  *
 579  * For P->M transfers, disable the peripheral first to stop it filling
 580  * the DMAC FIFO, and then pause the DMAC.
 581  */
 582 static void pl08x_pause_phy_chan(struct pl08x_phy_chan *ch)
 583 {
 584         u32 val;
 585         int timeout;
 586
 587         if (ch->ftdmac020) {
 588                 /* Use the enable bit on the FTDMAC020 */
 589                 val = readl(ch->reg_control);
 590                 val &= ~FTDMAC020_CH_CSR_EN;
 591                 writel(val, ch->reg_control);
 592                 return;
 593         }
 594
 595         /* Set the HALT bit and wait for the FIFO to drain */
 596         val = readl(ch->reg_config);
 597         val |= PL080_CONFIG_HALT;
 598         writel(val, ch->reg_config);
 599
 600         /* Wait for channel inactive */
 601         for (timeout = 1000; timeout; timeout--) {
 602                 if (!pl08x_phy_channel_busy(ch))
 603                         break;
 604                 udelay(1);
 605         }
 606         if (pl08x_phy_channel_busy(ch))
 607                 pr_err("pl08x: channel%u timeout waiting for pause\n", ch->id);
 608 }
 609
 610 static void pl08x_resume_phy_chan(struct pl08x_phy_chan *ch)
 611 {
 612         u32 val;
 613
 614         /* Use the enable bit on the FTDMAC020 */
 615         if (ch->ftdmac020) {
 616                 val = readl(ch->reg_control);
 617                 val |= FTDMAC020_CH_CSR_EN;
 618                 writel(val, ch->reg_control);
 619                 return;
 620         }
 621
 622         /* Clear the HALT bit */
 623         val = readl(ch->reg_config);
 624         val &= ~PL080_CONFIG_HALT;
 625         writel(val, ch->reg_config);
 626 }
 627
 628 /*
 629  * pl08x_terminate_phy_chan() stops the channel, clears the FIFO and
 630  * clears any pending interrupt status.  This should not be used for
 631  * an on-going transfer, but as a method of shutting down a channel
 632  * (eg, when it's no longer used) or terminating a transfer.
 633  */
 634 static void pl08x_terminate_phy_chan(struct pl08x_driver_data *pl08x,
 635         struct pl08x_phy_chan *ch)
 636 {
 637         u32 val;
 638
 639         /* The layout for the FTDMAC020 is different */
 640         if (ch->ftdmac020) {
 641                 /* Disable all interrupts */
 642                 val = readl(ch->reg_config);
 643                 val |= (FTDMAC020_CH_CFG_INT_ABT_MASK |
 644                         FTDMAC020_CH_CFG_INT_ERR_MASK |
 645                         FTDMAC020_CH_CFG_INT_TC_MASK);
 646                 writel(val, ch->reg_config);
 647
 648                 /* Abort and disable channel */
 649                 val = readl(ch->reg_control);
 650                 val &= ~FTDMAC020_CH_CSR_EN;
 651                 val |= FTDMAC020_CH_CSR_ABT;
 652                 writel(val, ch->reg_control);
 653
 654                 /* Clear ABT and ERR interrupt flags */
 655                 writel(BIT(ch->id) | BIT(ch->id + 16),
 656                        pl08x->base + PL080_ERR_CLEAR);
 657                 writel(BIT(ch->id), pl08x->base + PL080_TC_CLEAR);
 658
 659                 return;
 660         }
 661
 662         val = readl(ch->reg_config);
 663         val &= ~(PL080_CONFIG_ENABLE | PL080_CONFIG_ERR_IRQ_MASK |
 664                  PL080_CONFIG_TC_IRQ_MASK);
 665         writel(val, ch->reg_config);
 666
 667         writel(BIT(ch->id), pl08x->base + PL080_ERR_CLEAR);
 668         writel(BIT(ch->id), pl08x->base + PL080_TC_CLEAR);
 669 }
 670
 671 static u32 get_bytes_in_phy_channel(struct pl08x_phy_chan *ch)
 672 {
 673         u32 val;
 674         u32 bytes;
 675
 676         if (ch->ftdmac020) {
 677                 bytes = readl(ch->base + FTDMAC020_CH_SIZE);
 678
 679                 val = readl(ch->reg_control);
 680                 val &= FTDMAC020_CH_CSR_SRC_WIDTH_MSK;
 681                 val >>= FTDMAC020_CH_CSR_SRC_WIDTH_SHIFT;
 682         } else if (ch->pl080s) {
 683                 val = readl(ch->base + PL080S_CH_CONTROL2);
 684                 bytes = val & PL080S_CONTROL_TRANSFER_SIZE_MASK;
 685
 686                 val = readl(ch->reg_control);
 687                 val &= PL080_CONTROL_SWIDTH_MASK;
 688                 val >>= PL080_CONTROL_SWIDTH_SHIFT;
 689         } else {
 690                 /* Plain PL08x */
 691                 val = readl(ch->reg_control);
 692                 bytes = val & PL080_CONTROL_TRANSFER_SIZE_MASK;
 693
 694                 val &= PL080_CONTROL_SWIDTH_MASK;
 695                 val >>= PL080_CONTROL_SWIDTH_SHIFT;
 696         }
 697
 698         switch (val) {
 699         case PL080_WIDTH_8BIT:
 700                 break;
 701         case PL080_WIDTH_16BIT:
 702                 bytes *= 2;
 703                 break;
 704         case PL080_WIDTH_32BIT:
 705                 bytes *= 4;
 706                 break;
 707         }
 708         return bytes;
 709 }
 710
 711 static u32 get_bytes_in_lli(struct pl08x_phy_chan *ch, const u32 *llis_va)
 712 {
 713         u32 val;
 714         u32 bytes;
 715
 716         if (ch->ftdmac020) {
 717                 val = llis_va[PL080_LLI_CCTL];
 718                 bytes = val & FTDMAC020_LLI_TRANSFER_SIZE_MASK;
 719
 720                 val = llis_va[PL080_LLI_CCTL];
 721                 val &= FTDMAC020_LLI_SRC_WIDTH_MSK;
 722                 val >>= FTDMAC020_LLI_SRC_WIDTH_SHIFT;
 723         } else if (ch->pl080s) {
 724                 val = llis_va[PL080S_LLI_CCTL2];
 725                 bytes = val & PL080S_CONTROL_TRANSFER_SIZE_MASK;
 726
 727                 val = llis_va[PL080_LLI_CCTL];
 728                 val &= PL080_CONTROL_SWIDTH_MASK;
 729                 val >>= PL080_CONTROL_SWIDTH_SHIFT;
 730         } else {
 731                 /* Plain PL08x */
 732                 val = llis_va[PL080_LLI_CCTL];
 733                 bytes = val & PL080_CONTROL_TRANSFER_SIZE_MASK;
 734
 735                 val &= PL080_CONTROL_SWIDTH_MASK;
 736                 val >>= PL080_CONTROL_SWIDTH_SHIFT;
 737         }
 738
 739         switch (val) {
 740         case PL080_WIDTH_8BIT:
 741                 break;
 742         case PL080_WIDTH_16BIT:
 743                 bytes *= 2;
 744                 break;
 745         case PL080_WIDTH_32BIT:
 746                 bytes *= 4;
 747                 break;
 748         }
 749         return bytes;
 750 }
 751
 752 /* The channel should be paused when calling this */
 753 static u32 pl08x_getbytes_chan(struct pl08x_dma_chan *plchan)
 754 {
 755         struct pl08x_driver_data *pl08x = plchan->host;
 756         const u32 *llis_va, *llis_va_limit;
 757         struct pl08x_phy_chan *ch;
 758         dma_addr_t llis_bus;
 759         struct pl08x_txd *txd;
 760         u32 llis_max_words;
 761         size_t bytes;
 762         u32 clli;
 763
 764         ch = plchan->phychan;
 765         txd = plchan->at;
 766
 767         if (!ch || !txd)
 768                 return 0;
 769
 770         /*
 771          * Follow the LLIs to get the number of remaining
 772          * bytes in the currently active transaction.
 773          */
 774         clli = readl(ch->reg_lli) & ~PL080_LLI_LM_AHB2;
 775
 776         /* First get the remaining bytes in the active transfer */
 777         bytes = get_bytes_in_phy_channel(ch);
 778
 779         if (!clli)
 780                 return bytes;
 781
 782         llis_va = txd->llis_va;
 783         llis_bus = txd->llis_bus;
 784
 785         llis_max_words = pl08x->lli_words * MAX_NUM_TSFR_LLIS;
 786         BUG_ON(clli < llis_bus || clli >= llis_bus +
 787                                                 sizeof(u32) * llis_max_words);
 788
 789         /*
 790          * Locate the next LLI - as this is an array,
 791          * it's simple maths to find.
 792          */
 793         llis_va += (clli - llis_bus) / sizeof(u32);
 794
 795         llis_va_limit = llis_va + llis_max_words;
 796
 797         for (; llis_va < llis_va_limit; llis_va += pl08x->lli_words) {
 798                 bytes += get_bytes_in_lli(ch, llis_va);
 799
 800                 /*
 801                  * A LLI pointer going backward terminates the LLI list
 802                  */
 803                 if (llis_va[PL080_LLI_LLI] <= clli)
 804                         break;
 805         }
 806
 807         return bytes;
 808 }
 809
 810 /*
 811  * Allocate a physical channel for a virtual channel
 812  *
 813  * Try to locate a physical channel to be used for this transfer. If all
 814  * are taken return NULL and the requester will have to cope by using
 815  * some fallback PIO mode or retrying later.
 816  */
 817 static struct pl08x_phy_chan *
 818 pl08x_get_phy_channel(struct pl08x_driver_data *pl08x,
 819                       struct pl08x_dma_chan *virt_chan)
 820 {
 821         struct pl08x_phy_chan *ch = NULL;
 822         unsigned long flags;
 823         int i;
 824
 825         for (i = 0; i < pl08x->vd->channels; i++) {
 826                 ch = &pl08x->phy_chans[i];
 827
 828                 spin_lock_irqsave(&ch->lock, flags);
 829
 830                 if (!ch->locked && !ch->serving) {
 831                         ch->serving = virt_chan;
 832                         spin_unlock_irqrestore(&ch->lock, flags);
 833                         break;
 834                 }
 835
 836                 spin_unlock_irqrestore(&ch->lock, flags);
 837         }
 838
 839         if (i == pl08x->vd->channels) {
 840                 /* No physical channel available, cope with it */
 841                 return NULL;
 842         }
 843
 844         return ch;
 845 }
 846
 847 /* Mark the physical channel as free.  Note, this write is atomic. */
 848 static inline void pl08x_put_phy_channel(struct pl08x_driver_data *pl08x,
 849                                          struct pl08x_phy_chan *ch)
 850 {
 851         ch->serving = NULL;
 852 }
 853
 854 /*
 855  * Try to allocate a physical channel.  When successful, assign it to
 856  * this virtual channel, and initiate the next descriptor.  The
 857  * virtual channel lock must be held at this point.
 858  */
 859 static void pl08x_phy_alloc_and_start(struct pl08x_dma_chan *plchan)
 860 {
 861         struct pl08x_driver_data *pl08x = plchan->host;
 862         struct pl08x_phy_chan *ch;
 863
 864         ch = pl08x_get_phy_channel(pl08x, plchan);
 865         if (!ch) {
 866                 dev_dbg(&pl08x->adev->dev, "no physical channel available for xfer on %s\n", plchan->name);
 867                 plchan->state = PL08X_CHAN_WAITING;
 868                 plchan->waiting_at = jiffies;
 869                 return;
 870         }
 871
 872         dev_dbg(&pl08x->adev->dev, "allocated physical channel %d for xfer on %s\n",
 873                 ch->id, plchan->name);
 874
 875         plchan->phychan = ch;
 876         plchan->state = PL08X_CHAN_RUNNING;
 877         pl08x_start_next_txd(plchan);
 878 }
 879
 880 static void pl08x_phy_reassign_start(struct pl08x_phy_chan *ch,
 881         struct pl08x_dma_chan *plchan)
 882 {
 883         struct pl08x_driver_data *pl08x = plchan->host;
 884
 885         dev_dbg(&pl08x->adev->dev, "reassigned physical channel %d for xfer on %s\n",
 886                 ch->id, plchan->name);
 887
 888         /*
 889          * We do this without taking the lock; we're really only concerned
 890          * about whether this pointer is NULL or not, and we're guaranteed
 891          * that this will only be called when it _already_ is non-NULL.
 892          */
 893         ch->serving = plchan;
 894         plchan->phychan = ch;
 895         plchan->state = PL08X_CHAN_RUNNING;
 896         pl08x_start_next_txd(plchan);
 897 }
 898
 899 /*
 900  * Free a physical DMA channel, potentially reallocating it to another
 901  * virtual channel if we have any pending.
 902  */
 903 static void pl08x_phy_free(struct pl08x_dma_chan *plchan)
 904 {
 905         struct pl08x_driver_data *pl08x = plchan->host;
 906         struct pl08x_dma_chan *p, *next;
 907         unsigned long waiting_at;
 908  retry:
 909         next = NULL;
 910         waiting_at = jiffies;
 911
 912         /*
 913          * Find a waiting virtual channel for the next transfer.
 914          * To be fair, time when each channel reached waiting state is compared
 915          * to select channel that is waiting for the longest time.
 916          */
 917         list_for_each_entry(p, &pl08x->memcpy.channels, vc.chan.device_node)
 918                 if (p->state == PL08X_CHAN_WAITING &&
 919                     p->waiting_at <= waiting_at) {
 920                         next = p;
 921                         waiting_at = p->waiting_at;
 922                 }
 923
 924         if (!next && pl08x->has_slave) {
 925                 list_for_each_entry(p, &pl08x->slave.channels, vc.chan.device_node)
 926                         if (p->state == PL08X_CHAN_WAITING &&
 927                             p->waiting_at <= waiting_at) {
 928                                 next = p;
 929                                 waiting_at = p->waiting_at;
 930                         }
 931         }
 932
 933         /* Ensure that the physical channel is stopped */
 934         pl08x_terminate_phy_chan(pl08x, plchan->phychan);
 935
 936         if (next) {
 937                 bool success;
 938
 939                 /*
 940                  * Eww.  We know this isn't going to deadlock
 941                  * but lockdep probably doesn't.
 942                  */
 943                 spin_lock(&next->vc.lock);
 944                 /* Re-check the state now that we have the lock */
 945                 success = next->state == PL08X_CHAN_WAITING;
 946                 if (success)
 947                         pl08x_phy_reassign_start(plchan->phychan, next);
 948                 spin_unlock(&next->vc.lock);
 949
 950                 /* If the state changed, try to find another channel */
 951                 if (!success)
 952                         goto retry;
 953         } else {
 954                 /* No more jobs, so free up the physical channel */
 955                 pl08x_put_phy_channel(pl08x, plchan->phychan);
 956         }
 957
 958         plchan->phychan = NULL;
 959         plchan->state = PL08X_CHAN_IDLE;
 960 }
 961
 962 /*
 963  * LLI handling
 964  */
 965
 966 static inline unsigned int
 967 pl08x_get_bytes_for_lli(struct pl08x_driver_data *pl08x,
 968                         u32 cctl,
 969                         bool source)
 970 {
 971         u32 val;
 972
 973         if (pl08x->vd->ftdmac020) {
 974                 if (source)
 975                         val = (cctl & FTDMAC020_LLI_SRC_WIDTH_MSK) >>
 976                                 FTDMAC020_LLI_SRC_WIDTH_SHIFT;
 977                 else
 978                         val = (cctl & FTDMAC020_LLI_DST_WIDTH_MSK) >>
 979                                 FTDMAC020_LLI_DST_WIDTH_SHIFT;
 980         } else {
 981                 if (source)
 982                         val = (cctl & PL080_CONTROL_SWIDTH_MASK) >>
 983                                 PL080_CONTROL_SWIDTH_SHIFT;
 984                 else
 985                         val = (cctl & PL080_CONTROL_DWIDTH_MASK) >>
 986                                 PL080_CONTROL_DWIDTH_SHIFT;
 987         }
 988
 989         switch (val) {
 990         case PL080_WIDTH_8BIT:
 991                 return 1;
 992         case PL080_WIDTH_16BIT:
 993                 return 2;
 994         case PL080_WIDTH_32BIT:
 995                 return 4;
 996         default:
 997                 break;
 998         }
 999         BUG();
1000         return 0;
1001 }
1002
1003 static inline u32 pl08x_lli_control_bits(struct pl08x_driver_data *pl08x,
1004                                          u32 cctl,
1005                                          u8 srcwidth, u8 dstwidth,
1006                                          size_t tsize)
1007 {
1008         u32 retbits = cctl;
1009
1010         /*
1011          * Remove all src, dst and transfer size bits, then set the
1012          * width and size according to the parameters. The bit offsets
1013          * are different in the FTDMAC020 so we need to accound for this.
1014          */
1015         if (pl08x->vd->ftdmac020) {
1016                 retbits &= ~FTDMAC020_LLI_DST_WIDTH_MSK;
1017                 retbits &= ~FTDMAC020_LLI_SRC_WIDTH_MSK;
1018                 retbits &= ~FTDMAC020_LLI_TRANSFER_SIZE_MASK;
1019
1020                 switch (srcwidth) {
1021                 case 1:
1022                         retbits |= PL080_WIDTH_8BIT <<
1023                                 FTDMAC020_LLI_SRC_WIDTH_SHIFT;
1024                         break;
1025                 case 2:
1026                         retbits |= PL080_WIDTH_16BIT <<
1027                                 FTDMAC020_LLI_SRC_WIDTH_SHIFT;
1028                         break;
1029                 case 4:
1030                         retbits |= PL080_WIDTH_32BIT <<
1031                                 FTDMAC020_LLI_SRC_WIDTH_SHIFT;
1032                         break;
1033                 default:
1034                         BUG();
1035                         break;
1036                 }
1037
1038                 switch (dstwidth) {
1039                 case 1:
1040                         retbits |= PL080_WIDTH_8BIT <<
1041                                 FTDMAC020_LLI_DST_WIDTH_SHIFT;
1042                         break;
1043                 case 2:
1044                         retbits |= PL080_WIDTH_16BIT <<
1045                                 FTDMAC020_LLI_DST_WIDTH_SHIFT;
1046                         break;
1047                 case 4:
1048                         retbits |= PL080_WIDTH_32BIT <<
1049                                 FTDMAC020_LLI_DST_WIDTH_SHIFT;
1050                         break;
1051                 default:
1052                         BUG();
1053                         break;
1054                 }
1055
1056                 tsize &= FTDMAC020_LLI_TRANSFER_SIZE_MASK;
1057                 retbits |= tsize << FTDMAC020_LLI_TRANSFER_SIZE_SHIFT;
1058         } else {
1059                 retbits &= ~PL080_CONTROL_DWIDTH_MASK;
1060                 retbits &= ~PL080_CONTROL_SWIDTH_MASK;
1061                 retbits &= ~PL080_CONTROL_TRANSFER_SIZE_MASK;
1062
1063                 switch (srcwidth) {
1064                 case 1:
1065                         retbits |= PL080_WIDTH_8BIT <<
1066                                 PL080_CONTROL_SWIDTH_SHIFT;
1067                         break;
1068                 case 2:
1069                         retbits |= PL080_WIDTH_16BIT <<
1070                                 PL080_CONTROL_SWIDTH_SHIFT;
1071                         break;
1072                 case 4:
1073                         retbits |= PL080_WIDTH_32BIT <<
1074                                 PL080_CONTROL_SWIDTH_SHIFT;
1075                         break;
1076                 default:
1077                         BUG();
1078                         break;
1079                 }
1080
1081                 switch (dstwidth) {
1082                 case 1:
1083                         retbits |= PL080_WIDTH_8BIT <<
1084                                 PL080_CONTROL_DWIDTH_SHIFT;
1085                         break;
1086                 case 2:
1087                         retbits |= PL080_WIDTH_16BIT <<
1088                                 PL080_CONTROL_DWIDTH_SHIFT;
1089                         break;
1090                 case 4:
1091                         retbits |= PL080_WIDTH_32BIT <<
1092                                 PL080_CONTROL_DWIDTH_SHIFT;
1093                         break;
1094                 default:
1095                         BUG();
1096                         break;
1097                 }
1098
1099                 tsize &= PL080_CONTROL_TRANSFER_SIZE_MASK;
1100                 retbits |= tsize << PL080_CONTROL_TRANSFER_SIZE_SHIFT;
1101         }
1102
1103         return retbits;
1104 }
1105
1106 struct pl08x_lli_build_data {
1107         struct pl08x_txd *txd;
1108         struct pl08x_bus_data srcbus;
1109         struct pl08x_bus_data dstbus;
1110         size_t remainder;
1111         u32 lli_bus;
1112 };
1113
1114 /*
1115  * Autoselect a master bus to use for the transfer. Slave will be the chosen as
1116  * victim in case src & dest are not similarly aligned. i.e. If after aligning
1117  * masters address with width requirements of transfer (by sending few byte by
1118  * byte data), slave is still not aligned, then its width will be reduced to
1119  * BYTE.
1120  * - prefers the destination bus if both available
1121  * - prefers bus with fixed address (i.e. peripheral)
1122  */
1123 static void pl08x_choose_master_bus(struct pl08x_driver_data *pl08x,
1124                                     struct pl08x_lli_build_data *bd,
1125                                     struct pl08x_bus_data **mbus,
1126                                     struct pl08x_bus_data **sbus,
1127                                     u32 cctl)
1128 {
1129         bool dst_incr;
1130         bool src_incr;
1131
1132         /*
1133          * The FTDMAC020 only supports memory-to-memory transfer, so
1134          * source and destination always increase.
1135          */
1136         if (pl08x->vd->ftdmac020) {
1137                 dst_incr = true;
1138                 src_incr = true;
1139         } else {
1140                 dst_incr = !!(cctl & PL080_CONTROL_DST_INCR);
1141                 src_incr = !!(cctl & PL080_CONTROL_SRC_INCR);
1142         }
1143
1144         /*
1145          * If either bus is not advancing, i.e. it is a peripheral, that
1146          * one becomes master
1147          */
1148         if (!dst_incr) {
1149                 *mbus = &bd->dstbus;
1150                 *sbus = &bd->srcbus;
1151         } else if (!src_incr) {
1152                 *mbus = &bd->srcbus;
1153                 *sbus = &bd->dstbus;
1154         } else {
1155                 if (bd->dstbus.buswidth >= bd->srcbus.buswidth) {
1156                         *mbus = &bd->dstbus;
1157                         *sbus = &bd->srcbus;
1158                 } else {
1159                         *mbus = &bd->srcbus;
1160                         *sbus = &bd->dstbus;
1161                 }
1162         }
1163 }
1164
1165 /*
1166  * Fills in one LLI for a certain transfer descriptor and advance the counter
1167  */
1168 static void pl08x_fill_lli_for_desc(struct pl08x_driver_data *pl08x,
1169                                     struct pl08x_lli_build_data *bd,
1170                                     int num_llis, int len, u32 cctl, u32 cctl2)
1171 {
1172         u32 offset = num_llis * pl08x->lli_words;
1173         u32 *llis_va = bd->txd->llis_va + offset;
1174         dma_addr_t llis_bus = bd->txd->llis_bus;
1175
1176         BUG_ON(num_llis >= MAX_NUM_TSFR_LLIS);
1177
1178         /* Advance the offset to next LLI. */
1179         offset += pl08x->lli_words;
1180
1181         llis_va[PL080_LLI_SRC] = bd->srcbus.addr;
1182         llis_va[PL080_LLI_DST] = bd->dstbus.addr;
1183         llis_va[PL080_LLI_LLI] = (llis_bus + sizeof(u32) * offset);
1184         llis_va[PL080_LLI_LLI] |= bd->lli_bus;
1185         llis_va[PL080_LLI_CCTL] = cctl;
1186         if (pl08x->vd->pl080s)
1187                 llis_va[PL080S_LLI_CCTL2] = cctl2;
1188
1189         if (pl08x->vd->ftdmac020) {
1190                 /* FIXME: only memcpy so far so both increase */
1191                 bd->srcbus.addr += len;
1192                 bd->dstbus.addr += len;
1193         } else {
1194                 if (cctl & PL080_CONTROL_SRC_INCR)
1195                         bd->srcbus.addr += len;
1196                 if (cctl & PL080_CONTROL_DST_INCR)
1197                         bd->dstbus.addr += len;
1198         }
1199
1200         BUG_ON(bd->remainder < len);
1201
1202         bd->remainder -= len;
1203 }
1204
1205 static inline void prep_byte_width_lli(struct pl08x_driver_data *pl08x,
1206                         struct pl08x_lli_build_data *bd, u32 *cctl, u32 len,
1207                         int num_llis, size_t *total_bytes)
1208 {
1209         *cctl = pl08x_lli_control_bits(pl08x, *cctl, 1, 1, len);
1210         pl08x_fill_lli_for_desc(pl08x, bd, num_llis, len, *cctl, len);
1211         (*total_bytes) += len;
1212 }
1213
1214 #if 1
1215 static void pl08x_dump_lli(struct pl08x_driver_data *pl08x,
1216                            const u32 *llis_va, int num_llis)
1217 {
1218         int i;
1219
1220         if (pl08x->vd->pl080s) {
1221                 dev_vdbg(&pl08x->adev->dev,
1222                         "%-3s %-9s  %-10s %-10s %-10s %-10s %s\n",
1223                         "lli", "", "csrc", "cdst", "clli", "cctl", "cctl2");
1224                 for (i = 0; i < num_llis; i++) {
1225                         dev_vdbg(&pl08x->adev->dev,
1226                                 "%3d @%p: 0x%08x 0x%08x 0x%08x 0x%08x 0x%08x\n",
1227                                 i, llis_va, llis_va[PL080_LLI_SRC],
1228                                 llis_va[PL080_LLI_DST], llis_va[PL080_LLI_LLI],
1229                                 llis_va[PL080_LLI_CCTL],
1230                                 llis_va[PL080S_LLI_CCTL2]);
1231                         llis_va += pl08x->lli_words;
1232                 }
1233         } else {
1234                 dev_vdbg(&pl08x->adev->dev,
1235                         "%-3s %-9s  %-10s %-10s %-10s %s\n",
1236                         "lli", "", "csrc", "cdst", "clli", "cctl");
1237                 for (i = 0; i < num_llis; i++) {
1238                         dev_vdbg(&pl08x->adev->dev,
1239                                 "%3d @%p: 0x%08x 0x%08x 0x%08x 0x%08x\n",
1240                                 i, llis_va, llis_va[PL080_LLI_SRC],
1241                                 llis_va[PL080_LLI_DST], llis_va[PL080_LLI_LLI],
1242                                 llis_va[PL080_LLI_CCTL]);
1243                         llis_va += pl08x->lli_words;
1244                 }
1245         }
1246 }
1247 #else
1248 static inline void pl08x_dump_lli(struct pl08x_driver_data *pl08x,
1249                                   const u32 *llis_va, int num_llis) {}
1250 #endif
1251
1252 /*
1253  * This fills in the table of LLIs for the transfer descriptor
1254  * Note that we assume we never have to change the burst sizes
1255  * Return 0 for error
1256  */
1257 static int pl08x_fill_llis_for_desc(struct pl08x_driver_data *pl08x,
1258                               struct pl08x_txd *txd)
1259 {
1260         struct pl08x_bus_data *mbus, *sbus;
1261         struct pl08x_lli_build_data bd;
1262         int num_llis = 0;
1263         u32 cctl, early_bytes = 0;
1264         size_t max_bytes_per_lli, total_bytes;
1265         u32 *llis_va, *last_lli;
1266         struct pl08x_sg *dsg;
1267
1268         txd->llis_va = dma_pool_alloc(pl08x->pool, GFP_NOWAIT, &txd->llis_bus);
1269         if (!txd->llis_va) {
1270                 dev_err(&pl08x->adev->dev, "%s no memory for llis\n", __func__);
1271                 return 0;
1272         }
1273
1274         bd.txd = txd;
1275         bd.lli_bus = (pl08x->lli_buses & PL08X_AHB2) ? PL080_LLI_LM_AHB2 : 0;
1276         cctl = txd->cctl;
1277
1278         /* Find maximum width of the source bus */
1279         bd.srcbus.maxwidth = pl08x_get_bytes_for_lli(pl08x, cctl, true);
1280
1281         /* Find maximum width of the destination bus */
1282         bd.dstbus.maxwidth = pl08x_get_bytes_for_lli(pl08x, cctl, false);
1283
1284         list_for_each_entry(dsg, &txd->dsg_list, node) {
1285                 total_bytes = 0;
1286                 cctl = txd->cctl;
1287
1288                 bd.srcbus.addr = dsg->src_addr;
1289                 bd.dstbus.addr = dsg->dst_addr;
1290                 bd.remainder = dsg->len;
1291                 bd.srcbus.buswidth = bd.srcbus.maxwidth;
1292                 bd.dstbus.buswidth = bd.dstbus.maxwidth;
1293
1294                 pl08x_choose_master_bus(pl08x, &bd, &mbus, &sbus, cctl);
1295
1296                 dev_vdbg(&pl08x->adev->dev,
1297                         "src=0x%08llx%s/%u dst=0x%08llx%s/%u len=%zu\n",
1298                         (u64)bd.srcbus.addr,
1299                         cctl & PL080_CONTROL_SRC_INCR ? "+" : "",
1300                         bd.srcbus.buswidth,
1301                         (u64)bd.dstbus.addr,
1302                         cctl & PL080_CONTROL_DST_INCR ? "+" : "",
1303                         bd.dstbus.buswidth,
1304                         bd.remainder);
1305                 dev_vdbg(&pl08x->adev->dev, "mbus=%s sbus=%s\n",
1306                         mbus == &bd.srcbus ? "src" : "dst",
1307                         sbus == &bd.srcbus ? "src" : "dst");
1308
1309                 /*
1310                  * Zero length is only allowed if all these requirements are
1311                  * met:
1312                  * - flow controller is peripheral.
1313                  * - src.addr is aligned to src.width
1314                  * - dst.addr is aligned to dst.width
1315                  *
1316                  * sg_len == 1 should be true, as there can be two cases here:
1317                  *
1318                  * - Memory addresses are contiguous and are not scattered.
1319                  *   Here, Only one sg will be passed by user driver, with
1320                  *   memory address and zero length. We pass this to controller
1321                  *   and after the transfer it will receive the last burst
1322                  *   request from peripheral and so transfer finishes.
1323                  *
1324                  * - Memory addresses are scattered and are not contiguous.
1325                  *   Here, Obviously as DMA controller doesn't know when a lli's
1326                  *   transfer gets over, it can't load next lli. So in this
1327                  *   case, there has to be an assumption that only one lli is
1328                  *   supported. Thus, we can't have scattered addresses.
1329                  */
1330                 if (!bd.remainder) {
1331                         u32 fc;
1332
1333                         /* FTDMAC020 only does memory-to-memory */
1334                         if (pl08x->vd->ftdmac020)
1335                                 fc = PL080_FLOW_MEM2MEM;
1336                         else
1337                                 fc = (txd->ccfg & PL080_CONFIG_FLOW_CONTROL_MASK) >>
1338                                         PL080_CONFIG_FLOW_CONTROL_SHIFT;
1339                         if (!((fc >= PL080_FLOW_SRC2DST_DST) &&
1340                                         (fc <= PL080_FLOW_SRC2DST_SRC))) {
1341                                 dev_err(&pl08x->adev->dev, "%s sg len can't be zero",
1342                                         __func__);
1343                                 return 0;
1344                         }
1345
1346                         if (!IS_BUS_ALIGNED(&bd.srcbus) ||
1347                                 !IS_BUS_ALIGNED(&bd.dstbus)) {
1348                                 dev_err(&pl08x->adev->dev,
1349                                         "%s src & dst address must be aligned to src"
1350                                         " & dst width if peripheral is flow controller",
1351                                         __func__);
1352                                 return 0;
1353                         }
1354
1355                         cctl = pl08x_lli_control_bits(pl08x, cctl,
1356                                         bd.srcbus.buswidth, bd.dstbus.buswidth,
1357                                         0);
1358                         pl08x_fill_lli_for_desc(pl08x, &bd, num_llis++,
1359                                         0, cctl, 0);
1360                         break;
1361                 }
1362
1363                 /*
1364                  * Send byte by byte for following cases
1365                  * - Less than a bus width available
1366                  * - until master bus is aligned
1367                  */
1368                 if (bd.remainder < mbus->buswidth)
1369                         early_bytes = bd.remainder;
1370                 else if (!IS_BUS_ALIGNED(mbus)) {
1371                         early_bytes = mbus->buswidth -
1372                                 (mbus->addr & (mbus->buswidth - 1));
1373                         if ((bd.remainder - early_bytes) < mbus->buswidth)
1374                                 early_bytes = bd.remainder;
1375                 }
1376
1377                 if (early_bytes) {
1378                         dev_vdbg(&pl08x->adev->dev,
1379                                 "%s byte width LLIs (remain 0x%08zx)\n",
1380                                 __func__, bd.remainder);
1381                         prep_byte_width_lli(pl08x, &bd, &cctl, early_bytes,
1382                                 num_llis++, &total_bytes);
1383                 }
1384
1385                 if (bd.remainder) {
1386                         /*
1387                          * Master now aligned
1388                          * - if slave is not then we must set its width down
1389                          */
1390                         if (!IS_BUS_ALIGNED(sbus)) {
1391                                 dev_dbg(&pl08x->adev->dev,
1392                                         "%s set down bus width to one byte\n",
1393                                         __func__);
1394
1395                                 sbus->buswidth = 1;
1396                         }
1397
1398                         /*
1399                          * Bytes transferred = tsize * src width, not
1400                          * MIN(buswidths)
1401                          */
1402                         max_bytes_per_lli = bd.srcbus.buswidth *
1403                                                 pl08x->vd->max_transfer_size;
1404                         dev_vdbg(&pl08x->adev->dev,
1405                                 "%s max bytes per lli = %zu\n",
1406                                 __func__, max_bytes_per_lli);
1407
1408                         /*
1409                          * Make largest possible LLIs until less than one bus
1410                          * width left
1411                          */
1412                         while (bd.remainder > (mbus->buswidth - 1)) {
1413                                 size_t lli_len, tsize, width;
1414
1415                                 /*
1416                                  * If enough left try to send max possible,
1417                                  * otherwise try to send the remainder
1418                                  */
1419                                 lli_len = min(bd.remainder, max_bytes_per_lli);
1420
1421                                 /*
1422                                  * Check against maximum bus alignment:
1423                                  * Calculate actual transfer size in relation to
1424                                  * bus width an get a maximum remainder of the
1425                                  * highest bus width - 1
1426                                  */
1427                                 width = max(mbus->buswidth, sbus->buswidth);
1428                                 lli_len = (lli_len / width) * width;
1429                                 tsize = lli_len / bd.srcbus.buswidth;
1430
1431                                 dev_vdbg(&pl08x->adev->dev,
1432                                         "%s fill lli with single lli chunk of "
1433                                         "size 0x%08zx (remainder 0x%08zx)\n",
1434                                         __func__, lli_len, bd.remainder);
1435
1436                                 cctl = pl08x_lli_control_bits(pl08x, cctl,
1437                                         bd.srcbus.buswidth, bd.dstbus.buswidth,
1438                                         tsize);
1439                                 pl08x_fill_lli_for_desc(pl08x, &bd, num_llis++,
1440                                                 lli_len, cctl, tsize);
1441                                 total_bytes += lli_len;
1442                         }
1443
1444                         /*
1445                          * Send any odd bytes
1446                          */
1447                         if (bd.remainder) {
1448                                 dev_vdbg(&pl08x->adev->dev,
1449                                         "%s align with boundary, send odd bytes (remain %zu)\n",
1450                                         __func__, bd.remainder);
1451                                 prep_byte_width_lli(pl08x, &bd, &cctl,
1452                                         bd.remainder, num_llis++, &total_bytes);
1453                         }
1454                 }
1455
1456                 if (total_bytes != dsg->len) {
1457                         dev_err(&pl08x->adev->dev,
1458                                 "%s size of encoded lli:s don't match total txd, transferred 0x%08zx from size 0x%08zx\n",
1459                                 __func__, total_bytes, dsg->len);
1460                         return 0;
1461                 }
1462
1463                 if (num_llis >= MAX_NUM_TSFR_LLIS) {
1464                         dev_err(&pl08x->adev->dev,
1465                                 "%s need to increase MAX_NUM_TSFR_LLIS from 0x%08x\n",
1466                                 __func__, MAX_NUM_TSFR_LLIS);
1467                         return 0;
1468                 }
1469         }
1470
1471         llis_va = txd->llis_va;
1472         last_lli = llis_va + (num_llis - 1) * pl08x->lli_words;
1473
1474         if (txd->cyclic) {
1475                 /* Link back to the first LLI. */
1476                 last_lli[PL080_LLI_LLI] = txd->llis_bus | bd.lli_bus;
1477         } else {
1478                 /* The final LLI terminates the LLI. */
1479                 last_lli[PL080_LLI_LLI] = 0;
1480                 /* The final LLI element shall also fire an interrupt. */
1481                 if (pl08x->vd->ftdmac020)
1482                         last_lli[PL080_LLI_CCTL] &= ~FTDMAC020_LLI_TC_MSK;
1483                 else
1484                         last_lli[PL080_LLI_CCTL] |= PL080_CONTROL_TC_IRQ_EN;
1485         }
1486
1487         pl08x_dump_lli(pl08x, llis_va, num_llis);
1488
1489         return num_llis;
1490 }
1491
1492 static void pl08x_free_txd(struct pl08x_driver_data *pl08x,
1493                            struct pl08x_txd *txd)
1494 {
1495         struct pl08x_sg *dsg, *_dsg;
1496
1497         if (txd->llis_va)
1498                 dma_pool_free(pl08x->pool, txd->llis_va, txd->llis_bus);
1499
1500         list_for_each_entry_safe(dsg, _dsg, &txd->dsg_list, node) {
1501                 list_del(&dsg->node);
1502                 kfree(dsg);
1503         }
1504
1505         kfree(txd);
1506 }
1507
1508 static void pl08x_desc_free(struct virt_dma_desc *vd)
1509 {
1510         struct pl08x_txd *txd = to_pl08x_txd(&vd->tx);
1511         struct pl08x_dma_chan *plchan = to_pl08x_chan(vd->tx.chan);
1512
1513         dma_descriptor_unmap(&vd->tx);
1514         if (!txd->done)
1515                 pl08x_release_mux(plchan);
1516
1517         pl08x_free_txd(plchan->host, txd);
1518 }
1519
1520 static void pl08x_free_txd_list(struct pl08x_driver_data *pl08x,
1521                                 struct pl08x_dma_chan *plchan)
1522 {
1523         LIST_HEAD(head);
1524
1525         vchan_get_all_descriptors(&plchan->vc, &head);
1526         vchan_dma_desc_free_list(&plchan->vc, &head);
1527 }
1528
1529 /*
1530  * The DMA ENGINE API
1531  */
1532 static void pl08x_free_chan_resources(struct dma_chan *chan)
1533 {
1534         /* Ensure all queued descriptors are freed */
1535         vchan_free_chan_resources(to_virt_chan(chan));
1536 }
1537
1538 static struct dma_async_tx_descriptor *pl08x_prep_dma_interrupt(
1539                 struct dma_chan *chan, unsigned long flags)
1540 {
1541         struct dma_async_tx_descriptor *retval = NULL;
1542
1543         return retval;
1544 }
1545
1546 /*
1547  * Code accessing dma_async_is_complete() in a tight loop may give problems.
1548  * If slaves are relying on interrupts to signal completion this function
1549  * must not be called with interrupts disabled.
1550  */
1551 static enum dma_status pl08x_dma_tx_status(struct dma_chan *chan,
1552                 dma_cookie_t cookie, struct dma_tx_state *txstate)
1553 {
1554         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1555         struct virt_dma_desc *vd;
1556         unsigned long flags;
1557         enum dma_status ret;
1558         size_t bytes = 0;
1559
1560         ret = dma_cookie_status(chan, cookie, txstate);
1561         if (ret == DMA_COMPLETE)
1562                 return ret;
1563
1564         /*
1565          * There's no point calculating the residue if there's
1566          * no txstate to store the value.
1567          */
1568         if (!txstate) {
1569                 if (plchan->state == PL08X_CHAN_PAUSED)
1570                         ret = DMA_PAUSED;
1571                 return ret;
1572         }
1573
1574         spin_lock_irqsave(&plchan->vc.lock, flags);
1575         ret = dma_cookie_status(chan, cookie, txstate);
1576         if (ret != DMA_COMPLETE) {
1577                 vd = vchan_find_desc(&plchan->vc, cookie);
1578                 if (vd) {
1579                         /* On the issued list, so hasn't been processed yet */
1580                         struct pl08x_txd *txd = to_pl08x_txd(&vd->tx);
1581                         struct pl08x_sg *dsg;
1582
1583                         list_for_each_entry(dsg, &txd->dsg_list, node)
1584                                 bytes += dsg->len;
1585                 } else {
1586                         bytes = pl08x_getbytes_chan(plchan);
1587                 }
1588         }
1589         spin_unlock_irqrestore(&plchan->vc.lock, flags);
1590
1591         /*
1592          * This cookie not complete yet
1593          * Get number of bytes left in the active transactions and queue
1594          */
1595         dma_set_residue(txstate, bytes);
1596
1597         if (plchan->state == PL08X_CHAN_PAUSED && ret == DMA_IN_PROGRESS)
1598                 ret = DMA_PAUSED;
1599
1600         /* Whether waiting or running, we're in progress */
1601         return ret;
1602 }
1603
1604 /* PrimeCell DMA extension */
1605 struct burst_table {
1606         u32 burstwords;
1607         u32 reg;
1608 };
1609
1610 static const struct burst_table burst_sizes[] = {
1611         {
1612                 .burstwords = 256,
1613                 .reg = PL080_BSIZE_256,
1614         },
1615         {
1616                 .burstwords = 128,
1617                 .reg = PL080_BSIZE_128,
1618         },
1619         {
1620                 .burstwords = 64,
1621                 .reg = PL080_BSIZE_64,
1622         },
1623         {
1624                 .burstwords = 32,
1625                 .reg = PL080_BSIZE_32,
1626         },
1627         {
1628                 .burstwords = 16,
1629                 .reg = PL080_BSIZE_16,
1630         },
1631         {
1632                 .burstwords = 8,
1633                 .reg = PL080_BSIZE_8,
1634         },
1635         {
1636                 .burstwords = 4,
1637                 .reg = PL080_BSIZE_4,
1638         },
1639         {
1640                 .burstwords = 0,
1641                 .reg = PL080_BSIZE_1,
1642         },
1643 };
1644
1645 /*
1646  * Given the source and destination available bus masks, select which
1647  * will be routed to each port.  We try to have source and destination
1648  * on separate ports, but always respect the allowable settings.
1649  */
1650 static u32 pl08x_select_bus(bool ftdmac020, u8 src, u8 dst)
1651 {
1652         u32 cctl = 0;
1653         u32 dst_ahb2;
1654         u32 src_ahb2;
1655
1656         /* The FTDMAC020 use different bits to indicate src/dst bus */
1657         if (ftdmac020) {
1658                 dst_ahb2 = FTDMAC020_LLI_DST_SEL;
1659                 src_ahb2 = FTDMAC020_LLI_SRC_SEL;
1660         } else {
1661                 dst_ahb2 = PL080_CONTROL_DST_AHB2;
1662                 src_ahb2 = PL080_CONTROL_SRC_AHB2;
1663         }
1664
1665         if (!(dst & PL08X_AHB1) || ((dst & PL08X_AHB2) && (src & PL08X_AHB1)))
1666                 cctl |= dst_ahb2;
1667         if (!(src & PL08X_AHB1) || ((src & PL08X_AHB2) && !(dst & PL08X_AHB2)))
1668                 cctl |= src_ahb2;
1669
1670         return cctl;
1671 }
1672
1673 static u32 pl08x_cctl(u32 cctl)
1674 {
1675         cctl &= ~(PL080_CONTROL_SRC_AHB2 | PL080_CONTROL_DST_AHB2 |
1676                   PL080_CONTROL_SRC_INCR | PL080_CONTROL_DST_INCR |
1677                   PL080_CONTROL_PROT_MASK);
1678
1679         /* Access the cell in privileged mode, non-bufferable, non-cacheable */
1680         return cctl | PL080_CONTROL_PROT_SYS;
1681 }
1682
1683 static u32 pl08x_width(enum dma_slave_buswidth width)
1684 {
1685         switch (width) {
1686         case DMA_SLAVE_BUSWIDTH_1_BYTE:
1687                 return PL080_WIDTH_8BIT;
1688         case DMA_SLAVE_BUSWIDTH_2_BYTES:
1689                 return PL080_WIDTH_16BIT;
1690         case DMA_SLAVE_BUSWIDTH_4_BYTES:
1691                 return PL080_WIDTH_32BIT;
1692         default:
1693                 return ~0;
1694         }
1695 }
1696
1697 static u32 pl08x_burst(u32 maxburst)
1698 {
1699         int i;
1700
1701         for (i = 0; i < ARRAY_SIZE(burst_sizes); i++)
1702                 if (burst_sizes[i].burstwords <= maxburst)
1703                         break;
1704
1705         return burst_sizes[i].reg;
1706 }
1707
1708 static u32 pl08x_get_cctl(struct pl08x_dma_chan *plchan,
1709         enum dma_slave_buswidth addr_width, u32 maxburst)
1710 {
1711         u32 width, burst, cctl = 0;
1712
1713         width = pl08x_width(addr_width);
1714         if (width == ~0)
1715                 return ~0;
1716
1717         cctl |= width << PL080_CONTROL_SWIDTH_SHIFT;
1718         cctl |= width << PL080_CONTROL_DWIDTH_SHIFT;
1719
1720         /*
1721          * If this channel will only request single transfers, set this
1722          * down to ONE element.  Also select one element if no maxburst
1723          * is specified.
1724          */
1725         if (plchan->cd->single)
1726                 maxburst = 1;
1727
1728         burst = pl08x_burst(maxburst);
1729         cctl |= burst << PL080_CONTROL_SB_SIZE_SHIFT;
1730         cctl |= burst << PL080_CONTROL_DB_SIZE_SHIFT;
1731
1732         return pl08x_cctl(cctl);
1733 }
1734
1735 /*
1736  * Slave transactions callback to the slave device to allow
1737  * synchronization of slave DMA signals with the DMAC enable
1738  */
1739 static void pl08x_issue_pending(struct dma_chan *chan)
1740 {
1741         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1742         unsigned long flags;
1743
1744         spin_lock_irqsave(&plchan->vc.lock, flags);
1745         if (vchan_issue_pending(&plchan->vc)) {
1746                 if (!plchan->phychan && plchan->state != PL08X_CHAN_WAITING)
1747                         pl08x_phy_alloc_and_start(plchan);
1748         }
1749         spin_unlock_irqrestore(&plchan->vc.lock, flags);
1750 }
1751
1752 static struct pl08x_txd *pl08x_get_txd(struct pl08x_dma_chan *plchan)
1753 {
1754         struct pl08x_txd *txd = kzalloc(sizeof(*txd), GFP_NOWAIT);
1755
1756         if (txd)
1757                 INIT_LIST_HEAD(&txd->dsg_list);
1758         return txd;
1759 }
1760
1761 static u32 pl08x_memcpy_cctl(struct pl08x_driver_data *pl08x)
1762 {
1763         u32 cctl = 0;
1764
1765         /* Conjure cctl */
1766         switch (pl08x->pd->memcpy_burst_size) {
1767         default:
1768                 dev_err(&pl08x->adev->dev,
1769                         "illegal burst size for memcpy, set to 1\n");
1770                 /* Fall through */
1771         case PL08X_BURST_SZ_1:
1772                 cctl |= PL080_BSIZE_1 << PL080_CONTROL_SB_SIZE_SHIFT |
1773                         PL080_BSIZE_1 << PL080_CONTROL_DB_SIZE_SHIFT;
1774                 break;
1775         case PL08X_BURST_SZ_4:
1776                 cctl |= PL080_BSIZE_4 << PL080_CONTROL_SB_SIZE_SHIFT |
1777                         PL080_BSIZE_4 << PL080_CONTROL_DB_SIZE_SHIFT;
1778                 break;
1779         case PL08X_BURST_SZ_8:
1780                 cctl |= PL080_BSIZE_8 << PL080_CONTROL_SB_SIZE_SHIFT |
1781                         PL080_BSIZE_8 << PL080_CONTROL_DB_SIZE_SHIFT;
1782                 break;
1783         case PL08X_BURST_SZ_16:
1784                 cctl |= PL080_BSIZE_16 << PL080_CONTROL_SB_SIZE_SHIFT |
1785                         PL080_BSIZE_16 << PL080_CONTROL_DB_SIZE_SHIFT;
1786                 break;
1787         case PL08X_BURST_SZ_32:
1788                 cctl |= PL080_BSIZE_32 << PL080_CONTROL_SB_SIZE_SHIFT |
1789                         PL080_BSIZE_32 << PL080_CONTROL_DB_SIZE_SHIFT;
1790                 break;
1791         case PL08X_BURST_SZ_64:
1792                 cctl |= PL080_BSIZE_64 << PL080_CONTROL_SB_SIZE_SHIFT |
1793                         PL080_BSIZE_64 << PL080_CONTROL_DB_SIZE_SHIFT;
1794                 break;
1795         case PL08X_BURST_SZ_128:
1796                 cctl |= PL080_BSIZE_128 << PL080_CONTROL_SB_SIZE_SHIFT |
1797                         PL080_BSIZE_128 << PL080_CONTROL_DB_SIZE_SHIFT;
1798                 break;
1799         case PL08X_BURST_SZ_256:
1800                 cctl |= PL080_BSIZE_256 << PL080_CONTROL_SB_SIZE_SHIFT |
1801                         PL080_BSIZE_256 << PL080_CONTROL_DB_SIZE_SHIFT;
1802                 break;
1803         }
1804
1805         switch (pl08x->pd->memcpy_bus_width) {
1806         default:
1807                 dev_err(&pl08x->adev->dev,
1808                         "illegal bus width for memcpy, set to 8 bits\n");
1809                 /* Fall through */
1810         case PL08X_BUS_WIDTH_8_BITS:
1811                 cctl |= PL080_WIDTH_8BIT << PL080_CONTROL_SWIDTH_SHIFT |
1812                         PL080_WIDTH_8BIT << PL080_CONTROL_DWIDTH_SHIFT;
1813                 break;
1814         case PL08X_BUS_WIDTH_16_BITS:
1815                 cctl |= PL080_WIDTH_16BIT << PL080_CONTROL_SWIDTH_SHIFT |
1816                         PL080_WIDTH_16BIT << PL080_CONTROL_DWIDTH_SHIFT;
1817                 break;
1818         case PL08X_BUS_WIDTH_32_BITS:
1819                 cctl |= PL080_WIDTH_32BIT << PL080_CONTROL_SWIDTH_SHIFT |
1820                         PL080_WIDTH_32BIT << PL080_CONTROL_DWIDTH_SHIFT;
1821                 break;
1822         }
1823
1824         /* Protection flags */
1825         if (pl08x->pd->memcpy_prot_buff)
1826                 cctl |= PL080_CONTROL_PROT_BUFF;
1827         if (pl08x->pd->memcpy_prot_cache)
1828                 cctl |= PL080_CONTROL_PROT_CACHE;
1829
1830         /* We are the kernel, so we are in privileged mode */
1831         cctl |= PL080_CONTROL_PROT_SYS;
1832
1833         /* Both to be incremented or the code will break */
1834         cctl |= PL080_CONTROL_SRC_INCR | PL080_CONTROL_DST_INCR;
1835
1836         if (pl08x->vd->dualmaster)
1837                 cctl |= pl08x_select_bus(false,
1838                                          pl08x->mem_buses,
1839                                          pl08x->mem_buses);
1840
1841         return cctl;
1842 }
1843
1844 static u32 pl08x_ftdmac020_memcpy_cctl(struct pl08x_driver_data *pl08x)
1845 {
1846         u32 cctl = 0;
1847
1848         /* Conjure cctl */
1849         switch (pl08x->pd->memcpy_bus_width) {
1850         default:
1851                 dev_err(&pl08x->adev->dev,
1852                         "illegal bus width for memcpy, set to 8 bits\n");
1853                 /* Fall through */
1854         case PL08X_BUS_WIDTH_8_BITS:
1855                 cctl |= PL080_WIDTH_8BIT << FTDMAC020_LLI_SRC_WIDTH_SHIFT |
1856                         PL080_WIDTH_8BIT << FTDMAC020_LLI_DST_WIDTH_SHIFT;
1857                 break;
1858         case PL08X_BUS_WIDTH_16_BITS:
1859                 cctl |= PL080_WIDTH_16BIT << FTDMAC020_LLI_SRC_WIDTH_SHIFT |
1860                         PL080_WIDTH_16BIT << FTDMAC020_LLI_DST_WIDTH_SHIFT;
1861                 break;
1862         case PL08X_BUS_WIDTH_32_BITS:
1863                 cctl |= PL080_WIDTH_32BIT << FTDMAC020_LLI_SRC_WIDTH_SHIFT |
1864                         PL080_WIDTH_32BIT << FTDMAC020_LLI_DST_WIDTH_SHIFT;
1865                 break;
1866         }
1867
1868         /*
1869          * By default mask the TC IRQ on all LLIs, it will be unmasked on
1870          * the last LLI item by other code.
1871          */
1872         cctl |= FTDMAC020_LLI_TC_MSK;
1873
1874         /*
1875          * Both to be incremented so leave bits FTDMAC020_LLI_SRCAD_CTL
1876          * and FTDMAC020_LLI_DSTAD_CTL as zero
1877          */
1878         if (pl08x->vd->dualmaster)
1879                 cctl |= pl08x_select_bus(true,
1880                                          pl08x->mem_buses,
1881                                          pl08x->mem_buses);
1882
1883         return cctl;
1884 }
1885
1886 /*
1887  * Initialize a descriptor to be used by memcpy submit
1888  */
1889 static struct dma_async_tx_descriptor *pl08x_prep_dma_memcpy(
1890                 struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
1891                 size_t len, unsigned long flags)
1892 {
1893         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1894         struct pl08x_driver_data *pl08x = plchan->host;
1895         struct pl08x_txd *txd;
1896         struct pl08x_sg *dsg;
1897         int ret;
1898
1899         txd = pl08x_get_txd(plchan);
1900         if (!txd) {
1901                 dev_err(&pl08x->adev->dev,
1902                         "%s no memory for descriptor\n", __func__);
1903                 return NULL;
1904         }
1905
1906         dsg = kzalloc(sizeof(struct pl08x_sg), GFP_NOWAIT);
1907         if (!dsg) {
1908                 pl08x_free_txd(pl08x, txd);
1909                 return NULL;
1910         }
1911         list_add_tail(&dsg->node, &txd->dsg_list);
1912
1913         dsg->src_addr = src;
1914         dsg->dst_addr = dest;
1915         dsg->len = len;
1916         if (pl08x->vd->ftdmac020) {
1917                 /* Writing CCFG zero ENABLES all interrupts */
1918                 txd->ccfg = 0;
1919                 txd->cctl = pl08x_ftdmac020_memcpy_cctl(pl08x);
1920         } else {
1921                 txd->ccfg = PL080_CONFIG_ERR_IRQ_MASK |
1922                         PL080_CONFIG_TC_IRQ_MASK |
1923                         PL080_FLOW_MEM2MEM << PL080_CONFIG_FLOW_CONTROL_SHIFT;
1924                 txd->cctl = pl08x_memcpy_cctl(pl08x);
1925         }
1926
1927         ret = pl08x_fill_llis_for_desc(plchan->host, txd);
1928         if (!ret) {
1929                 pl08x_free_txd(pl08x, txd);
1930                 return NULL;
1931         }
1932
1933         return vchan_tx_prep(&plchan->vc, &txd->vd, flags);
1934 }
1935
1936 static struct pl08x_txd *pl08x_init_txd(
1937                 struct dma_chan *chan,
1938                 enum dma_transfer_direction direction,
1939                 dma_addr_t *slave_addr)
1940 {
1941         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
1942         struct pl08x_driver_data *pl08x = plchan->host;
1943         struct pl08x_txd *txd;
1944         enum dma_slave_buswidth addr_width;
1945         int ret, tmp;
1946         u8 src_buses, dst_buses;
1947         u32 maxburst, cctl;
1948
1949         txd = pl08x_get_txd(plchan);
1950         if (!txd) {
1951                 dev_err(&pl08x->adev->dev, "%s no txd\n", __func__);
1952                 return NULL;
1953         }
1954
1955         /*
1956          * Set up addresses, the PrimeCell configured address
1957          * will take precedence since this may configure the
1958          * channel target address dynamically at runtime.
1959          */
1960         if (direction == DMA_MEM_TO_DEV) {
1961                 cctl = PL080_CONTROL_SRC_INCR;
1962                 *slave_addr = plchan->cfg.dst_addr;
1963                 addr_width = plchan->cfg.dst_addr_width;
1964                 maxburst = plchan->cfg.dst_maxburst;
1965                 src_buses = pl08x->mem_buses;
1966                 dst_buses = plchan->cd->periph_buses;
1967         } else if (direction == DMA_DEV_TO_MEM) {
1968                 cctl = PL080_CONTROL_DST_INCR;
1969                 *slave_addr = plchan->cfg.src_addr;
1970                 addr_width = plchan->cfg.src_addr_width;
1971                 maxburst = plchan->cfg.src_maxburst;
1972                 src_buses = plchan->cd->periph_buses;
1973                 dst_buses = pl08x->mem_buses;
1974         } else {
1975                 pl08x_free_txd(pl08x, txd);
1976                 dev_err(&pl08x->adev->dev,
1977                         "%s direction unsupported\n", __func__);
1978                 return NULL;
1979         }
1980
1981         cctl |= pl08x_get_cctl(plchan, addr_width, maxburst);
1982         if (cctl == ~0) {
1983                 pl08x_free_txd(pl08x, txd);
1984                 dev_err(&pl08x->adev->dev,
1985                         "DMA slave configuration botched?\n");
1986                 return NULL;
1987         }
1988
1989         txd->cctl = cctl | pl08x_select_bus(false, src_buses, dst_buses);
1990
1991         if (plchan->cfg.device_fc)
1992                 tmp = (direction == DMA_MEM_TO_DEV) ? PL080_FLOW_MEM2PER_PER :
1993                         PL080_FLOW_PER2MEM_PER;
1994         else
1995                 tmp = (direction == DMA_MEM_TO_DEV) ? PL080_FLOW_MEM2PER :
1996                         PL080_FLOW_PER2MEM;
1997
1998         txd->ccfg = PL080_CONFIG_ERR_IRQ_MASK |
1999                 PL080_CONFIG_TC_IRQ_MASK |
2000                 tmp << PL080_CONFIG_FLOW_CONTROL_SHIFT;
2001
2002         ret = pl08x_request_mux(plchan);
2003         if (ret < 0) {
2004                 pl08x_free_txd(pl08x, txd);
2005                 dev_dbg(&pl08x->adev->dev,
2006                         "unable to mux for transfer on %s due to platform restrictions\n",
2007                         plchan->name);
2008                 return NULL;
2009         }
2010
2011         dev_dbg(&pl08x->adev->dev, "allocated DMA request signal %d for xfer on %s\n",
2012                  plchan->signal, plchan->name);
2013
2014         /* Assign the flow control signal to this channel */
2015         if (direction == DMA_MEM_TO_DEV)
2016                 txd->ccfg |= plchan->signal << PL080_CONFIG_DST_SEL_SHIFT;
2017         else
2018                 txd->ccfg |= plchan->signal << PL080_CONFIG_SRC_SEL_SHIFT;
2019
2020         return txd;
2021 }
2022
2023 static int pl08x_tx_add_sg(struct pl08x_txd *txd,
2024                            enum dma_transfer_direction direction,
2025                            dma_addr_t slave_addr,
2026                            dma_addr_t buf_addr,
2027                            unsigned int len)
2028 {
2029         struct pl08x_sg *dsg;
2030
2031         dsg = kzalloc(sizeof(struct pl08x_sg), GFP_NOWAIT);
2032         if (!dsg)
2033                 return -ENOMEM;
2034
2035         list_add_tail(&dsg->node, &txd->dsg_list);
2036
2037         dsg->len = len;
2038         if (direction == DMA_MEM_TO_DEV) {
2039                 dsg->src_addr = buf_addr;
2040                 dsg->dst_addr = slave_addr;
2041         } else {
2042                 dsg->src_addr = slave_addr;
2043                 dsg->dst_addr = buf_addr;
2044         }
2045
2046         return 0;
2047 }
2048
2049 static struct dma_async_tx_descriptor *pl08x_prep_slave_sg(
2050                 struct dma_chan *chan, struct scatterlist *sgl,
2051                 unsigned int sg_len, enum dma_transfer_direction direction,
2052                 unsigned long flags, void *context)
2053 {
2054         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
2055         struct pl08x_driver_data *pl08x = plchan->host;
2056         struct pl08x_txd *txd;
2057         struct scatterlist *sg;
2058         int ret, tmp;
2059         dma_addr_t slave_addr;
2060
2061         dev_dbg(&pl08x->adev->dev, "%s prepare transaction of %d bytes from %s\n",
2062                         __func__, sg_dma_len(sgl), plchan->name);
2063
2064         txd = pl08x_init_txd(chan, direction, &slave_addr);
2065         if (!txd)
2066                 return NULL;
2067
2068         for_each_sg(sgl, sg, sg_len, tmp) {
2069                 ret = pl08x_tx_add_sg(txd, direction, slave_addr,
2070                                       sg_dma_address(sg),
2071                                       sg_dma_len(sg));
2072                 if (ret) {
2073                         pl08x_release_mux(plchan);
2074                         pl08x_free_txd(pl08x, txd);
2075                         dev_err(&pl08x->adev->dev, "%s no mem for pl080 sg\n",
2076                                         __func__);
2077                         return NULL;
2078                 }
2079         }
2080
2081         ret = pl08x_fill_llis_for_desc(plchan->host, txd);
2082         if (!ret) {
2083                 pl08x_release_mux(plchan);
2084                 pl08x_free_txd(pl08x, txd);
2085                 return NULL;
2086         }
2087
2088         return vchan_tx_prep(&plchan->vc, &txd->vd, flags);
2089 }
2090
2091 static struct dma_async_tx_descriptor *pl08x_prep_dma_cyclic(
2092                 struct dma_chan *chan, dma_addr_t buf_addr, size_t buf_len,
2093                 size_t period_len, enum dma_transfer_direction direction,
2094                 unsigned long flags)
2095 {
2096         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
2097         struct pl08x_driver_data *pl08x = plchan->host;
2098         struct pl08x_txd *txd;
2099         int ret, tmp;
2100         dma_addr_t slave_addr;
2101
2102         dev_dbg(&pl08x->adev->dev,
2103                 "%s prepare cyclic transaction of %zd/%zd bytes %s %s\n",
2104                 __func__, period_len, buf_len,
2105                 direction == DMA_MEM_TO_DEV ? "to" : "from",
2106                 plchan->name);
2107
2108         txd = pl08x_init_txd(chan, direction, &slave_addr);
2109         if (!txd)
2110                 return NULL;
2111
2112         txd->cyclic = true;
2113         txd->cctl |= PL080_CONTROL_TC_IRQ_EN;
2114         for (tmp = 0; tmp < buf_len; tmp += period_len) {
2115                 ret = pl08x_tx_add_sg(txd, direction, slave_addr,
2116                                       buf_addr + tmp, period_len);
2117                 if (ret) {
2118                         pl08x_release_mux(plchan);
2119                         pl08x_free_txd(pl08x, txd);
2120                         return NULL;
2121                 }
2122         }
2123
2124         ret = pl08x_fill_llis_for_desc(plchan->host, txd);
2125         if (!ret) {
2126                 pl08x_release_mux(plchan);
2127                 pl08x_free_txd(pl08x, txd);
2128                 return NULL;
2129         }
2130
2131         return vchan_tx_prep(&plchan->vc, &txd->vd, flags);
2132 }
2133
2134 static int pl08x_config(struct dma_chan *chan,
2135                         struct dma_slave_config *config)
2136 {
2137         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
2138         struct pl08x_driver_data *pl08x = plchan->host;
2139
2140         if (!plchan->slave)
2141                 return -EINVAL;
2142
2143         /* Reject definitely invalid configurations */
2144         if (config->src_addr_width == DMA_SLAVE_BUSWIDTH_8_BYTES ||
2145             config->dst_addr_width == DMA_SLAVE_BUSWIDTH_8_BYTES)
2146                 return -EINVAL;
2147
2148         if (config->device_fc && pl08x->vd->pl080s) {
2149                 dev_err(&pl08x->adev->dev,
2150                         "%s: PL080S does not support peripheral flow control\n",
2151                         __func__);
2152                 return -EINVAL;
2153         }
2154
2155         plchan->cfg = *config;
2156
2157         return 0;
2158 }
2159
2160 static int pl08x_terminate_all(struct dma_chan *chan)
2161 {
2162         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
2163         struct pl08x_driver_data *pl08x = plchan->host;
2164         unsigned long flags;
2165
2166         spin_lock_irqsave(&plchan->vc.lock, flags);
2167         if (!plchan->phychan && !plchan->at) {
2168                 spin_unlock_irqrestore(&plchan->vc.lock, flags);
2169                 return 0;
2170         }
2171
2172         plchan->state = PL08X_CHAN_IDLE;
2173
2174         if (plchan->phychan) {
2175                 /*
2176                  * Mark physical channel as free and free any slave
2177                  * signal
2178                  */
2179                 pl08x_phy_free(plchan);
2180         }
2181         /* Dequeue jobs and free LLIs */
2182         if (plchan->at) {
2183                 vchan_terminate_vdesc(&plchan->at->vd);
2184                 plchan->at = NULL;
2185         }
2186         /* Dequeue jobs not yet fired as well */
2187         pl08x_free_txd_list(pl08x, plchan);
2188
2189         spin_unlock_irqrestore(&plchan->vc.lock, flags);
2190
2191         return 0;
2192 }
2193
2194 static void pl08x_synchronize(struct dma_chan *chan)
2195 {
2196         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
2197
2198         vchan_synchronize(&plchan->vc);
2199 }
2200
2201 static int pl08x_pause(struct dma_chan *chan)
2202 {
2203         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
2204         unsigned long flags;
2205
2206         /*
2207          * Anything succeeds on channels with no physical allocation and
2208          * no queued transfers.
2209          */
2210         spin_lock_irqsave(&plchan->vc.lock, flags);
2211         if (!plchan->phychan && !plchan->at) {
2212                 spin_unlock_irqrestore(&plchan->vc.lock, flags);
2213                 return 0;
2214         }
2215
2216         pl08x_pause_phy_chan(plchan->phychan);
2217         plchan->state = PL08X_CHAN_PAUSED;
2218
2219         spin_unlock_irqrestore(&plchan->vc.lock, flags);
2220
2221         return 0;
2222 }
2223
2224 static int pl08x_resume(struct dma_chan *chan)
2225 {
2226         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
2227         unsigned long flags;
2228
2229         /*
2230          * Anything succeeds on channels with no physical allocation and
2231          * no queued transfers.
2232          */
2233         spin_lock_irqsave(&plchan->vc.lock, flags);
2234         if (!plchan->phychan && !plchan->at) {
2235                 spin_unlock_irqrestore(&plchan->vc.lock, flags);
2236                 return 0;
2237         }
2238
2239         pl08x_resume_phy_chan(plchan->phychan);
2240         plchan->state = PL08X_CHAN_RUNNING;
2241
2242         spin_unlock_irqrestore(&plchan->vc.lock, flags);
2243
2244         return 0;
2245 }
2246
2247 bool pl08x_filter_id(struct dma_chan *chan, void *chan_id)
2248 {
2249         struct pl08x_dma_chan *plchan;
2250         char *name = chan_id;
2251
2252         /* Reject channels for devices not bound to this driver */
2253         if (chan->device->dev->driver != &pl08x_amba_driver.drv)
2254                 return false;
2255
2256         plchan = to_pl08x_chan(chan);
2257
2258         /* Check that the channel is not taken! */
2259         if (!strcmp(plchan->name, name))
2260                 return true;
2261
2262         return false;
2263 }
2264 EXPORT_SYMBOL_GPL(pl08x_filter_id);
2265
2266 static bool pl08x_filter_fn(struct dma_chan *chan, void *chan_id)
2267 {
2268         struct pl08x_dma_chan *plchan = to_pl08x_chan(chan);
2269
2270         return plchan->cd == chan_id;
2271 }
2272
2273 /*
2274  * Just check that the device is there and active
2275  * TODO: turn this bit on/off depending on the number of physical channels
2276  * actually used, if it is zero... well shut it off. That will save some
2277  * power. Cut the clock at the same time.
2278  */
2279 static void pl08x_ensure_on(struct pl08x_driver_data *pl08x)
2280 {
2281         /* The Nomadik variant does not have the config register */
2282         if (pl08x->vd->nomadik)
2283                 return;
2284         /* The FTDMAC020 variant does this in another register */
2285         if (pl08x->vd->ftdmac020) {
2286                 writel(PL080_CONFIG_ENABLE, pl08x->base + FTDMAC020_CSR);
2287                 return;
2288         }
2289         writel(PL080_CONFIG_ENABLE, pl08x->base + PL080_CONFIG);
2290 }
2291
2292 static irqreturn_t pl08x_irq(int irq, void *dev)
2293 {
2294         struct pl08x_driver_data *pl08x = dev;
2295         u32 mask = 0, err, tc, i;
2296
2297         /* check & clear - ERR & TC interrupts */
2298         err = readl(pl08x->base + PL080_ERR_STATUS);
2299         if (err) {
2300                 dev_err(&pl08x->adev->dev, "%s error interrupt, register value 0x%08x\n",
2301                         __func__, err);
2302                 writel(err, pl08x->base + PL080_ERR_CLEAR);
2303         }
2304         tc = readl(pl08x->base + PL080_TC_STATUS);
2305         if (tc)
2306                 writel(tc, pl08x->base + PL080_TC_CLEAR);
2307
2308         if (!err && !tc)
2309                 return IRQ_NONE;
2310
2311         for (i = 0; i < pl08x->vd->channels; i++) {
2312                 if ((BIT(i) & err) || (BIT(i) & tc)) {
2313                         /* Locate physical channel */
2314                         struct pl08x_phy_chan *phychan = &pl08x->phy_chans[i];
2315                         struct pl08x_dma_chan *plchan = phychan->serving;
2316                         struct pl08x_txd *tx;
2317
2318                         if (!plchan) {
2319                                 dev_err(&pl08x->adev->dev,
2320                                         "%s Error TC interrupt on unused channel: 0x%08x\n",
2321                                         __func__, i);
2322                                 continue;
2323                         }
2324
2325                         spin_lock(&plchan->vc.lock);
2326                         tx = plchan->at;
2327                         if (tx && tx->cyclic) {
2328                                 vchan_cyclic_callback(&tx->vd);
2329                         } else if (tx) {
2330                                 plchan->at = NULL;
2331                                 /*
2332                                  * This descriptor is done, release its mux
2333                                  * reservation.
2334                                  */
2335                                 pl08x_release_mux(plchan);
2336                                 tx->done = true;
2337                                 vchan_cookie_complete(&tx->vd);
2338
2339                                 /*
2340                                  * And start the next descriptor (if any),
2341                                  * otherwise free this channel.
2342                                  */
2343                                 if (vchan_next_desc(&plchan->vc))
2344                                         pl08x_start_next_txd(plchan);
2345                                 else
2346                                         pl08x_phy_free(plchan);
2347                         }
2348                         spin_unlock(&plchan->vc.lock);
2349
2350                         mask |= BIT(i);
2351                 }
2352         }
2353
2354         return mask ? IRQ_HANDLED : IRQ_NONE;
2355 }
2356
2357 static void pl08x_dma_slave_init(struct pl08x_dma_chan *chan)
2358 {
2359         chan->slave = true;
2360         chan->name = chan->cd->bus_id;
2361         chan->cfg.src_addr = chan->cd->addr;
2362         chan->cfg.dst_addr = chan->cd->addr;
2363 }
2364
2365 /*
2366  * Initialise the DMAC memcpy/slave channels.
2367  * Make a local wrapper to hold required data
2368  */
2369 static int pl08x_dma_init_virtual_channels(struct pl08x_driver_data *pl08x,
2370                 struct dma_device *dmadev, unsigned int channels, bool slave)
2371 {
2372         struct pl08x_dma_chan *chan;
2373         int i;
2374
2375         INIT_LIST_HEAD(&dmadev->channels);
2376
2377         /*
2378          * Register as many many memcpy as we have physical channels,
2379          * we won't always be able to use all but the code will have
2380          * to cope with that situation.
2381          */
2382         for (i = 0; i < channels; i++) {
2383                 chan = kzalloc(sizeof(*chan), GFP_KERNEL);
2384                 if (!chan)
2385                         return -ENOMEM;
2386
2387                 chan->host = pl08x;
2388                 chan->state = PL08X_CHAN_IDLE;
2389                 chan->signal = -1;
2390
2391                 if (slave) {
2392                         chan->cd = &pl08x->pd->slave_channels[i];
2393                         /*
2394                          * Some implementations have muxed signals, whereas some
2395                          * use a mux in front of the signals and need dynamic
2396                          * assignment of signals.
2397                          */
2398                         chan->signal = i;
2399                         pl08x_dma_slave_init(chan);
2400                 } else {
2401                         chan->cd = kzalloc(sizeof(*chan->cd), GFP_KERNEL);
2402                         if (!chan->cd) {
2403                                 kfree(chan);
2404                                 return -ENOMEM;
2405                         }
2406                         chan->cd->bus_id = "memcpy";
2407                         chan->cd->periph_buses = pl08x->pd->mem_buses;
2408                         chan->name = kasprintf(GFP_KERNEL, "memcpy%d", i);
2409                         if (!chan->name) {
2410                                 kfree(chan->cd);
2411                                 kfree(chan);
2412                                 return -ENOMEM;
2413                         }
2414                 }
2415                 dev_dbg(&pl08x->adev->dev,
2416                          "initialize virtual channel \"%s\"\n",
2417                          chan->name);
2418
2419                 chan->vc.desc_free = pl08x_desc_free;
2420                 vchan_init(&chan->vc, dmadev);
2421         }
2422         dev_info(&pl08x->adev->dev, "initialized %d virtual %s channels\n",
2423                  i, slave ? "slave" : "memcpy");
2424         return i;
2425 }
2426
2427 static void pl08x_free_virtual_channels(struct dma_device *dmadev)
2428 {
2429         struct pl08x_dma_chan *chan = NULL;
2430         struct pl08x_dma_chan *next;
2431
2432         list_for_each_entry_safe(chan,
2433                                  next, &dmadev->channels, vc.chan.device_node) {
2434                 list_del(&chan->vc.chan.device_node);
2435                 kfree(chan);
2436         }
2437 }
2438
2439 #ifdef CONFIG_DEBUG_FS
2440 static const char *pl08x_state_str(enum pl08x_dma_chan_state state)
2441 {
2442         switch (state) {
2443         case PL08X_CHAN_IDLE:
2444                 return "idle";
2445         case PL08X_CHAN_RUNNING:
2446                 return "running";
2447         case PL08X_CHAN_PAUSED:
2448                 return "paused";
2449         case PL08X_CHAN_WAITING:
2450                 return "waiting";
2451         default:
2452                 break;
2453         }
2454         return "UNKNOWN STATE";
2455 }
2456
2457 static int pl08x_debugfs_show(struct seq_file *s, void *data)
2458 {
2459         struct pl08x_driver_data *pl08x = s->private;
2460         struct pl08x_dma_chan *chan;
2461         struct pl08x_phy_chan *ch;
2462         unsigned long flags;
2463         int i;
2464
2465         seq_printf(s, "PL08x physical channels:\n");
2466         seq_printf(s, "CHANNEL:\tUSER:\n");
2467         seq_printf(s, "--------\t-----\n");
2468         for (i = 0; i < pl08x->vd->channels; i++) {
2469                 struct pl08x_dma_chan *virt_chan;
2470
2471                 ch = &pl08x->phy_chans[i];
2472
2473                 spin_lock_irqsave(&ch->lock, flags);
2474                 virt_chan = ch->serving;
2475
2476                 seq_printf(s, "%d\t\t%s%s\n",
2477                            ch->id,
2478                            virt_chan ? virt_chan->name : "(none)",
2479                            ch->locked ? " LOCKED" : "");
2480
2481                 spin_unlock_irqrestore(&ch->lock, flags);
2482         }
2483
2484         seq_printf(s, "\nPL08x virtual memcpy channels:\n");
2485         seq_printf(s, "CHANNEL:\tSTATE:\n");
2486         seq_printf(s, "--------\t------\n");
2487         list_for_each_entry(chan, &pl08x->memcpy.channels, vc.chan.device_node) {
2488                 seq_printf(s, "%s\t\t%s\n", chan->name,
2489                            pl08x_state_str(chan->state));
2490         }
2491
2492         if (pl08x->has_slave) {
2493                 seq_printf(s, "\nPL08x virtual slave channels:\n");
2494                 seq_printf(s, "CHANNEL:\tSTATE:\n");
2495                 seq_printf(s, "--------\t------\n");
2496                 list_for_each_entry(chan, &pl08x->slave.channels,
2497                                     vc.chan.device_node) {
2498                         seq_printf(s, "%s\t\t%s\n", chan->name,
2499                                    pl08x_state_str(chan->state));
2500                 }
2501         }
2502
2503         return 0;
2504 }
2505
2506 DEFINE_SHOW_ATTRIBUTE(pl08x_debugfs);
2507
2508 static void init_pl08x_debugfs(struct pl08x_driver_data *pl08x)
2509 {
2510         /* Expose a simple debugfs interface to view all clocks */
2511         debugfs_create_file(dev_name(&pl08x->adev->dev), S_IFREG | S_IRUGO,
2512                             NULL, pl08x, &pl08x_debugfs_fops);
2513 }
2514
2515 #else
2516 static inline void init_pl08x_debugfs(struct pl08x_driver_data *pl08x)
2517 {
2518 }
2519 #endif
2520
2521 #ifdef CONFIG_OF
2522 static struct dma_chan *pl08x_find_chan_id(struct pl08x_driver_data *pl08x,
2523                                          u32 id)
2524 {
2525         struct pl08x_dma_chan *chan;
2526
2527         /* Trying to get a slave channel from something with no slave support */
2528         if (!pl08x->has_slave)
2529                 return NULL;
2530
2531         list_for_each_entry(chan, &pl08x->slave.channels, vc.chan.device_node) {
2532                 if (chan->signal == id)
2533                         return &chan->vc.chan;
2534         }
2535
2536         return NULL;
2537 }
2538
2539 static struct dma_chan *pl08x_of_xlate(struct of_phandle_args *dma_spec,
2540                                        struct of_dma *ofdma)
2541 {
2542         struct pl08x_driver_data *pl08x = ofdma->of_dma_data;
2543         struct dma_chan *dma_chan;
2544         struct pl08x_dma_chan *plchan;
2545
2546         if (!pl08x)
2547                 return NULL;
2548
2549         if (dma_spec->args_count != 2) {
2550                 dev_err(&pl08x->adev->dev,
2551                         "DMA channel translation requires two cells\n");
2552                 return NULL;
2553         }
2554
2555         dma_chan = pl08x_find_chan_id(pl08x, dma_spec->args[0]);
2556         if (!dma_chan) {
2557                 dev_err(&pl08x->adev->dev,
2558                         "DMA slave channel not found\n");
2559                 return NULL;
2560         }
2561
2562         plchan = to_pl08x_chan(dma_chan);
2563         dev_dbg(&pl08x->adev->dev,
2564                 "translated channel for signal %d\n",
2565                 dma_spec->args[0]);
2566
2567         /* Augment channel data for applicable AHB buses */
2568         plchan->cd->periph_buses = dma_spec->args[1];
2569         return dma_get_slave_channel(dma_chan);
2570 }
2571
2572 static int pl08x_of_probe(struct amba_device *adev,
2573                           struct pl08x_driver_data *pl08x,
2574                           struct device_node *np)
2575 {
2576         struct pl08x_platform_data *pd;
2577         struct pl08x_channel_data *chanp = NULL;
2578         u32 val;
2579         int ret;
2580         int i;
2581
2582         pd = devm_kzalloc(&adev->dev, sizeof(*pd), GFP_KERNEL);
2583         if (!pd)
2584                 return -ENOMEM;
2585
2586         /* Eligible bus masters for fetching LLIs */
2587         if (of_property_read_bool(np, "lli-bus-interface-ahb1"))
2588                 pd->lli_buses |= PL08X_AHB1;
2589         if (of_property_read_bool(np, "lli-bus-interface-ahb2"))
2590                 pd->lli_buses |= PL08X_AHB2;
2591         if (!pd->lli_buses) {
2592                 dev_info(&adev->dev, "no bus masters for LLIs stated, assume all\n");
2593                 pd->lli_buses |= PL08X_AHB1 | PL08X_AHB2;
2594         }
2595
2596         /* Eligible bus masters for memory access */
2597         if (of_property_read_bool(np, "mem-bus-interface-ahb1"))
2598                 pd->mem_buses |= PL08X_AHB1;
2599         if (of_property_read_bool(np, "mem-bus-interface-ahb2"))
2600                 pd->mem_buses |= PL08X_AHB2;
2601         if (!pd->mem_buses) {
2602                 dev_info(&adev->dev, "no bus masters for memory stated, assume all\n");
2603                 pd->mem_buses |= PL08X_AHB1 | PL08X_AHB2;
2604         }
2605
2606         /* Parse the memcpy channel properties */
2607         ret = of_property_read_u32(np, "memcpy-burst-size", &val);
2608         if (ret) {
2609                 dev_info(&adev->dev, "no memcpy burst size specified, using 1 byte\n");
2610                 val = 1;
2611         }
2612         switch (val) {
2613         default:
2614                 dev_err(&adev->dev, "illegal burst size for memcpy, set to 1\n");
2615                 /* Fall through */
2616         case 1:
2617                 pd->memcpy_burst_size = PL08X_BURST_SZ_1;
2618                 break;
2619         case 4:
2620                 pd->memcpy_burst_size = PL08X_BURST_SZ_4;
2621                 break;
2622         case 8:
2623                 pd->memcpy_burst_size = PL08X_BURST_SZ_8;
2624                 break;
2625         case 16:
2626                 pd->memcpy_burst_size = PL08X_BURST_SZ_16;
2627                 break;
2628         case 32:
2629                 pd->memcpy_burst_size = PL08X_BURST_SZ_32;
2630                 break;
2631         case 64:
2632                 pd->memcpy_burst_size = PL08X_BURST_SZ_64;
2633                 break;
2634         case 128:
2635                 pd->memcpy_burst_size = PL08X_BURST_SZ_128;
2636                 break;
2637         case 256:
2638                 pd->memcpy_burst_size = PL08X_BURST_SZ_256;
2639                 break;
2640         }
2641
2642         ret = of_property_read_u32(np, "memcpy-bus-width", &val);
2643         if (ret) {
2644                 dev_info(&adev->dev, "no memcpy bus width specified, using 8 bits\n");
2645                 val = 8;
2646         }
2647         switch (val) {
2648         default:
2649                 dev_err(&adev->dev, "illegal bus width for memcpy, set to 8 bits\n");
2650                 /* Fall through */
2651         case 8:
2652                 pd->memcpy_bus_width = PL08X_BUS_WIDTH_8_BITS;
2653                 break;
2654         case 16:
2655                 pd->memcpy_bus_width = PL08X_BUS_WIDTH_16_BITS;
2656                 break;
2657         case 32:
2658                 pd->memcpy_bus_width = PL08X_BUS_WIDTH_32_BITS;
2659                 break;
2660         }
2661
2662         /*
2663          * Allocate channel data for all possible slave channels (one
2664          * for each possible signal), channels will then be allocated
2665          * for a device and have it's AHB interfaces set up at
2666          * translation time.
2667          */
2668         if (pl08x->vd->signals) {
2669                 chanp = devm_kcalloc(&adev->dev,
2670                                      pl08x->vd->signals,
2671                                      sizeof(struct pl08x_channel_data),
2672                                      GFP_KERNEL);
2673                 if (!chanp)
2674                         return -ENOMEM;
2675
2676                 pd->slave_channels = chanp;
2677                 for (i = 0; i < pl08x->vd->signals; i++) {
2678                         /*
2679                          * chanp->periph_buses will be assigned at translation
2680                          */
2681                         chanp->bus_id = kasprintf(GFP_KERNEL, "slave%d", i);
2682                         chanp++;
2683                 }
2684                 pd->num_slave_channels = pl08x->vd->signals;
2685         }
2686
2687         pl08x->pd = pd;
2688
2689         return of_dma_controller_register(adev->dev.of_node, pl08x_of_xlate,
2690                                           pl08x);
2691 }
2692 #else
2693 static inline int pl08x_of_probe(struct amba_device *adev,
2694                                  struct pl08x_driver_data *pl08x,
2695                                  struct device_node *np)
2696 {
2697         return -EINVAL;
2698 }
2699 #endif
2700
2701 static int pl08x_probe(struct amba_device *adev, const struct amba_id *id)
2702 {
2703         struct pl08x_driver_data *pl08x;
2704         struct vendor_data *vd = id->data;
2705         struct device_node *np = adev->dev.of_node;
2706         u32 tsfr_size;
2707         int ret = 0;
2708         int i;
2709
2710         ret = amba_request_regions(adev, NULL);
2711         if (ret)
2712                 return ret;
2713
2714         /* Ensure that we can do DMA */
2715         ret = dma_set_mask_and_coherent(&adev->dev, DMA_BIT_MASK(32));
2716         if (ret)
2717                 goto out_no_pl08x;
2718
2719         /* Create the driver state holder */
2720         pl08x = kzalloc(sizeof(*pl08x), GFP_KERNEL);
2721         if (!pl08x) {
2722                 ret = -ENOMEM;
2723                 goto out_no_pl08x;
2724         }
2725
2726         /* Assign useful pointers to the driver state */
2727         pl08x->adev = adev;
2728         pl08x->vd = vd;
2729
2730         pl08x->base = ioremap(adev->res.start, resource_size(&adev->res));
2731         if (!pl08x->base) {
2732                 ret = -ENOMEM;
2733                 goto out_no_ioremap;
2734         }
2735
2736         if (vd->ftdmac020) {
2737                 u32 val;
2738
2739                 val = readl(pl08x->base + FTDMAC020_REVISION);
2740                 dev_info(&pl08x->adev->dev, "FTDMAC020 %d.%d rel %d\n",
2741                          (val >> 16) & 0xff, (val >> 8) & 0xff, val & 0xff);
2742                 val = readl(pl08x->base + FTDMAC020_FEATURE);
2743                 dev_info(&pl08x->adev->dev, "FTDMAC020 %d channels, "
2744                          "%s built-in bridge, %s, %s linked lists\n",
2745                          (val >> 12) & 0x0f,
2746                          (val & BIT(10)) ? "no" : "has",
2747                          (val & BIT(9)) ? "AHB0 and AHB1" : "AHB0",
2748                          (val & BIT(8)) ? "supports" : "does not support");
2749
2750                 /* Vendor data from feature register */
2751                 if (!(val & BIT(8)))
2752                         dev_warn(&pl08x->adev->dev,
2753                                  "linked lists not supported, required\n");
2754                 vd->channels = (val >> 12) & 0x0f;
2755                 vd->dualmaster = !!(val & BIT(9));
2756         }
2757
2758         /* Initialize memcpy engine */
2759         dma_cap_set(DMA_MEMCPY, pl08x->memcpy.cap_mask);
2760         pl08x->memcpy.dev = &adev->dev;
2761         pl08x->memcpy.device_free_chan_resources = pl08x_free_chan_resources;
2762         pl08x->memcpy.device_prep_dma_memcpy = pl08x_prep_dma_memcpy;
2763         pl08x->memcpy.device_prep_dma_interrupt = pl08x_prep_dma_interrupt;
2764         pl08x->memcpy.device_tx_status = pl08x_dma_tx_status;
2765         pl08x->memcpy.device_issue_pending = pl08x_issue_pending;
2766         pl08x->memcpy.device_config = pl08x_config;
2767         pl08x->memcpy.device_pause = pl08x_pause;
2768         pl08x->memcpy.device_resume = pl08x_resume;
2769         pl08x->memcpy.device_terminate_all = pl08x_terminate_all;
2770         pl08x->memcpy.device_synchronize = pl08x_synchronize;
2771         pl08x->memcpy.src_addr_widths = PL80X_DMA_BUSWIDTHS;
2772         pl08x->memcpy.dst_addr_widths = PL80X_DMA_BUSWIDTHS;
2773         pl08x->memcpy.directions = BIT(DMA_MEM_TO_MEM);
2774         pl08x->memcpy.residue_granularity = DMA_RESIDUE_GRANULARITY_SEGMENT;
2775         if (vd->ftdmac020)
2776                 pl08x->memcpy.copy_align = DMAENGINE_ALIGN_4_BYTES;
2777
2778
2779         /*
2780          * Initialize slave engine, if the block has no signals, that means
2781          * we have no slave support.
2782          */
2783         if (vd->signals) {
2784                 pl08x->has_slave = true;
2785                 dma_cap_set(DMA_SLAVE, pl08x->slave.cap_mask);
2786                 dma_cap_set(DMA_CYCLIC, pl08x->slave.cap_mask);
2787                 pl08x->slave.dev = &adev->dev;
2788                 pl08x->slave.device_free_chan_resources =
2789                         pl08x_free_chan_resources;
2790                 pl08x->slave.device_prep_dma_interrupt =
2791                         pl08x_prep_dma_interrupt;
2792                 pl08x->slave.device_tx_status = pl08x_dma_tx_status;
2793                 pl08x->slave.device_issue_pending = pl08x_issue_pending;
2794                 pl08x->slave.device_prep_slave_sg = pl08x_prep_slave_sg;
2795                 pl08x->slave.device_prep_dma_cyclic = pl08x_prep_dma_cyclic;
2796                 pl08x->slave.device_config = pl08x_config;
2797                 pl08x->slave.device_pause = pl08x_pause;
2798                 pl08x->slave.device_resume = pl08x_resume;
2799                 pl08x->slave.device_terminate_all = pl08x_terminate_all;
2800                 pl08x->slave.device_synchronize = pl08x_synchronize;
2801                 pl08x->slave.src_addr_widths = PL80X_DMA_BUSWIDTHS;
2802                 pl08x->slave.dst_addr_widths = PL80X_DMA_BUSWIDTHS;
2803                 pl08x->slave.directions =
2804                         BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV);
2805                 pl08x->slave.residue_granularity =
2806                         DMA_RESIDUE_GRANULARITY_SEGMENT;
2807         }
2808
2809         /* Get the platform data */
2810         pl08x->pd = dev_get_platdata(&adev->dev);
2811         if (!pl08x->pd) {
2812                 if (np) {
2813                         ret = pl08x_of_probe(adev, pl08x, np);
2814                         if (ret)
2815                                 goto out_no_platdata;
2816                 } else {
2817                         dev_err(&adev->dev, "no platform data supplied\n");
2818                         ret = -EINVAL;
2819                         goto out_no_platdata;
2820                 }
2821         } else {
2822                 pl08x->slave.filter.map = pl08x->pd->slave_map;
2823                 pl08x->slave.filter.mapcnt = pl08x->pd->slave_map_len;
2824                 pl08x->slave.filter.fn = pl08x_filter_fn;
2825         }
2826
2827         /* By default, AHB1 only.  If dualmaster, from platform */
2828         pl08x->lli_buses = PL08X_AHB1;
2829         pl08x->mem_buses = PL08X_AHB1;
2830         if (pl08x->vd->dualmaster) {
2831                 pl08x->lli_buses = pl08x->pd->lli_buses;
2832                 pl08x->mem_buses = pl08x->pd->mem_buses;
2833         }
2834
2835         if (vd->pl080s)
2836                 pl08x->lli_words = PL080S_LLI_WORDS;
2837         else
2838                 pl08x->lli_words = PL080_LLI_WORDS;
2839         tsfr_size = MAX_NUM_TSFR_LLIS * pl08x->lli_words * sizeof(u32);
2840
2841         /* A DMA memory pool for LLIs, align on 1-byte boundary */
2842         pl08x->pool = dma_pool_create(DRIVER_NAME, &pl08x->adev->dev,
2843                                                 tsfr_size, PL08X_ALIGN, 0);
2844         if (!pl08x->pool) {
2845                 ret = -ENOMEM;
2846                 goto out_no_lli_pool;
2847         }
2848
2849         /* Turn on the PL08x */
2850         pl08x_ensure_on(pl08x);
2851
2852         /* Clear any pending interrupts */
2853         if (vd->ftdmac020)
2854                 /* This variant has error IRQs in bits 16-19 */
2855                 writel(0x0000FFFF, pl08x->base + PL080_ERR_CLEAR);
2856         else
2857                 writel(0x000000FF, pl08x->base + PL080_ERR_CLEAR);
2858         writel(0x000000FF, pl08x->base + PL080_TC_CLEAR);
2859
2860         /* Attach the interrupt handler */
2861         ret = request_irq(adev->irq[0], pl08x_irq, 0, DRIVER_NAME, pl08x);
2862         if (ret) {
2863                 dev_err(&adev->dev, "%s failed to request interrupt %d\n",
2864                         __func__, adev->irq[0]);
2865                 goto out_no_irq;
2866         }
2867
2868         /* Initialize physical channels */
2869         pl08x->phy_chans = kzalloc((vd->channels * sizeof(*pl08x->phy_chans)),
2870                         GFP_KERNEL);
2871         if (!pl08x->phy_chans) {
2872                 ret = -ENOMEM;
2873                 goto out_no_phychans;
2874         }
2875
2876         for (i = 0; i < vd->channels; i++) {
2877                 struct pl08x_phy_chan *ch = &pl08x->phy_chans[i];
2878
2879                 ch->id = i;
2880                 ch->base = pl08x->base + PL080_Cx_BASE(i);
2881                 if (vd->ftdmac020) {
2882                         /* FTDMA020 has a special channel busy register */
2883                         ch->reg_busy = ch->base + FTDMAC020_CH_BUSY;
2884                         ch->reg_config = ch->base + FTDMAC020_CH_CFG;
2885                         ch->reg_control = ch->base + FTDMAC020_CH_CSR;
2886                         ch->reg_src = ch->base + FTDMAC020_CH_SRC_ADDR;
2887                         ch->reg_dst = ch->base + FTDMAC020_CH_DST_ADDR;
2888                         ch->reg_lli = ch->base + FTDMAC020_CH_LLP;
2889                         ch->ftdmac020 = true;
2890                 } else {
2891                         ch->reg_config = ch->base + vd->config_offset;
2892                         ch->reg_control = ch->base + PL080_CH_CONTROL;
2893                         ch->reg_src = ch->base + PL080_CH_SRC_ADDR;
2894                         ch->reg_dst = ch->base + PL080_CH_DST_ADDR;
2895                         ch->reg_lli = ch->base + PL080_CH_LLI;
2896                 }
2897                 if (vd->pl080s)
2898                         ch->pl080s = true;
2899
2900                 spin_lock_init(&ch->lock);
2901
2902                 /*
2903                  * Nomadik variants can have channels that are locked
2904                  * down for the secure world only. Lock up these channels
2905                  * by perpetually serving a dummy virtual channel.
2906                  */
2907                 if (vd->nomadik) {
2908                         u32 val;
2909
2910                         val = readl(ch->reg_config);
2911                         if (val & (PL080N_CONFIG_ITPROT | PL080N_CONFIG_SECPROT)) {
2912                                 dev_info(&adev->dev, "physical channel %d reserved for secure access only\n", i);
2913                                 ch->locked = true;
2914                         }
2915                 }
2916
2917                 dev_dbg(&adev->dev, "physical channel %d is %s\n",
2918                         i, pl08x_phy_channel_busy(ch) ? "BUSY" : "FREE");
2919         }
2920
2921         /* Register as many memcpy channels as there are physical channels */
2922         ret = pl08x_dma_init_virtual_channels(pl08x, &pl08x->memcpy,
2923                                               pl08x->vd->channels, false);
2924         if (ret <= 0) {
2925                 dev_warn(&pl08x->adev->dev,
2926                          "%s failed to enumerate memcpy channels - %d\n",
2927                          __func__, ret);
2928                 goto out_no_memcpy;
2929         }
2930
2931         /* Register slave channels */
2932         if (pl08x->has_slave) {
2933                 ret = pl08x_dma_init_virtual_channels(pl08x, &pl08x->slave,
2934                                         pl08x->pd->num_slave_channels, true);
2935                 if (ret < 0) {
2936                         dev_warn(&pl08x->adev->dev,
2937                                  "%s failed to enumerate slave channels - %d\n",
2938                                  __func__, ret);
2939                         goto out_no_slave;
2940                 }
2941         }
2942
2943         ret = dma_async_device_register(&pl08x->memcpy);
2944         if (ret) {
2945                 dev_warn(&pl08x->adev->dev,
2946                         "%s failed to register memcpy as an async device - %d\n",
2947                         __func__, ret);
2948                 goto out_no_memcpy_reg;
2949         }
2950
2951         if (pl08x->has_slave) {
2952                 ret = dma_async_device_register(&pl08x->slave);
2953                 if (ret) {
2954                         dev_warn(&pl08x->adev->dev,
2955                         "%s failed to register slave as an async device - %d\n",
2956                         __func__, ret);
2957                         goto out_no_slave_reg;
2958                 }
2959         }
2960
2961         amba_set_drvdata(adev, pl08x);
2962         init_pl08x_debugfs(pl08x);
2963         dev_info(&pl08x->adev->dev, "DMA: PL%03x%s rev%u at 0x%08llx irq %d\n",
2964                  amba_part(adev), pl08x->vd->pl080s ? "s" : "", amba_rev(adev),
2965                  (unsigned long long)adev->res.start, adev->irq[0]);
2966
2967         return 0;
2968
2969 out_no_slave_reg:
2970         dma_async_device_unregister(&pl08x->memcpy);
2971 out_no_memcpy_reg:
2972         if (pl08x->has_slave)
2973                 pl08x_free_virtual_channels(&pl08x->slave);
2974 out_no_slave:
2975         pl08x_free_virtual_channels(&pl08x->memcpy);
2976 out_no_memcpy:
2977         kfree(pl08x->phy_chans);
2978 out_no_phychans:
2979         free_irq(adev->irq[0], pl08x);
2980 out_no_irq:
2981         dma_pool_destroy(pl08x->pool);
2982 out_no_lli_pool:
2983 out_no_platdata:
2984         iounmap(pl08x->base);
2985 out_no_ioremap:
2986         kfree(pl08x);
2987 out_no_pl08x:
2988         amba_release_regions(adev);
2989         return ret;
2990 }
2991
2992 /* PL080 has 8 channels and the PL080 have just 2 */
2993 static struct vendor_data vendor_pl080 = {
2994         .config_offset = PL080_CH_CONFIG,
2995         .channels = 8,
2996         .signals = 16,
2997         .dualmaster = true,
2998         .max_transfer_size = PL080_CONTROL_TRANSFER_SIZE_MASK,
2999 };
3000
3001 static struct vendor_data vendor_nomadik = {
3002         .config_offset = PL080_CH_CONFIG,
3003         .channels = 8,
3004         .signals = 32,
3005         .dualmaster = true,
3006         .nomadik = true,
3007         .max_transfer_size = PL080_CONTROL_TRANSFER_SIZE_MASK,
3008 };
3009
3010 static struct vendor_data vendor_pl080s = {
3011         .config_offset = PL080S_CH_CONFIG,
3012         .channels = 8,
3013         .signals = 32,
3014         .pl080s = true,
3015         .max_transfer_size = PL080S_CONTROL_TRANSFER_SIZE_MASK,
3016 };
3017
3018 static struct vendor_data vendor_pl081 = {
3019         .config_offset = PL080_CH_CONFIG,
3020         .channels = 2,
3021         .signals = 16,
3022         .dualmaster = false,
3023         .max_transfer_size = PL080_CONTROL_TRANSFER_SIZE_MASK,
3024 };
3025
3026 static struct vendor_data vendor_ftdmac020 = {
3027         .config_offset = PL080_CH_CONFIG,
3028         .ftdmac020 = true,
3029         .max_transfer_size = PL080_CONTROL_TRANSFER_SIZE_MASK,
3030 };
3031
3032 static const struct amba_id pl08x_ids[] = {
3033         /* Samsung PL080S variant */
3034         {
3035                 .id     = 0x0a141080,
3036                 .mask   = 0xffffffff,
3037                 .data   = &vendor_pl080s,
3038         },
3039         /* PL080 */
3040         {
3041                 .id     = 0x00041080,
3042                 .mask   = 0x000fffff,
3043                 .data   = &vendor_pl080,
3044         },
3045         /* PL081 */
3046         {
3047                 .id     = 0x00041081,
3048                 .mask   = 0x000fffff,
3049                 .data   = &vendor_pl081,
3050         },
3051         /* Nomadik 8815 PL080 variant */
3052         {
3053                 .id     = 0x00280080,
3054                 .mask   = 0x00ffffff,
3055                 .data   = &vendor_nomadik,
3056         },
3057         /* Faraday Technology FTDMAC020 */
3058         {
3059                 .id     = 0x0003b080,
3060                 .mask   = 0x000fffff,
3061                 .data   = &vendor_ftdmac020,
3062         },
3063         { 0, 0 },
3064 };
3065
3066 MODULE_DEVICE_TABLE(amba, pl08x_ids);
3067
3068 static struct amba_driver pl08x_amba_driver = {
3069         .drv.name       = DRIVER_NAME,
3070         .id_table       = pl08x_ids,
3071         .probe          = pl08x_probe,
3072 };
3073
3074 static int __init pl08x_init(void)
3075 {
3076         int retval;
3077         retval = amba_driver_register(&pl08x_amba_driver);
3078         if (retval)
3079                 printk(KERN_WARNING DRIVER_NAME
3080                        "failed to register as an AMBA device (%d)\n",
3081                        retval);
3082         return retval;
3083 }
3084 subsys_initcall(pl08x_init);