drivers/dma/pl330.c

   1 /*
   2  * Copyright (c) 2012 Samsung Electronics Co., Ltd.
   3  *              http://www.samsung.com
   4  *
   5  * Copyright (C) 2010 Samsung Electronics Co. Ltd.
   6  *      Jaswinder Singh <jassi.brar@samsung.com>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  */
  13
  14 #include <linux/kernel.h>
  15 #include <linux/io.h>
  16 #include <linux/init.h>
  17 #include <linux/slab.h>
  18 #include <linux/module.h>
  19 #include <linux/string.h>
  20 #include <linux/delay.h>
  21 #include <linux/interrupt.h>
  22 #include <linux/dma-mapping.h>
  23 #include <linux/dmaengine.h>
  24 #include <linux/amba/bus.h>
  25 #include <linux/scatterlist.h>
  26 #include <linux/of.h>
  27 #include <linux/of_dma.h>
  28 #include <linux/err.h>
  29 #include <linux/pm_runtime.h>
  30 #include <linux/bug.h>
  31
  32 #include "dmaengine.h"
  33 #define PL330_MAX_CHAN          8
  34 #define PL330_MAX_IRQS          32
  35 #define PL330_MAX_PERI          32
  36 #define PL330_MAX_BURST         16
  37
  38 #define PL330_QUIRK_BROKEN_NO_FLUSHP BIT(0)
  39
  40 enum pl330_cachectrl {
  41         CCTRL0,         /* Noncacheable and nonbufferable */
  42         CCTRL1,         /* Bufferable only */
  43         CCTRL2,         /* Cacheable, but do not allocate */
  44         CCTRL3,         /* Cacheable and bufferable, but do not allocate */
  45         INVALID1,       /* AWCACHE = 0x1000 */
  46         INVALID2,
  47         CCTRL6,         /* Cacheable write-through, allocate on writes only */
  48         CCTRL7,         /* Cacheable write-back, allocate on writes only */
  49 };
  50
  51 enum pl330_byteswap {
  52         SWAP_NO,
  53         SWAP_2,
  54         SWAP_4,
  55         SWAP_8,
  56         SWAP_16,
  57 };
  58
  59 /* Register and Bit field Definitions */
  60 #define DS                      0x0
  61 #define DS_ST_STOP              0x0
  62 #define DS_ST_EXEC              0x1
  63 #define DS_ST_CMISS             0x2
  64 #define DS_ST_UPDTPC            0x3
  65 #define DS_ST_WFE               0x4
  66 #define DS_ST_ATBRR             0x5
  67 #define DS_ST_QBUSY             0x6
  68 #define DS_ST_WFP               0x7
  69 #define DS_ST_KILL              0x8
  70 #define DS_ST_CMPLT             0x9
  71 #define DS_ST_FLTCMP            0xe
  72 #define DS_ST_FAULT             0xf
  73
  74 #define DPC                     0x4
  75 #define INTEN                   0x20
  76 #define ES                      0x24
  77 #define INTSTATUS               0x28
  78 #define INTCLR                  0x2c
  79 #define FSM                     0x30
  80 #define FSC                     0x34
  81 #define FTM                     0x38
  82
  83 #define _FTC                    0x40
  84 #define FTC(n)                  (_FTC + (n)*0x4)
  85
  86 #define _CS                     0x100
  87 #define CS(n)                   (_CS + (n)*0x8)
  88 #define CS_CNS                  (1 << 21)
  89
  90 #define _CPC                    0x104
  91 #define CPC(n)                  (_CPC + (n)*0x8)
  92
  93 #define _SA                     0x400
  94 #define SA(n)                   (_SA + (n)*0x20)
  95
  96 #define _DA                     0x404
  97 #define DA(n)                   (_DA + (n)*0x20)
  98
  99 #define _CC                     0x408
 100 #define CC(n)                   (_CC + (n)*0x20)
 101
 102 #define CC_SRCINC               (1 << 0)
 103 #define CC_DSTINC               (1 << 14)
 104 #define CC_SRCPRI               (1 << 8)
 105 #define CC_DSTPRI               (1 << 22)
 106 #define CC_SRCNS                (1 << 9)
 107 #define CC_DSTNS                (1 << 23)
 108 #define CC_SRCIA                (1 << 10)
 109 #define CC_DSTIA                (1 << 24)
 110 #define CC_SRCBRSTLEN_SHFT      4
 111 #define CC_DSTBRSTLEN_SHFT      18
 112 #define CC_SRCBRSTSIZE_SHFT     1
 113 #define CC_DSTBRSTSIZE_SHFT     15
 114 #define CC_SRCCCTRL_SHFT        11
 115 #define CC_SRCCCTRL_MASK        0x7
 116 #define CC_DSTCCTRL_SHFT        25
 117 #define CC_DRCCCTRL_MASK        0x7
 118 #define CC_SWAP_SHFT            28
 119
 120 #define _LC0                    0x40c
 121 #define LC0(n)                  (_LC0 + (n)*0x20)
 122
 123 #define _LC1                    0x410
 124 #define LC1(n)                  (_LC1 + (n)*0x20)
 125
 126 #define DBGSTATUS               0xd00
 127 #define DBG_BUSY                (1 << 0)
 128
 129 #define DBGCMD                  0xd04
 130 #define DBGINST0                0xd08
 131 #define DBGINST1                0xd0c
 132
 133 #define CR0                     0xe00
 134 #define CR1                     0xe04
 135 #define CR2                     0xe08
 136 #define CR3                     0xe0c
 137 #define CR4                     0xe10
 138 #define CRD                     0xe14
 139
 140 #define PERIPH_ID               0xfe0
 141 #define PERIPH_REV_SHIFT        20
 142 #define PERIPH_REV_MASK         0xf
 143 #define PERIPH_REV_R0P0         0
 144 #define PERIPH_REV_R1P0         1
 145 #define PERIPH_REV_R1P1         2
 146
 147 #define CR0_PERIPH_REQ_SET      (1 << 0)
 148 #define CR0_BOOT_EN_SET         (1 << 1)
 149 #define CR0_BOOT_MAN_NS         (1 << 2)
 150 #define CR0_NUM_CHANS_SHIFT     4
 151 #define CR0_NUM_CHANS_MASK      0x7
 152 #define CR0_NUM_PERIPH_SHIFT    12
 153 #define CR0_NUM_PERIPH_MASK     0x1f
 154 #define CR0_NUM_EVENTS_SHIFT    17
 155 #define CR0_NUM_EVENTS_MASK     0x1f
 156
 157 #define CR1_ICACHE_LEN_SHIFT    0
 158 #define CR1_ICACHE_LEN_MASK     0x7
 159 #define CR1_NUM_ICACHELINES_SHIFT       4
 160 #define CR1_NUM_ICACHELINES_MASK        0xf
 161
 162 #define CRD_DATA_WIDTH_SHIFT    0
 163 #define CRD_DATA_WIDTH_MASK     0x7
 164 #define CRD_WR_CAP_SHIFT        4
 165 #define CRD_WR_CAP_MASK         0x7
 166 #define CRD_WR_Q_DEP_SHIFT      8
 167 #define CRD_WR_Q_DEP_MASK       0xf
 168 #define CRD_RD_CAP_SHIFT        12
 169 #define CRD_RD_CAP_MASK         0x7
 170 #define CRD_RD_Q_DEP_SHIFT      16
 171 #define CRD_RD_Q_DEP_MASK       0xf
 172 #define CRD_DATA_BUFF_SHIFT     20
 173 #define CRD_DATA_BUFF_MASK      0x3ff
 174
 175 #define PART                    0x330
 176 #define DESIGNER                0x41
 177 #define REVISION                0x0
 178 #define INTEG_CFG               0x0
 179 #define PERIPH_ID_VAL           ((PART << 0) | (DESIGNER << 12))
 180
 181 #define PL330_STATE_STOPPED             (1 << 0)
 182 #define PL330_STATE_EXECUTING           (1 << 1)
 183 #define PL330_STATE_WFE                 (1 << 2)
 184 #define PL330_STATE_FAULTING            (1 << 3)
 185 #define PL330_STATE_COMPLETING          (1 << 4)
 186 #define PL330_STATE_WFP                 (1 << 5)
 187 #define PL330_STATE_KILLING             (1 << 6)
 188 #define PL330_STATE_FAULT_COMPLETING    (1 << 7)
 189 #define PL330_STATE_CACHEMISS           (1 << 8)
 190 #define PL330_STATE_UPDTPC              (1 << 9)
 191 #define PL330_STATE_ATBARRIER           (1 << 10)
 192 #define PL330_STATE_QUEUEBUSY           (1 << 11)
 193 #define PL330_STATE_INVALID             (1 << 15)
 194
 195 #define PL330_STABLE_STATES (PL330_STATE_STOPPED | PL330_STATE_EXECUTING \
 196                                 | PL330_STATE_WFE | PL330_STATE_FAULTING)
 197
 198 #define CMD_DMAADDH             0x54
 199 #define CMD_DMAEND              0x00
 200 #define CMD_DMAFLUSHP           0x35
 201 #define CMD_DMAGO               0xa0
 202 #define CMD_DMALD               0x04
 203 #define CMD_DMALDP              0x25
 204 #define CMD_DMALP               0x20
 205 #define CMD_DMALPEND            0x28
 206 #define CMD_DMAKILL             0x01
 207 #define CMD_DMAMOV              0xbc
 208 #define CMD_DMANOP              0x18
 209 #define CMD_DMARMB              0x12
 210 #define CMD_DMASEV              0x34
 211 #define CMD_DMAST               0x08
 212 #define CMD_DMASTP              0x29
 213 #define CMD_DMASTZ              0x0c
 214 #define CMD_DMAWFE              0x36
 215 #define CMD_DMAWFP              0x30
 216 #define CMD_DMAWMB              0x13
 217
 218 #define SZ_DMAADDH              3
 219 #define SZ_DMAEND               1
 220 #define SZ_DMAFLUSHP            2
 221 #define SZ_DMALD                1
 222 #define SZ_DMALDP               2
 223 #define SZ_DMALP                2
 224 #define SZ_DMALPEND             2
 225 #define SZ_DMAKILL              1
 226 #define SZ_DMAMOV               6
 227 #define SZ_DMANOP               1
 228 #define SZ_DMARMB               1
 229 #define SZ_DMASEV               2
 230 #define SZ_DMAST                1
 231 #define SZ_DMASTP               2
 232 #define SZ_DMASTZ               1
 233 #define SZ_DMAWFE               2
 234 #define SZ_DMAWFP               2
 235 #define SZ_DMAWMB               1
 236 #define SZ_DMAGO                6
 237
 238 #define BRST_LEN(ccr)           ((((ccr) >> CC_SRCBRSTLEN_SHFT) & 0xf) + 1)
 239 #define BRST_SIZE(ccr)          (1 << (((ccr) >> CC_SRCBRSTSIZE_SHFT) & 0x7))
 240
 241 #define BYTE_TO_BURST(b, ccr)   ((b) / BRST_SIZE(ccr) / BRST_LEN(ccr))
 242 #define BURST_TO_BYTE(c, ccr)   ((c) * BRST_SIZE(ccr) * BRST_LEN(ccr))
 243
 244 /*
 245  * With 256 bytes, we can do more than 2.5MB and 5MB xfers per req
 246  * at 1byte/burst for P<->M and M<->M respectively.
 247  * For typical scenario, at 1word/burst, 10MB and 20MB xfers per req
 248  * should be enough for P<->M and M<->M respectively.
 249  */
 250 #define MCODE_BUFF_PER_REQ      256
 251
 252 /* Use this _only_ to wait on transient states */
 253 #define UNTIL(t, s)     while (!(_state(t) & (s))) cpu_relax();
 254
 255 #ifdef PL330_DEBUG_MCGEN
 256 static unsigned cmd_line;
 257 #define PL330_DBGCMD_DUMP(off, x...)    do { \
 258                                                 printk("%x:", cmd_line); \
 259                                                 printk(x); \
 260                                                 cmd_line += off; \
 261                                         } while (0)
 262 #define PL330_DBGMC_START(addr)         (cmd_line = addr)
 263 #else
 264 #define PL330_DBGCMD_DUMP(off, x...)    do {} while (0)
 265 #define PL330_DBGMC_START(addr)         do {} while (0)
 266 #endif
 267
 268 /* The number of default descriptors */
 269
 270 #define NR_DEFAULT_DESC 16
 271
 272 /* Delay for runtime PM autosuspend, ms */
 273 #define PL330_AUTOSUSPEND_DELAY 20
 274
 275 /* Populated by the PL330 core driver for DMA API driver's info */
 276 struct pl330_config {
 277         u32     periph_id;
 278 #define DMAC_MODE_NS    (1 << 0)
 279         unsigned int    mode;
 280         unsigned int    data_bus_width:10; /* In number of bits */
 281         unsigned int    data_buf_dep:11;
 282         unsigned int    num_chan:4;
 283         unsigned int    num_peri:6;
 284         u32             peri_ns;
 285         unsigned int    num_events:6;
 286         u32             irq_ns;
 287 };
 288
 289 /**
 290  * Request Configuration.
 291  * The PL330 core does not modify this and uses the last
 292  * working configuration if the request doesn't provide any.
 293  *
 294  * The Client may want to provide this info only for the
 295  * first request and a request with new settings.
 296  */
 297 struct pl330_reqcfg {
 298         /* Address Incrementing */
 299         unsigned dst_inc:1;
 300         unsigned src_inc:1;
 301
 302         /*
 303          * For now, the SRC & DST protection levels
 304          * and burst size/length are assumed same.
 305          */
 306         bool nonsecure;
 307         bool privileged;
 308         bool insnaccess;
 309         unsigned brst_len:5;
 310         unsigned brst_size:3; /* in power of 2 */
 311
 312         enum pl330_cachectrl dcctl;
 313         enum pl330_cachectrl scctl;
 314         enum pl330_byteswap swap;
 315         struct pl330_config *pcfg;
 316 };
 317
 318 /*
 319  * One cycle of DMAC operation.
 320  * There may be more than one xfer in a request.
 321  */
 322 struct pl330_xfer {
 323         u32 src_addr;
 324         u32 dst_addr;
 325         /* Size to xfer */
 326         u32 bytes;
 327 };
 328
 329 /* The xfer callbacks are made with one of these arguments. */
 330 enum pl330_op_err {
 331         /* The all xfers in the request were success. */
 332         PL330_ERR_NONE,
 333         /* If req aborted due to global error. */
 334         PL330_ERR_ABORT,
 335         /* If req failed due to problem with Channel. */
 336         PL330_ERR_FAIL,
 337 };
 338
 339 enum dmamov_dst {
 340         SAR = 0,
 341         CCR,
 342         DAR,
 343 };
 344
 345 enum pl330_dst {
 346         SRC = 0,
 347         DST,
 348 };
 349
 350 enum pl330_cond {
 351         SINGLE,
 352         BURST,
 353         ALWAYS,
 354 };
 355
 356 struct dma_pl330_desc;
 357
 358 struct _pl330_req {
 359         u32 mc_bus;
 360         void *mc_cpu;
 361         struct dma_pl330_desc *desc;
 362 };
 363
 364 /* ToBeDone for tasklet */
 365 struct _pl330_tbd {
 366         bool reset_dmac;
 367         bool reset_mngr;
 368         u8 reset_chan;
 369 };
 370
 371 /* A DMAC Thread */
 372 struct pl330_thread {
 373         u8 id;
 374         int ev;
 375         /* If the channel is not yet acquired by any client */
 376         bool free;
 377         /* Parent DMAC */
 378         struct pl330_dmac *dmac;
 379         /* Only two at a time */
 380         struct _pl330_req req[2];
 381         /* Index of the last enqueued request */
 382         unsigned lstenq;
 383         /* Index of the last submitted request or -1 if the DMA is stopped */
 384         int req_running;
 385 };
 386
 387 enum pl330_dmac_state {
 388         UNINIT,
 389         INIT,
 390         DYING,
 391 };
 392
 393 enum desc_status {
 394         /* In the DMAC pool */
 395         FREE,
 396         /*
 397          * Allocated to some channel during prep_xxx
 398          * Also may be sitting on the work_list.
 399          */
 400         PREP,
 401         /*
 402          * Sitting on the work_list and already submitted
 403          * to the PL330 core. Not more than two descriptors
 404          * of a channel can be BUSY at any time.
 405          */
 406         BUSY,
 407         /*
 408          * Sitting on the channel work_list but xfer done
 409          * by PL330 core
 410          */
 411         DONE,
 412 };
 413
 414 struct dma_pl330_chan {
 415         /* Schedule desc completion */
 416         struct tasklet_struct task;
 417
 418         /* DMA-Engine Channel */
 419         struct dma_chan chan;
 420
 421         /* List of submitted descriptors */
 422         struct list_head submitted_list;
 423         /* List of issued descriptors */
 424         struct list_head work_list;
 425         /* List of completed descriptors */
 426         struct list_head completed_list;
 427
 428         /* Pointer to the DMAC that manages this channel,
 429          * NULL if the channel is available to be acquired.
 430          * As the parent, this DMAC also provides descriptors
 431          * to the channel.
 432          */
 433         struct pl330_dmac *dmac;
 434
 435         /* To protect channel manipulation */
 436         spinlock_t lock;
 437
 438         /*
 439          * Hardware channel thread of PL330 DMAC. NULL if the channel is
 440          * available.
 441          */
 442         struct pl330_thread *thread;
 443
 444         /* For D-to-M and M-to-D channels */
 445         int burst_sz; /* the peripheral fifo width */
 446         int burst_len; /* the number of burst */
 447         phys_addr_t fifo_addr;
 448         /* DMA-mapped view of the FIFO; may differ if an IOMMU is present */
 449         dma_addr_t fifo_dma;
 450         enum dma_data_direction dir;
 451         struct dma_slave_config slave_config;
 452
 453         /* for cyclic capability */
 454         bool cyclic;
 455
 456         /* for runtime pm tracking */
 457         bool active;
 458 };
 459
 460 struct pl330_dmac {
 461         /* DMA-Engine Device */
 462         struct dma_device ddma;
 463
 464         /* Holds info about sg limitations */
 465         struct device_dma_parameters dma_parms;
 466
 467         /* Pool of descriptors available for the DMAC's channels */
 468         struct list_head desc_pool;
 469         /* To protect desc_pool manipulation */
 470         spinlock_t pool_lock;
 471
 472         /* Size of MicroCode buffers for each channel. */
 473         unsigned mcbufsz;
 474         /* ioremap'ed address of PL330 registers. */
 475         void __iomem    *base;
 476         /* Populated by the PL330 core driver during pl330_add */
 477         struct pl330_config     pcfg;
 478
 479         spinlock_t              lock;
 480         /* Maximum possible events/irqs */
 481         int                     events[32];
 482         /* BUS address of MicroCode buffer */
 483         dma_addr_t              mcode_bus;
 484         /* CPU address of MicroCode buffer */
 485         void                    *mcode_cpu;
 486         /* List of all Channel threads */
 487         struct pl330_thread     *channels;
 488         /* Pointer to the MANAGER thread */
 489         struct pl330_thread     *manager;
 490         /* To handle bad news in interrupt */
 491         struct tasklet_struct   tasks;
 492         struct _pl330_tbd       dmac_tbd;
 493         /* State of DMAC operation */
 494         enum pl330_dmac_state   state;
 495         /* Holds list of reqs with due callbacks */
 496         struct list_head        req_done;
 497
 498         /* Peripheral channels connected to this DMAC */
 499         unsigned int num_peripherals;
 500         struct dma_pl330_chan *peripherals; /* keep at end */
 501         int quirks;
 502 };
 503
 504 static struct pl330_of_quirks {
 505         char *quirk;
 506         int id;
 507 } of_quirks[] = {
 508         {
 509                 .quirk = "arm,pl330-broken-no-flushp",
 510                 .id = PL330_QUIRK_BROKEN_NO_FLUSHP,
 511         }
 512 };
 513
 514 struct dma_pl330_desc {
 515         /* To attach to a queue as child */
 516         struct list_head node;
 517
 518         /* Descriptor for the DMA Engine API */
 519         struct dma_async_tx_descriptor txd;
 520
 521         /* Xfer for PL330 core */
 522         struct pl330_xfer px;
 523
 524         struct pl330_reqcfg rqcfg;
 525
 526         enum desc_status status;
 527
 528         int bytes_requested;
 529         bool last;
 530
 531         /* The channel which currently holds this desc */
 532         struct dma_pl330_chan *pchan;
 533
 534         enum dma_transfer_direction rqtype;
 535         /* Index of peripheral for the xfer. */
 536         unsigned peri:5;
 537         /* Hook to attach to DMAC's list of reqs with due callback */
 538         struct list_head rqd;
 539 };
 540
 541 struct _xfer_spec {
 542         u32 ccr;
 543         struct dma_pl330_desc *desc;
 544 };
 545
 546 static int pl330_config_write(struct dma_chan *chan,
 547                         struct dma_slave_config *slave_config,
 548                         enum dma_transfer_direction direction);
 549
 550 static inline bool _queue_full(struct pl330_thread *thrd)
 551 {
 552         return thrd->req[0].desc != NULL && thrd->req[1].desc != NULL;
 553 }
 554
 555 static inline bool is_manager(struct pl330_thread *thrd)
 556 {
 557         return thrd->dmac->manager == thrd;
 558 }
 559
 560 /* If manager of the thread is in Non-Secure mode */
 561 static inline bool _manager_ns(struct pl330_thread *thrd)
 562 {
 563         return (thrd->dmac->pcfg.mode & DMAC_MODE_NS) ? true : false;
 564 }
 565
 566 static inline u32 get_revision(u32 periph_id)
 567 {
 568         return (periph_id >> PERIPH_REV_SHIFT) & PERIPH_REV_MASK;
 569 }
 570
 571 static inline u32 _emit_END(unsigned dry_run, u8 buf[])
 572 {
 573         if (dry_run)
 574                 return SZ_DMAEND;
 575
 576         buf[0] = CMD_DMAEND;
 577
 578         PL330_DBGCMD_DUMP(SZ_DMAEND, "\tDMAEND\n");
 579
 580         return SZ_DMAEND;
 581 }
 582
 583 static inline u32 _emit_FLUSHP(unsigned dry_run, u8 buf[], u8 peri)
 584 {
 585         if (dry_run)
 586                 return SZ_DMAFLUSHP;
 587
 588         buf[0] = CMD_DMAFLUSHP;
 589
 590         peri &= 0x1f;
 591         peri <<= 3;
 592         buf[1] = peri;
 593
 594         PL330_DBGCMD_DUMP(SZ_DMAFLUSHP, "\tDMAFLUSHP %u\n", peri >> 3);
 595
 596         return SZ_DMAFLUSHP;
 597 }
 598
 599 static inline u32 _emit_LD(unsigned dry_run, u8 buf[],  enum pl330_cond cond)
 600 {
 601         if (dry_run)
 602                 return SZ_DMALD;
 603
 604         buf[0] = CMD_DMALD;
 605
 606         if (cond == SINGLE)
 607                 buf[0] |= (0 << 1) | (1 << 0);
 608         else if (cond == BURST)
 609                 buf[0] |= (1 << 1) | (1 << 0);
 610
 611         PL330_DBGCMD_DUMP(SZ_DMALD, "\tDMALD%c\n",
 612                 cond == SINGLE ? 'S' : (cond == BURST ? 'B' : 'A'));
 613
 614         return SZ_DMALD;
 615 }
 616
 617 static inline u32 _emit_LDP(unsigned dry_run, u8 buf[],
 618                 enum pl330_cond cond, u8 peri)
 619 {
 620         if (dry_run)
 621                 return SZ_DMALDP;
 622
 623         buf[0] = CMD_DMALDP;
 624
 625         if (cond == BURST)
 626                 buf[0] |= (1 << 1);
 627
 628         peri &= 0x1f;
 629         peri <<= 3;
 630         buf[1] = peri;
 631
 632         PL330_DBGCMD_DUMP(SZ_DMALDP, "\tDMALDP%c %u\n",
 633                 cond == SINGLE ? 'S' : 'B', peri >> 3);
 634
 635         return SZ_DMALDP;
 636 }
 637
 638 static inline u32 _emit_LP(unsigned dry_run, u8 buf[],
 639                 unsigned loop, u8 cnt)
 640 {
 641         if (dry_run)
 642                 return SZ_DMALP;
 643
 644         buf[0] = CMD_DMALP;
 645
 646         if (loop)
 647                 buf[0] |= (1 << 1);
 648
 649         cnt--; /* DMAC increments by 1 internally */
 650         buf[1] = cnt;
 651
 652         PL330_DBGCMD_DUMP(SZ_DMALP, "\tDMALP_%c %u\n", loop ? '1' : '0', cnt);
 653
 654         return SZ_DMALP;
 655 }
 656
 657 struct _arg_LPEND {
 658         enum pl330_cond cond;
 659         bool forever;
 660         unsigned loop;
 661         u8 bjump;
 662 };
 663
 664 static inline u32 _emit_LPEND(unsigned dry_run, u8 buf[],
 665                 const struct _arg_LPEND *arg)
 666 {
 667         enum pl330_cond cond = arg->cond;
 668         bool forever = arg->forever;
 669         unsigned loop = arg->loop;
 670         u8 bjump = arg->bjump;
 671
 672         if (dry_run)
 673                 return SZ_DMALPEND;
 674
 675         buf[0] = CMD_DMALPEND;
 676
 677         if (loop)
 678                 buf[0] |= (1 << 2);
 679
 680         if (!forever)
 681                 buf[0] |= (1 << 4);
 682
 683         if (cond == SINGLE)
 684                 buf[0] |= (0 << 1) | (1 << 0);
 685         else if (cond == BURST)
 686                 buf[0] |= (1 << 1) | (1 << 0);
 687
 688         buf[1] = bjump;
 689
 690         PL330_DBGCMD_DUMP(SZ_DMALPEND, "\tDMALP%s%c_%c bjmpto_%x\n",
 691                         forever ? "FE" : "END",
 692                         cond == SINGLE ? 'S' : (cond == BURST ? 'B' : 'A'),
 693                         loop ? '1' : '0',
 694                         bjump);
 695
 696         return SZ_DMALPEND;
 697 }
 698
 699 static inline u32 _emit_KILL(unsigned dry_run, u8 buf[])
 700 {
 701         if (dry_run)
 702                 return SZ_DMAKILL;
 703
 704         buf[0] = CMD_DMAKILL;
 705
 706         return SZ_DMAKILL;
 707 }
 708
 709 static inline u32 _emit_MOV(unsigned dry_run, u8 buf[],
 710                 enum dmamov_dst dst, u32 val)
 711 {
 712         if (dry_run)
 713                 return SZ_DMAMOV;
 714
 715         buf[0] = CMD_DMAMOV;
 716         buf[1] = dst;
 717         buf[2] = val;
 718         buf[3] = val >> 8;
 719         buf[4] = val >> 16;
 720         buf[5] = val >> 24;
 721
 722         PL330_DBGCMD_DUMP(SZ_DMAMOV, "\tDMAMOV %s 0x%x\n",
 723                 dst == SAR ? "SAR" : (dst == DAR ? "DAR" : "CCR"), val);
 724
 725         return SZ_DMAMOV;
 726 }
 727
 728 static inline u32 _emit_RMB(unsigned dry_run, u8 buf[])
 729 {
 730         if (dry_run)
 731                 return SZ_DMARMB;
 732
 733         buf[0] = CMD_DMARMB;
 734
 735         PL330_DBGCMD_DUMP(SZ_DMARMB, "\tDMARMB\n");
 736
 737         return SZ_DMARMB;
 738 }
 739
 740 static inline u32 _emit_SEV(unsigned dry_run, u8 buf[], u8 ev)
 741 {
 742         if (dry_run)
 743                 return SZ_DMASEV;
 744
 745         buf[0] = CMD_DMASEV;
 746
 747         ev &= 0x1f;
 748         ev <<= 3;
 749         buf[1] = ev;
 750
 751         PL330_DBGCMD_DUMP(SZ_DMASEV, "\tDMASEV %u\n", ev >> 3);
 752
 753         return SZ_DMASEV;
 754 }
 755
 756 static inline u32 _emit_ST(unsigned dry_run, u8 buf[], enum pl330_cond cond)
 757 {
 758         if (dry_run)
 759                 return SZ_DMAST;
 760
 761         buf[0] = CMD_DMAST;
 762
 763         if (cond == SINGLE)
 764                 buf[0] |= (0 << 1) | (1 << 0);
 765         else if (cond == BURST)
 766                 buf[0] |= (1 << 1) | (1 << 0);
 767
 768         PL330_DBGCMD_DUMP(SZ_DMAST, "\tDMAST%c\n",
 769                 cond == SINGLE ? 'S' : (cond == BURST ? 'B' : 'A'));
 770
 771         return SZ_DMAST;
 772 }
 773
 774 static inline u32 _emit_STP(unsigned dry_run, u8 buf[],
 775                 enum pl330_cond cond, u8 peri)
 776 {
 777         if (dry_run)
 778                 return SZ_DMASTP;
 779
 780         buf[0] = CMD_DMASTP;
 781
 782         if (cond == BURST)
 783                 buf[0] |= (1 << 1);
 784
 785         peri &= 0x1f;
 786         peri <<= 3;
 787         buf[1] = peri;
 788
 789         PL330_DBGCMD_DUMP(SZ_DMASTP, "\tDMASTP%c %u\n",
 790                 cond == SINGLE ? 'S' : 'B', peri >> 3);
 791
 792         return SZ_DMASTP;
 793 }
 794
 795 static inline u32 _emit_WFP(unsigned dry_run, u8 buf[],
 796                 enum pl330_cond cond, u8 peri)
 797 {
 798         if (dry_run)
 799                 return SZ_DMAWFP;
 800
 801         buf[0] = CMD_DMAWFP;
 802
 803         if (cond == SINGLE)
 804                 buf[0] |= (0 << 1) | (0 << 0);
 805         else if (cond == BURST)
 806                 buf[0] |= (1 << 1) | (0 << 0);
 807         else
 808                 buf[0] |= (0 << 1) | (1 << 0);
 809
 810         peri &= 0x1f;
 811         peri <<= 3;
 812         buf[1] = peri;
 813
 814         PL330_DBGCMD_DUMP(SZ_DMAWFP, "\tDMAWFP%c %u\n",
 815                 cond == SINGLE ? 'S' : (cond == BURST ? 'B' : 'P'), peri >> 3);
 816
 817         return SZ_DMAWFP;
 818 }
 819
 820 static inline u32 _emit_WMB(unsigned dry_run, u8 buf[])
 821 {
 822         if (dry_run)
 823                 return SZ_DMAWMB;
 824
 825         buf[0] = CMD_DMAWMB;
 826
 827         PL330_DBGCMD_DUMP(SZ_DMAWMB, "\tDMAWMB\n");
 828
 829         return SZ_DMAWMB;
 830 }
 831
 832 struct _arg_GO {
 833         u8 chan;
 834         u32 addr;
 835         unsigned ns;
 836 };
 837
 838 static inline u32 _emit_GO(unsigned dry_run, u8 buf[],
 839                 const struct _arg_GO *arg)
 840 {
 841         u8 chan = arg->chan;
 842         u32 addr = arg->addr;
 843         unsigned ns = arg->ns;
 844
 845         if (dry_run)
 846                 return SZ_DMAGO;
 847
 848         buf[0] = CMD_DMAGO;
 849         buf[0] |= (ns << 1);
 850         buf[1] = chan & 0x7;
 851         buf[2] = addr;
 852         buf[3] = addr >> 8;
 853         buf[4] = addr >> 16;
 854         buf[5] = addr >> 24;
 855
 856         return SZ_DMAGO;
 857 }
 858
 859 #define msecs_to_loops(t) (loops_per_jiffy / 1000 * HZ * t)
 860
 861 /* Returns Time-Out */
 862 static bool _until_dmac_idle(struct pl330_thread *thrd)
 863 {
 864         void __iomem *regs = thrd->dmac->base;
 865         unsigned long loops = msecs_to_loops(5);
 866
 867         do {
 868                 /* Until Manager is Idle */
 869                 if (!(readl(regs + DBGSTATUS) & DBG_BUSY))
 870                         break;
 871
 872                 cpu_relax();
 873         } while (--loops);
 874
 875         if (!loops)
 876                 return true;
 877
 878         return false;
 879 }
 880
 881 static inline void _execute_DBGINSN(struct pl330_thread *thrd,
 882                 u8 insn[], bool as_manager)
 883 {
 884         void __iomem *regs = thrd->dmac->base;
 885         u32 val;
 886
 887         val = (insn[0] << 16) | (insn[1] << 24);
 888         if (!as_manager) {
 889                 val |= (1 << 0);
 890                 val |= (thrd->id << 8); /* Channel Number */
 891         }
 892         writel(val, regs + DBGINST0);
 893
 894         val = le32_to_cpu(*((__le32 *)&insn[2]));
 895         writel(val, regs + DBGINST1);
 896
 897         /* If timed out due to halted state-machine */
 898         if (_until_dmac_idle(thrd)) {
 899                 dev_err(thrd->dmac->ddma.dev, "DMAC halted!\n");
 900                 return;
 901         }
 902
 903         /* Get going */
 904         writel(0, regs + DBGCMD);
 905 }
 906
 907 static inline u32 _state(struct pl330_thread *thrd)
 908 {
 909         void __iomem *regs = thrd->dmac->base;
 910         u32 val;
 911
 912         if (is_manager(thrd))
 913                 val = readl(regs + DS) & 0xf;
 914         else
 915                 val = readl(regs + CS(thrd->id)) & 0xf;
 916
 917         switch (val) {
 918         case DS_ST_STOP:
 919                 return PL330_STATE_STOPPED;
 920         case DS_ST_EXEC:
 921                 return PL330_STATE_EXECUTING;
 922         case DS_ST_CMISS:
 923                 return PL330_STATE_CACHEMISS;
 924         case DS_ST_UPDTPC:
 925                 return PL330_STATE_UPDTPC;
 926         case DS_ST_WFE:
 927                 return PL330_STATE_WFE;
 928         case DS_ST_FAULT:
 929                 return PL330_STATE_FAULTING;
 930         case DS_ST_ATBRR:
 931                 if (is_manager(thrd))
 932                         return PL330_STATE_INVALID;
 933                 else
 934                         return PL330_STATE_ATBARRIER;
 935         case DS_ST_QBUSY:
 936                 if (is_manager(thrd))
 937                         return PL330_STATE_INVALID;
 938                 else
 939                         return PL330_STATE_QUEUEBUSY;
 940         case DS_ST_WFP:
 941                 if (is_manager(thrd))
 942                         return PL330_STATE_INVALID;
 943                 else
 944                         return PL330_STATE_WFP;
 945         case DS_ST_KILL:
 946                 if (is_manager(thrd))
 947                         return PL330_STATE_INVALID;
 948                 else
 949                         return PL330_STATE_KILLING;
 950         case DS_ST_CMPLT:
 951                 if (is_manager(thrd))
 952                         return PL330_STATE_INVALID;
 953                 else
 954                         return PL330_STATE_COMPLETING;
 955         case DS_ST_FLTCMP:
 956                 if (is_manager(thrd))
 957                         return PL330_STATE_INVALID;
 958                 else
 959                         return PL330_STATE_FAULT_COMPLETING;
 960         default:
 961                 return PL330_STATE_INVALID;
 962         }
 963 }
 964
 965 static void _stop(struct pl330_thread *thrd)
 966 {
 967         void __iomem *regs = thrd->dmac->base;
 968         u8 insn[6] = {0, 0, 0, 0, 0, 0};
 969
 970         if (_state(thrd) == PL330_STATE_FAULT_COMPLETING)
 971                 UNTIL(thrd, PL330_STATE_FAULTING | PL330_STATE_KILLING);
 972
 973         /* Return if nothing needs to be done */
 974         if (_state(thrd) == PL330_STATE_COMPLETING
 975                   || _state(thrd) == PL330_STATE_KILLING
 976                   || _state(thrd) == PL330_STATE_STOPPED)
 977                 return;
 978
 979         _emit_KILL(0, insn);
 980
 981         /* Stop generating interrupts for SEV */
 982         writel(readl(regs + INTEN) & ~(1 << thrd->ev), regs + INTEN);
 983
 984         _execute_DBGINSN(thrd, insn, is_manager(thrd));
 985 }
 986
 987 /* Start doing req 'idx' of thread 'thrd' */
 988 static bool _trigger(struct pl330_thread *thrd)
 989 {
 990         void __iomem *regs = thrd->dmac->base;
 991         struct _pl330_req *req;
 992         struct dma_pl330_desc *desc;
 993         struct _arg_GO go;
 994         unsigned ns;
 995         u8 insn[6] = {0, 0, 0, 0, 0, 0};
 996         int idx;
 997
 998         /* Return if already ACTIVE */
 999         if (_state(thrd) != PL330_STATE_STOPPED)
1000                 return true;
1001
1002         idx = 1 - thrd->lstenq;
1003         if (thrd->req[idx].desc != NULL) {
1004                 req = &thrd->req[idx];
1005         } else {
1006                 idx = thrd->lstenq;
1007                 if (thrd->req[idx].desc != NULL)
1008                         req = &thrd->req[idx];
1009                 else
1010                         req = NULL;
1011         }
1012
1013         /* Return if no request */
1014         if (!req)
1015                 return true;
1016
1017         /* Return if req is running */
1018         if (idx == thrd->req_running)
1019                 return true;
1020
1021         desc = req->desc;
1022
1023         ns = desc->rqcfg.nonsecure ? 1 : 0;
1024
1025         /* See 'Abort Sources' point-4 at Page 2-25 */
1026         if (_manager_ns(thrd) && !ns)
1027                 dev_info(thrd->dmac->ddma.dev, "%s:%d Recipe for ABORT!\n",
1028                         __func__, __LINE__);
1029
1030         go.chan = thrd->id;
1031         go.addr = req->mc_bus;
1032         go.ns = ns;
1033         _emit_GO(0, insn, &go);
1034
1035         /* Set to generate interrupts for SEV */
1036         writel(readl(regs + INTEN) | (1 << thrd->ev), regs + INTEN);
1037
1038         /* Only manager can execute GO */
1039         _execute_DBGINSN(thrd, insn, true);
1040
1041         thrd->req_running = idx;
1042
1043         return true;
1044 }
1045
1046 static bool _start(struct pl330_thread *thrd)
1047 {
1048         switch (_state(thrd)) {
1049         case PL330_STATE_FAULT_COMPLETING:
1050                 UNTIL(thrd, PL330_STATE_FAULTING | PL330_STATE_KILLING);
1051
1052                 if (_state(thrd) == PL330_STATE_KILLING)
1053                         UNTIL(thrd, PL330_STATE_STOPPED)
1054                 /* fall through */
1055
1056         case PL330_STATE_FAULTING:
1057                 _stop(thrd);
1058                 /* fall through */
1059
1060         case PL330_STATE_KILLING:
1061         case PL330_STATE_COMPLETING:
1062                 UNTIL(thrd, PL330_STATE_STOPPED)
1063                 /* fall through */
1064
1065         case PL330_STATE_STOPPED:
1066                 return _trigger(thrd);
1067
1068         case PL330_STATE_WFP:
1069         case PL330_STATE_QUEUEBUSY:
1070         case PL330_STATE_ATBARRIER:
1071         case PL330_STATE_UPDTPC:
1072         case PL330_STATE_CACHEMISS:
1073         case PL330_STATE_EXECUTING:
1074                 return true;
1075
1076         case PL330_STATE_WFE: /* For RESUME, nothing yet */
1077         default:
1078                 return false;
1079         }
1080 }
1081
1082 static inline int _ldst_memtomem(unsigned dry_run, u8 buf[],
1083                 const struct _xfer_spec *pxs, int cyc)
1084 {
1085         int off = 0;
1086         struct pl330_config *pcfg = pxs->desc->rqcfg.pcfg;
1087
1088         /* check lock-up free version */
1089         if (get_revision(pcfg->periph_id) >= PERIPH_REV_R1P0) {
1090                 while (cyc--) {
1091                         off += _emit_LD(dry_run, &buf[off], ALWAYS);
1092                         off += _emit_ST(dry_run, &buf[off], ALWAYS);
1093                 }
1094         } else {
1095                 while (cyc--) {
1096                         off += _emit_LD(dry_run, &buf[off], ALWAYS);
1097                         off += _emit_RMB(dry_run, &buf[off]);
1098                         off += _emit_ST(dry_run, &buf[off], ALWAYS);
1099                         off += _emit_WMB(dry_run, &buf[off]);
1100                 }
1101         }
1102
1103         return off;
1104 }
1105
1106 static u32 _emit_load(unsigned int dry_run, u8 buf[],
1107         enum pl330_cond cond, enum dma_transfer_direction direction,
1108         u8 peri)
1109 {
1110         int off = 0;
1111
1112         switch (direction) {
1113         case DMA_MEM_TO_MEM:
1114                 /* fall through */
1115         case DMA_MEM_TO_DEV:
1116                 off += _emit_LD(dry_run, &buf[off], cond);
1117                 break;
1118
1119         case DMA_DEV_TO_MEM:
1120                 if (cond == ALWAYS) {
1121                         off += _emit_LDP(dry_run, &buf[off], SINGLE,
1122                                 peri);
1123                         off += _emit_LDP(dry_run, &buf[off], BURST,
1124                                 peri);
1125                 } else {
1126                         off += _emit_LDP(dry_run, &buf[off], cond,
1127                                 peri);
1128                 }
1129                 break;
1130
1131         default:
1132                 /* this code should be unreachable */
1133                 WARN_ON(1);
1134                 break;
1135         }
1136
1137         return off;
1138 }
1139
1140 static inline u32 _emit_store(unsigned int dry_run, u8 buf[],
1141         enum pl330_cond cond, enum dma_transfer_direction direction,
1142         u8 peri)
1143 {
1144         int off = 0;
1145
1146         switch (direction) {
1147         case DMA_MEM_TO_MEM:
1148                 /* fall through */
1149         case DMA_DEV_TO_MEM:
1150                 off += _emit_ST(dry_run, &buf[off], cond);
1151                 break;
1152
1153         case DMA_MEM_TO_DEV:
1154                 if (cond == ALWAYS) {
1155                         off += _emit_STP(dry_run, &buf[off], SINGLE,
1156                                 peri);
1157                         off += _emit_STP(dry_run, &buf[off], BURST,
1158                                 peri);
1159                 } else {
1160                         off += _emit_STP(dry_run, &buf[off], cond,
1161                                 peri);
1162                 }
1163                 break;
1164
1165         default:
1166                 /* this code should be unreachable */
1167                 WARN_ON(1);
1168                 break;
1169         }
1170
1171         return off;
1172 }
1173
1174 static inline int _ldst_peripheral(struct pl330_dmac *pl330,
1175                                  unsigned dry_run, u8 buf[],
1176                                  const struct _xfer_spec *pxs, int cyc,
1177                                  enum pl330_cond cond)
1178 {
1179         int off = 0;
1180
1181         if (pl330->quirks & PL330_QUIRK_BROKEN_NO_FLUSHP)
1182                 cond = BURST;
1183
1184         /*
1185          * do FLUSHP at beginning to clear any stale dma requests before the
1186          * first WFP.
1187          */
1188         if (!(pl330->quirks & PL330_QUIRK_BROKEN_NO_FLUSHP))
1189                 off += _emit_FLUSHP(dry_run, &buf[off], pxs->desc->peri);
1190         while (cyc--) {
1191                 off += _emit_WFP(dry_run, &buf[off], cond, pxs->desc->peri);
1192                 off += _emit_load(dry_run, &buf[off], cond, pxs->desc->rqtype,
1193                         pxs->desc->peri);
1194                 off += _emit_store(dry_run, &buf[off], cond, pxs->desc->rqtype,
1195                         pxs->desc->peri);
1196         }
1197
1198         return off;
1199 }
1200
1201 static int _bursts(struct pl330_dmac *pl330, unsigned dry_run, u8 buf[],
1202                 const struct _xfer_spec *pxs, int cyc)
1203 {
1204         int off = 0;
1205         enum pl330_cond cond = BRST_LEN(pxs->ccr) > 1 ? BURST : SINGLE;
1206
1207         switch (pxs->desc->rqtype) {
1208         case DMA_MEM_TO_DEV:
1209                 /* fall through */
1210         case DMA_DEV_TO_MEM:
1211                 off += _ldst_peripheral(pl330, dry_run, &buf[off], pxs, cyc,
1212                         cond);
1213                 break;
1214
1215         case DMA_MEM_TO_MEM:
1216                 off += _ldst_memtomem(dry_run, &buf[off], pxs, cyc);
1217                 break;
1218
1219         default:
1220                 /* this code should be unreachable */
1221                 WARN_ON(1);
1222                 break;
1223         }
1224
1225         return off;
1226 }
1227
1228 /*
1229  * transfer dregs with single transfers to peripheral, or a reduced size burst
1230  * for mem-to-mem.
1231  */
1232 static int _dregs(struct pl330_dmac *pl330, unsigned int dry_run, u8 buf[],
1233                 const struct _xfer_spec *pxs, int transfer_length)
1234 {
1235         int off = 0;
1236         int dregs_ccr;
1237
1238         if (transfer_length == 0)
1239                 return off;
1240
1241         switch (pxs->desc->rqtype) {
1242         case DMA_MEM_TO_DEV:
1243                 /* fall through */
1244         case DMA_DEV_TO_MEM:
1245                 off += _ldst_peripheral(pl330, dry_run, &buf[off], pxs,
1246                         transfer_length, SINGLE);
1247                 break;
1248
1249         case DMA_MEM_TO_MEM:
1250                 dregs_ccr = pxs->ccr;
1251                 dregs_ccr &= ~((0xf << CC_SRCBRSTLEN_SHFT) |
1252                         (0xf << CC_DSTBRSTLEN_SHFT));
1253                 dregs_ccr |= (((transfer_length - 1) & 0xf) <<
1254                         CC_SRCBRSTLEN_SHFT);
1255                 dregs_ccr |= (((transfer_length - 1) & 0xf) <<
1256                         CC_DSTBRSTLEN_SHFT);
1257                 off += _emit_MOV(dry_run, &buf[off], CCR, dregs_ccr);
1258                 off += _ldst_memtomem(dry_run, &buf[off], pxs, 1);
1259                 break;
1260
1261         default:
1262                 /* this code should be unreachable */
1263                 WARN_ON(1);
1264                 break;
1265         }
1266
1267         return off;
1268 }
1269
1270 /* Returns bytes consumed and updates bursts */
1271 static inline int _loop(struct pl330_dmac *pl330, unsigned dry_run, u8 buf[],
1272                 unsigned long *bursts, const struct _xfer_spec *pxs)
1273 {
1274         int cyc, cycmax, szlp, szlpend, szbrst, off;
1275         unsigned lcnt0, lcnt1, ljmp0, ljmp1;
1276         struct _arg_LPEND lpend;
1277
1278         if (*bursts == 1)
1279                 return _bursts(pl330, dry_run, buf, pxs, 1);
1280
1281         /* Max iterations possible in DMALP is 256 */
1282         if (*bursts >= 256*256) {
1283                 lcnt1 = 256;
1284                 lcnt0 = 256;
1285                 cyc = *bursts / lcnt1 / lcnt0;
1286         } else if (*bursts > 256) {
1287                 lcnt1 = 256;
1288                 lcnt0 = *bursts / lcnt1;
1289                 cyc = 1;
1290         } else {
1291                 lcnt1 = *bursts;
1292                 lcnt0 = 0;
1293                 cyc = 1;
1294         }
1295
1296         szlp = _emit_LP(1, buf, 0, 0);
1297         szbrst = _bursts(pl330, 1, buf, pxs, 1);
1298
1299         lpend.cond = ALWAYS;
1300         lpend.forever = false;
1301         lpend.loop = 0;
1302         lpend.bjump = 0;
1303         szlpend = _emit_LPEND(1, buf, &lpend);
1304
1305         if (lcnt0) {
1306                 szlp *= 2;
1307                 szlpend *= 2;
1308         }
1309
1310         /*
1311          * Max bursts that we can unroll due to limit on the
1312          * size of backward jump that can be encoded in DMALPEND
1313          * which is 8-bits and hence 255
1314          */
1315         cycmax = (255 - (szlp + szlpend)) / szbrst;
1316
1317         cyc = (cycmax < cyc) ? cycmax : cyc;
1318
1319         off = 0;
1320
1321         if (lcnt0) {
1322                 off += _emit_LP(dry_run, &buf[off], 0, lcnt0);
1323                 ljmp0 = off;
1324         }
1325
1326         off += _emit_LP(dry_run, &buf[off], 1, lcnt1);
1327         ljmp1 = off;
1328
1329         off += _bursts(pl330, dry_run, &buf[off], pxs, cyc);
1330
1331         lpend.cond = ALWAYS;
1332         lpend.forever = false;
1333         lpend.loop = 1;
1334         lpend.bjump = off - ljmp1;
1335         off += _emit_LPEND(dry_run, &buf[off], &lpend);
1336
1337         if (lcnt0) {
1338                 lpend.cond = ALWAYS;
1339                 lpend.forever = false;
1340                 lpend.loop = 0;
1341                 lpend.bjump = off - ljmp0;
1342                 off += _emit_LPEND(dry_run, &buf[off], &lpend);
1343         }
1344
1345         *bursts = lcnt1 * cyc;
1346         if (lcnt0)
1347                 *bursts *= lcnt0;
1348
1349         return off;
1350 }
1351
1352 static inline int _setup_loops(struct pl330_dmac *pl330,
1353                                unsigned dry_run, u8 buf[],
1354                                const struct _xfer_spec *pxs)
1355 {
1356         struct pl330_xfer *x = &pxs->desc->px;
1357         u32 ccr = pxs->ccr;
1358         unsigned long c, bursts = BYTE_TO_BURST(x->bytes, ccr);
1359         int num_dregs = (x->bytes - BURST_TO_BYTE(bursts, ccr)) /
1360                 BRST_SIZE(ccr);
1361         int off = 0;
1362
1363         while (bursts) {
1364                 c = bursts;
1365                 off += _loop(pl330, dry_run, &buf[off], &c, pxs);
1366                 bursts -= c;
1367         }
1368         off += _dregs(pl330, dry_run, &buf[off], pxs, num_dregs);
1369
1370         return off;
1371 }
1372
1373 static inline int _setup_xfer(struct pl330_dmac *pl330,
1374                               unsigned dry_run, u8 buf[],
1375                               const struct _xfer_spec *pxs)
1376 {
1377         struct pl330_xfer *x = &pxs->desc->px;
1378         int off = 0;
1379
1380         /* DMAMOV SAR, x->src_addr */
1381         off += _emit_MOV(dry_run, &buf[off], SAR, x->src_addr);
1382         /* DMAMOV DAR, x->dst_addr */
1383         off += _emit_MOV(dry_run, &buf[off], DAR, x->dst_addr);
1384
1385         /* Setup Loop(s) */
1386         off += _setup_loops(pl330, dry_run, &buf[off], pxs);
1387
1388         return off;
1389 }
1390
1391 /*
1392  * A req is a sequence of one or more xfer units.
1393  * Returns the number of bytes taken to setup the MC for the req.
1394  */
1395 static int _setup_req(struct pl330_dmac *pl330, unsigned dry_run,
1396                       struct pl330_thread *thrd, unsigned index,
1397                       struct _xfer_spec *pxs)
1398 {
1399         struct _pl330_req *req = &thrd->req[index];
1400         u8 *buf = req->mc_cpu;
1401         int off = 0;
1402
1403         PL330_DBGMC_START(req->mc_bus);
1404
1405         /* DMAMOV CCR, ccr */
1406         off += _emit_MOV(dry_run, &buf[off], CCR, pxs->ccr);
1407
1408         off += _setup_xfer(pl330, dry_run, &buf[off], pxs);
1409
1410         /* DMASEV peripheral/event */
1411         off += _emit_SEV(dry_run, &buf[off], thrd->ev);
1412         /* DMAEND */
1413         off += _emit_END(dry_run, &buf[off]);
1414
1415         return off;
1416 }
1417
1418 static inline u32 _prepare_ccr(const struct pl330_reqcfg *rqc)
1419 {
1420         u32 ccr = 0;
1421
1422         if (rqc->src_inc)
1423                 ccr |= CC_SRCINC;
1424
1425         if (rqc->dst_inc)
1426                 ccr |= CC_DSTINC;
1427
1428         /* We set same protection levels for Src and DST for now */
1429         if (rqc->privileged)
1430                 ccr |= CC_SRCPRI | CC_DSTPRI;
1431         if (rqc->nonsecure)
1432                 ccr |= CC_SRCNS | CC_DSTNS;
1433         if (rqc->insnaccess)
1434                 ccr |= CC_SRCIA | CC_DSTIA;
1435
1436         ccr |= (((rqc->brst_len - 1) & 0xf) << CC_SRCBRSTLEN_SHFT);
1437         ccr |= (((rqc->brst_len - 1) & 0xf) << CC_DSTBRSTLEN_SHFT);
1438
1439         ccr |= (rqc->brst_size << CC_SRCBRSTSIZE_SHFT);
1440         ccr |= (rqc->brst_size << CC_DSTBRSTSIZE_SHFT);
1441
1442         ccr |= (rqc->scctl << CC_SRCCCTRL_SHFT);
1443         ccr |= (rqc->dcctl << CC_DSTCCTRL_SHFT);
1444
1445         ccr |= (rqc->swap << CC_SWAP_SHFT);
1446
1447         return ccr;
1448 }
1449
1450 /*
1451  * Submit a list of xfers after which the client wants notification.
1452  * Client is not notified after each xfer unit, just once after all
1453  * xfer units are done or some error occurs.
1454  */
1455 static int pl330_submit_req(struct pl330_thread *thrd,
1456         struct dma_pl330_desc *desc)
1457 {
1458         struct pl330_dmac *pl330 = thrd->dmac;
1459         struct _xfer_spec xs;
1460         unsigned long flags;
1461         unsigned idx;
1462         u32 ccr;
1463         int ret = 0;
1464
1465         switch (desc->rqtype) {
1466         case DMA_MEM_TO_DEV:
1467                 break;
1468
1469         case DMA_DEV_TO_MEM:
1470                 break;
1471
1472         case DMA_MEM_TO_MEM:
1473                 break;
1474
1475         default:
1476                 return -ENOTSUPP;
1477         }
1478
1479         if (pl330->state == DYING
1480                 || pl330->dmac_tbd.reset_chan & (1 << thrd->id)) {
1481                 dev_info(thrd->dmac->ddma.dev, "%s:%d\n",
1482                         __func__, __LINE__);
1483                 return -EAGAIN;
1484         }
1485
1486         /* If request for non-existing peripheral */
1487         if (desc->rqtype != DMA_MEM_TO_MEM &&
1488             desc->peri >= pl330->pcfg.num_peri) {
1489                 dev_info(thrd->dmac->ddma.dev,
1490                                 "%s:%d Invalid peripheral(%u)!\n",
1491                                 __func__, __LINE__, desc->peri);
1492                 return -EINVAL;
1493         }
1494
1495         spin_lock_irqsave(&pl330->lock, flags);
1496
1497         if (_queue_full(thrd)) {
1498                 ret = -EAGAIN;
1499                 goto xfer_exit;
1500         }
1501
1502         /* Prefer Secure Channel */
1503         if (!_manager_ns(thrd))
1504                 desc->rqcfg.nonsecure = 0;
1505         else
1506                 desc->rqcfg.nonsecure = 1;
1507
1508         ccr = _prepare_ccr(&desc->rqcfg);
1509
1510         idx = thrd->req[0].desc == NULL ? 0 : 1;
1511
1512         xs.ccr = ccr;
1513         xs.desc = desc;
1514
1515         /* First dry run to check if req is acceptable */
1516         ret = _setup_req(pl330, 1, thrd, idx, &xs);
1517         if (ret < 0)
1518                 goto xfer_exit;
1519
1520         if (ret > pl330->mcbufsz / 2) {
1521                 dev_info(pl330->ddma.dev, "%s:%d Try increasing mcbufsz (%i/%i)\n",
1522                                 __func__, __LINE__, ret, pl330->mcbufsz / 2);
1523                 ret = -ENOMEM;
1524                 goto xfer_exit;
1525         }
1526
1527         /* Hook the request */
1528         thrd->lstenq = idx;
1529         thrd->req[idx].desc = desc;
1530         _setup_req(pl330, 0, thrd, idx, &xs);
1531
1532         ret = 0;
1533
1534 xfer_exit:
1535         spin_unlock_irqrestore(&pl330->lock, flags);
1536
1537         return ret;
1538 }
1539
1540 static void dma_pl330_rqcb(struct dma_pl330_desc *desc, enum pl330_op_err err)
1541 {
1542         struct dma_pl330_chan *pch;
1543         unsigned long flags;
1544
1545         if (!desc)
1546                 return;
1547
1548         pch = desc->pchan;
1549
1550         /* If desc aborted */
1551         if (!pch)
1552                 return;
1553
1554         spin_lock_irqsave(&pch->lock, flags);
1555
1556         desc->status = DONE;
1557
1558         spin_unlock_irqrestore(&pch->lock, flags);
1559
1560         tasklet_schedule(&pch->task);
1561 }
1562
1563 static void pl330_dotask(unsigned long data)
1564 {
1565         struct pl330_dmac *pl330 = (struct pl330_dmac *) data;
1566         unsigned long flags;
1567         int i;
1568
1569         spin_lock_irqsave(&pl330->lock, flags);
1570
1571         /* The DMAC itself gone nuts */
1572         if (pl330->dmac_tbd.reset_dmac) {
1573                 pl330->state = DYING;
1574                 /* Reset the manager too */
1575                 pl330->dmac_tbd.reset_mngr = true;
1576                 /* Clear the reset flag */
1577                 pl330->dmac_tbd.reset_dmac = false;
1578         }
1579
1580         if (pl330->dmac_tbd.reset_mngr) {
1581                 _stop(pl330->manager);
1582                 /* Reset all channels */
1583                 pl330->dmac_tbd.reset_chan = (1 << pl330->pcfg.num_chan) - 1;
1584                 /* Clear the reset flag */
1585                 pl330->dmac_tbd.reset_mngr = false;
1586         }
1587
1588         for (i = 0; i < pl330->pcfg.num_chan; i++) {
1589
1590                 if (pl330->dmac_tbd.reset_chan & (1 << i)) {
1591                         struct pl330_thread *thrd = &pl330->channels[i];
1592                         void __iomem *regs = pl330->base;
1593                         enum pl330_op_err err;
1594
1595                         _stop(thrd);
1596
1597                         if (readl(regs + FSC) & (1 << thrd->id))
1598                                 err = PL330_ERR_FAIL;
1599                         else
1600                                 err = PL330_ERR_ABORT;
1601
1602                         spin_unlock_irqrestore(&pl330->lock, flags);
1603                         dma_pl330_rqcb(thrd->req[1 - thrd->lstenq].desc, err);
1604                         dma_pl330_rqcb(thrd->req[thrd->lstenq].desc, err);
1605                         spin_lock_irqsave(&pl330->lock, flags);
1606
1607                         thrd->req[0].desc = NULL;
1608                         thrd->req[1].desc = NULL;
1609                         thrd->req_running = -1;
1610
1611                         /* Clear the reset flag */
1612                         pl330->dmac_tbd.reset_chan &= ~(1 << i);
1613                 }
1614         }
1615
1616         spin_unlock_irqrestore(&pl330->lock, flags);
1617
1618         return;
1619 }
1620
1621 /* Returns 1 if state was updated, 0 otherwise */
1622 static int pl330_update(struct pl330_dmac *pl330)
1623 {
1624         struct dma_pl330_desc *descdone;
1625         unsigned long flags;
1626         void __iomem *regs;
1627         u32 val;
1628         int id, ev, ret = 0;
1629
1630         regs = pl330->base;
1631
1632         spin_lock_irqsave(&pl330->lock, flags);
1633
1634         val = readl(regs + FSM) & 0x1;
1635         if (val)
1636                 pl330->dmac_tbd.reset_mngr = true;
1637         else
1638                 pl330->dmac_tbd.reset_mngr = false;
1639
1640         val = readl(regs + FSC) & ((1 << pl330->pcfg.num_chan) - 1);
1641         pl330->dmac_tbd.reset_chan |= val;
1642         if (val) {
1643                 int i = 0;
1644                 while (i < pl330->pcfg.num_chan) {
1645                         if (val & (1 << i)) {
1646                                 dev_info(pl330->ddma.dev,
1647                                         "Reset Channel-%d\t CS-%x FTC-%x\n",
1648                                                 i, readl(regs + CS(i)),
1649                                                 readl(regs + FTC(i)));
1650                                 _stop(&pl330->channels[i]);
1651                         }
1652                         i++;
1653                 }
1654         }
1655
1656         /* Check which event happened i.e, thread notified */
1657         val = readl(regs + ES);
1658         if (pl330->pcfg.num_events < 32
1659                         && val & ~((1 << pl330->pcfg.num_events) - 1)) {
1660                 pl330->dmac_tbd.reset_dmac = true;
1661                 dev_err(pl330->ddma.dev, "%s:%d Unexpected!\n", __func__,
1662                         __LINE__);
1663                 ret = 1;
1664                 goto updt_exit;
1665         }
1666
1667         for (ev = 0; ev < pl330->pcfg.num_events; ev++) {
1668                 if (val & (1 << ev)) { /* Event occurred */
1669                         struct pl330_thread *thrd;
1670                         u32 inten = readl(regs + INTEN);
1671                         int active;
1672
1673                         /* Clear the event */
1674                         if (inten & (1 << ev))
1675                                 writel(1 << ev, regs + INTCLR);
1676
1677                         ret = 1;
1678
1679                         id = pl330->events[ev];
1680
1681                         thrd = &pl330->channels[id];
1682
1683                         active = thrd->req_running;
1684                         if (active == -1) /* Aborted */
1685                                 continue;
1686
1687                         /* Detach the req */
1688                         descdone = thrd->req[active].desc;
1689                         thrd->req[active].desc = NULL;
1690
1691                         thrd->req_running = -1;
1692
1693                         /* Get going again ASAP */
1694                         _start(thrd);
1695
1696                         /* For now, just make a list of callbacks to be done */
1697                         list_add_tail(&descdone->rqd, &pl330->req_done);
1698                 }
1699         }
1700
1701         /* Now that we are in no hurry, do the callbacks */
1702         while (!list_empty(&pl330->req_done)) {
1703                 descdone = list_first_entry(&pl330->req_done,
1704                                             struct dma_pl330_desc, rqd);
1705                 list_del(&descdone->rqd);
1706                 spin_unlock_irqrestore(&pl330->lock, flags);
1707                 dma_pl330_rqcb(descdone, PL330_ERR_NONE);
1708                 spin_lock_irqsave(&pl330->lock, flags);
1709         }
1710
1711 updt_exit:
1712         spin_unlock_irqrestore(&pl330->lock, flags);
1713
1714         if (pl330->dmac_tbd.reset_dmac
1715                         || pl330->dmac_tbd.reset_mngr
1716                         || pl330->dmac_tbd.reset_chan) {
1717                 ret = 1;
1718                 tasklet_schedule(&pl330->tasks);
1719         }
1720
1721         return ret;
1722 }
1723
1724 /* Reserve an event */
1725 static inline int _alloc_event(struct pl330_thread *thrd)
1726 {
1727         struct pl330_dmac *pl330 = thrd->dmac;
1728         int ev;
1729
1730         for (ev = 0; ev < pl330->pcfg.num_events; ev++)
1731                 if (pl330->events[ev] == -1) {
1732                         pl330->events[ev] = thrd->id;
1733                         return ev;
1734                 }
1735
1736         return -1;
1737 }
1738
1739 static bool _chan_ns(const struct pl330_dmac *pl330, int i)
1740 {
1741         return pl330->pcfg.irq_ns & (1 << i);
1742 }
1743
1744 /* Upon success, returns IdentityToken for the
1745  * allocated channel, NULL otherwise.
1746  */
1747 static struct pl330_thread *pl330_request_channel(struct pl330_dmac *pl330)
1748 {
1749         struct pl330_thread *thrd = NULL;
1750         int chans, i;
1751
1752         if (pl330->state == DYING)
1753                 return NULL;
1754
1755         chans = pl330->pcfg.num_chan;
1756
1757         for (i = 0; i < chans; i++) {
1758                 thrd = &pl330->channels[i];
1759                 if ((thrd->free) && (!_manager_ns(thrd) ||
1760                                         _chan_ns(pl330, i))) {
1761                         thrd->ev = _alloc_event(thrd);
1762                         if (thrd->ev >= 0) {
1763                                 thrd->free = false;
1764                                 thrd->lstenq = 1;
1765                                 thrd->req[0].desc = NULL;
1766                                 thrd->req[1].desc = NULL;
1767                                 thrd->req_running = -1;
1768                                 break;
1769                         }
1770                 }
1771                 thrd = NULL;
1772         }
1773
1774         return thrd;
1775 }
1776
1777 /* Release an event */
1778 static inline void _free_event(struct pl330_thread *thrd, int ev)
1779 {
1780         struct pl330_dmac *pl330 = thrd->dmac;
1781
1782         /* If the event is valid and was held by the thread */
1783         if (ev >= 0 && ev < pl330->pcfg.num_events
1784                         && pl330->events[ev] == thrd->id)
1785                 pl330->events[ev] = -1;
1786 }
1787
1788 static void pl330_release_channel(struct pl330_thread *thrd)
1789 {
1790         if (!thrd || thrd->free)
1791                 return;
1792
1793         _stop(thrd);
1794
1795         dma_pl330_rqcb(thrd->req[1 - thrd->lstenq].desc, PL330_ERR_ABORT);
1796         dma_pl330_rqcb(thrd->req[thrd->lstenq].desc, PL330_ERR_ABORT);
1797
1798         _free_event(thrd, thrd->ev);
1799         thrd->free = true;
1800 }
1801
1802 /* Initialize the structure for PL330 configuration, that can be used
1803  * by the client driver the make best use of the DMAC
1804  */
1805 static void read_dmac_config(struct pl330_dmac *pl330)
1806 {
1807         void __iomem *regs = pl330->base;
1808         u32 val;
1809
1810         val = readl(regs + CRD) >> CRD_DATA_WIDTH_SHIFT;
1811         val &= CRD_DATA_WIDTH_MASK;
1812         pl330->pcfg.data_bus_width = 8 * (1 << val);
1813
1814         val = readl(regs + CRD) >> CRD_DATA_BUFF_SHIFT;
1815         val &= CRD_DATA_BUFF_MASK;
1816         pl330->pcfg.data_buf_dep = val + 1;
1817
1818         val = readl(regs + CR0) >> CR0_NUM_CHANS_SHIFT;
1819         val &= CR0_NUM_CHANS_MASK;
1820         val += 1;
1821         pl330->pcfg.num_chan = val;
1822
1823         val = readl(regs + CR0);
1824         if (val & CR0_PERIPH_REQ_SET) {
1825                 val = (val >> CR0_NUM_PERIPH_SHIFT) & CR0_NUM_PERIPH_MASK;
1826                 val += 1;
1827                 pl330->pcfg.num_peri = val;
1828                 pl330->pcfg.peri_ns = readl(regs + CR4);
1829         } else {
1830                 pl330->pcfg.num_peri = 0;
1831         }
1832
1833         val = readl(regs + CR0);
1834         if (val & CR0_BOOT_MAN_NS)
1835                 pl330->pcfg.mode |= DMAC_MODE_NS;
1836         else
1837                 pl330->pcfg.mode &= ~DMAC_MODE_NS;
1838
1839         val = readl(regs + CR0) >> CR0_NUM_EVENTS_SHIFT;
1840         val &= CR0_NUM_EVENTS_MASK;
1841         val += 1;
1842         pl330->pcfg.num_events = val;
1843
1844         pl330->pcfg.irq_ns = readl(regs + CR3);
1845 }
1846
1847 static inline void _reset_thread(struct pl330_thread *thrd)
1848 {
1849         struct pl330_dmac *pl330 = thrd->dmac;
1850
1851         thrd->req[0].mc_cpu = pl330->mcode_cpu
1852                                 + (thrd->id * pl330->mcbufsz);
1853         thrd->req[0].mc_bus = pl330->mcode_bus
1854                                 + (thrd->id * pl330->mcbufsz);
1855         thrd->req[0].desc = NULL;
1856
1857         thrd->req[1].mc_cpu = thrd->req[0].mc_cpu
1858                                 + pl330->mcbufsz / 2;
1859         thrd->req[1].mc_bus = thrd->req[0].mc_bus
1860                                 + pl330->mcbufsz / 2;
1861         thrd->req[1].desc = NULL;
1862
1863         thrd->req_running = -1;
1864 }
1865
1866 static int dmac_alloc_threads(struct pl330_dmac *pl330)
1867 {
1868         int chans = pl330->pcfg.num_chan;
1869         struct pl330_thread *thrd;
1870         int i;
1871
1872         /* Allocate 1 Manager and 'chans' Channel threads */
1873         pl330->channels = kcalloc(1 + chans, sizeof(*thrd),
1874                                         GFP_KERNEL);
1875         if (!pl330->channels)
1876                 return -ENOMEM;
1877
1878         /* Init Channel threads */
1879         for (i = 0; i < chans; i++) {
1880                 thrd = &pl330->channels[i];
1881                 thrd->id = i;
1882                 thrd->dmac = pl330;
1883                 _reset_thread(thrd);
1884                 thrd->free = true;
1885         }
1886
1887         /* MANAGER is indexed at the end */
1888         thrd = &pl330->channels[chans];
1889         thrd->id = chans;
1890         thrd->dmac = pl330;
1891         thrd->free = false;
1892         pl330->manager = thrd;
1893
1894         return 0;
1895 }
1896
1897 static int dmac_alloc_resources(struct pl330_dmac *pl330)
1898 {
1899         int chans = pl330->pcfg.num_chan;
1900         int ret;
1901
1902         /*
1903          * Alloc MicroCode buffer for 'chans' Channel threads.
1904          * A channel's buffer offset is (Channel_Id * MCODE_BUFF_PERCHAN)
1905          */
1906         pl330->mcode_cpu = dma_alloc_attrs(pl330->ddma.dev,
1907                                 chans * pl330->mcbufsz,
1908                                 &pl330->mcode_bus, GFP_KERNEL,
1909                                 DMA_ATTR_PRIVILEGED);
1910         if (!pl330->mcode_cpu) {
1911                 dev_err(pl330->ddma.dev, "%s:%d Can't allocate memory!\n",
1912                         __func__, __LINE__);
1913                 return -ENOMEM;
1914         }
1915
1916         ret = dmac_alloc_threads(pl330);
1917         if (ret) {
1918                 dev_err(pl330->ddma.dev, "%s:%d Can't to create channels for DMAC!\n",
1919                         __func__, __LINE__);
1920                 dma_free_coherent(pl330->ddma.dev,
1921                                 chans * pl330->mcbufsz,
1922                                 pl330->mcode_cpu, pl330->mcode_bus);
1923                 return ret;
1924         }
1925
1926         return 0;
1927 }
1928
1929 static int pl330_add(struct pl330_dmac *pl330)
1930 {
1931         int i, ret;
1932
1933         /* Check if we can handle this DMAC */
1934         if ((pl330->pcfg.periph_id & 0xfffff) != PERIPH_ID_VAL) {
1935                 dev_err(pl330->ddma.dev, "PERIPH_ID 0x%x !\n",
1936                         pl330->pcfg.periph_id);
1937                 return -EINVAL;
1938         }
1939
1940         /* Read the configuration of the DMAC */
1941         read_dmac_config(pl330);
1942
1943         if (pl330->pcfg.num_events == 0) {
1944                 dev_err(pl330->ddma.dev, "%s:%d Can't work without events!\n",
1945                         __func__, __LINE__);
1946                 return -EINVAL;
1947         }
1948
1949         spin_lock_init(&pl330->lock);
1950
1951         INIT_LIST_HEAD(&pl330->req_done);
1952
1953         /* Use default MC buffer size if not provided */
1954         if (!pl330->mcbufsz)
1955                 pl330->mcbufsz = MCODE_BUFF_PER_REQ * 2;
1956
1957         /* Mark all events as free */
1958         for (i = 0; i < pl330->pcfg.num_events; i++)
1959                 pl330->events[i] = -1;
1960
1961         /* Allocate resources needed by the DMAC */
1962         ret = dmac_alloc_resources(pl330);
1963         if (ret) {
1964                 dev_err(pl330->ddma.dev, "Unable to create channels for DMAC\n");
1965                 return ret;
1966         }
1967
1968         tasklet_init(&pl330->tasks, pl330_dotask, (unsigned long) pl330);
1969
1970         pl330->state = INIT;
1971
1972         return 0;
1973 }
1974
1975 static int dmac_free_threads(struct pl330_dmac *pl330)
1976 {
1977         struct pl330_thread *thrd;
1978         int i;
1979
1980         /* Release Channel threads */
1981         for (i = 0; i < pl330->pcfg.num_chan; i++) {
1982                 thrd = &pl330->channels[i];
1983                 pl330_release_channel(thrd);
1984         }
1985
1986         /* Free memory */
1987         kfree(pl330->channels);
1988
1989         return 0;
1990 }
1991
1992 static void pl330_del(struct pl330_dmac *pl330)
1993 {
1994         pl330->state = UNINIT;
1995
1996         tasklet_kill(&pl330->tasks);
1997
1998         /* Free DMAC resources */
1999         dmac_free_threads(pl330);
2000
2001         dma_free_coherent(pl330->ddma.dev,
2002                 pl330->pcfg.num_chan * pl330->mcbufsz, pl330->mcode_cpu,
2003                 pl330->mcode_bus);
2004 }
2005
2006 /* forward declaration */
2007 static struct amba_driver pl330_driver;
2008
2009 static inline struct dma_pl330_chan *
2010 to_pchan(struct dma_chan *ch)
2011 {
2012         if (!ch)
2013                 return NULL;
2014
2015         return container_of(ch, struct dma_pl330_chan, chan);
2016 }
2017
2018 static inline struct dma_pl330_desc *
2019 to_desc(struct dma_async_tx_descriptor *tx)
2020 {
2021         return container_of(tx, struct dma_pl330_desc, txd);
2022 }
2023
2024 static inline void fill_queue(struct dma_pl330_chan *pch)
2025 {
2026         struct dma_pl330_desc *desc;
2027         int ret;
2028
2029         list_for_each_entry(desc, &pch->work_list, node) {
2030
2031                 /* If already submitted */
2032                 if (desc->status == BUSY)
2033                         continue;
2034
2035                 ret = pl330_submit_req(pch->thread, desc);
2036                 if (!ret) {
2037                         desc->status = BUSY;
2038                 } else if (ret == -EAGAIN) {
2039                         /* QFull or DMAC Dying */
2040                         break;
2041                 } else {
2042                         /* Unacceptable request */
2043                         desc->status = DONE;
2044                         dev_err(pch->dmac->ddma.dev, "%s:%d Bad Desc(%d)\n",
2045                                         __func__, __LINE__, desc->txd.cookie);
2046                         tasklet_schedule(&pch->task);
2047                 }
2048         }
2049 }
2050
2051 static void pl330_tasklet(unsigned long data)
2052 {
2053         struct dma_pl330_chan *pch = (struct dma_pl330_chan *)data;
2054         struct dma_pl330_desc *desc, *_dt;
2055         unsigned long flags;
2056         bool power_down = false;
2057
2058         spin_lock_irqsave(&pch->lock, flags);
2059
2060         /* Pick up ripe tomatoes */
2061         list_for_each_entry_safe(desc, _dt, &pch->work_list, node)
2062                 if (desc->status == DONE) {
2063                         if (!pch->cyclic)
2064                                 dma_cookie_complete(&desc->txd);
2065                         list_move_tail(&desc->node, &pch->completed_list);
2066                 }
2067
2068         /* Try to submit a req imm. next to the last completed cookie */
2069         fill_queue(pch);
2070
2071         if (list_empty(&pch->work_list)) {
2072                 spin_lock(&pch->thread->dmac->lock);
2073                 _stop(pch->thread);
2074                 spin_unlock(&pch->thread->dmac->lock);
2075                 power_down = true;
2076                 pch->active = false;
2077         } else {
2078                 /* Make sure the PL330 Channel thread is active */
2079                 spin_lock(&pch->thread->dmac->lock);
2080                 _start(pch->thread);
2081                 spin_unlock(&pch->thread->dmac->lock);
2082         }
2083
2084         while (!list_empty(&pch->completed_list)) {
2085                 struct dmaengine_desc_callback cb;
2086
2087                 desc = list_first_entry(&pch->completed_list,
2088                                         struct dma_pl330_desc, node);
2089
2090                 dmaengine_desc_get_callback(&desc->txd, &cb);
2091
2092                 if (pch->cyclic) {
2093                         desc->status = PREP;
2094                         list_move_tail(&desc->node, &pch->work_list);
2095                         if (power_down) {
2096                                 pch->active = true;
2097                                 spin_lock(&pch->thread->dmac->lock);
2098                                 _start(pch->thread);
2099                                 spin_unlock(&pch->thread->dmac->lock);
2100                                 power_down = false;
2101                         }
2102                 } else {
2103                         desc->status = FREE;
2104                         list_move_tail(&desc->node, &pch->dmac->desc_pool);
2105                 }
2106
2107                 dma_descriptor_unmap(&desc->txd);
2108
2109                 if (dmaengine_desc_callback_valid(&cb)) {
2110                         spin_unlock_irqrestore(&pch->lock, flags);
2111                         dmaengine_desc_callback_invoke(&cb, NULL);
2112                         spin_lock_irqsave(&pch->lock, flags);
2113                 }
2114         }
2115         spin_unlock_irqrestore(&pch->lock, flags);
2116
2117         /* If work list empty, power down */
2118         if (power_down) {
2119                 pm_runtime_mark_last_busy(pch->dmac->ddma.dev);
2120                 pm_runtime_put_autosuspend(pch->dmac->ddma.dev);
2121         }
2122 }
2123
2124 static struct dma_chan *of_dma_pl330_xlate(struct of_phandle_args *dma_spec,
2125                                                 struct of_dma *ofdma)
2126 {
2127         int count = dma_spec->args_count;
2128         struct pl330_dmac *pl330 = ofdma->of_dma_data;
2129         unsigned int chan_id;
2130
2131         if (!pl330)
2132                 return NULL;
2133
2134         if (count != 1)
2135                 return NULL;
2136
2137         chan_id = dma_spec->args[0];
2138         if (chan_id >= pl330->num_peripherals)
2139                 return NULL;
2140
2141         return dma_get_slave_channel(&pl330->peripherals[chan_id].chan);
2142 }
2143
2144 static int pl330_alloc_chan_resources(struct dma_chan *chan)
2145 {
2146         struct dma_pl330_chan *pch = to_pchan(chan);
2147         struct pl330_dmac *pl330 = pch->dmac;
2148         unsigned long flags;
2149
2150         spin_lock_irqsave(&pl330->lock, flags);
2151
2152         dma_cookie_init(chan);
2153         pch->cyclic = false;
2154
2155         pch->thread = pl330_request_channel(pl330);
2156         if (!pch->thread) {
2157                 spin_unlock_irqrestore(&pl330->lock, flags);
2158                 return -ENOMEM;
2159         }
2160
2161         tasklet_init(&pch->task, pl330_tasklet, (unsigned long) pch);
2162
2163         spin_unlock_irqrestore(&pl330->lock, flags);
2164
2165         return 1;
2166 }
2167
2168 /*
2169  * We need the data direction between the DMAC (the dma-mapping "device") and
2170  * the FIFO (the dmaengine "dev"), from the FIFO's point of view. Confusing!
2171  */
2172 static enum dma_data_direction
2173 pl330_dma_slave_map_dir(enum dma_transfer_direction dir)
2174 {
2175         switch (dir) {
2176         case DMA_MEM_TO_DEV:
2177                 return DMA_FROM_DEVICE;
2178         case DMA_DEV_TO_MEM:
2179                 return DMA_TO_DEVICE;
2180         case DMA_DEV_TO_DEV:
2181                 return DMA_BIDIRECTIONAL;
2182         default:
2183                 return DMA_NONE;
2184         }
2185 }
2186
2187 static void pl330_unprep_slave_fifo(struct dma_pl330_chan *pch)
2188 {
2189         if (pch->dir != DMA_NONE)
2190                 dma_unmap_resource(pch->chan.device->dev, pch->fifo_dma,
2191                                    1 << pch->burst_sz, pch->dir, 0);
2192         pch->dir = DMA_NONE;
2193 }
2194
2195
2196 static bool pl330_prep_slave_fifo(struct dma_pl330_chan *pch,
2197                                   enum dma_transfer_direction dir)
2198 {
2199         struct device *dev = pch->chan.device->dev;
2200         enum dma_data_direction dma_dir = pl330_dma_slave_map_dir(dir);
2201
2202         /* Already mapped for this config? */
2203         if (pch->dir == dma_dir)
2204                 return true;
2205
2206         pl330_unprep_slave_fifo(pch);
2207         pch->fifo_dma = dma_map_resource(dev, pch->fifo_addr,
2208                                          1 << pch->burst_sz, dma_dir, 0);
2209         if (dma_mapping_error(dev, pch->fifo_dma))
2210                 return false;
2211
2212         pch->dir = dma_dir;
2213         return true;
2214 }
2215
2216 static int fixup_burst_len(int max_burst_len, int quirks)
2217 {
2218         if (quirks & PL330_QUIRK_BROKEN_NO_FLUSHP)
2219                 return 1;
2220         else if (max_burst_len > PL330_MAX_BURST)
2221                 return PL330_MAX_BURST;
2222         else if (max_burst_len < 1)
2223                 return 1;
2224         else
2225                 return max_burst_len;
2226 }
2227
2228 static int pl330_config_write(struct dma_chan *chan,
2229                         struct dma_slave_config *slave_config,
2230                         enum dma_transfer_direction direction)
2231 {
2232         struct dma_pl330_chan *pch = to_pchan(chan);
2233
2234         pl330_unprep_slave_fifo(pch);
2235         if (direction == DMA_MEM_TO_DEV) {
2236                 if (slave_config->dst_addr)
2237                         pch->fifo_addr = slave_config->dst_addr;
2238                 if (slave_config->dst_addr_width)
2239                         pch->burst_sz = __ffs(slave_config->dst_addr_width);
2240                 pch->burst_len = fixup_burst_len(slave_config->dst_maxburst,
2241                         pch->dmac->quirks);
2242         } else if (direction == DMA_DEV_TO_MEM) {
2243                 if (slave_config->src_addr)
2244                         pch->fifo_addr = slave_config->src_addr;
2245                 if (slave_config->src_addr_width)
2246                         pch->burst_sz = __ffs(slave_config->src_addr_width);
2247                 pch->burst_len = fixup_burst_len(slave_config->src_maxburst,
2248                         pch->dmac->quirks);
2249         }
2250
2251         return 0;
2252 }
2253
2254 static int pl330_config(struct dma_chan *chan,
2255                         struct dma_slave_config *slave_config)
2256 {
2257         struct dma_pl330_chan *pch = to_pchan(chan);
2258
2259         memcpy(&pch->slave_config, slave_config, sizeof(*slave_config));
2260
2261         return 0;
2262 }
2263
2264 static int pl330_terminate_all(struct dma_chan *chan)
2265 {
2266         struct dma_pl330_chan *pch = to_pchan(chan);
2267         struct dma_pl330_desc *desc;
2268         unsigned long flags;
2269         struct pl330_dmac *pl330 = pch->dmac;
2270         bool power_down = false;
2271
2272         pm_runtime_get_sync(pl330->ddma.dev);
2273         spin_lock_irqsave(&pch->lock, flags);
2274
2275         spin_lock(&pl330->lock);
2276         _stop(pch->thread);
2277         pch->thread->req[0].desc = NULL;
2278         pch->thread->req[1].desc = NULL;
2279         pch->thread->req_running = -1;
2280         spin_unlock(&pl330->lock);
2281
2282         power_down = pch->active;
2283         pch->active = false;
2284
2285         /* Mark all desc done */
2286         list_for_each_entry(desc, &pch->submitted_list, node) {
2287                 desc->status = FREE;
2288                 dma_cookie_complete(&desc->txd);
2289         }
2290
2291         list_for_each_entry(desc, &pch->work_list , node) {
2292                 desc->status = FREE;
2293                 dma_cookie_complete(&desc->txd);
2294         }
2295
2296         list_splice_tail_init(&pch->submitted_list, &pl330->desc_pool);
2297         list_splice_tail_init(&pch->work_list, &pl330->desc_pool);
2298         list_splice_tail_init(&pch->completed_list, &pl330->desc_pool);
2299         spin_unlock_irqrestore(&pch->lock, flags);
2300         pm_runtime_mark_last_busy(pl330->ddma.dev);
2301         if (power_down)
2302                 pm_runtime_put_autosuspend(pl330->ddma.dev);
2303         pm_runtime_put_autosuspend(pl330->ddma.dev);
2304
2305         return 0;
2306 }
2307
2308 /*
2309  * We don't support DMA_RESUME command because of hardware
2310  * limitations, so after pausing the channel we cannot restore
2311  * it to active state. We have to terminate channel and setup
2312  * DMA transfer again. This pause feature was implemented to
2313  * allow safely read residue before channel termination.
2314  */
2315 static int pl330_pause(struct dma_chan *chan)
2316 {
2317         struct dma_pl330_chan *pch = to_pchan(chan);
2318         struct pl330_dmac *pl330 = pch->dmac;
2319         unsigned long flags;
2320
2321         pm_runtime_get_sync(pl330->ddma.dev);
2322         spin_lock_irqsave(&pch->lock, flags);
2323
2324         spin_lock(&pl330->lock);
2325         _stop(pch->thread);
2326         spin_unlock(&pl330->lock);
2327
2328         spin_unlock_irqrestore(&pch->lock, flags);
2329         pm_runtime_mark_last_busy(pl330->ddma.dev);
2330         pm_runtime_put_autosuspend(pl330->ddma.dev);
2331
2332         return 0;
2333 }
2334
2335 static void pl330_free_chan_resources(struct dma_chan *chan)
2336 {
2337         struct dma_pl330_chan *pch = to_pchan(chan);
2338         struct pl330_dmac *pl330 = pch->dmac;
2339         unsigned long flags;
2340
2341         tasklet_kill(&pch->task);
2342
2343         pm_runtime_get_sync(pch->dmac->ddma.dev);
2344         spin_lock_irqsave(&pl330->lock, flags);
2345
2346         pl330_release_channel(pch->thread);
2347         pch->thread = NULL;
2348
2349         if (pch->cyclic)
2350                 list_splice_tail_init(&pch->work_list, &pch->dmac->desc_pool);
2351
2352         spin_unlock_irqrestore(&pl330->lock, flags);
2353         pm_runtime_mark_last_busy(pch->dmac->ddma.dev);
2354         pm_runtime_put_autosuspend(pch->dmac->ddma.dev);
2355         pl330_unprep_slave_fifo(pch);
2356 }
2357
2358 static int pl330_get_current_xferred_count(struct dma_pl330_chan *pch,
2359                                            struct dma_pl330_desc *desc)
2360 {
2361         struct pl330_thread *thrd = pch->thread;
2362         struct pl330_dmac *pl330 = pch->dmac;
2363         void __iomem *regs = thrd->dmac->base;
2364         u32 val, addr;
2365
2366         pm_runtime_get_sync(pl330->ddma.dev);
2367         val = addr = 0;
2368         if (desc->rqcfg.src_inc) {
2369                 val = readl(regs + SA(thrd->id));
2370                 addr = desc->px.src_addr;
2371         } else {
2372                 val = readl(regs + DA(thrd->id));
2373                 addr = desc->px.dst_addr;
2374         }
2375         pm_runtime_mark_last_busy(pch->dmac->ddma.dev);
2376         pm_runtime_put_autosuspend(pl330->ddma.dev);
2377
2378         /* If DMAMOV hasn't finished yet, SAR/DAR can be zero */
2379         if (!val)
2380                 return 0;
2381
2382         return val - addr;
2383 }
2384
2385 static enum dma_status
2386 pl330_tx_status(struct dma_chan *chan, dma_cookie_t cookie,
2387                  struct dma_tx_state *txstate)
2388 {
2389         enum dma_status ret;
2390         unsigned long flags;
2391         struct dma_pl330_desc *desc, *running = NULL, *last_enq = NULL;
2392         struct dma_pl330_chan *pch = to_pchan(chan);
2393         unsigned int transferred, residual = 0;
2394
2395         ret = dma_cookie_status(chan, cookie, txstate);
2396
2397         if (!txstate)
2398                 return ret;
2399
2400         if (ret == DMA_COMPLETE)
2401                 goto out;
2402
2403         spin_lock_irqsave(&pch->lock, flags);
2404         spin_lock(&pch->thread->dmac->lock);
2405
2406         if (pch->thread->req_running != -1)
2407                 running = pch->thread->req[pch->thread->req_running].desc;
2408
2409         last_enq = pch->thread->req[pch->thread->lstenq].desc;
2410
2411         /* Check in pending list */
2412         list_for_each_entry(desc, &pch->work_list, node) {
2413                 if (desc->status == DONE)
2414                         transferred = desc->bytes_requested;
2415                 else if (running && desc == running)
2416                         transferred =
2417                                 pl330_get_current_xferred_count(pch, desc);
2418                 else if (desc->status == BUSY)
2419                         /*
2420                          * Busy but not running means either just enqueued,
2421                          * or finished and not yet marked done
2422                          */
2423                         if (desc == last_enq)
2424                                 transferred = 0;
2425                         else
2426                                 transferred = desc->bytes_requested;
2427                 else
2428                         transferred = 0;
2429                 residual += desc->bytes_requested - transferred;
2430                 if (desc->txd.cookie == cookie) {
2431                         switch (desc->status) {
2432                         case DONE:
2433                                 ret = DMA_COMPLETE;
2434                                 break;
2435                         case PREP:
2436                         case BUSY:
2437                                 ret = DMA_IN_PROGRESS;
2438                                 break;
2439                         default:
2440                                 WARN_ON(1);
2441                         }
2442                         break;
2443                 }
2444                 if (desc->last)
2445                         residual = 0;
2446         }
2447         spin_unlock(&pch->thread->dmac->lock);
2448         spin_unlock_irqrestore(&pch->lock, flags);
2449
2450 out:
2451         dma_set_residue(txstate, residual);
2452
2453         return ret;
2454 }
2455
2456 static void pl330_issue_pending(struct dma_chan *chan)
2457 {
2458         struct dma_pl330_chan *pch = to_pchan(chan);
2459         unsigned long flags;
2460
2461         spin_lock_irqsave(&pch->lock, flags);
2462         if (list_empty(&pch->work_list)) {
2463                 /*
2464                  * Warn on nothing pending. Empty submitted_list may
2465                  * break our pm_runtime usage counter as it is
2466                  * updated on work_list emptiness status.
2467                  */
2468                 WARN_ON(list_empty(&pch->submitted_list));
2469                 pch->active = true;
2470                 pm_runtime_get_sync(pch->dmac->ddma.dev);
2471         }
2472         list_splice_tail_init(&pch->submitted_list, &pch->work_list);
2473         spin_unlock_irqrestore(&pch->lock, flags);
2474
2475         pl330_tasklet((unsigned long)pch);
2476 }
2477
2478 /*
2479  * We returned the last one of the circular list of descriptor(s)
2480  * from prep_xxx, so the argument to submit corresponds to the last
2481  * descriptor of the list.
2482  */
2483 static dma_cookie_t pl330_tx_submit(struct dma_async_tx_descriptor *tx)
2484 {
2485         struct dma_pl330_desc *desc, *last = to_desc(tx);
2486         struct dma_pl330_chan *pch = to_pchan(tx->chan);
2487         dma_cookie_t cookie;
2488         unsigned long flags;
2489
2490         spin_lock_irqsave(&pch->lock, flags);
2491
2492         /* Assign cookies to all nodes */
2493         while (!list_empty(&last->node)) {
2494                 desc = list_entry(last->node.next, struct dma_pl330_desc, node);
2495                 if (pch->cyclic) {
2496                         desc->txd.callback = last->txd.callback;
2497                         desc->txd.callback_param = last->txd.callback_param;
2498                 }
2499                 desc->last = false;
2500
2501                 dma_cookie_assign(&desc->txd);
2502
2503                 list_move_tail(&desc->node, &pch->submitted_list);
2504         }
2505
2506         last->last = true;
2507         cookie = dma_cookie_assign(&last->txd);
2508         list_add_tail(&last->node, &pch->submitted_list);
2509         spin_unlock_irqrestore(&pch->lock, flags);
2510
2511         return cookie;
2512 }
2513
2514 static inline void _init_desc(struct dma_pl330_desc *desc)
2515 {
2516         desc->rqcfg.swap = SWAP_NO;
2517         desc->rqcfg.scctl = CCTRL0;
2518         desc->rqcfg.dcctl = CCTRL0;
2519         desc->txd.tx_submit = pl330_tx_submit;
2520
2521         INIT_LIST_HEAD(&desc->node);
2522 }
2523
2524 /* Returns the number of descriptors added to the DMAC pool */
2525 static int add_desc(struct list_head *pool, spinlock_t *lock,
2526                     gfp_t flg, int count)
2527 {
2528         struct dma_pl330_desc *desc;
2529         unsigned long flags;
2530         int i;
2531
2532         desc = kcalloc(count, sizeof(*desc), flg);
2533         if (!desc)
2534                 return 0;
2535
2536         spin_lock_irqsave(lock, flags);
2537
2538         for (i = 0; i < count; i++) {
2539                 _init_desc(&desc[i]);
2540                 list_add_tail(&desc[i].node, pool);
2541         }
2542
2543         spin_unlock_irqrestore(lock, flags);
2544
2545         return count;
2546 }
2547
2548 static struct dma_pl330_desc *pluck_desc(struct list_head *pool,
2549                                          spinlock_t *lock)
2550 {
2551         struct dma_pl330_desc *desc = NULL;
2552         unsigned long flags;
2553
2554         spin_lock_irqsave(lock, flags);
2555
2556         if (!list_empty(pool)) {
2557                 desc = list_entry(pool->next,
2558                                 struct dma_pl330_desc, node);
2559
2560                 list_del_init(&desc->node);
2561
2562                 desc->status = PREP;
2563                 desc->txd.callback = NULL;
2564         }
2565
2566         spin_unlock_irqrestore(lock, flags);
2567
2568         return desc;
2569 }
2570
2571 static struct dma_pl330_desc *pl330_get_desc(struct dma_pl330_chan *pch)
2572 {
2573         struct pl330_dmac *pl330 = pch->dmac;
2574         u8 *peri_id = pch->chan.private;
2575         struct dma_pl330_desc *desc;
2576
2577         /* Pluck one desc from the pool of DMAC */
2578         desc = pluck_desc(&pl330->desc_pool, &pl330->pool_lock);
2579
2580         /* If the DMAC pool is empty, alloc new */
2581         if (!desc) {
2582                 DEFINE_SPINLOCK(lock);
2583                 LIST_HEAD(pool);
2584
2585                 if (!add_desc(&pool, &lock, GFP_ATOMIC, 1))
2586                         return NULL;
2587
2588                 desc = pluck_desc(&pool, &lock);
2589                 WARN_ON(!desc || !list_empty(&pool));
2590         }
2591
2592         /* Initialize the descriptor */
2593         desc->pchan = pch;
2594         desc->txd.cookie = 0;
2595         async_tx_ack(&desc->txd);
2596
2597         desc->peri = peri_id ? pch->chan.chan_id : 0;
2598         desc->rqcfg.pcfg = &pch->dmac->pcfg;
2599
2600         dma_async_tx_descriptor_init(&desc->txd, &pch->chan);
2601
2602         return desc;
2603 }
2604
2605 static inline void fill_px(struct pl330_xfer *px,
2606                 dma_addr_t dst, dma_addr_t src, size_t len)
2607 {
2608         px->bytes = len;
2609         px->dst_addr = dst;
2610         px->src_addr = src;
2611 }
2612
2613 static struct dma_pl330_desc *
2614 __pl330_prep_dma_memcpy(struct dma_pl330_chan *pch, dma_addr_t dst,
2615                 dma_addr_t src, size_t len)
2616 {
2617         struct dma_pl330_desc *desc = pl330_get_desc(pch);
2618
2619         if (!desc) {
2620                 dev_err(pch->dmac->ddma.dev, "%s:%d Unable to fetch desc\n",
2621                         __func__, __LINE__);
2622                 return NULL;
2623         }
2624
2625         /*
2626          * Ideally we should lookout for reqs bigger than
2627          * those that can be programmed with 256 bytes of
2628          * MC buffer, but considering a req size is seldom
2629          * going to be word-unaligned and more than 200MB,
2630          * we take it easy.
2631          * Also, should the limit is reached we'd rather
2632          * have the platform increase MC buffer size than
2633          * complicating this API driver.
2634          */
2635         fill_px(&desc->px, dst, src, len);
2636
2637         return desc;
2638 }
2639
2640 /* Call after fixing burst size */
2641 static inline int get_burst_len(struct dma_pl330_desc *desc, size_t len)
2642 {
2643         struct dma_pl330_chan *pch = desc->pchan;
2644         struct pl330_dmac *pl330 = pch->dmac;
2645         int burst_len;
2646
2647         burst_len = pl330->pcfg.data_bus_width / 8;
2648         burst_len *= pl330->pcfg.data_buf_dep / pl330->pcfg.num_chan;
2649         burst_len >>= desc->rqcfg.brst_size;
2650
2651         /* src/dst_burst_len can't be more than 16 */
2652         if (burst_len > PL330_MAX_BURST)
2653                 burst_len = PL330_MAX_BURST;
2654
2655         return burst_len;
2656 }
2657
2658 static struct dma_async_tx_descriptor *pl330_prep_dma_cyclic(
2659                 struct dma_chan *chan, dma_addr_t dma_addr, size_t len,
2660                 size_t period_len, enum dma_transfer_direction direction,
2661                 unsigned long flags)
2662 {
2663         struct dma_pl330_desc *desc = NULL, *first = NULL;
2664         struct dma_pl330_chan *pch = to_pchan(chan);
2665         struct pl330_dmac *pl330 = pch->dmac;
2666         unsigned int i;
2667         dma_addr_t dst;
2668         dma_addr_t src;
2669
2670         if (len % period_len != 0)
2671                 return NULL;
2672
2673         if (!is_slave_direction(direction)) {
2674                 dev_err(pch->dmac->ddma.dev, "%s:%d Invalid dma direction\n",
2675                 __func__, __LINE__);
2676                 return NULL;
2677         }
2678
2679         pl330_config_write(chan, &pch->slave_config, direction);
2680
2681         if (!pl330_prep_slave_fifo(pch, direction))
2682                 return NULL;
2683
2684         for (i = 0; i < len / period_len; i++) {
2685                 desc = pl330_get_desc(pch);
2686                 if (!desc) {
2687                         dev_err(pch->dmac->ddma.dev, "%s:%d Unable to fetch desc\n",
2688                                 __func__, __LINE__);
2689
2690                         if (!first)
2691                                 return NULL;
2692
2693                         spin_lock_irqsave(&pl330->pool_lock, flags);
2694
2695                         while (!list_empty(&first->node)) {
2696                                 desc = list_entry(first->node.next,
2697                                                 struct dma_pl330_desc, node);
2698                                 list_move_tail(&desc->node, &pl330->desc_pool);
2699                         }
2700
2701                         list_move_tail(&first->node, &pl330->desc_pool);
2702
2703                         spin_unlock_irqrestore(&pl330->pool_lock, flags);
2704
2705                         return NULL;
2706                 }
2707
2708                 switch (direction) {
2709                 case DMA_MEM_TO_DEV:
2710                         desc->rqcfg.src_inc = 1;
2711                         desc->rqcfg.dst_inc = 0;
2712                         src = dma_addr;
2713                         dst = pch->fifo_dma;
2714                         break;
2715                 case DMA_DEV_TO_MEM:
2716                         desc->rqcfg.src_inc = 0;
2717                         desc->rqcfg.dst_inc = 1;
2718                         src = pch->fifo_dma;
2719                         dst = dma_addr;
2720                         break;
2721                 default:
2722                         break;
2723                 }
2724
2725                 desc->rqtype = direction;
2726                 desc->rqcfg.brst_size = pch->burst_sz;
2727                 desc->rqcfg.brst_len = pch->burst_len;
2728                 desc->bytes_requested = period_len;
2729                 fill_px(&desc->px, dst, src, period_len);
2730
2731                 if (!first)
2732                         first = desc;
2733                 else
2734                         list_add_tail(&desc->node, &first->node);
2735
2736                 dma_addr += period_len;
2737         }
2738
2739         if (!desc)
2740                 return NULL;
2741
2742         pch->cyclic = true;
2743         desc->txd.flags = flags;
2744
2745         return &desc->txd;
2746 }
2747
2748 static struct dma_async_tx_descriptor *
2749 pl330_prep_dma_memcpy(struct dma_chan *chan, dma_addr_t dst,
2750                 dma_addr_t src, size_t len, unsigned long flags)
2751 {
2752         struct dma_pl330_desc *desc;
2753         struct dma_pl330_chan *pch = to_pchan(chan);
2754         struct pl330_dmac *pl330;
2755         int burst;
2756
2757         if (unlikely(!pch || !len))
2758                 return NULL;
2759
2760         pl330 = pch->dmac;
2761
2762         desc = __pl330_prep_dma_memcpy(pch, dst, src, len);
2763         if (!desc)
2764                 return NULL;
2765
2766         desc->rqcfg.src_inc = 1;
2767         desc->rqcfg.dst_inc = 1;
2768         desc->rqtype = DMA_MEM_TO_MEM;
2769
2770         /* Select max possible burst size */
2771         burst = pl330->pcfg.data_bus_width / 8;
2772
2773         /*
2774          * Make sure we use a burst size that aligns with all the memcpy
2775          * parameters because our DMA programming algorithm doesn't cope with
2776          * transfers which straddle an entry in the DMA device's MFIFO.
2777          */
2778         while ((src | dst | len) & (burst - 1))
2779                 burst /= 2;
2780
2781         desc->rqcfg.brst_size = 0;
2782         while (burst != (1 << desc->rqcfg.brst_size))
2783                 desc->rqcfg.brst_size++;
2784
2785         /*
2786          * If burst size is smaller than bus width then make sure we only
2787          * transfer one at a time to avoid a burst stradling an MFIFO entry.
2788          */
2789         if (desc->rqcfg.brst_size * 8 < pl330->pcfg.data_bus_width)
2790                 desc->rqcfg.brst_len = 1;
2791
2792         desc->rqcfg.brst_len = get_burst_len(desc, len);
2793         desc->bytes_requested = len;
2794
2795         desc->txd.flags = flags;
2796
2797         return &desc->txd;
2798 }
2799
2800 static void __pl330_giveback_desc(struct pl330_dmac *pl330,
2801                                   struct dma_pl330_desc *first)
2802 {
2803         unsigned long flags;
2804         struct dma_pl330_desc *desc;
2805
2806         if (!first)
2807                 return;
2808
2809         spin_lock_irqsave(&pl330->pool_lock, flags);
2810
2811         while (!list_empty(&first->node)) {
2812                 desc = list_entry(first->node.next,
2813                                 struct dma_pl330_desc, node);
2814                 list_move_tail(&desc->node, &pl330->desc_pool);
2815         }
2816
2817         list_move_tail(&first->node, &pl330->desc_pool);
2818
2819         spin_unlock_irqrestore(&pl330->pool_lock, flags);
2820 }
2821
2822 static struct dma_async_tx_descriptor *
2823 pl330_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl,
2824                 unsigned int sg_len, enum dma_transfer_direction direction,
2825                 unsigned long flg, void *context)
2826 {
2827         struct dma_pl330_desc *first, *desc = NULL;
2828         struct dma_pl330_chan *pch = to_pchan(chan);
2829         struct scatterlist *sg;
2830         int i;
2831
2832         if (unlikely(!pch || !sgl || !sg_len))
2833                 return NULL;
2834
2835         pl330_config_write(chan, &pch->slave_config, direction);
2836
2837         if (!pl330_prep_slave_fifo(pch, direction))
2838                 return NULL;
2839
2840         first = NULL;
2841
2842         for_each_sg(sgl, sg, sg_len, i) {
2843
2844                 desc = pl330_get_desc(pch);
2845                 if (!desc) {
2846                         struct pl330_dmac *pl330 = pch->dmac;
2847
2848                         dev_err(pch->dmac->ddma.dev,
2849                                 "%s:%d Unable to fetch desc\n",
2850                                 __func__, __LINE__);
2851                         __pl330_giveback_desc(pl330, first);
2852
2853                         return NULL;
2854                 }
2855
2856                 if (!first)
2857                         first = desc;
2858                 else
2859                         list_add_tail(&desc->node, &first->node);
2860
2861                 if (direction == DMA_MEM_TO_DEV) {
2862                         desc->rqcfg.src_inc = 1;
2863                         desc->rqcfg.dst_inc = 0;
2864                         fill_px(&desc->px, pch->fifo_dma, sg_dma_address(sg),
2865                                 sg_dma_len(sg));
2866                 } else {
2867                         desc->rqcfg.src_inc = 0;
2868                         desc->rqcfg.dst_inc = 1;
2869                         fill_px(&desc->px, sg_dma_address(sg), pch->fifo_dma,
2870                                 sg_dma_len(sg));
2871                 }
2872
2873                 desc->rqcfg.brst_size = pch->burst_sz;
2874                 desc->rqcfg.brst_len = pch->burst_len;
2875                 desc->rqtype = direction;
2876                 desc->bytes_requested = sg_dma_len(sg);
2877         }
2878
2879         /* Return the last desc in the chain */
2880         desc->txd.flags = flg;
2881         return &desc->txd;
2882 }
2883
2884 static irqreturn_t pl330_irq_handler(int irq, void *data)
2885 {
2886         if (pl330_update(data))
2887                 return IRQ_HANDLED;
2888         else
2889                 return IRQ_NONE;
2890 }
2891
2892 #define PL330_DMA_BUSWIDTHS \
2893         BIT(DMA_SLAVE_BUSWIDTH_UNDEFINED) | \
2894         BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) | \
2895         BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) | \
2896         BIT(DMA_SLAVE_BUSWIDTH_4_BYTES) | \
2897         BIT(DMA_SLAVE_BUSWIDTH_8_BYTES)
2898
2899 /*
2900  * Runtime PM callbacks are provided by amba/bus.c driver.
2901  *
2902  * It is assumed here that IRQ safe runtime PM is chosen in probe and amba
2903  * bus driver will only disable/enable the clock in runtime PM callbacks.
2904  */
2905 static int __maybe_unused pl330_suspend(struct device *dev)
2906 {
2907         struct amba_device *pcdev = to_amba_device(dev);
2908
2909         pm_runtime_disable(dev);
2910
2911         if (!pm_runtime_status_suspended(dev)) {
2912                 /* amba did not disable the clock */
2913                 amba_pclk_disable(pcdev);
2914         }
2915         amba_pclk_unprepare(pcdev);
2916
2917         return 0;
2918 }
2919
2920 static int __maybe_unused pl330_resume(struct device *dev)
2921 {
2922         struct amba_device *pcdev = to_amba_device(dev);
2923         int ret;
2924
2925         ret = amba_pclk_prepare(pcdev);
2926         if (ret)
2927                 return ret;
2928
2929         if (!pm_runtime_status_suspended(dev))
2930                 ret = amba_pclk_enable(pcdev);
2931
2932         pm_runtime_enable(dev);
2933
2934         return ret;
2935 }
2936
2937 static SIMPLE_DEV_PM_OPS(pl330_pm, pl330_suspend, pl330_resume);
2938
2939 static int
2940 pl330_probe(struct amba_device *adev, const struct amba_id *id)
2941 {
2942         struct pl330_config *pcfg;
2943         struct pl330_dmac *pl330;
2944         struct dma_pl330_chan *pch, *_p;
2945         struct dma_device *pd;
2946         struct resource *res;
2947         int i, ret, irq;
2948         int num_chan;
2949         struct device_node *np = adev->dev.of_node;
2950
2951         ret = dma_set_mask_and_coherent(&adev->dev, DMA_BIT_MASK(32));
2952         if (ret)
2953                 return ret;
2954
2955         /* Allocate a new DMAC and its Channels */
2956         pl330 = devm_kzalloc(&adev->dev, sizeof(*pl330), GFP_KERNEL);
2957         if (!pl330)
2958                 return -ENOMEM;
2959
2960         pd = &pl330->ddma;
2961         pd->dev = &adev->dev;
2962
2963         pl330->mcbufsz = 0;
2964
2965         /* get quirk */
2966         for (i = 0; i < ARRAY_SIZE(of_quirks); i++)
2967                 if (of_property_read_bool(np, of_quirks[i].quirk))
2968                         pl330->quirks |= of_quirks[i].id;
2969
2970         res = &adev->res;
2971         pl330->base = devm_ioremap_resource(&adev->dev, res);
2972         if (IS_ERR(pl330->base))
2973                 return PTR_ERR(pl330->base);
2974
2975         amba_set_drvdata(adev, pl330);
2976
2977         for (i = 0; i < AMBA_NR_IRQS; i++) {
2978                 irq = adev->irq[i];
2979                 if (irq) {
2980                         ret = devm_request_irq(&adev->dev, irq,
2981                                                pl330_irq_handler, 0,
2982                                                dev_name(&adev->dev), pl330);
2983                         if (ret)
2984                                 return ret;
2985                 } else {
2986                         break;
2987                 }
2988         }
2989
2990         pcfg = &pl330->pcfg;
2991
2992         pcfg->periph_id = adev->periphid;
2993         ret = pl330_add(pl330);
2994         if (ret)
2995                 return ret;
2996
2997         INIT_LIST_HEAD(&pl330->desc_pool);
2998         spin_lock_init(&pl330->pool_lock);
2999
3000         /* Create a descriptor pool of default size */
3001         if (!add_desc(&pl330->desc_pool, &pl330->pool_lock,
3002                       GFP_KERNEL, NR_DEFAULT_DESC))
3003                 dev_warn(&adev->dev, "unable to allocate desc\n");
3004
3005         INIT_LIST_HEAD(&pd->channels);
3006
3007         /* Initialize channel parameters */
3008         num_chan = max_t(int, pcfg->num_peri, pcfg->num_chan);
3009
3010         pl330->num_peripherals = num_chan;
3011
3012         pl330->peripherals = kcalloc(num_chan, sizeof(*pch), GFP_KERNEL);
3013         if (!pl330->peripherals) {
3014                 ret = -ENOMEM;
3015                 goto probe_err2;
3016         }
3017
3018         for (i = 0; i < num_chan; i++) {
3019                 pch = &pl330->peripherals[i];
3020
3021                 pch->chan.private = adev->dev.of_node;
3022                 INIT_LIST_HEAD(&pch->submitted_list);
3023                 INIT_LIST_HEAD(&pch->work_list);
3024                 INIT_LIST_HEAD(&pch->completed_list);
3025                 spin_lock_init(&pch->lock);
3026                 pch->thread = NULL;
3027                 pch->chan.device = pd;
3028                 pch->dmac = pl330;
3029                 pch->dir = DMA_NONE;
3030
3031                 /* Add the channel to the DMAC list */
3032                 list_add_tail(&pch->chan.device_node, &pd->channels);
3033         }
3034
3035         dma_cap_set(DMA_MEMCPY, pd->cap_mask);
3036         if (pcfg->num_peri) {
3037                 dma_cap_set(DMA_SLAVE, pd->cap_mask);
3038                 dma_cap_set(DMA_CYCLIC, pd->cap_mask);
3039                 dma_cap_set(DMA_PRIVATE, pd->cap_mask);
3040         }
3041
3042         pd->device_alloc_chan_resources = pl330_alloc_chan_resources;
3043         pd->device_free_chan_resources = pl330_free_chan_resources;
3044         pd->device_prep_dma_memcpy = pl330_prep_dma_memcpy;
3045         pd->device_prep_dma_cyclic = pl330_prep_dma_cyclic;
3046         pd->device_tx_status = pl330_tx_status;
3047         pd->device_prep_slave_sg = pl330_prep_slave_sg;
3048         pd->device_config = pl330_config;
3049         pd->device_pause = pl330_pause;
3050         pd->device_terminate_all = pl330_terminate_all;
3051         pd->device_issue_pending = pl330_issue_pending;
3052         pd->src_addr_widths = PL330_DMA_BUSWIDTHS;
3053         pd->dst_addr_widths = PL330_DMA_BUSWIDTHS;
3054         pd->directions = BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV);
3055         pd->residue_granularity = DMA_RESIDUE_GRANULARITY_BURST;
3056         pd->max_burst = ((pl330->quirks & PL330_QUIRK_BROKEN_NO_FLUSHP) ?
3057                          1 : PL330_MAX_BURST);
3058
3059         ret = dma_async_device_register(pd);
3060         if (ret) {
3061                 dev_err(&adev->dev, "unable to register DMAC\n");
3062                 goto probe_err3;
3063         }
3064
3065         if (adev->dev.of_node) {
3066                 ret = of_dma_controller_register(adev->dev.of_node,
3067                                          of_dma_pl330_xlate, pl330);
3068                 if (ret) {
3069                         dev_err(&adev->dev,
3070                         "unable to register DMA to the generic DT DMA helpers\n");
3071                 }
3072         }
3073
3074         adev->dev.dma_parms = &pl330->dma_parms;
3075
3076         /*
3077          * This is the limit for transfers with a buswidth of 1, larger
3078          * buswidths will have larger limits.
3079          */
3080         ret = dma_set_max_seg_size(&adev->dev, 1900800);
3081         if (ret)
3082                 dev_err(&adev->dev, "unable to set the seg size\n");
3083
3084
3085         dev_info(&adev->dev,
3086                 "Loaded driver for PL330 DMAC-%x\n", adev->periphid);
3087         dev_info(&adev->dev,
3088                 "\tDBUFF-%ux%ubytes Num_Chans-%u Num_Peri-%u Num_Events-%u\n",
3089                 pcfg->data_buf_dep, pcfg->data_bus_width / 8, pcfg->num_chan,
3090                 pcfg->num_peri, pcfg->num_events);
3091
3092         pm_runtime_irq_safe(&adev->dev);
3093         pm_runtime_use_autosuspend(&adev->dev);
3094         pm_runtime_set_autosuspend_delay(&adev->dev, PL330_AUTOSUSPEND_DELAY);
3095         pm_runtime_mark_last_busy(&adev->dev);
3096         pm_runtime_put_autosuspend(&adev->dev);
3097
3098         return 0;
3099 probe_err3:
3100         /* Idle the DMAC */
3101         list_for_each_entry_safe(pch, _p, &pl330->ddma.channels,
3102                         chan.device_node) {
3103
3104                 /* Remove the channel */
3105                 list_del(&pch->chan.device_node);
3106
3107                 /* Flush the channel */
3108                 if (pch->thread) {
3109                         pl330_terminate_all(&pch->chan);
3110                         pl330_free_chan_resources(&pch->chan);
3111                 }
3112         }
3113 probe_err2:
3114         pl330_del(pl330);
3115
3116         return ret;
3117 }
3118
3119 static int pl330_remove(struct amba_device *adev)
3120 {
3121         struct pl330_dmac *pl330 = amba_get_drvdata(adev);
3122         struct dma_pl330_chan *pch, *_p;
3123         int i, irq;
3124
3125         pm_runtime_get_noresume(pl330->ddma.dev);
3126
3127         if (adev->dev.of_node)
3128                 of_dma_controller_free(adev->dev.of_node);
3129
3130         for (i = 0; i < AMBA_NR_IRQS; i++) {
3131                 irq = adev->irq[i];
3132                 if (irq)
3133                         devm_free_irq(&adev->dev, irq, pl330);
3134         }
3135
3136         dma_async_device_unregister(&pl330->ddma);
3137
3138         /* Idle the DMAC */
3139         list_for_each_entry_safe(pch, _p, &pl330->ddma.channels,
3140                         chan.device_node) {
3141
3142                 /* Remove the channel */
3143                 list_del(&pch->chan.device_node);
3144
3145                 /* Flush the channel */
3146                 if (pch->thread) {
3147                         pl330_terminate_all(&pch->chan);
3148                         pl330_free_chan_resources(&pch->chan);
3149                 }
3150         }
3151
3152         pl330_del(pl330);
3153
3154         return 0;
3155 }
3156
3157 static const struct amba_id pl330_ids[] = {
3158         {
3159                 .id     = 0x00041330,
3160                 .mask   = 0x000fffff,
3161         },
3162         { 0, 0 },
3163 };
3164
3165 MODULE_DEVICE_TABLE(amba, pl330_ids);
3166
3167 static struct amba_driver pl330_driver = {
3168         .drv = {
3169                 .owner = THIS_MODULE,
3170                 .name = "dma-pl330",
3171                 .pm = &pl330_pm,
3172         },
3173         .id_table = pl330_ids,
3174         .probe = pl330_probe,
3175         .remove = pl330_remove,
3176 };
3177
3178 module_amba_driver(pl330_driver);
3179
3180 MODULE_AUTHOR("Jaswinder Singh <jassisinghbrar@gmail.com>");
3181 MODULE_DESCRIPTION("API Driver for PL330 DMAC");
3182 MODULE_LICENSE("GPL");