drivers/firewire/ohci.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Driver for OHCI 1394 controllers
   4  *
   5  * Copyright (C) 2003-2006 Kristian Hoegsberg <krh@bitplanet.net>
   6  */
   7
   8 #include <linux/bitops.h>
   9 #include <linux/bug.h>
  10 #include <linux/compiler.h>
  11 #include <linux/delay.h>
  12 #include <linux/device.h>
  13 #include <linux/dma-mapping.h>
  14 #include <linux/firewire.h>
  15 #include <linux/firewire-constants.h>
  16 #include <linux/init.h>
  17 #include <linux/interrupt.h>
  18 #include <linux/io.h>
  19 #include <linux/kernel.h>
  20 #include <linux/list.h>
  21 #include <linux/mm.h>
  22 #include <linux/module.h>
  23 #include <linux/moduleparam.h>
  24 #include <linux/mutex.h>
  25 #include <linux/pci.h>
  26 #include <linux/pci_ids.h>
  27 #include <linux/slab.h>
  28 #include <linux/spinlock.h>
  29 #include <linux/string.h>
  30 #include <linux/time.h>
  31 #include <linux/vmalloc.h>
  32 #include <linux/workqueue.h>
  33
  34 #include <asm/byteorder.h>
  35 #include <asm/page.h>
  36
  37 #ifdef CONFIG_PPC_PMAC
  38 #include <asm/pmac_feature.h>
  39 #endif
  40
  41 #include "core.h"
  42 #include "ohci.h"
  43
  44 #define ohci_info(ohci, f, args...)     dev_info(ohci->card.device, f, ##args)
  45 #define ohci_notice(ohci, f, args...)   dev_notice(ohci->card.device, f, ##args)
  46 #define ohci_err(ohci, f, args...)      dev_err(ohci->card.device, f, ##args)
  47
  48 #define DESCRIPTOR_OUTPUT_MORE          0
  49 #define DESCRIPTOR_OUTPUT_LAST          (1 << 12)
  50 #define DESCRIPTOR_INPUT_MORE           (2 << 12)
  51 #define DESCRIPTOR_INPUT_LAST           (3 << 12)
  52 #define DESCRIPTOR_STATUS               (1 << 11)
  53 #define DESCRIPTOR_KEY_IMMEDIATE        (2 << 8)
  54 #define DESCRIPTOR_PING                 (1 << 7)
  55 #define DESCRIPTOR_YY                   (1 << 6)
  56 #define DESCRIPTOR_NO_IRQ               (0 << 4)
  57 #define DESCRIPTOR_IRQ_ERROR            (1 << 4)
  58 #define DESCRIPTOR_IRQ_ALWAYS           (3 << 4)
  59 #define DESCRIPTOR_BRANCH_ALWAYS        (3 << 2)
  60 #define DESCRIPTOR_WAIT                 (3 << 0)
  61
  62 #define DESCRIPTOR_CMD                  (0xf << 12)
  63
  64 struct descriptor {
  65         __le16 req_count;
  66         __le16 control;
  67         __le32 data_address;
  68         __le32 branch_address;
  69         __le16 res_count;
  70         __le16 transfer_status;
  71 } __attribute__((aligned(16)));
  72
  73 #define CONTROL_SET(regs)       (regs)
  74 #define CONTROL_CLEAR(regs)     ((regs) + 4)
  75 #define COMMAND_PTR(regs)       ((regs) + 12)
  76 #define CONTEXT_MATCH(regs)     ((regs) + 16)
  77
  78 #define AR_BUFFER_SIZE  (32*1024)
  79 #define AR_BUFFERS_MIN  DIV_ROUND_UP(AR_BUFFER_SIZE, PAGE_SIZE)
  80 /* we need at least two pages for proper list management */
  81 #define AR_BUFFERS      (AR_BUFFERS_MIN >= 2 ? AR_BUFFERS_MIN : 2)
  82
  83 #define MAX_ASYNC_PAYLOAD       4096
  84 #define MAX_AR_PACKET_SIZE      (16 + MAX_ASYNC_PAYLOAD + 4)
  85 #define AR_WRAPAROUND_PAGES     DIV_ROUND_UP(MAX_AR_PACKET_SIZE, PAGE_SIZE)
  86
  87 struct ar_context {
  88         struct fw_ohci *ohci;
  89         struct page *pages[AR_BUFFERS];
  90         void *buffer;
  91         struct descriptor *descriptors;
  92         dma_addr_t descriptors_bus;
  93         void *pointer;
  94         unsigned int last_buffer_index;
  95         u32 regs;
  96         struct tasklet_struct tasklet;
  97 };
  98
  99 struct context;
 100
 101 typedef int (*descriptor_callback_t)(struct context *ctx,
 102                                      struct descriptor *d,
 103                                      struct descriptor *last);
 104
 105 /*
 106  * A buffer that contains a block of DMA-able coherent memory used for
 107  * storing a portion of a DMA descriptor program.
 108  */
 109 struct descriptor_buffer {
 110         struct list_head list;
 111         dma_addr_t buffer_bus;
 112         size_t buffer_size;
 113         size_t used;
 114         struct descriptor buffer[];
 115 };
 116
 117 struct context {
 118         struct fw_ohci *ohci;
 119         u32 regs;
 120         int total_allocation;
 121         u32 current_bus;
 122         bool running;
 123         bool flushing;
 124
 125         /*
 126          * List of page-sized buffers for storing DMA descriptors.
 127          * Head of list contains buffers in use and tail of list contains
 128          * free buffers.
 129          */
 130         struct list_head buffer_list;
 131
 132         /*
 133          * Pointer to a buffer inside buffer_list that contains the tail
 134          * end of the current DMA program.
 135          */
 136         struct descriptor_buffer *buffer_tail;
 137
 138         /*
 139          * The descriptor containing the branch address of the first
 140          * descriptor that has not yet been filled by the device.
 141          */
 142         struct descriptor *last;
 143
 144         /*
 145          * The last descriptor block in the DMA program. It contains the branch
 146          * address that must be updated upon appending a new descriptor.
 147          */
 148         struct descriptor *prev;
 149         int prev_z;
 150
 151         descriptor_callback_t callback;
 152
 153         struct tasklet_struct tasklet;
 154 };
 155
 156 #define IT_HEADER_SY(v)          ((v) <<  0)
 157 #define IT_HEADER_TCODE(v)       ((v) <<  4)
 158 #define IT_HEADER_CHANNEL(v)     ((v) <<  8)
 159 #define IT_HEADER_TAG(v)         ((v) << 14)
 160 #define IT_HEADER_SPEED(v)       ((v) << 16)
 161 #define IT_HEADER_DATA_LENGTH(v) ((v) << 16)
 162
 163 struct iso_context {
 164         struct fw_iso_context base;
 165         struct context context;
 166         void *header;
 167         size_t header_length;
 168         unsigned long flushing_completions;
 169         u32 mc_buffer_bus;
 170         u16 mc_completed;
 171         u16 last_timestamp;
 172         u8 sync;
 173         u8 tags;
 174 };
 175
 176 #define CONFIG_ROM_SIZE 1024
 177
 178 struct fw_ohci {
 179         struct fw_card card;
 180
 181         __iomem char *registers;
 182         int node_id;
 183         int generation;
 184         int request_generation; /* for timestamping incoming requests */
 185         unsigned quirks;
 186         unsigned int pri_req_max;
 187         u32 bus_time;
 188         bool bus_time_running;
 189         bool is_root;
 190         bool csr_state_setclear_abdicate;
 191         int n_ir;
 192         int n_it;
 193         /*
 194          * Spinlock for accessing fw_ohci data.  Never call out of
 195          * this driver with this lock held.
 196          */
 197         spinlock_t lock;
 198
 199         struct mutex phy_reg_mutex;
 200
 201         void *misc_buffer;
 202         dma_addr_t misc_buffer_bus;
 203
 204         struct ar_context ar_request_ctx;
 205         struct ar_context ar_response_ctx;
 206         struct context at_request_ctx;
 207         struct context at_response_ctx;
 208
 209         u32 it_context_support;
 210         u32 it_context_mask;     /* unoccupied IT contexts */
 211         struct iso_context *it_context_list;
 212         u64 ir_context_channels; /* unoccupied channels */
 213         u32 ir_context_support;
 214         u32 ir_context_mask;     /* unoccupied IR contexts */
 215         struct iso_context *ir_context_list;
 216         u64 mc_channels; /* channels in use by the multichannel IR context */
 217         bool mc_allocated;
 218
 219         __be32    *config_rom;
 220         dma_addr_t config_rom_bus;
 221         __be32    *next_config_rom;
 222         dma_addr_t next_config_rom_bus;
 223         __be32     next_header;
 224
 225         __le32    *self_id;
 226         dma_addr_t self_id_bus;
 227         struct work_struct bus_reset_work;
 228
 229         u32 self_id_buffer[512];
 230 };
 231
 232 static struct workqueue_struct *selfid_workqueue;
 233
 234 static inline struct fw_ohci *fw_ohci(struct fw_card *card)
 235 {
 236         return container_of(card, struct fw_ohci, card);
 237 }
 238
 239 #define IT_CONTEXT_CYCLE_MATCH_ENABLE   0x80000000
 240 #define IR_CONTEXT_BUFFER_FILL          0x80000000
 241 #define IR_CONTEXT_ISOCH_HEADER         0x40000000
 242 #define IR_CONTEXT_CYCLE_MATCH_ENABLE   0x20000000
 243 #define IR_CONTEXT_MULTI_CHANNEL_MODE   0x10000000
 244 #define IR_CONTEXT_DUAL_BUFFER_MODE     0x08000000
 245
 246 #define CONTEXT_RUN     0x8000
 247 #define CONTEXT_WAKE    0x1000
 248 #define CONTEXT_DEAD    0x0800
 249 #define CONTEXT_ACTIVE  0x0400
 250
 251 #define OHCI1394_MAX_AT_REQ_RETRIES     0xf
 252 #define OHCI1394_MAX_AT_RESP_RETRIES    0x2
 253 #define OHCI1394_MAX_PHYS_RESP_RETRIES  0x8
 254
 255 #define OHCI1394_REGISTER_SIZE          0x800
 256 #define OHCI1394_PCI_HCI_Control        0x40
 257 #define SELF_ID_BUF_SIZE                0x800
 258 #define OHCI_TCODE_PHY_PACKET           0x0e
 259 #define OHCI_VERSION_1_1                0x010010
 260
 261 static char ohci_driver_name[] = KBUILD_MODNAME;
 262
 263 #define PCI_VENDOR_ID_PINNACLE_SYSTEMS  0x11bd
 264 #define PCI_DEVICE_ID_AGERE_FW643       0x5901
 265 #define PCI_DEVICE_ID_CREATIVE_SB1394   0x4001
 266 #define PCI_DEVICE_ID_JMICRON_JMB38X_FW 0x2380
 267 #define PCI_DEVICE_ID_TI_TSB12LV22      0x8009
 268 #define PCI_DEVICE_ID_TI_TSB12LV26      0x8020
 269 #define PCI_DEVICE_ID_TI_TSB82AA2       0x8025
 270 #define PCI_DEVICE_ID_VIA_VT630X        0x3044
 271 #define PCI_REV_ID_VIA_VT6306           0x46
 272 #define PCI_DEVICE_ID_VIA_VT6315        0x3403
 273
 274 #define QUIRK_CYCLE_TIMER               0x1
 275 #define QUIRK_RESET_PACKET              0x2
 276 #define QUIRK_BE_HEADERS                0x4
 277 #define QUIRK_NO_1394A                  0x8
 278 #define QUIRK_NO_MSI                    0x10
 279 #define QUIRK_TI_SLLZ059                0x20
 280 #define QUIRK_IR_WAKE                   0x40
 281
 282 // On PCI Express Root Complex in any type of AMD Ryzen machine, VIA VT6306/6307/6308 with Asmedia
 283 // ASM1083/1085 brings an inconvenience that the read accesses to 'Isochronous Cycle Timer' register
 284 // (at offset 0xf0 in PCI I/O space) often causes unexpected system reboot. The mechanism is not
 285 // clear, since the read access to the other registers is enough safe; e.g. 'Node ID' register,
 286 // while it is probable due to detection of any type of PCIe error.
 287 #define QUIRK_REBOOT_BY_CYCLE_TIMER_READ        0x80000000
 288
 289 #if IS_ENABLED(CONFIG_X86)
 290
 291 static bool has_reboot_by_cycle_timer_read_quirk(const struct fw_ohci *ohci)
 292 {
 293         return !!(ohci->quirks & QUIRK_REBOOT_BY_CYCLE_TIMER_READ);
 294 }
 295
 296 #define PCI_DEVICE_ID_ASMEDIA_ASM108X   0x1080
 297
 298 static bool detect_vt630x_with_asm1083_on_amd_ryzen_machine(const struct pci_dev *pdev)
 299 {
 300         const struct pci_dev *pcie_to_pci_bridge;
 301
 302         // Detect any type of AMD Ryzen machine.
 303         if (!static_cpu_has(X86_FEATURE_ZEN))
 304                 return false;
 305
 306         // Detect VIA VT6306/6307/6308.
 307         if (pdev->vendor != PCI_VENDOR_ID_VIA)
 308                 return false;
 309         if (pdev->device != PCI_DEVICE_ID_VIA_VT630X)
 310                 return false;
 311
 312         // Detect Asmedia ASM1083/1085.
 313         pcie_to_pci_bridge = pdev->bus->self;
 314         if (pcie_to_pci_bridge->vendor != PCI_VENDOR_ID_ASMEDIA)
 315                 return false;
 316         if (pcie_to_pci_bridge->device != PCI_DEVICE_ID_ASMEDIA_ASM108X)
 317                 return false;
 318
 319         return true;
 320 }
 321
 322 #else
 323 #define has_reboot_by_cycle_timer_read_quirk(ohci) false
 324 #define detect_vt630x_with_asm1083_on_amd_ryzen_machine(pdev)   false
 325 #endif
 326
 327 /* In case of multiple matches in ohci_quirks[], only the first one is used. */
 328 static const struct {
 329         unsigned short vendor, device, revision, flags;
 330 } ohci_quirks[] = {
 331         {PCI_VENDOR_ID_AL, PCI_ANY_ID, PCI_ANY_ID,
 332                 QUIRK_CYCLE_TIMER},
 333
 334         {PCI_VENDOR_ID_APPLE, PCI_DEVICE_ID_APPLE_UNI_N_FW, PCI_ANY_ID,
 335                 QUIRK_BE_HEADERS},
 336
 337         {PCI_VENDOR_ID_ATT, PCI_DEVICE_ID_AGERE_FW643, 6,
 338                 QUIRK_NO_MSI},
 339
 340         {PCI_VENDOR_ID_CREATIVE, PCI_DEVICE_ID_CREATIVE_SB1394, PCI_ANY_ID,
 341                 QUIRK_RESET_PACKET},
 342
 343         {PCI_VENDOR_ID_JMICRON, PCI_DEVICE_ID_JMICRON_JMB38X_FW, PCI_ANY_ID,
 344                 QUIRK_NO_MSI},
 345
 346         {PCI_VENDOR_ID_NEC, PCI_ANY_ID, PCI_ANY_ID,
 347                 QUIRK_CYCLE_TIMER},
 348
 349         {PCI_VENDOR_ID_O2, PCI_ANY_ID, PCI_ANY_ID,
 350                 QUIRK_NO_MSI},
 351
 352         {PCI_VENDOR_ID_RICOH, PCI_ANY_ID, PCI_ANY_ID,
 353                 QUIRK_CYCLE_TIMER | QUIRK_NO_MSI},
 354
 355         {PCI_VENDOR_ID_TI, PCI_DEVICE_ID_TI_TSB12LV22, PCI_ANY_ID,
 356                 QUIRK_CYCLE_TIMER | QUIRK_RESET_PACKET | QUIRK_NO_1394A},
 357
 358         {PCI_VENDOR_ID_TI, PCI_DEVICE_ID_TI_TSB12LV26, PCI_ANY_ID,
 359                 QUIRK_RESET_PACKET | QUIRK_TI_SLLZ059},
 360
 361         {PCI_VENDOR_ID_TI, PCI_DEVICE_ID_TI_TSB82AA2, PCI_ANY_ID,
 362                 QUIRK_RESET_PACKET | QUIRK_TI_SLLZ059},
 363
 364         {PCI_VENDOR_ID_TI, PCI_ANY_ID, PCI_ANY_ID,
 365                 QUIRK_RESET_PACKET},
 366
 367         {PCI_VENDOR_ID_VIA, PCI_DEVICE_ID_VIA_VT630X, PCI_REV_ID_VIA_VT6306,
 368                 QUIRK_CYCLE_TIMER | QUIRK_IR_WAKE},
 369
 370         {PCI_VENDOR_ID_VIA, PCI_DEVICE_ID_VIA_VT6315, 0,
 371                 QUIRK_CYCLE_TIMER /* FIXME: necessary? */ | QUIRK_NO_MSI},
 372
 373         {PCI_VENDOR_ID_VIA, PCI_DEVICE_ID_VIA_VT6315, PCI_ANY_ID,
 374                 QUIRK_NO_MSI},
 375
 376         {PCI_VENDOR_ID_VIA, PCI_ANY_ID, PCI_ANY_ID,
 377                 QUIRK_CYCLE_TIMER | QUIRK_NO_MSI},
 378 };
 379
 380 /* This overrides anything that was found in ohci_quirks[]. */
 381 static int param_quirks;
 382 module_param_named(quirks, param_quirks, int, 0644);
 383 MODULE_PARM_DESC(quirks, "Chip quirks (default = 0"
 384         ", nonatomic cycle timer = "    __stringify(QUIRK_CYCLE_TIMER)
 385         ", reset packet generation = "  __stringify(QUIRK_RESET_PACKET)
 386         ", AR/selfID endianness = "     __stringify(QUIRK_BE_HEADERS)
 387         ", no 1394a enhancements = "    __stringify(QUIRK_NO_1394A)
 388         ", disable MSI = "              __stringify(QUIRK_NO_MSI)
 389         ", TI SLLZ059 erratum = "       __stringify(QUIRK_TI_SLLZ059)
 390         ", IR wake unreliable = "       __stringify(QUIRK_IR_WAKE)
 391         ")");
 392
 393 #define OHCI_PARAM_DEBUG_AT_AR          1
 394 #define OHCI_PARAM_DEBUG_SELFIDS        2
 395 #define OHCI_PARAM_DEBUG_IRQS           4
 396 #define OHCI_PARAM_DEBUG_BUSRESETS      8 /* only effective before chip init */
 397
 398 static int param_debug;
 399 module_param_named(debug, param_debug, int, 0644);
 400 MODULE_PARM_DESC(debug, "Verbose logging (default = 0"
 401         ", AT/AR events = "     __stringify(OHCI_PARAM_DEBUG_AT_AR)
 402         ", self-IDs = "         __stringify(OHCI_PARAM_DEBUG_SELFIDS)
 403         ", IRQs = "             __stringify(OHCI_PARAM_DEBUG_IRQS)
 404         ", busReset events = "  __stringify(OHCI_PARAM_DEBUG_BUSRESETS)
 405         ", or a combination, or all = -1)");
 406
 407 static bool param_remote_dma;
 408 module_param_named(remote_dma, param_remote_dma, bool, 0444);
 409 MODULE_PARM_DESC(remote_dma, "Enable unfiltered remote DMA (default = N)");
 410
 411 static void log_irqs(struct fw_ohci *ohci, u32 evt)
 412 {
 413         if (likely(!(param_debug &
 414                         (OHCI_PARAM_DEBUG_IRQS | OHCI_PARAM_DEBUG_BUSRESETS))))
 415                 return;
 416
 417         if (!(param_debug & OHCI_PARAM_DEBUG_IRQS) &&
 418             !(evt & OHCI1394_busReset))
 419                 return;
 420
 421         ohci_notice(ohci, "IRQ %08x%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s\n", evt,
 422             evt & OHCI1394_selfIDComplete       ? " selfID"             : "",
 423             evt & OHCI1394_RQPkt                ? " AR_req"             : "",
 424             evt & OHCI1394_RSPkt                ? " AR_resp"            : "",
 425             evt & OHCI1394_reqTxComplete        ? " AT_req"             : "",
 426             evt & OHCI1394_respTxComplete       ? " AT_resp"            : "",
 427             evt & OHCI1394_isochRx              ? " IR"                 : "",
 428             evt & OHCI1394_isochTx              ? " IT"                 : "",
 429             evt & OHCI1394_postedWriteErr       ? " postedWriteErr"     : "",
 430             evt & OHCI1394_cycleTooLong         ? " cycleTooLong"       : "",
 431             evt & OHCI1394_cycle64Seconds       ? " cycle64Seconds"     : "",
 432             evt & OHCI1394_cycleInconsistent    ? " cycleInconsistent"  : "",
 433             evt & OHCI1394_regAccessFail        ? " regAccessFail"      : "",
 434             evt & OHCI1394_unrecoverableError   ? " unrecoverableError" : "",
 435             evt & OHCI1394_busReset             ? " busReset"           : "",
 436             evt & ~(OHCI1394_selfIDComplete | OHCI1394_RQPkt |
 437                     OHCI1394_RSPkt | OHCI1394_reqTxComplete |
 438                     OHCI1394_respTxComplete | OHCI1394_isochRx |
 439                     OHCI1394_isochTx | OHCI1394_postedWriteErr |
 440                     OHCI1394_cycleTooLong | OHCI1394_cycle64Seconds |
 441                     OHCI1394_cycleInconsistent |
 442                     OHCI1394_regAccessFail | OHCI1394_busReset)
 443                                                 ? " ?"                  : "");
 444 }
 445
 446 static const char *speed[] = {
 447         [0] = "S100", [1] = "S200", [2] = "S400",    [3] = "beta",
 448 };
 449 static const char *power[] = {
 450         [0] = "+0W",  [1] = "+15W", [2] = "+30W",    [3] = "+45W",
 451         [4] = "-3W",  [5] = " ?W",  [6] = "-3..-6W", [7] = "-3..-10W",
 452 };
 453 static const char port[] = { '.', '-', 'p', 'c', };
 454
 455 static char _p(u32 *s, int shift)
 456 {
 457         return port[*s >> shift & 3];
 458 }
 459
 460 static void log_selfids(struct fw_ohci *ohci, int generation, int self_id_count)
 461 {
 462         u32 *s;
 463
 464         if (likely(!(param_debug & OHCI_PARAM_DEBUG_SELFIDS)))
 465                 return;
 466
 467         ohci_notice(ohci, "%d selfIDs, generation %d, local node ID %04x\n",
 468                     self_id_count, generation, ohci->node_id);
 469
 470         for (s = ohci->self_id_buffer; self_id_count--; ++s)
 471                 if ((*s & 1 << 23) == 0)
 472                         ohci_notice(ohci,
 473                             "selfID 0: %08x, phy %d [%c%c%c] %s gc=%d %s %s%s%s\n",
 474                             *s, *s >> 24 & 63, _p(s, 6), _p(s, 4), _p(s, 2),
 475                             speed[*s >> 14 & 3], *s >> 16 & 63,
 476                             power[*s >> 8 & 7], *s >> 22 & 1 ? "L" : "",
 477                             *s >> 11 & 1 ? "c" : "", *s & 2 ? "i" : "");
 478                 else
 479                         ohci_notice(ohci,
 480                             "selfID n: %08x, phy %d [%c%c%c%c%c%c%c%c]\n",
 481                             *s, *s >> 24 & 63,
 482                             _p(s, 16), _p(s, 14), _p(s, 12), _p(s, 10),
 483                             _p(s,  8), _p(s,  6), _p(s,  4), _p(s,  2));
 484 }
 485
 486 static const char *evts[] = {
 487         [0x00] = "evt_no_status",       [0x01] = "-reserved-",
 488         [0x02] = "evt_long_packet",     [0x03] = "evt_missing_ack",
 489         [0x04] = "evt_underrun",        [0x05] = "evt_overrun",
 490         [0x06] = "evt_descriptor_read", [0x07] = "evt_data_read",
 491         [0x08] = "evt_data_write",      [0x09] = "evt_bus_reset",
 492         [0x0a] = "evt_timeout",         [0x0b] = "evt_tcode_err",
 493         [0x0c] = "-reserved-",          [0x0d] = "-reserved-",
 494         [0x0e] = "evt_unknown",         [0x0f] = "evt_flushed",
 495         [0x10] = "-reserved-",          [0x11] = "ack_complete",
 496         [0x12] = "ack_pending ",        [0x13] = "-reserved-",
 497         [0x14] = "ack_busy_X",          [0x15] = "ack_busy_A",
 498         [0x16] = "ack_busy_B",          [0x17] = "-reserved-",
 499         [0x18] = "-reserved-",          [0x19] = "-reserved-",
 500         [0x1a] = "-reserved-",          [0x1b] = "ack_tardy",
 501         [0x1c] = "-reserved-",          [0x1d] = "ack_data_error",
 502         [0x1e] = "ack_type_error",      [0x1f] = "-reserved-",
 503         [0x20] = "pending/cancelled",
 504 };
 505 static const char *tcodes[] = {
 506         [0x0] = "QW req",               [0x1] = "BW req",
 507         [0x2] = "W resp",               [0x3] = "-reserved-",
 508         [0x4] = "QR req",               [0x5] = "BR req",
 509         [0x6] = "QR resp",              [0x7] = "BR resp",
 510         [0x8] = "cycle start",          [0x9] = "Lk req",
 511         [0xa] = "async stream packet",  [0xb] = "Lk resp",
 512         [0xc] = "-reserved-",           [0xd] = "-reserved-",
 513         [0xe] = "link internal",        [0xf] = "-reserved-",
 514 };
 515
 516 static void log_ar_at_event(struct fw_ohci *ohci,
 517                             char dir, int speed, u32 *header, int evt)
 518 {
 519         int tcode = header[0] >> 4 & 0xf;
 520         char specific[12];
 521
 522         if (likely(!(param_debug & OHCI_PARAM_DEBUG_AT_AR)))
 523                 return;
 524
 525         if (unlikely(evt >= ARRAY_SIZE(evts)))
 526                         evt = 0x1f;
 527
 528         if (evt == OHCI1394_evt_bus_reset) {
 529                 ohci_notice(ohci, "A%c evt_bus_reset, generation %d\n",
 530                             dir, (header[2] >> 16) & 0xff);
 531                 return;
 532         }
 533
 534         switch (tcode) {
 535         case 0x0: case 0x6: case 0x8:
 536                 snprintf(specific, sizeof(specific), " = %08x",
 537                          be32_to_cpu((__force __be32)header[3]));
 538                 break;
 539         case 0x1: case 0x5: case 0x7: case 0x9: case 0xb:
 540                 snprintf(specific, sizeof(specific), " %x,%x",
 541                          header[3] >> 16, header[3] & 0xffff);
 542                 break;
 543         default:
 544                 specific[0] = '\0';
 545         }
 546
 547         switch (tcode) {
 548         case 0xa:
 549                 ohci_notice(ohci, "A%c %s, %s\n",
 550                             dir, evts[evt], tcodes[tcode]);
 551                 break;
 552         case 0xe:
 553                 ohci_notice(ohci, "A%c %s, PHY %08x %08x\n",
 554                             dir, evts[evt], header[1], header[2]);
 555                 break;
 556         case 0x0: case 0x1: case 0x4: case 0x5: case 0x9:
 557                 ohci_notice(ohci,
 558                             "A%c spd %x tl %02x, %04x -> %04x, %s, %s, %04x%08x%s\n",
 559                             dir, speed, header[0] >> 10 & 0x3f,
 560                             header[1] >> 16, header[0] >> 16, evts[evt],
 561                             tcodes[tcode], header[1] & 0xffff, header[2], specific);
 562                 break;
 563         default:
 564                 ohci_notice(ohci,
 565                             "A%c spd %x tl %02x, %04x -> %04x, %s, %s%s\n",
 566                             dir, speed, header[0] >> 10 & 0x3f,
 567                             header[1] >> 16, header[0] >> 16, evts[evt],
 568                             tcodes[tcode], specific);
 569         }
 570 }
 571
 572 static inline void reg_write(const struct fw_ohci *ohci, int offset, u32 data)
 573 {
 574         writel(data, ohci->registers + offset);
 575 }
 576
 577 static inline u32 reg_read(const struct fw_ohci *ohci, int offset)
 578 {
 579         return readl(ohci->registers + offset);
 580 }
 581
 582 static inline void flush_writes(const struct fw_ohci *ohci)
 583 {
 584         /* Do a dummy read to flush writes. */
 585         reg_read(ohci, OHCI1394_Version);
 586 }
 587
 588 /*
 589  * Beware!  read_phy_reg(), write_phy_reg(), update_phy_reg(), and
 590  * read_paged_phy_reg() require the caller to hold ohci->phy_reg_mutex.
 591  * In other words, only use ohci_read_phy_reg() and ohci_update_phy_reg()
 592  * directly.  Exceptions are intrinsically serialized contexts like pci_probe.
 593  */
 594 static int read_phy_reg(struct fw_ohci *ohci, int addr)
 595 {
 596         u32 val;
 597         int i;
 598
 599         reg_write(ohci, OHCI1394_PhyControl, OHCI1394_PhyControl_Read(addr));
 600         for (i = 0; i < 3 + 100; i++) {
 601                 val = reg_read(ohci, OHCI1394_PhyControl);
 602                 if (!~val)
 603                         return -ENODEV; /* Card was ejected. */
 604
 605                 if (val & OHCI1394_PhyControl_ReadDone)
 606                         return OHCI1394_PhyControl_ReadData(val);
 607
 608                 /*
 609                  * Try a few times without waiting.  Sleeping is necessary
 610                  * only when the link/PHY interface is busy.
 611                  */
 612                 if (i >= 3)
 613                         msleep(1);
 614         }
 615         ohci_err(ohci, "failed to read phy reg %d\n", addr);
 616         dump_stack();
 617
 618         return -EBUSY;
 619 }
 620
 621 static int write_phy_reg(const struct fw_ohci *ohci, int addr, u32 val)
 622 {
 623         int i;
 624
 625         reg_write(ohci, OHCI1394_PhyControl,
 626                   OHCI1394_PhyControl_Write(addr, val));
 627         for (i = 0; i < 3 + 100; i++) {
 628                 val = reg_read(ohci, OHCI1394_PhyControl);
 629                 if (!~val)
 630                         return -ENODEV; /* Card was ejected. */
 631
 632                 if (!(val & OHCI1394_PhyControl_WritePending))
 633                         return 0;
 634
 635                 if (i >= 3)
 636                         msleep(1);
 637         }
 638         ohci_err(ohci, "failed to write phy reg %d, val %u\n", addr, val);
 639         dump_stack();
 640
 641         return -EBUSY;
 642 }
 643
 644 static int update_phy_reg(struct fw_ohci *ohci, int addr,
 645                           int clear_bits, int set_bits)
 646 {
 647         int ret = read_phy_reg(ohci, addr);
 648         if (ret < 0)
 649                 return ret;
 650
 651         /*
 652          * The interrupt status bits are cleared by writing a one bit.
 653          * Avoid clearing them unless explicitly requested in set_bits.
 654          */
 655         if (addr == 5)
 656                 clear_bits |= PHY_INT_STATUS_BITS;
 657
 658         return write_phy_reg(ohci, addr, (ret & ~clear_bits) | set_bits);
 659 }
 660
 661 static int read_paged_phy_reg(struct fw_ohci *ohci, int page, int addr)
 662 {
 663         int ret;
 664
 665         ret = update_phy_reg(ohci, 7, PHY_PAGE_SELECT, page << 5);
 666         if (ret < 0)
 667                 return ret;
 668
 669         return read_phy_reg(ohci, addr);
 670 }
 671
 672 static int ohci_read_phy_reg(struct fw_card *card, int addr)
 673 {
 674         struct fw_ohci *ohci = fw_ohci(card);
 675         int ret;
 676
 677         mutex_lock(&ohci->phy_reg_mutex);
 678         ret = read_phy_reg(ohci, addr);
 679         mutex_unlock(&ohci->phy_reg_mutex);
 680
 681         return ret;
 682 }
 683
 684 static int ohci_update_phy_reg(struct fw_card *card, int addr,
 685                                int clear_bits, int set_bits)
 686 {
 687         struct fw_ohci *ohci = fw_ohci(card);
 688         int ret;
 689
 690         mutex_lock(&ohci->phy_reg_mutex);
 691         ret = update_phy_reg(ohci, addr, clear_bits, set_bits);
 692         mutex_unlock(&ohci->phy_reg_mutex);
 693
 694         return ret;
 695 }
 696
 697 static inline dma_addr_t ar_buffer_bus(struct ar_context *ctx, unsigned int i)
 698 {
 699         return page_private(ctx->pages[i]);
 700 }
 701
 702 static void ar_context_link_page(struct ar_context *ctx, unsigned int index)
 703 {
 704         struct descriptor *d;
 705
 706         d = &ctx->descriptors[index];
 707         d->branch_address  &= cpu_to_le32(~0xf);
 708         d->res_count       =  cpu_to_le16(PAGE_SIZE);
 709         d->transfer_status =  0;
 710
 711         wmb(); /* finish init of new descriptors before branch_address update */
 712         d = &ctx->descriptors[ctx->last_buffer_index];
 713         d->branch_address  |= cpu_to_le32(1);
 714
 715         ctx->last_buffer_index = index;
 716
 717         reg_write(ctx->ohci, CONTROL_SET(ctx->regs), CONTEXT_WAKE);
 718 }
 719
 720 static void ar_context_release(struct ar_context *ctx)
 721 {
 722         struct device *dev = ctx->ohci->card.device;
 723         unsigned int i;
 724
 725         vunmap(ctx->buffer);
 726
 727         for (i = 0; i < AR_BUFFERS; i++) {
 728                 if (ctx->pages[i])
 729                         dma_free_pages(dev, PAGE_SIZE, ctx->pages[i],
 730                                        ar_buffer_bus(ctx, i), DMA_FROM_DEVICE);
 731         }
 732 }
 733
 734 static void ar_context_abort(struct ar_context *ctx, const char *error_msg)
 735 {
 736         struct fw_ohci *ohci = ctx->ohci;
 737
 738         if (reg_read(ohci, CONTROL_CLEAR(ctx->regs)) & CONTEXT_RUN) {
 739                 reg_write(ohci, CONTROL_CLEAR(ctx->regs), CONTEXT_RUN);
 740                 flush_writes(ohci);
 741
 742                 ohci_err(ohci, "AR error: %s; DMA stopped\n", error_msg);
 743         }
 744         /* FIXME: restart? */
 745 }
 746
 747 static inline unsigned int ar_next_buffer_index(unsigned int index)
 748 {
 749         return (index + 1) % AR_BUFFERS;
 750 }
 751
 752 static inline unsigned int ar_first_buffer_index(struct ar_context *ctx)
 753 {
 754         return ar_next_buffer_index(ctx->last_buffer_index);
 755 }
 756
 757 /*
 758  * We search for the buffer that contains the last AR packet DMA data written
 759  * by the controller.
 760  */
 761 static unsigned int ar_search_last_active_buffer(struct ar_context *ctx,
 762                                                  unsigned int *buffer_offset)
 763 {
 764         unsigned int i, next_i, last = ctx->last_buffer_index;
 765         __le16 res_count, next_res_count;
 766
 767         i = ar_first_buffer_index(ctx);
 768         res_count = READ_ONCE(ctx->descriptors[i].res_count);
 769
 770         /* A buffer that is not yet completely filled must be the last one. */
 771         while (i != last && res_count == 0) {
 772
 773                 /* Peek at the next descriptor. */
 774                 next_i = ar_next_buffer_index(i);
 775                 rmb(); /* read descriptors in order */
 776                 next_res_count = READ_ONCE(ctx->descriptors[next_i].res_count);
 777                 /*
 778                  * If the next descriptor is still empty, we must stop at this
 779                  * descriptor.
 780                  */
 781                 if (next_res_count == cpu_to_le16(PAGE_SIZE)) {
 782                         /*
 783                          * The exception is when the DMA data for one packet is
 784                          * split over three buffers; in this case, the middle
 785                          * buffer's descriptor might be never updated by the
 786                          * controller and look still empty, and we have to peek
 787                          * at the third one.
 788                          */
 789                         if (MAX_AR_PACKET_SIZE > PAGE_SIZE && i != last) {
 790                                 next_i = ar_next_buffer_index(next_i);
 791                                 rmb();
 792                                 next_res_count = READ_ONCE(ctx->descriptors[next_i].res_count);
 793                                 if (next_res_count != cpu_to_le16(PAGE_SIZE))
 794                                         goto next_buffer_is_active;
 795                         }
 796
 797                         break;
 798                 }
 799
 800 next_buffer_is_active:
 801                 i = next_i;
 802                 res_count = next_res_count;
 803         }
 804
 805         rmb(); /* read res_count before the DMA data */
 806
 807         *buffer_offset = PAGE_SIZE - le16_to_cpu(res_count);
 808         if (*buffer_offset > PAGE_SIZE) {
 809                 *buffer_offset = 0;
 810                 ar_context_abort(ctx, "corrupted descriptor");
 811         }
 812
 813         return i;
 814 }
 815
 816 static void ar_sync_buffers_for_cpu(struct ar_context *ctx,
 817                                     unsigned int end_buffer_index,
 818                                     unsigned int end_buffer_offset)
 819 {
 820         unsigned int i;
 821
 822         i = ar_first_buffer_index(ctx);
 823         while (i != end_buffer_index) {
 824                 dma_sync_single_for_cpu(ctx->ohci->card.device,
 825                                         ar_buffer_bus(ctx, i),
 826                                         PAGE_SIZE, DMA_FROM_DEVICE);
 827                 i = ar_next_buffer_index(i);
 828         }
 829         if (end_buffer_offset > 0)
 830                 dma_sync_single_for_cpu(ctx->ohci->card.device,
 831                                         ar_buffer_bus(ctx, i),
 832                                         end_buffer_offset, DMA_FROM_DEVICE);
 833 }
 834
 835 #if defined(CONFIG_PPC_PMAC) && defined(CONFIG_PPC32)
 836 #define cond_le32_to_cpu(v) \
 837         (ohci->quirks & QUIRK_BE_HEADERS ? (__force __u32)(v) : le32_to_cpu(v))
 838 #else
 839 #define cond_le32_to_cpu(v) le32_to_cpu(v)
 840 #endif
 841
 842 static __le32 *handle_ar_packet(struct ar_context *ctx, __le32 *buffer)
 843 {
 844         struct fw_ohci *ohci = ctx->ohci;
 845         struct fw_packet p;
 846         u32 status, length, tcode;
 847         int evt;
 848
 849         p.header[0] = cond_le32_to_cpu(buffer[0]);
 850         p.header[1] = cond_le32_to_cpu(buffer[1]);
 851         p.header[2] = cond_le32_to_cpu(buffer[2]);
 852
 853         tcode = (p.header[0] >> 4) & 0x0f;
 854         switch (tcode) {
 855         case TCODE_WRITE_QUADLET_REQUEST:
 856         case TCODE_READ_QUADLET_RESPONSE:
 857                 p.header[3] = (__force __u32) buffer[3];
 858                 p.header_length = 16;
 859                 p.payload_length = 0;
 860                 break;
 861
 862         case TCODE_READ_BLOCK_REQUEST :
 863                 p.header[3] = cond_le32_to_cpu(buffer[3]);
 864                 p.header_length = 16;
 865                 p.payload_length = 0;
 866                 break;
 867
 868         case TCODE_WRITE_BLOCK_REQUEST:
 869         case TCODE_READ_BLOCK_RESPONSE:
 870         case TCODE_LOCK_REQUEST:
 871         case TCODE_LOCK_RESPONSE:
 872                 p.header[3] = cond_le32_to_cpu(buffer[3]);
 873                 p.header_length = 16;
 874                 p.payload_length = p.header[3] >> 16;
 875                 if (p.payload_length > MAX_ASYNC_PAYLOAD) {
 876                         ar_context_abort(ctx, "invalid packet length");
 877                         return NULL;
 878                 }
 879                 break;
 880
 881         case TCODE_WRITE_RESPONSE:
 882         case TCODE_READ_QUADLET_REQUEST:
 883         case OHCI_TCODE_PHY_PACKET:
 884                 p.header_length = 12;
 885                 p.payload_length = 0;
 886                 break;
 887
 888         default:
 889                 ar_context_abort(ctx, "invalid tcode");
 890                 return NULL;
 891         }
 892
 893         p.payload = (void *) buffer + p.header_length;
 894
 895         /* FIXME: What to do about evt_* errors? */
 896         length = (p.header_length + p.payload_length + 3) / 4;
 897         status = cond_le32_to_cpu(buffer[length]);
 898         evt    = (status >> 16) & 0x1f;
 899
 900         p.ack        = evt - 16;
 901         p.speed      = (status >> 21) & 0x7;
 902         p.timestamp  = status & 0xffff;
 903         p.generation = ohci->request_generation;
 904
 905         log_ar_at_event(ohci, 'R', p.speed, p.header, evt);
 906
 907         /*
 908          * Several controllers, notably from NEC and VIA, forget to
 909          * write ack_complete status at PHY packet reception.
 910          */
 911         if (evt == OHCI1394_evt_no_status &&
 912             (p.header[0] & 0xff) == (OHCI1394_phy_tcode << 4))
 913                 p.ack = ACK_COMPLETE;
 914
 915         /*
 916          * The OHCI bus reset handler synthesizes a PHY packet with
 917          * the new generation number when a bus reset happens (see
 918          * section 8.4.2.3).  This helps us determine when a request
 919          * was received and make sure we send the response in the same
 920          * generation.  We only need this for requests; for responses
 921          * we use the unique tlabel for finding the matching
 922          * request.
 923          *
 924          * Alas some chips sometimes emit bus reset packets with a
 925          * wrong generation.  We set the correct generation for these
 926          * at a slightly incorrect time (in bus_reset_work).
 927          */
 928         if (evt == OHCI1394_evt_bus_reset) {
 929                 if (!(ohci->quirks & QUIRK_RESET_PACKET))
 930                         ohci->request_generation = (p.header[2] >> 16) & 0xff;
 931         } else if (ctx == &ohci->ar_request_ctx) {
 932                 fw_core_handle_request(&ohci->card, &p);
 933         } else {
 934                 fw_core_handle_response(&ohci->card, &p);
 935         }
 936
 937         return buffer + length + 1;
 938 }
 939
 940 static void *handle_ar_packets(struct ar_context *ctx, void *p, void *end)
 941 {
 942         void *next;
 943
 944         while (p < end) {
 945                 next = handle_ar_packet(ctx, p);
 946                 if (!next)
 947                         return p;
 948                 p = next;
 949         }
 950
 951         return p;
 952 }
 953
 954 static void ar_recycle_buffers(struct ar_context *ctx, unsigned int end_buffer)
 955 {
 956         unsigned int i;
 957
 958         i = ar_first_buffer_index(ctx);
 959         while (i != end_buffer) {
 960                 dma_sync_single_for_device(ctx->ohci->card.device,
 961                                            ar_buffer_bus(ctx, i),
 962                                            PAGE_SIZE, DMA_FROM_DEVICE);
 963                 ar_context_link_page(ctx, i);
 964                 i = ar_next_buffer_index(i);
 965         }
 966 }
 967
 968 static void ar_context_tasklet(unsigned long data)
 969 {
 970         struct ar_context *ctx = (struct ar_context *)data;
 971         unsigned int end_buffer_index, end_buffer_offset;
 972         void *p, *end;
 973
 974         p = ctx->pointer;
 975         if (!p)
 976                 return;
 977
 978         end_buffer_index = ar_search_last_active_buffer(ctx,
 979                                                         &end_buffer_offset);
 980         ar_sync_buffers_for_cpu(ctx, end_buffer_index, end_buffer_offset);
 981         end = ctx->buffer + end_buffer_index * PAGE_SIZE + end_buffer_offset;
 982
 983         if (end_buffer_index < ar_first_buffer_index(ctx)) {
 984                 /*
 985                  * The filled part of the overall buffer wraps around; handle
 986                  * all packets up to the buffer end here.  If the last packet
 987                  * wraps around, its tail will be visible after the buffer end
 988                  * because the buffer start pages are mapped there again.
 989                  */
 990                 void *buffer_end = ctx->buffer + AR_BUFFERS * PAGE_SIZE;
 991                 p = handle_ar_packets(ctx, p, buffer_end);
 992                 if (p < buffer_end)
 993                         goto error;
 994                 /* adjust p to point back into the actual buffer */
 995                 p -= AR_BUFFERS * PAGE_SIZE;
 996         }
 997
 998         p = handle_ar_packets(ctx, p, end);
 999         if (p != end) {
1000                 if (p > end)
1001                         ar_context_abort(ctx, "inconsistent descriptor");
1002                 goto error;
1003         }
1004
1005         ctx->pointer = p;
1006         ar_recycle_buffers(ctx, end_buffer_index);
1007
1008         return;
1009
1010 error:
1011         ctx->pointer = NULL;
1012 }
1013
1014 static int ar_context_init(struct ar_context *ctx, struct fw_ohci *ohci,
1015                            unsigned int descriptors_offset, u32 regs)
1016 {
1017         struct device *dev = ohci->card.device;
1018         unsigned int i;
1019         dma_addr_t dma_addr;
1020         struct page *pages[AR_BUFFERS + AR_WRAPAROUND_PAGES];
1021         struct descriptor *d;
1022
1023         ctx->regs        = regs;
1024         ctx->ohci        = ohci;
1025         tasklet_init(&ctx->tasklet, ar_context_tasklet, (unsigned long)ctx);
1026
1027         for (i = 0; i < AR_BUFFERS; i++) {
1028                 ctx->pages[i] = dma_alloc_pages(dev, PAGE_SIZE, &dma_addr,
1029                                                 DMA_FROM_DEVICE, GFP_KERNEL);
1030                 if (!ctx->pages[i])
1031                         goto out_of_memory;
1032                 set_page_private(ctx->pages[i], dma_addr);
1033                 dma_sync_single_for_device(dev, dma_addr, PAGE_SIZE,
1034                                            DMA_FROM_DEVICE);
1035         }
1036
1037         for (i = 0; i < AR_BUFFERS; i++)
1038                 pages[i]              = ctx->pages[i];
1039         for (i = 0; i < AR_WRAPAROUND_PAGES; i++)
1040                 pages[AR_BUFFERS + i] = ctx->pages[i];
1041         ctx->buffer = vmap(pages, ARRAY_SIZE(pages), VM_MAP, PAGE_KERNEL);
1042         if (!ctx->buffer)
1043                 goto out_of_memory;
1044
1045         ctx->descriptors     = ohci->misc_buffer     + descriptors_offset;
1046         ctx->descriptors_bus = ohci->misc_buffer_bus + descriptors_offset;
1047
1048         for (i = 0; i < AR_BUFFERS; i++) {
1049                 d = &ctx->descriptors[i];
1050                 d->req_count      = cpu_to_le16(PAGE_SIZE);
1051                 d->control        = cpu_to_le16(DESCRIPTOR_INPUT_MORE |
1052                                                 DESCRIPTOR_STATUS |
1053                                                 DESCRIPTOR_BRANCH_ALWAYS);
1054                 d->data_address   = cpu_to_le32(ar_buffer_bus(ctx, i));
1055                 d->branch_address = cpu_to_le32(ctx->descriptors_bus +
1056                         ar_next_buffer_index(i) * sizeof(struct descriptor));
1057         }
1058
1059         return 0;
1060
1061 out_of_memory:
1062         ar_context_release(ctx);
1063
1064         return -ENOMEM;
1065 }
1066
1067 static void ar_context_run(struct ar_context *ctx)
1068 {
1069         unsigned int i;
1070
1071         for (i = 0; i < AR_BUFFERS; i++)
1072                 ar_context_link_page(ctx, i);
1073
1074         ctx->pointer = ctx->buffer;
1075
1076         reg_write(ctx->ohci, COMMAND_PTR(ctx->regs), ctx->descriptors_bus | 1);
1077         reg_write(ctx->ohci, CONTROL_SET(ctx->regs), CONTEXT_RUN);
1078 }
1079
1080 static struct descriptor *find_branch_descriptor(struct descriptor *d, int z)
1081 {
1082         __le16 branch;
1083
1084         branch = d->control & cpu_to_le16(DESCRIPTOR_BRANCH_ALWAYS);
1085
1086         /* figure out which descriptor the branch address goes in */
1087         if (z == 2 && branch == cpu_to_le16(DESCRIPTOR_BRANCH_ALWAYS))
1088                 return d;
1089         else
1090                 return d + z - 1;
1091 }
1092
1093 static void context_tasklet(unsigned long data)
1094 {
1095         struct context *ctx = (struct context *) data;
1096         struct descriptor *d, *last;
1097         u32 address;
1098         int z;
1099         struct descriptor_buffer *desc;
1100
1101         desc = list_entry(ctx->buffer_list.next,
1102                         struct descriptor_buffer, list);
1103         last = ctx->last;
1104         while (last->branch_address != 0) {
1105                 struct descriptor_buffer *old_desc = desc;
1106                 address = le32_to_cpu(last->branch_address);
1107                 z = address & 0xf;
1108                 address &= ~0xf;
1109                 ctx->current_bus = address;
1110
1111                 /* If the branch address points to a buffer outside of the
1112                  * current buffer, advance to the next buffer. */
1113                 if (address < desc->buffer_bus ||
1114                                 address >= desc->buffer_bus + desc->used)
1115                         desc = list_entry(desc->list.next,
1116                                         struct descriptor_buffer, list);
1117                 d = desc->buffer + (address - desc->buffer_bus) / sizeof(*d);
1118                 last = find_branch_descriptor(d, z);
1119
1120                 if (!ctx->callback(ctx, d, last))
1121                         break;
1122
1123                 if (old_desc != desc) {
1124                         /* If we've advanced to the next buffer, move the
1125                          * previous buffer to the free list. */
1126                         unsigned long flags;
1127                         old_desc->used = 0;
1128                         spin_lock_irqsave(&ctx->ohci->lock, flags);
1129                         list_move_tail(&old_desc->list, &ctx->buffer_list);
1130                         spin_unlock_irqrestore(&ctx->ohci->lock, flags);
1131                 }
1132                 ctx->last = last;
1133         }
1134 }
1135
1136 /*
1137  * Allocate a new buffer and add it to the list of free buffers for this
1138  * context.  Must be called with ohci->lock held.
1139  */
1140 static int context_add_buffer(struct context *ctx)
1141 {
1142         struct descriptor_buffer *desc;
1143         dma_addr_t bus_addr;
1144         int offset;
1145
1146         /*
1147          * 16MB of descriptors should be far more than enough for any DMA
1148          * program.  This will catch run-away userspace or DoS attacks.
1149          */
1150         if (ctx->total_allocation >= 16*1024*1024)
1151                 return -ENOMEM;
1152
1153         desc = dma_alloc_coherent(ctx->ohci->card.device, PAGE_SIZE,
1154                         &bus_addr, GFP_ATOMIC);
1155         if (!desc)
1156                 return -ENOMEM;
1157
1158         offset = (void *)&desc->buffer - (void *)desc;
1159         /*
1160          * Some controllers, like JMicron ones, always issue 0x20-byte DMA reads
1161          * for descriptors, even 0x10-byte ones. This can cause page faults when
1162          * an IOMMU is in use and the oversized read crosses a page boundary.
1163          * Work around this by always leaving at least 0x10 bytes of padding.
1164          */
1165         desc->buffer_size = PAGE_SIZE - offset - 0x10;
1166         desc->buffer_bus = bus_addr + offset;
1167         desc->used = 0;
1168
1169         list_add_tail(&desc->list, &ctx->buffer_list);
1170         ctx->total_allocation += PAGE_SIZE;
1171
1172         return 0;
1173 }
1174
1175 static int context_init(struct context *ctx, struct fw_ohci *ohci,
1176                         u32 regs, descriptor_callback_t callback)
1177 {
1178         ctx->ohci = ohci;
1179         ctx->regs = regs;
1180         ctx->total_allocation = 0;
1181
1182         INIT_LIST_HEAD(&ctx->buffer_list);
1183         if (context_add_buffer(ctx) < 0)
1184                 return -ENOMEM;
1185
1186         ctx->buffer_tail = list_entry(ctx->buffer_list.next,
1187                         struct descriptor_buffer, list);
1188
1189         tasklet_init(&ctx->tasklet, context_tasklet, (unsigned long)ctx);
1190         ctx->callback = callback;
1191
1192         /*
1193          * We put a dummy descriptor in the buffer that has a NULL
1194          * branch address and looks like it's been sent.  That way we
1195          * have a descriptor to append DMA programs to.
1196          */
1197         memset(ctx->buffer_tail->buffer, 0, sizeof(*ctx->buffer_tail->buffer));
1198         ctx->buffer_tail->buffer->control = cpu_to_le16(DESCRIPTOR_OUTPUT_LAST);
1199         ctx->buffer_tail->buffer->transfer_status = cpu_to_le16(0x8011);
1200         ctx->buffer_tail->used += sizeof(*ctx->buffer_tail->buffer);
1201         ctx->last = ctx->buffer_tail->buffer;
1202         ctx->prev = ctx->buffer_tail->buffer;
1203         ctx->prev_z = 1;
1204
1205         return 0;
1206 }
1207
1208 static void context_release(struct context *ctx)
1209 {
1210         struct fw_card *card = &ctx->ohci->card;
1211         struct descriptor_buffer *desc, *tmp;
1212
1213         list_for_each_entry_safe(desc, tmp, &ctx->buffer_list, list)
1214                 dma_free_coherent(card->device, PAGE_SIZE, desc,
1215                         desc->buffer_bus -
1216                         ((void *)&desc->buffer - (void *)desc));
1217 }
1218
1219 /* Must be called with ohci->lock held */
1220 static struct descriptor *context_get_descriptors(struct context *ctx,
1221                                                   int z, dma_addr_t *d_bus)
1222 {
1223         struct descriptor *d = NULL;
1224         struct descriptor_buffer *desc = ctx->buffer_tail;
1225
1226         if (z * sizeof(*d) > desc->buffer_size)
1227                 return NULL;
1228
1229         if (z * sizeof(*d) > desc->buffer_size - desc->used) {
1230                 /* No room for the descriptor in this buffer, so advance to the
1231                  * next one. */
1232
1233                 if (desc->list.next == &ctx->buffer_list) {
1234                         /* If there is no free buffer next in the list,
1235                          * allocate one. */
1236                         if (context_add_buffer(ctx) < 0)
1237                                 return NULL;
1238                 }
1239                 desc = list_entry(desc->list.next,
1240                                 struct descriptor_buffer, list);
1241                 ctx->buffer_tail = desc;
1242         }
1243
1244         d = desc->buffer + desc->used / sizeof(*d);
1245         memset(d, 0, z * sizeof(*d));
1246         *d_bus = desc->buffer_bus + desc->used;
1247
1248         return d;
1249 }
1250
1251 static void context_run(struct context *ctx, u32 extra)
1252 {
1253         struct fw_ohci *ohci = ctx->ohci;
1254
1255         reg_write(ohci, COMMAND_PTR(ctx->regs),
1256                   le32_to_cpu(ctx->last->branch_address));
1257         reg_write(ohci, CONTROL_CLEAR(ctx->regs), ~0);
1258         reg_write(ohci, CONTROL_SET(ctx->regs), CONTEXT_RUN | extra);
1259         ctx->running = true;
1260         flush_writes(ohci);
1261 }
1262
1263 static void context_append(struct context *ctx,
1264                            struct descriptor *d, int z, int extra)
1265 {
1266         dma_addr_t d_bus;
1267         struct descriptor_buffer *desc = ctx->buffer_tail;
1268         struct descriptor *d_branch;
1269
1270         d_bus = desc->buffer_bus + (d - desc->buffer) * sizeof(*d);
1271
1272         desc->used += (z + extra) * sizeof(*d);
1273
1274         wmb(); /* finish init of new descriptors before branch_address update */
1275
1276         d_branch = find_branch_descriptor(ctx->prev, ctx->prev_z);
1277         d_branch->branch_address = cpu_to_le32(d_bus | z);
1278
1279         /*
1280          * VT6306 incorrectly checks only the single descriptor at the
1281          * CommandPtr when the wake bit is written, so if it's a
1282          * multi-descriptor block starting with an INPUT_MORE, put a copy of
1283          * the branch address in the first descriptor.
1284          *
1285          * Not doing this for transmit contexts since not sure how it interacts
1286          * with skip addresses.
1287          */
1288         if (unlikely(ctx->ohci->quirks & QUIRK_IR_WAKE) &&
1289             d_branch != ctx->prev &&
1290             (ctx->prev->control & cpu_to_le16(DESCRIPTOR_CMD)) ==
1291              cpu_to_le16(DESCRIPTOR_INPUT_MORE)) {
1292                 ctx->prev->branch_address = cpu_to_le32(d_bus | z);
1293         }
1294
1295         ctx->prev = d;
1296         ctx->prev_z = z;
1297 }
1298
1299 static void context_stop(struct context *ctx)
1300 {
1301         struct fw_ohci *ohci = ctx->ohci;
1302         u32 reg;
1303         int i;
1304
1305         reg_write(ohci, CONTROL_CLEAR(ctx->regs), CONTEXT_RUN);
1306         ctx->running = false;
1307
1308         for (i = 0; i < 1000; i++) {
1309                 reg = reg_read(ohci, CONTROL_SET(ctx->regs));
1310                 if ((reg & CONTEXT_ACTIVE) == 0)
1311                         return;
1312
1313                 if (i)
1314                         udelay(10);
1315         }
1316         ohci_err(ohci, "DMA context still active (0x%08x)\n", reg);
1317 }
1318
1319 struct driver_data {
1320         u8 inline_data[8];
1321         struct fw_packet *packet;
1322 };
1323
1324 /*
1325  * This function apppends a packet to the DMA queue for transmission.
1326  * Must always be called with the ochi->lock held to ensure proper
1327  * generation handling and locking around packet queue manipulation.
1328  */
1329 static int at_context_queue_packet(struct context *ctx,
1330                                    struct fw_packet *packet)
1331 {
1332         struct fw_ohci *ohci = ctx->ohci;
1333         dma_addr_t d_bus, payload_bus;
1334         struct driver_data *driver_data;
1335         struct descriptor *d, *last;
1336         __le32 *header;
1337         int z, tcode;
1338
1339         d = context_get_descriptors(ctx, 4, &d_bus);
1340         if (d == NULL) {
1341                 packet->ack = RCODE_SEND_ERROR;
1342                 return -1;
1343         }
1344
1345         d[0].control   = cpu_to_le16(DESCRIPTOR_KEY_IMMEDIATE);
1346         d[0].res_count = cpu_to_le16(packet->timestamp);
1347
1348         /*
1349          * The DMA format for asynchronous link packets is different
1350          * from the IEEE1394 layout, so shift the fields around
1351          * accordingly.
1352          */
1353
1354         tcode = (packet->header[0] >> 4) & 0x0f;
1355         header = (__le32 *) &d[1];
1356         switch (tcode) {
1357         case TCODE_WRITE_QUADLET_REQUEST:
1358         case TCODE_WRITE_BLOCK_REQUEST:
1359         case TCODE_WRITE_RESPONSE:
1360         case TCODE_READ_QUADLET_REQUEST:
1361         case TCODE_READ_BLOCK_REQUEST:
1362         case TCODE_READ_QUADLET_RESPONSE:
1363         case TCODE_READ_BLOCK_RESPONSE:
1364         case TCODE_LOCK_REQUEST:
1365         case TCODE_LOCK_RESPONSE:
1366                 header[0] = cpu_to_le32((packet->header[0] & 0xffff) |
1367                                         (packet->speed << 16));
1368                 header[1] = cpu_to_le32((packet->header[1] & 0xffff) |
1369                                         (packet->header[0] & 0xffff0000));
1370                 header[2] = cpu_to_le32(packet->header[2]);
1371
1372                 if (TCODE_IS_BLOCK_PACKET(tcode))
1373                         header[3] = cpu_to_le32(packet->header[3]);
1374                 else
1375                         header[3] = (__force __le32) packet->header[3];
1376
1377                 d[0].req_count = cpu_to_le16(packet->header_length);
1378                 break;
1379
1380         case TCODE_LINK_INTERNAL:
1381                 header[0] = cpu_to_le32((OHCI1394_phy_tcode << 4) |
1382                                         (packet->speed << 16));
1383                 header[1] = cpu_to_le32(packet->header[1]);
1384                 header[2] = cpu_to_le32(packet->header[2]);
1385                 d[0].req_count = cpu_to_le16(12);
1386
1387                 if (is_ping_packet(&packet->header[1]))
1388                         d[0].control |= cpu_to_le16(DESCRIPTOR_PING);
1389                 break;
1390
1391         case TCODE_STREAM_DATA:
1392                 header[0] = cpu_to_le32((packet->header[0] & 0xffff) |
1393                                         (packet->speed << 16));
1394                 header[1] = cpu_to_le32(packet->header[0] & 0xffff0000);
1395                 d[0].req_count = cpu_to_le16(8);
1396                 break;
1397
1398         default:
1399                 /* BUG(); */
1400                 packet->ack = RCODE_SEND_ERROR;
1401                 return -1;
1402         }
1403
1404         BUILD_BUG_ON(sizeof(struct driver_data) > sizeof(struct descriptor));
1405         driver_data = (struct driver_data *) &d[3];
1406         driver_data->packet = packet;
1407         packet->driver_data = driver_data;
1408
1409         if (packet->payload_length > 0) {
1410                 if (packet->payload_length > sizeof(driver_data->inline_data)) {
1411                         payload_bus = dma_map_single(ohci->card.device,
1412                                                      packet->payload,
1413                                                      packet->payload_length,
1414                                                      DMA_TO_DEVICE);
1415                         if (dma_mapping_error(ohci->card.device, payload_bus)) {
1416                                 packet->ack = RCODE_SEND_ERROR;
1417                                 return -1;
1418                         }
1419                         packet->payload_bus     = payload_bus;
1420                         packet->payload_mapped  = true;
1421                 } else {
1422                         memcpy(driver_data->inline_data, packet->payload,
1423                                packet->payload_length);
1424                         payload_bus = d_bus + 3 * sizeof(*d);
1425                 }
1426
1427                 d[2].req_count    = cpu_to_le16(packet->payload_length);
1428                 d[2].data_address = cpu_to_le32(payload_bus);
1429                 last = &d[2];
1430                 z = 3;
1431         } else {
1432                 last = &d[0];
1433                 z = 2;
1434         }
1435
1436         last->control |= cpu_to_le16(DESCRIPTOR_OUTPUT_LAST |
1437                                      DESCRIPTOR_IRQ_ALWAYS |
1438                                      DESCRIPTOR_BRANCH_ALWAYS);
1439
1440         /* FIXME: Document how the locking works. */
1441         if (ohci->generation != packet->generation) {
1442                 if (packet->payload_mapped)
1443                         dma_unmap_single(ohci->card.device, payload_bus,
1444                                          packet->payload_length, DMA_TO_DEVICE);
1445                 packet->ack = RCODE_GENERATION;
1446                 return -1;
1447         }
1448
1449         context_append(ctx, d, z, 4 - z);
1450
1451         if (ctx->running)
1452                 reg_write(ohci, CONTROL_SET(ctx->regs), CONTEXT_WAKE);
1453         else
1454                 context_run(ctx, 0);
1455
1456         return 0;
1457 }
1458
1459 static void at_context_flush(struct context *ctx)
1460 {
1461         tasklet_disable(&ctx->tasklet);
1462
1463         ctx->flushing = true;
1464         context_tasklet((unsigned long)ctx);
1465         ctx->flushing = false;
1466
1467         tasklet_enable(&ctx->tasklet);
1468 }
1469
1470 static int handle_at_packet(struct context *context,
1471                             struct descriptor *d,
1472                             struct descriptor *last)
1473 {
1474         struct driver_data *driver_data;
1475         struct fw_packet *packet;
1476         struct fw_ohci *ohci = context->ohci;
1477         int evt;
1478
1479         if (last->transfer_status == 0 && !context->flushing)
1480                 /* This descriptor isn't done yet, stop iteration. */
1481                 return 0;
1482
1483         driver_data = (struct driver_data *) &d[3];
1484         packet = driver_data->packet;
1485         if (packet == NULL)
1486                 /* This packet was cancelled, just continue. */
1487                 return 1;
1488
1489         if (packet->payload_mapped)
1490                 dma_unmap_single(ohci->card.device, packet->payload_bus,
1491                                  packet->payload_length, DMA_TO_DEVICE);
1492
1493         evt = le16_to_cpu(last->transfer_status) & 0x1f;
1494         packet->timestamp = le16_to_cpu(last->res_count);
1495
1496         log_ar_at_event(ohci, 'T', packet->speed, packet->header, evt);
1497
1498         switch (evt) {
1499         case OHCI1394_evt_timeout:
1500                 /* Async response transmit timed out. */
1501                 packet->ack = RCODE_CANCELLED;
1502                 break;
1503
1504         case OHCI1394_evt_flushed:
1505                 /*
1506                  * The packet was flushed should give same error as
1507                  * when we try to use a stale generation count.
1508                  */
1509                 packet->ack = RCODE_GENERATION;
1510                 break;
1511
1512         case OHCI1394_evt_missing_ack:
1513                 if (context->flushing)
1514                         packet->ack = RCODE_GENERATION;
1515                 else {
1516                         /*
1517                          * Using a valid (current) generation count, but the
1518                          * node is not on the bus or not sending acks.
1519                          */
1520                         packet->ack = RCODE_NO_ACK;
1521                 }
1522                 break;
1523
1524         case ACK_COMPLETE + 0x10:
1525         case ACK_PENDING + 0x10:
1526         case ACK_BUSY_X + 0x10:
1527         case ACK_BUSY_A + 0x10:
1528         case ACK_BUSY_B + 0x10:
1529         case ACK_DATA_ERROR + 0x10:
1530         case ACK_TYPE_ERROR + 0x10:
1531                 packet->ack = evt - 0x10;
1532                 break;
1533
1534         case OHCI1394_evt_no_status:
1535                 if (context->flushing) {
1536                         packet->ack = RCODE_GENERATION;
1537                         break;
1538                 }
1539                 fallthrough;
1540
1541         default:
1542                 packet->ack = RCODE_SEND_ERROR;
1543                 break;
1544         }
1545
1546         packet->callback(packet, &ohci->card, packet->ack);
1547
1548         return 1;
1549 }
1550
1551 #define HEADER_GET_DESTINATION(q)       (((q) >> 16) & 0xffff)
1552 #define HEADER_GET_TCODE(q)             (((q) >> 4) & 0x0f)
1553 #define HEADER_GET_OFFSET_HIGH(q)       (((q) >> 0) & 0xffff)
1554 #define HEADER_GET_DATA_LENGTH(q)       (((q) >> 16) & 0xffff)
1555 #define HEADER_GET_EXTENDED_TCODE(q)    (((q) >> 0) & 0xffff)
1556
1557 static void handle_local_rom(struct fw_ohci *ohci,
1558                              struct fw_packet *packet, u32 csr)
1559 {
1560         struct fw_packet response;
1561         int tcode, length, i;
1562
1563         tcode = HEADER_GET_TCODE(packet->header[0]);
1564         if (TCODE_IS_BLOCK_PACKET(tcode))
1565                 length = HEADER_GET_DATA_LENGTH(packet->header[3]);
1566         else
1567                 length = 4;
1568
1569         i = csr - CSR_CONFIG_ROM;
1570         if (i + length > CONFIG_ROM_SIZE) {
1571                 fw_fill_response(&response, packet->header,
1572                                  RCODE_ADDRESS_ERROR, NULL, 0);
1573         } else if (!TCODE_IS_READ_REQUEST(tcode)) {
1574                 fw_fill_response(&response, packet->header,
1575                                  RCODE_TYPE_ERROR, NULL, 0);
1576         } else {
1577                 fw_fill_response(&response, packet->header, RCODE_COMPLETE,
1578                                  (void *) ohci->config_rom + i, length);
1579         }
1580
1581         fw_core_handle_response(&ohci->card, &response);
1582 }
1583
1584 static void handle_local_lock(struct fw_ohci *ohci,
1585                               struct fw_packet *packet, u32 csr)
1586 {
1587         struct fw_packet response;
1588         int tcode, length, ext_tcode, sel, try;
1589         __be32 *payload, lock_old;
1590         u32 lock_arg, lock_data;
1591
1592         tcode = HEADER_GET_TCODE(packet->header[0]);
1593         length = HEADER_GET_DATA_LENGTH(packet->header[3]);
1594         payload = packet->payload;
1595         ext_tcode = HEADER_GET_EXTENDED_TCODE(packet->header[3]);
1596
1597         if (tcode == TCODE_LOCK_REQUEST &&
1598             ext_tcode == EXTCODE_COMPARE_SWAP && length == 8) {
1599                 lock_arg = be32_to_cpu(payload[0]);
1600                 lock_data = be32_to_cpu(payload[1]);
1601         } else if (tcode == TCODE_READ_QUADLET_REQUEST) {
1602                 lock_arg = 0;
1603                 lock_data = 0;
1604         } else {
1605                 fw_fill_response(&response, packet->header,
1606                                  RCODE_TYPE_ERROR, NULL, 0);
1607                 goto out;
1608         }
1609
1610         sel = (csr - CSR_BUS_MANAGER_ID) / 4;
1611         reg_write(ohci, OHCI1394_CSRData, lock_data);
1612         reg_write(ohci, OHCI1394_CSRCompareData, lock_arg);
1613         reg_write(ohci, OHCI1394_CSRControl, sel);
1614
1615         for (try = 0; try < 20; try++)
1616                 if (reg_read(ohci, OHCI1394_CSRControl) & 0x80000000) {
1617                         lock_old = cpu_to_be32(reg_read(ohci,
1618                                                         OHCI1394_CSRData));
1619                         fw_fill_response(&response, packet->header,
1620                                          RCODE_COMPLETE,
1621                                          &lock_old, sizeof(lock_old));
1622                         goto out;
1623                 }
1624
1625         ohci_err(ohci, "swap not done (CSR lock timeout)\n");
1626         fw_fill_response(&response, packet->header, RCODE_BUSY, NULL, 0);
1627
1628  out:
1629         fw_core_handle_response(&ohci->card, &response);
1630 }
1631
1632 static void handle_local_request(struct context *ctx, struct fw_packet *packet)
1633 {
1634         u64 offset, csr;
1635
1636         if (ctx == &ctx->ohci->at_request_ctx) {
1637                 packet->ack = ACK_PENDING;
1638                 packet->callback(packet, &ctx->ohci->card, packet->ack);
1639         }
1640
1641         offset =
1642                 ((unsigned long long)
1643                  HEADER_GET_OFFSET_HIGH(packet->header[1]) << 32) |
1644                 packet->header[2];
1645         csr = offset - CSR_REGISTER_BASE;
1646
1647         /* Handle config rom reads. */
1648         if (csr >= CSR_CONFIG_ROM && csr < CSR_CONFIG_ROM_END)
1649                 handle_local_rom(ctx->ohci, packet, csr);
1650         else switch (csr) {
1651         case CSR_BUS_MANAGER_ID:
1652         case CSR_BANDWIDTH_AVAILABLE:
1653         case CSR_CHANNELS_AVAILABLE_HI:
1654         case CSR_CHANNELS_AVAILABLE_LO:
1655                 handle_local_lock(ctx->ohci, packet, csr);
1656                 break;
1657         default:
1658                 if (ctx == &ctx->ohci->at_request_ctx)
1659                         fw_core_handle_request(&ctx->ohci->card, packet);
1660                 else
1661                         fw_core_handle_response(&ctx->ohci->card, packet);
1662                 break;
1663         }
1664
1665         if (ctx == &ctx->ohci->at_response_ctx) {
1666                 packet->ack = ACK_COMPLETE;
1667                 packet->callback(packet, &ctx->ohci->card, packet->ack);
1668         }
1669 }
1670
1671 static void at_context_transmit(struct context *ctx, struct fw_packet *packet)
1672 {
1673         unsigned long flags;
1674         int ret;
1675
1676         spin_lock_irqsave(&ctx->ohci->lock, flags);
1677
1678         if (HEADER_GET_DESTINATION(packet->header[0]) == ctx->ohci->node_id &&
1679             ctx->ohci->generation == packet->generation) {
1680                 spin_unlock_irqrestore(&ctx->ohci->lock, flags);
1681                 handle_local_request(ctx, packet);
1682                 return;
1683         }
1684
1685         ret = at_context_queue_packet(ctx, packet);
1686         spin_unlock_irqrestore(&ctx->ohci->lock, flags);
1687
1688         if (ret < 0)
1689                 packet->callback(packet, &ctx->ohci->card, packet->ack);
1690
1691 }
1692
1693 static void detect_dead_context(struct fw_ohci *ohci,
1694                                 const char *name, unsigned int regs)
1695 {
1696         u32 ctl;
1697
1698         ctl = reg_read(ohci, CONTROL_SET(regs));
1699         if (ctl & CONTEXT_DEAD)
1700                 ohci_err(ohci, "DMA context %s has stopped, error code: %s\n",
1701                         name, evts[ctl & 0x1f]);
1702 }
1703
1704 static void handle_dead_contexts(struct fw_ohci *ohci)
1705 {
1706         unsigned int i;
1707         char name[8];
1708
1709         detect_dead_context(ohci, "ATReq", OHCI1394_AsReqTrContextBase);
1710         detect_dead_context(ohci, "ATRsp", OHCI1394_AsRspTrContextBase);
1711         detect_dead_context(ohci, "ARReq", OHCI1394_AsReqRcvContextBase);
1712         detect_dead_context(ohci, "ARRsp", OHCI1394_AsRspRcvContextBase);
1713         for (i = 0; i < 32; ++i) {
1714                 if (!(ohci->it_context_support & (1 << i)))
1715                         continue;
1716                 sprintf(name, "IT%u", i);
1717                 detect_dead_context(ohci, name, OHCI1394_IsoXmitContextBase(i));
1718         }
1719         for (i = 0; i < 32; ++i) {
1720                 if (!(ohci->ir_context_support & (1 << i)))
1721                         continue;
1722                 sprintf(name, "IR%u", i);
1723                 detect_dead_context(ohci, name, OHCI1394_IsoRcvContextBase(i));
1724         }
1725         /* TODO: maybe try to flush and restart the dead contexts */
1726 }
1727
1728 static u32 cycle_timer_ticks(u32 cycle_timer)
1729 {
1730         u32 ticks;
1731
1732         ticks = cycle_timer & 0xfff;
1733         ticks += 3072 * ((cycle_timer >> 12) & 0x1fff);
1734         ticks += (3072 * 8000) * (cycle_timer >> 25);
1735
1736         return ticks;
1737 }
1738
1739 /*
1740  * Some controllers exhibit one or more of the following bugs when updating the
1741  * iso cycle timer register:
1742  *  - When the lowest six bits are wrapping around to zero, a read that happens
1743  *    at the same time will return garbage in the lowest ten bits.
1744  *  - When the cycleOffset field wraps around to zero, the cycleCount field is
1745  *    not incremented for about 60 ns.
1746  *  - Occasionally, the entire register reads zero.
1747  *
1748  * To catch these, we read the register three times and ensure that the
1749  * difference between each two consecutive reads is approximately the same, i.e.
1750  * less than twice the other.  Furthermore, any negative difference indicates an
1751  * error.  (A PCI read should take at least 20 ticks of the 24.576 MHz timer to
1752  * execute, so we have enough precision to compute the ratio of the differences.)
1753  */
1754 static u32 get_cycle_time(struct fw_ohci *ohci)
1755 {
1756         u32 c0, c1, c2;
1757         u32 t0, t1, t2;
1758         s32 diff01, diff12;
1759         int i;
1760
1761         if (has_reboot_by_cycle_timer_read_quirk(ohci))
1762                 return 0;
1763
1764         c2 = reg_read(ohci, OHCI1394_IsochronousCycleTimer);
1765
1766         if (ohci->quirks & QUIRK_CYCLE_TIMER) {
1767                 i = 0;
1768                 c1 = c2;
1769                 c2 = reg_read(ohci, OHCI1394_IsochronousCycleTimer);
1770                 do {
1771                         c0 = c1;
1772                         c1 = c2;
1773                         c2 = reg_read(ohci, OHCI1394_IsochronousCycleTimer);
1774                         t0 = cycle_timer_ticks(c0);
1775                         t1 = cycle_timer_ticks(c1);
1776                         t2 = cycle_timer_ticks(c2);
1777                         diff01 = t1 - t0;
1778                         diff12 = t2 - t1;
1779                 } while ((diff01 <= 0 || diff12 <= 0 ||
1780                           diff01 / diff12 >= 2 || diff12 / diff01 >= 2)
1781                          && i++ < 20);
1782         }
1783
1784         return c2;
1785 }
1786
1787 /*
1788  * This function has to be called at least every 64 seconds.  The bus_time
1789  * field stores not only the upper 25 bits of the BUS_TIME register but also
1790  * the most significant bit of the cycle timer in bit 6 so that we can detect
1791  * changes in this bit.
1792  */
1793 static u32 update_bus_time(struct fw_ohci *ohci)
1794 {
1795         u32 cycle_time_seconds = get_cycle_time(ohci) >> 25;
1796
1797         if (unlikely(!ohci->bus_time_running)) {
1798                 reg_write(ohci, OHCI1394_IntMaskSet, OHCI1394_cycle64Seconds);
1799                 ohci->bus_time = (lower_32_bits(ktime_get_seconds()) & ~0x7f) |
1800                                  (cycle_time_seconds & 0x40);
1801                 ohci->bus_time_running = true;
1802         }
1803
1804         if ((ohci->bus_time & 0x40) != (cycle_time_seconds & 0x40))
1805                 ohci->bus_time += 0x40;
1806
1807         return ohci->bus_time | cycle_time_seconds;
1808 }
1809
1810 static int get_status_for_port(struct fw_ohci *ohci, int port_index)
1811 {
1812         int reg;
1813
1814         mutex_lock(&ohci->phy_reg_mutex);
1815         reg = write_phy_reg(ohci, 7, port_index);
1816         if (reg >= 0)
1817                 reg = read_phy_reg(ohci, 8);
1818         mutex_unlock(&ohci->phy_reg_mutex);
1819         if (reg < 0)
1820                 return reg;
1821
1822         switch (reg & 0x0f) {
1823         case 0x06:
1824                 return 2;       /* is child node (connected to parent node) */
1825         case 0x0e:
1826                 return 3;       /* is parent node (connected to child node) */
1827         }
1828         return 1;               /* not connected */
1829 }
1830
1831 static int get_self_id_pos(struct fw_ohci *ohci, u32 self_id,
1832         int self_id_count)
1833 {
1834         int i;
1835         u32 entry;
1836
1837         for (i = 0; i < self_id_count; i++) {
1838                 entry = ohci->self_id_buffer[i];
1839                 if ((self_id & 0xff000000) == (entry & 0xff000000))
1840                         return -1;
1841                 if ((self_id & 0xff000000) < (entry & 0xff000000))
1842                         return i;
1843         }
1844         return i;
1845 }
1846
1847 static int initiated_reset(struct fw_ohci *ohci)
1848 {
1849         int reg;
1850         int ret = 0;
1851
1852         mutex_lock(&ohci->phy_reg_mutex);
1853         reg = write_phy_reg(ohci, 7, 0xe0); /* Select page 7 */
1854         if (reg >= 0) {
1855                 reg = read_phy_reg(ohci, 8);
1856                 reg |= 0x40;
1857                 reg = write_phy_reg(ohci, 8, reg); /* set PMODE bit */
1858                 if (reg >= 0) {
1859                         reg = read_phy_reg(ohci, 12); /* read register 12 */
1860                         if (reg >= 0) {
1861                                 if ((reg & 0x08) == 0x08) {
1862                                         /* bit 3 indicates "initiated reset" */
1863                                         ret = 0x2;
1864                                 }
1865                         }
1866                 }
1867         }
1868         mutex_unlock(&ohci->phy_reg_mutex);
1869         return ret;
1870 }
1871
1872 /*
1873  * TI TSB82AA2B and TSB12LV26 do not receive the selfID of a locally
1874  * attached TSB41BA3D phy; see http://www.ti.com/litv/pdf/sllz059.
1875  * Construct the selfID from phy register contents.
1876  */
1877 static int find_and_insert_self_id(struct fw_ohci *ohci, int self_id_count)
1878 {
1879         int reg, i, pos, status;
1880         /* link active 1, speed 3, bridge 0, contender 1, more packets 0 */
1881         u32 self_id = 0x8040c800;
1882
1883         reg = reg_read(ohci, OHCI1394_NodeID);
1884         if (!(reg & OHCI1394_NodeID_idValid)) {
1885                 ohci_notice(ohci,
1886                             "node ID not valid, new bus reset in progress\n");
1887                 return -EBUSY;
1888         }
1889         self_id |= ((reg & 0x3f) << 24); /* phy ID */
1890
1891         reg = ohci_read_phy_reg(&ohci->card, 4);
1892         if (reg < 0)
1893                 return reg;
1894         self_id |= ((reg & 0x07) << 8); /* power class */
1895
1896         reg = ohci_read_phy_reg(&ohci->card, 1);
1897         if (reg < 0)
1898                 return reg;
1899         self_id |= ((reg & 0x3f) << 16); /* gap count */
1900
1901         for (i = 0; i < 3; i++) {
1902                 status = get_status_for_port(ohci, i);
1903                 if (status < 0)
1904                         return status;
1905                 self_id |= ((status & 0x3) << (6 - (i * 2)));
1906         }
1907
1908         self_id |= initiated_reset(ohci);
1909
1910         pos = get_self_id_pos(ohci, self_id, self_id_count);
1911         if (pos >= 0) {
1912                 memmove(&(ohci->self_id_buffer[pos+1]),
1913                         &(ohci->self_id_buffer[pos]),
1914                         (self_id_count - pos) * sizeof(*ohci->self_id_buffer));
1915                 ohci->self_id_buffer[pos] = self_id;
1916                 self_id_count++;
1917         }
1918         return self_id_count;
1919 }
1920
1921 static void bus_reset_work(struct work_struct *work)
1922 {
1923         struct fw_ohci *ohci =
1924                 container_of(work, struct fw_ohci, bus_reset_work);
1925         int self_id_count, generation, new_generation, i, j;
1926         u32 reg;
1927         void *free_rom = NULL;
1928         dma_addr_t free_rom_bus = 0;
1929         bool is_new_root;
1930
1931         reg = reg_read(ohci, OHCI1394_NodeID);
1932         if (!(reg & OHCI1394_NodeID_idValid)) {
1933                 ohci_notice(ohci,
1934                             "node ID not valid, new bus reset in progress\n");
1935                 return;
1936         }
1937         if ((reg & OHCI1394_NodeID_nodeNumber) == 63) {
1938                 ohci_notice(ohci, "malconfigured bus\n");
1939                 return;
1940         }
1941         ohci->node_id = reg & (OHCI1394_NodeID_busNumber |
1942                                OHCI1394_NodeID_nodeNumber);
1943
1944         is_new_root = (reg & OHCI1394_NodeID_root) != 0;
1945         if (!(ohci->is_root && is_new_root))
1946                 reg_write(ohci, OHCI1394_LinkControlSet,
1947                           OHCI1394_LinkControl_cycleMaster);
1948         ohci->is_root = is_new_root;
1949
1950         reg = reg_read(ohci, OHCI1394_SelfIDCount);
1951         if (reg & OHCI1394_SelfIDCount_selfIDError) {
1952                 ohci_notice(ohci, "self ID receive error\n");
1953                 return;
1954         }
1955         /*
1956          * The count in the SelfIDCount register is the number of
1957          * bytes in the self ID receive buffer.  Since we also receive
1958          * the inverted quadlets and a header quadlet, we shift one
1959          * bit extra to get the actual number of self IDs.
1960          */
1961         self_id_count = (reg >> 3) & 0xff;
1962
1963         if (self_id_count > 252) {
1964                 ohci_notice(ohci, "bad selfIDSize (%08x)\n", reg);
1965                 return;
1966         }
1967
1968         generation = (cond_le32_to_cpu(ohci->self_id[0]) >> 16) & 0xff;
1969         rmb();
1970
1971         for (i = 1, j = 0; j < self_id_count; i += 2, j++) {
1972                 u32 id  = cond_le32_to_cpu(ohci->self_id[i]);
1973                 u32 id2 = cond_le32_to_cpu(ohci->self_id[i + 1]);
1974
1975                 if (id != ~id2) {
1976                         /*
1977                          * If the invalid data looks like a cycle start packet,
1978                          * it's likely to be the result of the cycle master
1979                          * having a wrong gap count.  In this case, the self IDs
1980                          * so far are valid and should be processed so that the
1981                          * bus manager can then correct the gap count.
1982                          */
1983                         if (id == 0xffff008f) {
1984                                 ohci_notice(ohci, "ignoring spurious self IDs\n");
1985                                 self_id_count = j;
1986                                 break;
1987                         }
1988
1989                         ohci_notice(ohci, "bad self ID %d/%d (%08x != ~%08x)\n",
1990                                     j, self_id_count, id, id2);
1991                         return;
1992                 }
1993                 ohci->self_id_buffer[j] = id;
1994         }
1995
1996         if (ohci->quirks & QUIRK_TI_SLLZ059) {
1997                 self_id_count = find_and_insert_self_id(ohci, self_id_count);
1998                 if (self_id_count < 0) {
1999                         ohci_notice(ohci,
2000                                     "could not construct local self ID\n");
2001                         return;
2002                 }
2003         }
2004
2005         if (self_id_count == 0) {
2006                 ohci_notice(ohci, "no self IDs\n");
2007                 return;
2008         }
2009         rmb();
2010
2011         /*
2012          * Check the consistency of the self IDs we just read.  The
2013          * problem we face is that a new bus reset can start while we
2014          * read out the self IDs from the DMA buffer. If this happens,
2015          * the DMA buffer will be overwritten with new self IDs and we
2016          * will read out inconsistent data.  The OHCI specification
2017          * (section 11.2) recommends a technique similar to
2018          * linux/seqlock.h, where we remember the generation of the
2019          * self IDs in the buffer before reading them out and compare
2020          * it to the current generation after reading them out.  If
2021          * the two generations match we know we have a consistent set
2022          * of self IDs.
2023          */
2024
2025         new_generation = (reg_read(ohci, OHCI1394_SelfIDCount) >> 16) & 0xff;
2026         if (new_generation != generation) {
2027                 ohci_notice(ohci, "new bus reset, discarding self ids\n");
2028                 return;
2029         }
2030
2031         /* FIXME: Document how the locking works. */
2032         spin_lock_irq(&ohci->lock);
2033
2034         ohci->generation = -1; /* prevent AT packet queueing */
2035         context_stop(&ohci->at_request_ctx);
2036         context_stop(&ohci->at_response_ctx);
2037
2038         spin_unlock_irq(&ohci->lock);
2039
2040         /*
2041          * Per OHCI 1.2 draft, clause 7.2.3.3, hardware may leave unsent
2042          * packets in the AT queues and software needs to drain them.
2043          * Some OHCI 1.1 controllers (JMicron) apparently require this too.
2044          */
2045         at_context_flush(&ohci->at_request_ctx);
2046         at_context_flush(&ohci->at_response_ctx);
2047
2048         spin_lock_irq(&ohci->lock);
2049
2050         ohci->generation = generation;
2051         reg_write(ohci, OHCI1394_IntEventClear, OHCI1394_busReset);
2052         if (param_debug & OHCI_PARAM_DEBUG_BUSRESETS)
2053                 reg_write(ohci, OHCI1394_IntMaskSet, OHCI1394_busReset);
2054
2055         if (ohci->quirks & QUIRK_RESET_PACKET)
2056                 ohci->request_generation = generation;
2057
2058         /*
2059          * This next bit is unrelated to the AT context stuff but we
2060          * have to do it under the spinlock also.  If a new config rom
2061          * was set up before this reset, the old one is now no longer
2062          * in use and we can free it. Update the config rom pointers
2063          * to point to the current config rom and clear the
2064          * next_config_rom pointer so a new update can take place.
2065          */
2066
2067         if (ohci->next_config_rom != NULL) {
2068                 if (ohci->next_config_rom != ohci->config_rom) {
2069                         free_rom      = ohci->config_rom;
2070                         free_rom_bus  = ohci->config_rom_bus;
2071                 }
2072                 ohci->config_rom      = ohci->next_config_rom;
2073                 ohci->config_rom_bus  = ohci->next_config_rom_bus;
2074                 ohci->next_config_rom = NULL;
2075
2076                 /*
2077                  * Restore config_rom image and manually update
2078                  * config_rom registers.  Writing the header quadlet
2079                  * will indicate that the config rom is ready, so we
2080                  * do that last.
2081                  */
2082                 reg_write(ohci, OHCI1394_BusOptions,
2083                           be32_to_cpu(ohci->config_rom[2]));
2084                 ohci->config_rom[0] = ohci->next_header;
2085                 reg_write(ohci, OHCI1394_ConfigROMhdr,
2086                           be32_to_cpu(ohci->next_header));
2087         }
2088
2089         if (param_remote_dma) {
2090                 reg_write(ohci, OHCI1394_PhyReqFilterHiSet, ~0);
2091                 reg_write(ohci, OHCI1394_PhyReqFilterLoSet, ~0);
2092         }
2093
2094         spin_unlock_irq(&ohci->lock);
2095
2096         if (free_rom)
2097                 dma_free_coherent(ohci->card.device, CONFIG_ROM_SIZE,
2098                                   free_rom, free_rom_bus);
2099
2100         log_selfids(ohci, generation, self_id_count);
2101
2102         fw_core_handle_bus_reset(&ohci->card, ohci->node_id, generation,
2103                                  self_id_count, ohci->self_id_buffer,
2104                                  ohci->csr_state_setclear_abdicate);
2105         ohci->csr_state_setclear_abdicate = false;
2106 }
2107
2108 static irqreturn_t irq_handler(int irq, void *data)
2109 {
2110         struct fw_ohci *ohci = data;
2111         u32 event, iso_event;
2112         int i;
2113
2114         event = reg_read(ohci, OHCI1394_IntEventClear);
2115
2116         if (!event || !~event)
2117                 return IRQ_NONE;
2118
2119         /*
2120          * busReset and postedWriteErr events must not be cleared yet
2121          * (OHCI 1.1 clauses 7.2.3.2 and 13.2.8.1)
2122          */
2123         reg_write(ohci, OHCI1394_IntEventClear,
2124                   event & ~(OHCI1394_busReset | OHCI1394_postedWriteErr));
2125         log_irqs(ohci, event);
2126         if (event & OHCI1394_busReset)
2127                 reg_write(ohci, OHCI1394_IntMaskClear, OHCI1394_busReset);
2128
2129         if (event & OHCI1394_selfIDComplete)
2130                 queue_work(selfid_workqueue, &ohci->bus_reset_work);
2131
2132         if (event & OHCI1394_RQPkt)
2133                 tasklet_schedule(&ohci->ar_request_ctx.tasklet);
2134
2135         if (event & OHCI1394_RSPkt)
2136                 tasklet_schedule(&ohci->ar_response_ctx.tasklet);
2137
2138         if (event & OHCI1394_reqTxComplete)
2139                 tasklet_schedule(&ohci->at_request_ctx.tasklet);
2140
2141         if (event & OHCI1394_respTxComplete)
2142                 tasklet_schedule(&ohci->at_response_ctx.tasklet);
2143
2144         if (event & OHCI1394_isochRx) {
2145                 iso_event = reg_read(ohci, OHCI1394_IsoRecvIntEventClear);
2146                 reg_write(ohci, OHCI1394_IsoRecvIntEventClear, iso_event);
2147
2148                 while (iso_event) {
2149                         i = ffs(iso_event) - 1;
2150                         tasklet_schedule(
2151                                 &ohci->ir_context_list[i].context.tasklet);
2152                         iso_event &= ~(1 << i);
2153                 }
2154         }
2155
2156         if (event & OHCI1394_isochTx) {
2157                 iso_event = reg_read(ohci, OHCI1394_IsoXmitIntEventClear);
2158                 reg_write(ohci, OHCI1394_IsoXmitIntEventClear, iso_event);
2159
2160                 while (iso_event) {
2161                         i = ffs(iso_event) - 1;
2162                         tasklet_schedule(
2163                                 &ohci->it_context_list[i].context.tasklet);
2164                         iso_event &= ~(1 << i);
2165                 }
2166         }
2167
2168         if (unlikely(event & OHCI1394_regAccessFail))
2169                 ohci_err(ohci, "register access failure\n");
2170
2171         if (unlikely(event & OHCI1394_postedWriteErr)) {
2172                 reg_read(ohci, OHCI1394_PostedWriteAddressHi);
2173                 reg_read(ohci, OHCI1394_PostedWriteAddressLo);
2174                 reg_write(ohci, OHCI1394_IntEventClear,
2175                           OHCI1394_postedWriteErr);
2176                 if (printk_ratelimit())
2177                         ohci_err(ohci, "PCI posted write error\n");
2178         }
2179
2180         if (unlikely(event & OHCI1394_cycleTooLong)) {
2181                 if (printk_ratelimit())
2182                         ohci_notice(ohci, "isochronous cycle too long\n");
2183                 reg_write(ohci, OHCI1394_LinkControlSet,
2184                           OHCI1394_LinkControl_cycleMaster);
2185         }
2186
2187         if (unlikely(event & OHCI1394_cycleInconsistent)) {
2188                 /*
2189                  * We need to clear this event bit in order to make
2190                  * cycleMatch isochronous I/O work.  In theory we should
2191                  * stop active cycleMatch iso contexts now and restart
2192                  * them at least two cycles later.  (FIXME?)
2193                  */
2194                 if (printk_ratelimit())
2195                         ohci_notice(ohci, "isochronous cycle inconsistent\n");
2196         }
2197
2198         if (unlikely(event & OHCI1394_unrecoverableError))
2199                 handle_dead_contexts(ohci);
2200
2201         if (event & OHCI1394_cycle64Seconds) {
2202                 spin_lock(&ohci->lock);
2203                 update_bus_time(ohci);
2204                 spin_unlock(&ohci->lock);
2205         } else
2206                 flush_writes(ohci);
2207
2208         return IRQ_HANDLED;
2209 }
2210
2211 static int software_reset(struct fw_ohci *ohci)
2212 {
2213         u32 val;
2214         int i;
2215
2216         reg_write(ohci, OHCI1394_HCControlSet, OHCI1394_HCControl_softReset);
2217         for (i = 0; i < 500; i++) {
2218                 val = reg_read(ohci, OHCI1394_HCControlSet);
2219                 if (!~val)
2220                         return -ENODEV; /* Card was ejected. */
2221
2222                 if (!(val & OHCI1394_HCControl_softReset))
2223                         return 0;
2224
2225                 msleep(1);
2226         }
2227
2228         return -EBUSY;
2229 }
2230
2231 static void copy_config_rom(__be32 *dest, const __be32 *src, size_t length)
2232 {
2233         size_t size = length * 4;
2234
2235         memcpy(dest, src, size);
2236         if (size < CONFIG_ROM_SIZE)
2237                 memset(&dest[length], 0, CONFIG_ROM_SIZE - size);
2238 }
2239
2240 static int configure_1394a_enhancements(struct fw_ohci *ohci)
2241 {
2242         bool enable_1394a;
2243         int ret, clear, set, offset;
2244
2245         /* Check if the driver should configure link and PHY. */
2246         if (!(reg_read(ohci, OHCI1394_HCControlSet) &
2247               OHCI1394_HCControl_programPhyEnable))
2248                 return 0;
2249
2250         /* Paranoia: check whether the PHY supports 1394a, too. */
2251         enable_1394a = false;
2252         ret = read_phy_reg(ohci, 2);
2253         if (ret < 0)
2254                 return ret;
2255         if ((ret & PHY_EXTENDED_REGISTERS) == PHY_EXTENDED_REGISTERS) {
2256                 ret = read_paged_phy_reg(ohci, 1, 8);
2257                 if (ret < 0)
2258                         return ret;
2259                 if (ret >= 1)
2260                         enable_1394a = true;
2261         }
2262
2263         if (ohci->quirks & QUIRK_NO_1394A)
2264                 enable_1394a = false;
2265
2266         /* Configure PHY and link consistently. */
2267         if (enable_1394a) {
2268                 clear = 0;
2269                 set = PHY_ENABLE_ACCEL | PHY_ENABLE_MULTI;
2270         } else {
2271                 clear = PHY_ENABLE_ACCEL | PHY_ENABLE_MULTI;
2272                 set = 0;
2273         }
2274         ret = update_phy_reg(ohci, 5, clear, set);
2275         if (ret < 0)
2276                 return ret;
2277
2278         if (enable_1394a)
2279                 offset = OHCI1394_HCControlSet;
2280         else
2281                 offset = OHCI1394_HCControlClear;
2282         reg_write(ohci, offset, OHCI1394_HCControl_aPhyEnhanceEnable);
2283
2284         /* Clean up: configuration has been taken care of. */
2285         reg_write(ohci, OHCI1394_HCControlClear,
2286                   OHCI1394_HCControl_programPhyEnable);
2287
2288         return 0;
2289 }
2290
2291 static int probe_tsb41ba3d(struct fw_ohci *ohci)
2292 {
2293         /* TI vendor ID = 0x080028, TSB41BA3D product ID = 0x833005 (sic) */
2294         static const u8 id[] = { 0x08, 0x00, 0x28, 0x83, 0x30, 0x05, };
2295         int reg, i;
2296
2297         reg = read_phy_reg(ohci, 2);
2298         if (reg < 0)
2299                 return reg;
2300         if ((reg & PHY_EXTENDED_REGISTERS) != PHY_EXTENDED_REGISTERS)
2301                 return 0;
2302
2303         for (i = ARRAY_SIZE(id) - 1; i >= 0; i--) {
2304                 reg = read_paged_phy_reg(ohci, 1, i + 10);
2305                 if (reg < 0)
2306                         return reg;
2307                 if (reg != id[i])
2308                         return 0;
2309         }
2310         return 1;
2311 }
2312
2313 static int ohci_enable(struct fw_card *card,
2314                        const __be32 *config_rom, size_t length)
2315 {
2316         struct fw_ohci *ohci = fw_ohci(card);
2317         u32 lps, version, irqs;
2318         int i, ret;
2319
2320         ret = software_reset(ohci);
2321         if (ret < 0) {
2322                 ohci_err(ohci, "failed to reset ohci card\n");
2323                 return ret;
2324         }
2325
2326         /*
2327          * Now enable LPS, which we need in order to start accessing
2328          * most of the registers.  In fact, on some cards (ALI M5251),
2329          * accessing registers in the SClk domain without LPS enabled
2330          * will lock up the machine.  Wait 50msec to make sure we have
2331          * full link enabled.  However, with some cards (well, at least
2332          * a JMicron PCIe card), we have to try again sometimes.
2333          *
2334          * TI TSB82AA2 + TSB81BA3(A) cards signal LPS enabled early but
2335          * cannot actually use the phy at that time.  These need tens of
2336          * millisecods pause between LPS write and first phy access too.
2337          */
2338
2339         reg_write(ohci, OHCI1394_HCControlSet,
2340                   OHCI1394_HCControl_LPS |
2341                   OHCI1394_HCControl_postedWriteEnable);
2342         flush_writes(ohci);
2343
2344         for (lps = 0, i = 0; !lps && i < 3; i++) {
2345                 msleep(50);
2346                 lps = reg_read(ohci, OHCI1394_HCControlSet) &
2347                       OHCI1394_HCControl_LPS;
2348         }
2349
2350         if (!lps) {
2351                 ohci_err(ohci, "failed to set Link Power Status\n");
2352                 return -EIO;
2353         }
2354
2355         if (ohci->quirks & QUIRK_TI_SLLZ059) {
2356                 ret = probe_tsb41ba3d(ohci);
2357                 if (ret < 0)
2358                         return ret;
2359                 if (ret)
2360                         ohci_notice(ohci, "local TSB41BA3D phy\n");
2361                 else
2362                         ohci->quirks &= ~QUIRK_TI_SLLZ059;
2363         }
2364
2365         reg_write(ohci, OHCI1394_HCControlClear,
2366                   OHCI1394_HCControl_noByteSwapData);
2367
2368         reg_write(ohci, OHCI1394_SelfIDBuffer, ohci->self_id_bus);
2369         reg_write(ohci, OHCI1394_LinkControlSet,
2370                   OHCI1394_LinkControl_cycleTimerEnable |
2371                   OHCI1394_LinkControl_cycleMaster);
2372
2373         reg_write(ohci, OHCI1394_ATRetries,
2374                   OHCI1394_MAX_AT_REQ_RETRIES |
2375                   (OHCI1394_MAX_AT_RESP_RETRIES << 4) |
2376                   (OHCI1394_MAX_PHYS_RESP_RETRIES << 8) |
2377                   (200 << 16));
2378
2379         ohci->bus_time_running = false;
2380
2381         for (i = 0; i < 32; i++)
2382                 if (ohci->ir_context_support & (1 << i))
2383                         reg_write(ohci, OHCI1394_IsoRcvContextControlClear(i),
2384                                   IR_CONTEXT_MULTI_CHANNEL_MODE);
2385
2386         version = reg_read(ohci, OHCI1394_Version) & 0x00ff00ff;
2387         if (version >= OHCI_VERSION_1_1) {
2388                 reg_write(ohci, OHCI1394_InitialChannelsAvailableHi,
2389                           0xfffffffe);
2390                 card->broadcast_channel_auto_allocated = true;
2391         }
2392
2393         /* Get implemented bits of the priority arbitration request counter. */
2394         reg_write(ohci, OHCI1394_FairnessControl, 0x3f);
2395         ohci->pri_req_max = reg_read(ohci, OHCI1394_FairnessControl) & 0x3f;
2396         reg_write(ohci, OHCI1394_FairnessControl, 0);
2397         card->priority_budget_implemented = ohci->pri_req_max != 0;
2398
2399         reg_write(ohci, OHCI1394_PhyUpperBound, FW_MAX_PHYSICAL_RANGE >> 16);
2400         reg_write(ohci, OHCI1394_IntEventClear, ~0);
2401         reg_write(ohci, OHCI1394_IntMaskClear, ~0);
2402
2403         ret = configure_1394a_enhancements(ohci);
2404         if (ret < 0)
2405                 return ret;
2406
2407         /* Activate link_on bit and contender bit in our self ID packets.*/
2408         ret = ohci_update_phy_reg(card, 4, 0, PHY_LINK_ACTIVE | PHY_CONTENDER);
2409         if (ret < 0)
2410                 return ret;
2411
2412         /*
2413          * When the link is not yet enabled, the atomic config rom
2414          * update mechanism described below in ohci_set_config_rom()
2415          * is not active.  We have to update ConfigRomHeader and
2416          * BusOptions manually, and the write to ConfigROMmap takes
2417          * effect immediately.  We tie this to the enabling of the
2418          * link, so we have a valid config rom before enabling - the
2419          * OHCI requires that ConfigROMhdr and BusOptions have valid
2420          * values before enabling.
2421          *
2422          * However, when the ConfigROMmap is written, some controllers
2423          * always read back quadlets 0 and 2 from the config rom to
2424          * the ConfigRomHeader and BusOptions registers on bus reset.
2425          * They shouldn't do that in this initial case where the link
2426          * isn't enabled.  This means we have to use the same
2427          * workaround here, setting the bus header to 0 and then write
2428          * the right values in the bus reset tasklet.
2429          */
2430
2431         if (config_rom) {
2432                 ohci->next_config_rom =
2433                         dma_alloc_coherent(ohci->card.device, CONFIG_ROM_SIZE,
2434                                            &ohci->next_config_rom_bus,
2435                                            GFP_KERNEL);
2436                 if (ohci->next_config_rom == NULL)
2437                         return -ENOMEM;
2438
2439                 copy_config_rom(ohci->next_config_rom, config_rom, length);
2440         } else {
2441                 /*
2442                  * In the suspend case, config_rom is NULL, which
2443                  * means that we just reuse the old config rom.
2444                  */
2445                 ohci->next_config_rom = ohci->config_rom;
2446                 ohci->next_config_rom_bus = ohci->config_rom_bus;
2447         }
2448
2449         ohci->next_header = ohci->next_config_rom[0];
2450         ohci->next_config_rom[0] = 0;
2451         reg_write(ohci, OHCI1394_ConfigROMhdr, 0);
2452         reg_write(ohci, OHCI1394_BusOptions,
2453                   be32_to_cpu(ohci->next_config_rom[2]));
2454         reg_write(ohci, OHCI1394_ConfigROMmap, ohci->next_config_rom_bus);
2455
2456         reg_write(ohci, OHCI1394_AsReqFilterHiSet, 0x80000000);
2457
2458         irqs =  OHCI1394_reqTxComplete | OHCI1394_respTxComplete |
2459                 OHCI1394_RQPkt | OHCI1394_RSPkt |
2460                 OHCI1394_isochTx | OHCI1394_isochRx |
2461                 OHCI1394_postedWriteErr |
2462                 OHCI1394_selfIDComplete |
2463                 OHCI1394_regAccessFail |
2464                 OHCI1394_cycleInconsistent |
2465                 OHCI1394_unrecoverableError |
2466                 OHCI1394_cycleTooLong |
2467                 OHCI1394_masterIntEnable;
2468         if (param_debug & OHCI_PARAM_DEBUG_BUSRESETS)
2469                 irqs |= OHCI1394_busReset;
2470         reg_write(ohci, OHCI1394_IntMaskSet, irqs);
2471
2472         reg_write(ohci, OHCI1394_HCControlSet,
2473                   OHCI1394_HCControl_linkEnable |
2474                   OHCI1394_HCControl_BIBimageValid);
2475
2476         reg_write(ohci, OHCI1394_LinkControlSet,
2477                   OHCI1394_LinkControl_rcvSelfID |
2478                   OHCI1394_LinkControl_rcvPhyPkt);
2479
2480         ar_context_run(&ohci->ar_request_ctx);
2481         ar_context_run(&ohci->ar_response_ctx);
2482
2483         flush_writes(ohci);
2484
2485         /* We are ready to go, reset bus to finish initialization. */
2486         fw_schedule_bus_reset(&ohci->card, false, true);
2487
2488         return 0;
2489 }
2490
2491 static int ohci_set_config_rom(struct fw_card *card,
2492                                const __be32 *config_rom, size_t length)
2493 {
2494         struct fw_ohci *ohci;
2495         __be32 *next_config_rom;
2496         dma_addr_t next_config_rom_bus;
2497
2498         ohci = fw_ohci(card);
2499
2500         /*
2501          * When the OHCI controller is enabled, the config rom update
2502          * mechanism is a bit tricky, but easy enough to use.  See
2503          * section 5.5.6 in the OHCI specification.
2504          *
2505          * The OHCI controller caches the new config rom address in a
2506          * shadow register (ConfigROMmapNext) and needs a bus reset
2507          * for the changes to take place.  When the bus reset is
2508          * detected, the controller loads the new values for the
2509          * ConfigRomHeader and BusOptions registers from the specified
2510          * config rom and loads ConfigROMmap from the ConfigROMmapNext
2511          * shadow register. All automatically and atomically.
2512          *
2513          * Now, there's a twist to this story.  The automatic load of
2514          * ConfigRomHeader and BusOptions doesn't honor the
2515          * noByteSwapData bit, so with a be32 config rom, the
2516          * controller will load be32 values in to these registers
2517          * during the atomic update, even on litte endian
2518          * architectures.  The workaround we use is to put a 0 in the
2519          * header quadlet; 0 is endian agnostic and means that the
2520          * config rom isn't ready yet.  In the bus reset tasklet we
2521          * then set up the real values for the two registers.
2522          *
2523          * We use ohci->lock to avoid racing with the code that sets
2524          * ohci->next_config_rom to NULL (see bus_reset_work).
2525          */
2526
2527         next_config_rom =
2528                 dma_alloc_coherent(ohci->card.device, CONFIG_ROM_SIZE,
2529                                    &next_config_rom_bus, GFP_KERNEL);
2530         if (next_config_rom == NULL)
2531                 return -ENOMEM;
2532
2533         spin_lock_irq(&ohci->lock);
2534
2535         /*
2536          * If there is not an already pending config_rom update,
2537          * push our new allocation into the ohci->next_config_rom
2538          * and then mark the local variable as null so that we
2539          * won't deallocate the new buffer.
2540          *
2541          * OTOH, if there is a pending config_rom update, just
2542          * use that buffer with the new config_rom data, and
2543          * let this routine free the unused DMA allocation.
2544          */
2545
2546         if (ohci->next_config_rom == NULL) {
2547                 ohci->next_config_rom = next_config_rom;
2548                 ohci->next_config_rom_bus = next_config_rom_bus;
2549                 next_config_rom = NULL;
2550         }
2551
2552         copy_config_rom(ohci->next_config_rom, config_rom, length);
2553
2554         ohci->next_header = config_rom[0];
2555         ohci->next_config_rom[0] = 0;
2556
2557         reg_write(ohci, OHCI1394_ConfigROMmap, ohci->next_config_rom_bus);
2558
2559         spin_unlock_irq(&ohci->lock);
2560
2561         /* If we didn't use the DMA allocation, delete it. */
2562         if (next_config_rom != NULL)
2563                 dma_free_coherent(ohci->card.device, CONFIG_ROM_SIZE,
2564                                   next_config_rom, next_config_rom_bus);
2565
2566         /*
2567          * Now initiate a bus reset to have the changes take
2568          * effect. We clean up the old config rom memory and DMA
2569          * mappings in the bus reset tasklet, since the OHCI
2570          * controller could need to access it before the bus reset
2571          * takes effect.
2572          */
2573
2574         fw_schedule_bus_reset(&ohci->card, true, true);
2575
2576         return 0;
2577 }
2578
2579 static void ohci_send_request(struct fw_card *card, struct fw_packet *packet)
2580 {
2581         struct fw_ohci *ohci = fw_ohci(card);
2582
2583         at_context_transmit(&ohci->at_request_ctx, packet);
2584 }
2585
2586 static void ohci_send_response(struct fw_card *card, struct fw_packet *packet)
2587 {
2588         struct fw_ohci *ohci = fw_ohci(card);
2589
2590         at_context_transmit(&ohci->at_response_ctx, packet);
2591 }
2592
2593 static int ohci_cancel_packet(struct fw_card *card, struct fw_packet *packet)
2594 {
2595         struct fw_ohci *ohci = fw_ohci(card);
2596         struct context *ctx = &ohci->at_request_ctx;
2597         struct driver_data *driver_data = packet->driver_data;
2598         int ret = -ENOENT;
2599
2600         tasklet_disable(&ctx->tasklet);
2601
2602         if (packet->ack != 0)
2603                 goto out;
2604
2605         if (packet->payload_mapped)
2606                 dma_unmap_single(ohci->card.device, packet->payload_bus,
2607                                  packet->payload_length, DMA_TO_DEVICE);
2608
2609         log_ar_at_event(ohci, 'T', packet->speed, packet->header, 0x20);
2610         driver_data->packet = NULL;
2611         packet->ack = RCODE_CANCELLED;
2612         packet->callback(packet, &ohci->card, packet->ack);
2613         ret = 0;
2614  out:
2615         tasklet_enable(&ctx->tasklet);
2616
2617         return ret;
2618 }
2619
2620 static int ohci_enable_phys_dma(struct fw_card *card,
2621                                 int node_id, int generation)
2622 {
2623         struct fw_ohci *ohci = fw_ohci(card);
2624         unsigned long flags;
2625         int n, ret = 0;
2626
2627         if (param_remote_dma)
2628                 return 0;
2629
2630         /*
2631          * FIXME:  Make sure this bitmask is cleared when we clear the busReset
2632          * interrupt bit.  Clear physReqResourceAllBuses on bus reset.
2633          */
2634
2635         spin_lock_irqsave(&ohci->lock, flags);
2636
2637         if (ohci->generation != generation) {
2638                 ret = -ESTALE;
2639                 goto out;
2640         }
2641
2642         /*
2643          * Note, if the node ID contains a non-local bus ID, physical DMA is
2644          * enabled for _all_ nodes on remote buses.
2645          */
2646
2647         n = (node_id & 0xffc0) == LOCAL_BUS ? node_id & 0x3f : 63;
2648         if (n < 32)
2649                 reg_write(ohci, OHCI1394_PhyReqFilterLoSet, 1 << n);
2650         else
2651                 reg_write(ohci, OHCI1394_PhyReqFilterHiSet, 1 << (n - 32));
2652
2653         flush_writes(ohci);
2654  out:
2655         spin_unlock_irqrestore(&ohci->lock, flags);
2656
2657         return ret;
2658 }
2659
2660 static u32 ohci_read_csr(struct fw_card *card, int csr_offset)
2661 {
2662         struct fw_ohci *ohci = fw_ohci(card);
2663         unsigned long flags;
2664         u32 value;
2665
2666         switch (csr_offset) {
2667         case CSR_STATE_CLEAR:
2668         case CSR_STATE_SET:
2669                 if (ohci->is_root &&
2670                     (reg_read(ohci, OHCI1394_LinkControlSet) &
2671                      OHCI1394_LinkControl_cycleMaster))
2672                         value = CSR_STATE_BIT_CMSTR;
2673                 else
2674                         value = 0;
2675                 if (ohci->csr_state_setclear_abdicate)
2676                         value |= CSR_STATE_BIT_ABDICATE;
2677
2678                 return value;
2679
2680         case CSR_NODE_IDS:
2681                 return reg_read(ohci, OHCI1394_NodeID) << 16;
2682
2683         case CSR_CYCLE_TIME:
2684                 return get_cycle_time(ohci);
2685
2686         case CSR_BUS_TIME:
2687                 /*
2688                  * We might be called just after the cycle timer has wrapped
2689                  * around but just before the cycle64Seconds handler, so we
2690                  * better check here, too, if the bus time needs to be updated.
2691                  */
2692                 spin_lock_irqsave(&ohci->lock, flags);
2693                 value = update_bus_time(ohci);
2694                 spin_unlock_irqrestore(&ohci->lock, flags);
2695                 return value;
2696
2697         case CSR_BUSY_TIMEOUT:
2698                 value = reg_read(ohci, OHCI1394_ATRetries);
2699                 return (value >> 4) & 0x0ffff00f;
2700
2701         case CSR_PRIORITY_BUDGET:
2702                 return (reg_read(ohci, OHCI1394_FairnessControl) & 0x3f) |
2703                         (ohci->pri_req_max << 8);
2704
2705         default:
2706                 WARN_ON(1);
2707                 return 0;
2708         }
2709 }
2710
2711 static void ohci_write_csr(struct fw_card *card, int csr_offset, u32 value)
2712 {
2713         struct fw_ohci *ohci = fw_ohci(card);
2714         unsigned long flags;
2715
2716         switch (csr_offset) {
2717         case CSR_STATE_CLEAR:
2718                 if ((value & CSR_STATE_BIT_CMSTR) && ohci->is_root) {
2719                         reg_write(ohci, OHCI1394_LinkControlClear,
2720                                   OHCI1394_LinkControl_cycleMaster);
2721                         flush_writes(ohci);
2722                 }
2723                 if (value & CSR_STATE_BIT_ABDICATE)
2724                         ohci->csr_state_setclear_abdicate = false;
2725                 break;
2726
2727         case CSR_STATE_SET:
2728                 if ((value & CSR_STATE_BIT_CMSTR) && ohci->is_root) {
2729                         reg_write(ohci, OHCI1394_LinkControlSet,
2730                                   OHCI1394_LinkControl_cycleMaster);
2731                         flush_writes(ohci);
2732                 }
2733                 if (value & CSR_STATE_BIT_ABDICATE)
2734                         ohci->csr_state_setclear_abdicate = true;
2735                 break;
2736
2737         case CSR_NODE_IDS:
2738                 reg_write(ohci, OHCI1394_NodeID, value >> 16);
2739                 flush_writes(ohci);
2740                 break;
2741
2742         case CSR_CYCLE_TIME:
2743                 reg_write(ohci, OHCI1394_IsochronousCycleTimer, value);
2744                 reg_write(ohci, OHCI1394_IntEventSet,
2745                           OHCI1394_cycleInconsistent);
2746                 flush_writes(ohci);
2747                 break;
2748
2749         case CSR_BUS_TIME:
2750                 spin_lock_irqsave(&ohci->lock, flags);
2751                 ohci->bus_time = (update_bus_time(ohci) & 0x40) |
2752                                  (value & ~0x7f);
2753                 spin_unlock_irqrestore(&ohci->lock, flags);
2754                 break;
2755
2756         case CSR_BUSY_TIMEOUT:
2757                 value = (value & 0xf) | ((value & 0xf) << 4) |
2758                         ((value & 0xf) << 8) | ((value & 0x0ffff000) << 4);
2759                 reg_write(ohci, OHCI1394_ATRetries, value);
2760                 flush_writes(ohci);
2761                 break;
2762
2763         case CSR_PRIORITY_BUDGET:
2764                 reg_write(ohci, OHCI1394_FairnessControl, value & 0x3f);
2765                 flush_writes(ohci);
2766                 break;
2767
2768         default:
2769                 WARN_ON(1);
2770                 break;
2771         }
2772 }
2773
2774 static void flush_iso_completions(struct iso_context *ctx)
2775 {
2776         ctx->base.callback.sc(&ctx->base, ctx->last_timestamp,
2777                               ctx->header_length, ctx->header,
2778                               ctx->base.callback_data);
2779         ctx->header_length = 0;
2780 }
2781
2782 static void copy_iso_headers(struct iso_context *ctx, const u32 *dma_hdr)
2783 {
2784         u32 *ctx_hdr;
2785
2786         if (ctx->header_length + ctx->base.header_size > PAGE_SIZE) {
2787                 if (ctx->base.drop_overflow_headers)
2788                         return;
2789                 flush_iso_completions(ctx);
2790         }
2791
2792         ctx_hdr = ctx->header + ctx->header_length;
2793         ctx->last_timestamp = (u16)le32_to_cpu((__force __le32)dma_hdr[0]);
2794
2795         /*
2796          * The two iso header quadlets are byteswapped to little
2797          * endian by the controller, but we want to present them
2798          * as big endian for consistency with the bus endianness.
2799          */
2800         if (ctx->base.header_size > 0)
2801                 ctx_hdr[0] = swab32(dma_hdr[1]); /* iso packet header */
2802         if (ctx->base.header_size > 4)
2803                 ctx_hdr[1] = swab32(dma_hdr[0]); /* timestamp */
2804         if (ctx->base.header_size > 8)
2805                 memcpy(&ctx_hdr[2], &dma_hdr[2], ctx->base.header_size - 8);
2806         ctx->header_length += ctx->base.header_size;
2807 }
2808
2809 static int handle_ir_packet_per_buffer(struct context *context,
2810                                        struct descriptor *d,
2811                                        struct descriptor *last)
2812 {
2813         struct iso_context *ctx =
2814                 container_of(context, struct iso_context, context);
2815         struct descriptor *pd;
2816         u32 buffer_dma;
2817
2818         for (pd = d; pd <= last; pd++)
2819                 if (pd->transfer_status)
2820                         break;
2821         if (pd > last)
2822                 /* Descriptor(s) not done yet, stop iteration */
2823                 return 0;
2824
2825         while (!(d->control & cpu_to_le16(DESCRIPTOR_BRANCH_ALWAYS))) {
2826                 d++;
2827                 buffer_dma = le32_to_cpu(d->data_address);
2828                 dma_sync_single_range_for_cpu(context->ohci->card.device,
2829                                               buffer_dma & PAGE_MASK,
2830                                               buffer_dma & ~PAGE_MASK,
2831                                               le16_to_cpu(d->req_count),
2832                                               DMA_FROM_DEVICE);
2833         }
2834
2835         copy_iso_headers(ctx, (u32 *) (last + 1));
2836
2837         if (last->control & cpu_to_le16(DESCRIPTOR_IRQ_ALWAYS))
2838                 flush_iso_completions(ctx);
2839
2840         return 1;
2841 }
2842
2843 /* d == last because each descriptor block is only a single descriptor. */
2844 static int handle_ir_buffer_fill(struct context *context,
2845                                  struct descriptor *d,
2846                                  struct descriptor *last)
2847 {
2848         struct iso_context *ctx =
2849                 container_of(context, struct iso_context, context);
2850         unsigned int req_count, res_count, completed;
2851         u32 buffer_dma;
2852
2853         req_count = le16_to_cpu(last->req_count);
2854         res_count = le16_to_cpu(READ_ONCE(last->res_count));
2855         completed = req_count - res_count;
2856         buffer_dma = le32_to_cpu(last->data_address);
2857
2858         if (completed > 0) {
2859                 ctx->mc_buffer_bus = buffer_dma;
2860                 ctx->mc_completed = completed;
2861         }
2862
2863         if (res_count != 0)
2864                 /* Descriptor(s) not done yet, stop iteration */
2865                 return 0;
2866
2867         dma_sync_single_range_for_cpu(context->ohci->card.device,
2868                                       buffer_dma & PAGE_MASK,
2869                                       buffer_dma & ~PAGE_MASK,
2870                                       completed, DMA_FROM_DEVICE);
2871
2872         if (last->control & cpu_to_le16(DESCRIPTOR_IRQ_ALWAYS)) {
2873                 ctx->base.callback.mc(&ctx->base,
2874                                       buffer_dma + completed,
2875                                       ctx->base.callback_data);
2876                 ctx->mc_completed = 0;
2877         }
2878
2879         return 1;
2880 }
2881
2882 static void flush_ir_buffer_fill(struct iso_context *ctx)
2883 {
2884         dma_sync_single_range_for_cpu(ctx->context.ohci->card.device,
2885                                       ctx->mc_buffer_bus & PAGE_MASK,
2886                                       ctx->mc_buffer_bus & ~PAGE_MASK,
2887                                       ctx->mc_completed, DMA_FROM_DEVICE);
2888
2889         ctx->base.callback.mc(&ctx->base,
2890                               ctx->mc_buffer_bus + ctx->mc_completed,
2891                               ctx->base.callback_data);
2892         ctx->mc_completed = 0;
2893 }
2894
2895 static inline void sync_it_packet_for_cpu(struct context *context,
2896                                           struct descriptor *pd)
2897 {
2898         __le16 control;
2899         u32 buffer_dma;
2900
2901         /* only packets beginning with OUTPUT_MORE* have data buffers */
2902         if (pd->control & cpu_to_le16(DESCRIPTOR_BRANCH_ALWAYS))
2903                 return;
2904
2905         /* skip over the OUTPUT_MORE_IMMEDIATE descriptor */
2906         pd += 2;
2907
2908         /*
2909          * If the packet has a header, the first OUTPUT_MORE/LAST descriptor's
2910          * data buffer is in the context program's coherent page and must not
2911          * be synced.
2912          */
2913         if ((le32_to_cpu(pd->data_address) & PAGE_MASK) ==
2914             (context->current_bus          & PAGE_MASK)) {
2915                 if (pd->control & cpu_to_le16(DESCRIPTOR_BRANCH_ALWAYS))
2916                         return;
2917                 pd++;
2918         }
2919
2920         do {
2921                 buffer_dma = le32_to_cpu(pd->data_address);
2922                 dma_sync_single_range_for_cpu(context->ohci->card.device,
2923                                               buffer_dma & PAGE_MASK,
2924                                               buffer_dma & ~PAGE_MASK,
2925                                               le16_to_cpu(pd->req_count),
2926                                               DMA_TO_DEVICE);
2927                 control = pd->control;
2928                 pd++;
2929         } while (!(control & cpu_to_le16(DESCRIPTOR_BRANCH_ALWAYS)));
2930 }
2931
2932 static int handle_it_packet(struct context *context,
2933                             struct descriptor *d,
2934                             struct descriptor *last)
2935 {
2936         struct iso_context *ctx =
2937                 container_of(context, struct iso_context, context);
2938         struct descriptor *pd;
2939         __be32 *ctx_hdr;
2940
2941         for (pd = d; pd <= last; pd++)
2942                 if (pd->transfer_status)
2943                         break;
2944         if (pd > last)
2945                 /* Descriptor(s) not done yet, stop iteration */
2946                 return 0;
2947
2948         sync_it_packet_for_cpu(context, d);
2949
2950         if (ctx->header_length + 4 > PAGE_SIZE) {
2951                 if (ctx->base.drop_overflow_headers)
2952                         return 1;
2953                 flush_iso_completions(ctx);
2954         }
2955
2956         ctx_hdr = ctx->header + ctx->header_length;
2957         ctx->last_timestamp = le16_to_cpu(last->res_count);
2958         /* Present this value as big-endian to match the receive code */
2959         *ctx_hdr = cpu_to_be32((le16_to_cpu(pd->transfer_status) << 16) |
2960                                le16_to_cpu(pd->res_count));
2961         ctx->header_length += 4;
2962
2963         if (last->control & cpu_to_le16(DESCRIPTOR_IRQ_ALWAYS))
2964                 flush_iso_completions(ctx);
2965
2966         return 1;
2967 }
2968
2969 static void set_multichannel_mask(struct fw_ohci *ohci, u64 channels)
2970 {
2971         u32 hi = channels >> 32, lo = channels;
2972
2973         reg_write(ohci, OHCI1394_IRMultiChanMaskHiClear, ~hi);
2974         reg_write(ohci, OHCI1394_IRMultiChanMaskLoClear, ~lo);
2975         reg_write(ohci, OHCI1394_IRMultiChanMaskHiSet, hi);
2976         reg_write(ohci, OHCI1394_IRMultiChanMaskLoSet, lo);
2977         ohci->mc_channels = channels;
2978 }
2979
2980 static struct fw_iso_context *ohci_allocate_iso_context(struct fw_card *card,
2981                                 int type, int channel, size_t header_size)
2982 {
2983         struct fw_ohci *ohci = fw_ohci(card);
2984         struct iso_context *ctx;
2985         descriptor_callback_t callback;
2986         u64 *channels;
2987         u32 *mask, regs;
2988         int index, ret = -EBUSY;
2989
2990         spin_lock_irq(&ohci->lock);
2991
2992         switch (type) {
2993         case FW_ISO_CONTEXT_TRANSMIT:
2994                 mask     = &ohci->it_context_mask;
2995                 callback = handle_it_packet;
2996                 index    = ffs(*mask) - 1;
2997                 if (index >= 0) {
2998                         *mask &= ~(1 << index);
2999                         regs = OHCI1394_IsoXmitContextBase(index);
3000                         ctx  = &ohci->it_context_list[index];
3001                 }
3002                 break;
3003
3004         case FW_ISO_CONTEXT_RECEIVE:
3005                 channels = &ohci->ir_context_channels;
3006                 mask     = &ohci->ir_context_mask;
3007                 callback = handle_ir_packet_per_buffer;
3008                 index    = *channels & 1ULL << channel ? ffs(*mask) - 1 : -1;
3009                 if (index >= 0) {
3010                         *channels &= ~(1ULL << channel);
3011                         *mask     &= ~(1 << index);
3012                         regs = OHCI1394_IsoRcvContextBase(index);
3013                         ctx  = &ohci->ir_context_list[index];
3014                 }
3015                 break;
3016
3017         case FW_ISO_CONTEXT_RECEIVE_MULTICHANNEL:
3018                 mask     = &ohci->ir_context_mask;
3019                 callback = handle_ir_buffer_fill;
3020                 index    = !ohci->mc_allocated ? ffs(*mask) - 1 : -1;
3021                 if (index >= 0) {
3022                         ohci->mc_allocated = true;
3023                         *mask &= ~(1 << index);
3024                         regs = OHCI1394_IsoRcvContextBase(index);
3025                         ctx  = &ohci->ir_context_list[index];
3026                 }
3027                 break;
3028
3029         default:
3030                 index = -1;
3031                 ret = -ENOSYS;
3032         }
3033
3034         spin_unlock_irq(&ohci->lock);
3035
3036         if (index < 0)
3037                 return ERR_PTR(ret);
3038
3039         memset(ctx, 0, sizeof(*ctx));
3040         ctx->header_length = 0;
3041         ctx->header = (void *) __get_free_page(GFP_KERNEL);
3042         if (ctx->header == NULL) {
3043                 ret = -ENOMEM;
3044                 goto out;
3045         }
3046         ret = context_init(&ctx->context, ohci, regs, callback);
3047         if (ret < 0)
3048                 goto out_with_header;
3049
3050         if (type == FW_ISO_CONTEXT_RECEIVE_MULTICHANNEL) {
3051                 set_multichannel_mask(ohci, 0);
3052                 ctx->mc_completed = 0;
3053         }
3054
3055         return &ctx->base;
3056
3057  out_with_header:
3058         free_page((unsigned long)ctx->header);
3059  out:
3060         spin_lock_irq(&ohci->lock);
3061
3062         switch (type) {
3063         case FW_ISO_CONTEXT_RECEIVE:
3064                 *channels |= 1ULL << channel;
3065                 break;
3066
3067         case FW_ISO_CONTEXT_RECEIVE_MULTICHANNEL:
3068                 ohci->mc_allocated = false;
3069                 break;
3070         }
3071         *mask |= 1 << index;
3072
3073         spin_unlock_irq(&ohci->lock);
3074
3075         return ERR_PTR(ret);
3076 }
3077
3078 static int ohci_start_iso(struct fw_iso_context *base,
3079                           s32 cycle, u32 sync, u32 tags)
3080 {
3081         struct iso_context *ctx = container_of(base, struct iso_context, base);
3082         struct fw_ohci *ohci = ctx->context.ohci;
3083         u32 control = IR_CONTEXT_ISOCH_HEADER, match;
3084         int index;
3085
3086         /* the controller cannot start without any queued packets */
3087         if (ctx->context.last->branch_address == 0)
3088                 return -ENODATA;
3089
3090         switch (ctx->base.type) {
3091         case FW_ISO_CONTEXT_TRANSMIT:
3092                 index = ctx - ohci->it_context_list;
3093                 match = 0;
3094                 if (cycle >= 0)
3095                         match = IT_CONTEXT_CYCLE_MATCH_ENABLE |
3096                                 (cycle & 0x7fff) << 16;
3097
3098                 reg_write(ohci, OHCI1394_IsoXmitIntEventClear, 1 << index);
3099                 reg_write(ohci, OHCI1394_IsoXmitIntMaskSet, 1 << index);
3100                 context_run(&ctx->context, match);
3101                 break;
3102
3103         case FW_ISO_CONTEXT_RECEIVE_MULTICHANNEL:
3104                 control |= IR_CONTEXT_BUFFER_FILL|IR_CONTEXT_MULTI_CHANNEL_MODE;
3105                 fallthrough;
3106         case FW_ISO_CONTEXT_RECEIVE:
3107                 index = ctx - ohci->ir_context_list;
3108                 match = (tags << 28) | (sync << 8) | ctx->base.channel;
3109                 if (cycle >= 0) {
3110                         match |= (cycle & 0x07fff) << 12;
3111                         control |= IR_CONTEXT_CYCLE_MATCH_ENABLE;
3112                 }
3113
3114                 reg_write(ohci, OHCI1394_IsoRecvIntEventClear, 1 << index);
3115                 reg_write(ohci, OHCI1394_IsoRecvIntMaskSet, 1 << index);
3116                 reg_write(ohci, CONTEXT_MATCH(ctx->context.regs), match);
3117                 context_run(&ctx->context, control);
3118
3119                 ctx->sync = sync;
3120                 ctx->tags = tags;
3121
3122                 break;
3123         }
3124
3125         return 0;
3126 }
3127
3128 static int ohci_stop_iso(struct fw_iso_context *base)
3129 {
3130         struct fw_ohci *ohci = fw_ohci(base->card);
3131         struct iso_context *ctx = container_of(base, struct iso_context, base);
3132         int index;
3133
3134         switch (ctx->base.type) {
3135         case FW_ISO_CONTEXT_TRANSMIT:
3136                 index = ctx - ohci->it_context_list;
3137                 reg_write(ohci, OHCI1394_IsoXmitIntMaskClear, 1 << index);
3138                 break;
3139
3140         case FW_ISO_CONTEXT_RECEIVE:
3141         case FW_ISO_CONTEXT_RECEIVE_MULTICHANNEL:
3142                 index = ctx - ohci->ir_context_list;
3143                 reg_write(ohci, OHCI1394_IsoRecvIntMaskClear, 1 << index);
3144                 break;
3145         }
3146         flush_writes(ohci);
3147         context_stop(&ctx->context);
3148         tasklet_kill(&ctx->context.tasklet);
3149
3150         return 0;
3151 }
3152
3153 static void ohci_free_iso_context(struct fw_iso_context *base)
3154 {
3155         struct fw_ohci *ohci = fw_ohci(base->card);
3156         struct iso_context *ctx = container_of(base, struct iso_context, base);
3157         unsigned long flags;
3158         int index;
3159
3160         ohci_stop_iso(base);
3161         context_release(&ctx->context);
3162         free_page((unsigned long)ctx->header);
3163
3164         spin_lock_irqsave(&ohci->lock, flags);
3165
3166         switch (base->type) {
3167         case FW_ISO_CONTEXT_TRANSMIT:
3168                 index = ctx - ohci->it_context_list;
3169                 ohci->it_context_mask |= 1 << index;
3170                 break;
3171
3172         case FW_ISO_CONTEXT_RECEIVE:
3173                 index = ctx - ohci->ir_context_list;
3174                 ohci->ir_context_mask |= 1 << index;
3175                 ohci->ir_context_channels |= 1ULL << base->channel;
3176                 break;
3177
3178         case FW_ISO_CONTEXT_RECEIVE_MULTICHANNEL:
3179                 index = ctx - ohci->ir_context_list;
3180                 ohci->ir_context_mask |= 1 << index;
3181                 ohci->ir_context_channels |= ohci->mc_channels;
3182                 ohci->mc_channels = 0;
3183                 ohci->mc_allocated = false;
3184                 break;
3185         }
3186
3187         spin_unlock_irqrestore(&ohci->lock, flags);
3188 }
3189
3190 static int ohci_set_iso_channels(struct fw_iso_context *base, u64 *channels)
3191 {
3192         struct fw_ohci *ohci = fw_ohci(base->card);
3193         unsigned long flags;
3194         int ret;
3195
3196         switch (base->type) {
3197         case FW_ISO_CONTEXT_RECEIVE_MULTICHANNEL:
3198
3199                 spin_lock_irqsave(&ohci->lock, flags);
3200
3201                 /* Don't allow multichannel to grab other contexts' channels. */
3202                 if (~ohci->ir_context_channels & ~ohci->mc_channels & *channels) {
3203                         *channels = ohci->ir_context_channels;
3204                         ret = -EBUSY;
3205                 } else {
3206                         set_multichannel_mask(ohci, *channels);
3207                         ret = 0;
3208                 }
3209
3210                 spin_unlock_irqrestore(&ohci->lock, flags);
3211
3212                 break;
3213         default:
3214                 ret = -EINVAL;
3215         }
3216
3217         return ret;
3218 }
3219
3220 #ifdef CONFIG_PM
3221 static void ohci_resume_iso_dma(struct fw_ohci *ohci)
3222 {
3223         int i;
3224         struct iso_context *ctx;
3225
3226         for (i = 0 ; i < ohci->n_ir ; i++) {
3227                 ctx = &ohci->ir_context_list[i];
3228                 if (ctx->context.running)
3229                         ohci_start_iso(&ctx->base, 0, ctx->sync, ctx->tags);
3230         }
3231
3232         for (i = 0 ; i < ohci->n_it ; i++) {
3233                 ctx = &ohci->it_context_list[i];
3234                 if (ctx->context.running)
3235                         ohci_start_iso(&ctx->base, 0, ctx->sync, ctx->tags);
3236         }
3237 }
3238 #endif
3239
3240 static int queue_iso_transmit(struct iso_context *ctx,
3241                               struct fw_iso_packet *packet,
3242                               struct fw_iso_buffer *buffer,
3243                               unsigned long payload)
3244 {
3245         struct descriptor *d, *last, *pd;
3246         struct fw_iso_packet *p;
3247         __le32 *header;
3248         dma_addr_t d_bus, page_bus;
3249         u32 z, header_z, payload_z, irq;
3250         u32 payload_index, payload_end_index, next_page_index;
3251         int page, end_page, i, length, offset;
3252
3253         p = packet;
3254         payload_index = payload;
3255
3256         if (p->skip)
3257                 z = 1;
3258         else
3259                 z = 2;
3260         if (p->header_length > 0)
3261                 z++;
3262
3263         /* Determine the first page the payload isn't contained in. */
3264         end_page = PAGE_ALIGN(payload_index + p->payload_length) >> PAGE_SHIFT;
3265         if (p->payload_length > 0)
3266                 payload_z = end_page - (payload_index >> PAGE_SHIFT);
3267         else
3268                 payload_z = 0;
3269
3270         z += payload_z;
3271
3272         /* Get header size in number of descriptors. */
3273         header_z = DIV_ROUND_UP(p->header_length, sizeof(*d));
3274
3275         d = context_get_descriptors(&ctx->context, z + header_z, &d_bus);
3276         if (d == NULL)
3277                 return -ENOMEM;
3278
3279         if (!p->skip) {
3280                 d[0].control   = cpu_to_le16(DESCRIPTOR_KEY_IMMEDIATE);
3281                 d[0].req_count = cpu_to_le16(8);
3282                 /*
3283                  * Link the skip address to this descriptor itself.  This causes
3284                  * a context to skip a cycle whenever lost cycles or FIFO
3285                  * overruns occur, without dropping the data.  The application
3286                  * should then decide whether this is an error condition or not.
3287                  * FIXME:  Make the context's cycle-lost behaviour configurable?
3288                  */
3289                 d[0].branch_address = cpu_to_le32(d_bus | z);
3290
3291                 header = (__le32 *) &d[1];
3292                 header[0] = cpu_to_le32(IT_HEADER_SY(p->sy) |
3293                                         IT_HEADER_TAG(p->tag) |
3294                                         IT_HEADER_TCODE(TCODE_STREAM_DATA) |
3295                                         IT_HEADER_CHANNEL(ctx->base.channel) |
3296                                         IT_HEADER_SPEED(ctx->base.speed));
3297                 header[1] =
3298                         cpu_to_le32(IT_HEADER_DATA_LENGTH(p->header_length +
3299                                                           p->payload_length));
3300         }
3301
3302         if (p->header_length > 0) {
3303                 d[2].req_count    = cpu_to_le16(p->header_length);
3304                 d[2].data_address = cpu_to_le32(d_bus + z * sizeof(*d));
3305                 memcpy(&d[z], p->header, p->header_length);
3306         }
3307
3308         pd = d + z - payload_z;
3309         payload_end_index = payload_index + p->payload_length;
3310         for (i = 0; i < payload_z; i++) {
3311                 page               = payload_index >> PAGE_SHIFT;
3312                 offset             = payload_index & ~PAGE_MASK;
3313                 next_page_index    = (page + 1) << PAGE_SHIFT;
3314                 length             =
3315                         min(next_page_index, payload_end_index) - payload_index;
3316                 pd[i].req_count    = cpu_to_le16(length);
3317
3318                 page_bus = page_private(buffer->pages[page]);
3319                 pd[i].data_address = cpu_to_le32(page_bus + offset);
3320
3321                 dma_sync_single_range_for_device(ctx->context.ohci->card.device,
3322                                                  page_bus, offset, length,
3323                                                  DMA_TO_DEVICE);
3324
3325                 payload_index += length;
3326         }
3327
3328         if (p->interrupt)
3329                 irq = DESCRIPTOR_IRQ_ALWAYS;
3330         else
3331                 irq = DESCRIPTOR_NO_IRQ;
3332
3333         last = z == 2 ? d : d + z - 1;
3334         last->control |= cpu_to_le16(DESCRIPTOR_OUTPUT_LAST |
3335                                      DESCRIPTOR_STATUS |
3336                                      DESCRIPTOR_BRANCH_ALWAYS |
3337                                      irq);
3338
3339         context_append(&ctx->context, d, z, header_z);
3340
3341         return 0;
3342 }
3343
3344 static int queue_iso_packet_per_buffer(struct iso_context *ctx,
3345                                        struct fw_iso_packet *packet,
3346                                        struct fw_iso_buffer *buffer,
3347                                        unsigned long payload)
3348 {
3349         struct device *device = ctx->context.ohci->card.device;
3350         struct descriptor *d, *pd;
3351         dma_addr_t d_bus, page_bus;
3352         u32 z, header_z, rest;
3353         int i, j, length;
3354         int page, offset, packet_count, header_size, payload_per_buffer;
3355
3356         /*
3357          * The OHCI controller puts the isochronous header and trailer in the
3358          * buffer, so we need at least 8 bytes.
3359          */
3360         packet_count = packet->header_length / ctx->base.header_size;
3361         header_size  = max(ctx->base.header_size, (size_t)8);
3362
3363         /* Get header size in number of descriptors. */
3364         header_z = DIV_ROUND_UP(header_size, sizeof(*d));
3365         page     = payload >> PAGE_SHIFT;
3366         offset   = payload & ~PAGE_MASK;
3367         payload_per_buffer = packet->payload_length / packet_count;
3368
3369         for (i = 0; i < packet_count; i++) {
3370                 /* d points to the header descriptor */
3371                 z = DIV_ROUND_UP(payload_per_buffer + offset, PAGE_SIZE) + 1;
3372                 d = context_get_descriptors(&ctx->context,
3373                                 z + header_z, &d_bus);
3374                 if (d == NULL)
3375                         return -ENOMEM;
3376
3377                 d->control      = cpu_to_le16(DESCRIPTOR_STATUS |
3378                                               DESCRIPTOR_INPUT_MORE);
3379                 if (packet->skip && i == 0)
3380                         d->control |= cpu_to_le16(DESCRIPTOR_WAIT);
3381                 d->req_count    = cpu_to_le16(header_size);
3382                 d->res_count    = d->req_count;
3383                 d->transfer_status = 0;
3384                 d->data_address = cpu_to_le32(d_bus + (z * sizeof(*d)));
3385
3386                 rest = payload_per_buffer;
3387                 pd = d;
3388                 for (j = 1; j < z; j++) {
3389                         pd++;
3390                         pd->control = cpu_to_le16(DESCRIPTOR_STATUS |
3391                                                   DESCRIPTOR_INPUT_MORE);
3392
3393                         if (offset + rest < PAGE_SIZE)
3394                                 length = rest;
3395                         else
3396                                 length = PAGE_SIZE - offset;
3397                         pd->req_count = cpu_to_le16(length);
3398                         pd->res_count = pd->req_count;
3399                         pd->transfer_status = 0;
3400
3401                         page_bus = page_private(buffer->pages[page]);
3402                         pd->data_address = cpu_to_le32(page_bus + offset);
3403
3404                         dma_sync_single_range_for_device(device, page_bus,
3405                                                          offset, length,
3406                                                          DMA_FROM_DEVICE);
3407
3408                         offset = (offset + length) & ~PAGE_MASK;
3409                         rest -= length;
3410                         if (offset == 0)
3411                                 page++;
3412                 }
3413                 pd->control = cpu_to_le16(DESCRIPTOR_STATUS |
3414                                           DESCRIPTOR_INPUT_LAST |
3415                                           DESCRIPTOR_BRANCH_ALWAYS);
3416                 if (packet->interrupt && i == packet_count - 1)
3417                         pd->control |= cpu_to_le16(DESCRIPTOR_IRQ_ALWAYS);
3418
3419                 context_append(&ctx->context, d, z, header_z);
3420         }
3421
3422         return 0;
3423 }
3424
3425 static int queue_iso_buffer_fill(struct iso_context *ctx,
3426                                  struct fw_iso_packet *packet,
3427                                  struct fw_iso_buffer *buffer,
3428                                  unsigned long payload)
3429 {
3430         struct descriptor *d;
3431         dma_addr_t d_bus, page_bus;
3432         int page, offset, rest, z, i, length;
3433
3434         page   = payload >> PAGE_SHIFT;
3435         offset = payload & ~PAGE_MASK;
3436         rest   = packet->payload_length;
3437
3438         /* We need one descriptor for each page in the buffer. */
3439         z = DIV_ROUND_UP(offset + rest, PAGE_SIZE);
3440
3441         if (WARN_ON(offset & 3 || rest & 3 || page + z > buffer->page_count))
3442                 return -EFAULT;
3443
3444         for (i = 0; i < z; i++) {
3445                 d = context_get_descriptors(&ctx->context, 1, &d_bus);
3446                 if (d == NULL)
3447                         return -ENOMEM;
3448
3449                 d->control = cpu_to_le16(DESCRIPTOR_INPUT_MORE |
3450                                          DESCRIPTOR_BRANCH_ALWAYS);
3451                 if (packet->skip && i == 0)
3452                         d->control |= cpu_to_le16(DESCRIPTOR_WAIT);
3453                 if (packet->interrupt && i == z - 1)
3454                         d->control |= cpu_to_le16(DESCRIPTOR_IRQ_ALWAYS);
3455
3456                 if (offset + rest < PAGE_SIZE)
3457                         length = rest;
3458                 else
3459                         length = PAGE_SIZE - offset;
3460                 d->req_count = cpu_to_le16(length);
3461                 d->res_count = d->req_count;
3462                 d->transfer_status = 0;
3463
3464                 page_bus = page_private(buffer->pages[page]);
3465                 d->data_address = cpu_to_le32(page_bus + offset);
3466
3467                 dma_sync_single_range_for_device(ctx->context.ohci->card.device,
3468                                                  page_bus, offset, length,
3469                                                  DMA_FROM_DEVICE);
3470
3471                 rest -= length;
3472                 offset = 0;
3473                 page++;
3474
3475                 context_append(&ctx->context, d, 1, 0);
3476         }
3477
3478         return 0;
3479 }
3480
3481 static int ohci_queue_iso(struct fw_iso_context *base,
3482                           struct fw_iso_packet *packet,
3483                           struct fw_iso_buffer *buffer,
3484                           unsigned long payload)
3485 {
3486         struct iso_context *ctx = container_of(base, struct iso_context, base);
3487         unsigned long flags;
3488         int ret = -ENOSYS;
3489
3490         spin_lock_irqsave(&ctx->context.ohci->lock, flags);
3491         switch (base->type) {
3492         case FW_ISO_CONTEXT_TRANSMIT:
3493                 ret = queue_iso_transmit(ctx, packet, buffer, payload);
3494                 break;
3495         case FW_ISO_CONTEXT_RECEIVE:
3496                 ret = queue_iso_packet_per_buffer(ctx, packet, buffer, payload);
3497                 break;
3498         case FW_ISO_CONTEXT_RECEIVE_MULTICHANNEL:
3499                 ret = queue_iso_buffer_fill(ctx, packet, buffer, payload);
3500                 break;
3501         }
3502         spin_unlock_irqrestore(&ctx->context.ohci->lock, flags);
3503
3504         return ret;
3505 }
3506
3507 static void ohci_flush_queue_iso(struct fw_iso_context *base)
3508 {
3509         struct context *ctx =
3510                         &container_of(base, struct iso_context, base)->context;
3511
3512         reg_write(ctx->ohci, CONTROL_SET(ctx->regs), CONTEXT_WAKE);
3513 }
3514
3515 static int ohci_flush_iso_completions(struct fw_iso_context *base)
3516 {
3517         struct iso_context *ctx = container_of(base, struct iso_context, base);
3518         int ret = 0;
3519
3520         tasklet_disable(&ctx->context.tasklet);
3521
3522         if (!test_and_set_bit_lock(0, &ctx->flushing_completions)) {
3523                 context_tasklet((unsigned long)&ctx->context);
3524
3525                 switch (base->type) {
3526                 case FW_ISO_CONTEXT_TRANSMIT:
3527                 case FW_ISO_CONTEXT_RECEIVE:
3528                         if (ctx->header_length != 0)
3529                                 flush_iso_completions(ctx);
3530                         break;
3531                 case FW_ISO_CONTEXT_RECEIVE_MULTICHANNEL:
3532                         if (ctx->mc_completed != 0)
3533                                 flush_ir_buffer_fill(ctx);
3534                         break;
3535                 default:
3536                         ret = -ENOSYS;
3537                 }
3538
3539                 clear_bit_unlock(0, &ctx->flushing_completions);
3540                 smp_mb__after_atomic();
3541         }
3542
3543         tasklet_enable(&ctx->context.tasklet);
3544
3545         return ret;
3546 }
3547
3548 static const struct fw_card_driver ohci_driver = {
3549         .enable                 = ohci_enable,
3550         .read_phy_reg           = ohci_read_phy_reg,
3551         .update_phy_reg         = ohci_update_phy_reg,
3552         .set_config_rom         = ohci_set_config_rom,
3553         .send_request           = ohci_send_request,
3554         .send_response          = ohci_send_response,
3555         .cancel_packet          = ohci_cancel_packet,
3556         .enable_phys_dma        = ohci_enable_phys_dma,
3557         .read_csr               = ohci_read_csr,
3558         .write_csr              = ohci_write_csr,
3559
3560         .allocate_iso_context   = ohci_allocate_iso_context,
3561         .free_iso_context       = ohci_free_iso_context,
3562         .set_iso_channels       = ohci_set_iso_channels,
3563         .queue_iso              = ohci_queue_iso,
3564         .flush_queue_iso        = ohci_flush_queue_iso,
3565         .flush_iso_completions  = ohci_flush_iso_completions,
3566         .start_iso              = ohci_start_iso,
3567         .stop_iso               = ohci_stop_iso,
3568 };
3569
3570 #ifdef CONFIG_PPC_PMAC
3571 static void pmac_ohci_on(struct pci_dev *dev)
3572 {
3573         if (machine_is(powermac)) {
3574                 struct device_node *ofn = pci_device_to_OF_node(dev);
3575
3576                 if (ofn) {
3577                         pmac_call_feature(PMAC_FTR_1394_CABLE_POWER, ofn, 0, 1);
3578                         pmac_call_feature(PMAC_FTR_1394_ENABLE, ofn, 0, 1);
3579                 }
3580         }
3581 }
3582
3583 static void pmac_ohci_off(struct pci_dev *dev)
3584 {
3585         if (machine_is(powermac)) {
3586                 struct device_node *ofn = pci_device_to_OF_node(dev);
3587
3588                 if (ofn) {
3589                         pmac_call_feature(PMAC_FTR_1394_ENABLE, ofn, 0, 0);
3590                         pmac_call_feature(PMAC_FTR_1394_CABLE_POWER, ofn, 0, 0);
3591                 }
3592         }
3593 }
3594 #else
3595 static inline void pmac_ohci_on(struct pci_dev *dev) {}
3596 static inline void pmac_ohci_off(struct pci_dev *dev) {}
3597 #endif /* CONFIG_PPC_PMAC */
3598
3599 static int pci_probe(struct pci_dev *dev,
3600                                const struct pci_device_id *ent)
3601 {
3602         struct fw_ohci *ohci;
3603         u32 bus_options, max_receive, link_speed, version;
3604         u64 guid;
3605         int i, err;
3606         size_t size;
3607
3608         if (dev->vendor == PCI_VENDOR_ID_PINNACLE_SYSTEMS) {
3609                 dev_err(&dev->dev, "Pinnacle MovieBoard is not yet supported\n");
3610                 return -ENOSYS;
3611         }
3612
3613         ohci = kzalloc(sizeof(*ohci), GFP_KERNEL);
3614         if (ohci == NULL) {
3615                 err = -ENOMEM;
3616                 goto fail;
3617         }
3618
3619         fw_card_initialize(&ohci->card, &ohci_driver, &dev->dev);
3620
3621         pmac_ohci_on(dev);
3622
3623         err = pci_enable_device(dev);
3624         if (err) {
3625                 dev_err(&dev->dev, "failed to enable OHCI hardware\n");
3626                 goto fail_free;
3627         }
3628
3629         pci_set_master(dev);
3630         pci_write_config_dword(dev, OHCI1394_PCI_HCI_Control, 0);
3631         pci_set_drvdata(dev, ohci);
3632
3633         spin_lock_init(&ohci->lock);
3634         mutex_init(&ohci->phy_reg_mutex);
3635
3636         INIT_WORK(&ohci->bus_reset_work, bus_reset_work);
3637
3638         if (!(pci_resource_flags(dev, 0) & IORESOURCE_MEM) ||
3639             pci_resource_len(dev, 0) < OHCI1394_REGISTER_SIZE) {
3640                 ohci_err(ohci, "invalid MMIO resource\n");
3641                 err = -ENXIO;
3642                 goto fail_disable;
3643         }
3644
3645         err = pci_request_region(dev, 0, ohci_driver_name);
3646         if (err) {
3647                 ohci_err(ohci, "MMIO resource unavailable\n");
3648                 goto fail_disable;
3649         }
3650
3651         ohci->registers = pci_iomap(dev, 0, OHCI1394_REGISTER_SIZE);
3652         if (ohci->registers == NULL) {
3653                 ohci_err(ohci, "failed to remap registers\n");
3654                 err = -ENXIO;
3655                 goto fail_iomem;
3656         }
3657
3658         for (i = 0; i < ARRAY_SIZE(ohci_quirks); i++)
3659                 if ((ohci_quirks[i].vendor == dev->vendor) &&
3660                     (ohci_quirks[i].device == (unsigned short)PCI_ANY_ID ||
3661                      ohci_quirks[i].device == dev->device) &&
3662                     (ohci_quirks[i].revision == (unsigned short)PCI_ANY_ID ||
3663                      ohci_quirks[i].revision >= dev->revision)) {
3664                         ohci->quirks = ohci_quirks[i].flags;
3665                         break;
3666                 }
3667         if (param_quirks)
3668                 ohci->quirks = param_quirks;
3669
3670         if (detect_vt630x_with_asm1083_on_amd_ryzen_machine(dev))
3671                 ohci->quirks |= QUIRK_REBOOT_BY_CYCLE_TIMER_READ;
3672
3673         /*
3674          * Because dma_alloc_coherent() allocates at least one page,
3675          * we save space by using a common buffer for the AR request/
3676          * response descriptors and the self IDs buffer.
3677          */
3678         BUILD_BUG_ON(AR_BUFFERS * sizeof(struct descriptor) > PAGE_SIZE/4);
3679         BUILD_BUG_ON(SELF_ID_BUF_SIZE > PAGE_SIZE/2);
3680         ohci->misc_buffer = dma_alloc_coherent(ohci->card.device,
3681                                                PAGE_SIZE,
3682                                                &ohci->misc_buffer_bus,
3683                                                GFP_KERNEL);
3684         if (!ohci->misc_buffer) {
3685                 err = -ENOMEM;
3686                 goto fail_iounmap;
3687         }
3688
3689         err = ar_context_init(&ohci->ar_request_ctx, ohci, 0,
3690                               OHCI1394_AsReqRcvContextControlSet);
3691         if (err < 0)
3692                 goto fail_misc_buf;
3693
3694         err = ar_context_init(&ohci->ar_response_ctx, ohci, PAGE_SIZE/4,
3695                               OHCI1394_AsRspRcvContextControlSet);
3696         if (err < 0)
3697                 goto fail_arreq_ctx;
3698
3699         err = context_init(&ohci->at_request_ctx, ohci,
3700                            OHCI1394_AsReqTrContextControlSet, handle_at_packet);
3701         if (err < 0)
3702                 goto fail_arrsp_ctx;
3703
3704         err = context_init(&ohci->at_response_ctx, ohci,
3705                            OHCI1394_AsRspTrContextControlSet, handle_at_packet);
3706         if (err < 0)
3707                 goto fail_atreq_ctx;
3708
3709         reg_write(ohci, OHCI1394_IsoRecvIntMaskSet, ~0);
3710         ohci->ir_context_channels = ~0ULL;
3711         ohci->ir_context_support = reg_read(ohci, OHCI1394_IsoRecvIntMaskSet);
3712         reg_write(ohci, OHCI1394_IsoRecvIntMaskClear, ~0);
3713         ohci->ir_context_mask = ohci->ir_context_support;
3714         ohci->n_ir = hweight32(ohci->ir_context_mask);
3715         size = sizeof(struct iso_context) * ohci->n_ir;
3716         ohci->ir_context_list = kzalloc(size, GFP_KERNEL);
3717
3718         reg_write(ohci, OHCI1394_IsoXmitIntMaskSet, ~0);
3719         ohci->it_context_support = reg_read(ohci, OHCI1394_IsoXmitIntMaskSet);
3720         /* JMicron JMB38x often shows 0 at first read, just ignore it */
3721         if (!ohci->it_context_support) {
3722                 ohci_notice(ohci, "overriding IsoXmitIntMask\n");
3723                 ohci->it_context_support = 0xf;
3724         }
3725         reg_write(ohci, OHCI1394_IsoXmitIntMaskClear, ~0);
3726         ohci->it_context_mask = ohci->it_context_support;
3727         ohci->n_it = hweight32(ohci->it_context_mask);
3728         size = sizeof(struct iso_context) * ohci->n_it;
3729         ohci->it_context_list = kzalloc(size, GFP_KERNEL);
3730
3731         if (ohci->it_context_list == NULL || ohci->ir_context_list == NULL) {
3732                 err = -ENOMEM;
3733                 goto fail_contexts;
3734         }
3735
3736         ohci->self_id     = ohci->misc_buffer     + PAGE_SIZE/2;
3737         ohci->self_id_bus = ohci->misc_buffer_bus + PAGE_SIZE/2;
3738
3739         bus_options = reg_read(ohci, OHCI1394_BusOptions);
3740         max_receive = (bus_options >> 12) & 0xf;
3741         link_speed = bus_options & 0x7;
3742         guid = ((u64) reg_read(ohci, OHCI1394_GUIDHi) << 32) |
3743                 reg_read(ohci, OHCI1394_GUIDLo);
3744
3745         if (!(ohci->quirks & QUIRK_NO_MSI))
3746                 pci_enable_msi(dev);
3747         if (request_irq(dev->irq, irq_handler,
3748                         pci_dev_msi_enabled(dev) ? 0 : IRQF_SHARED,
3749                         ohci_driver_name, ohci)) {
3750                 ohci_err(ohci, "failed to allocate interrupt %d\n", dev->irq);
3751                 err = -EIO;
3752                 goto fail_msi;
3753         }
3754
3755         err = fw_card_add(&ohci->card, max_receive, link_speed, guid);
3756         if (err)
3757                 goto fail_irq;
3758
3759         version = reg_read(ohci, OHCI1394_Version) & 0x00ff00ff;
3760         ohci_notice(ohci,
3761                     "added OHCI v%x.%x device as card %d, "
3762                     "%d IR + %d IT contexts, quirks 0x%x%s\n",
3763                     version >> 16, version & 0xff, ohci->card.index,
3764                     ohci->n_ir, ohci->n_it, ohci->quirks,
3765                     reg_read(ohci, OHCI1394_PhyUpperBound) ?
3766                         ", physUB" : "");
3767
3768         return 0;
3769
3770  fail_irq:
3771         free_irq(dev->irq, ohci);
3772  fail_msi:
3773         pci_disable_msi(dev);
3774  fail_contexts:
3775         kfree(ohci->ir_context_list);
3776         kfree(ohci->it_context_list);
3777         context_release(&ohci->at_response_ctx);
3778  fail_atreq_ctx:
3779         context_release(&ohci->at_request_ctx);
3780  fail_arrsp_ctx:
3781         ar_context_release(&ohci->ar_response_ctx);
3782  fail_arreq_ctx:
3783         ar_context_release(&ohci->ar_request_ctx);
3784  fail_misc_buf:
3785         dma_free_coherent(ohci->card.device, PAGE_SIZE,
3786                           ohci->misc_buffer, ohci->misc_buffer_bus);
3787  fail_iounmap:
3788         pci_iounmap(dev, ohci->registers);
3789  fail_iomem:
3790         pci_release_region(dev, 0);
3791  fail_disable:
3792         pci_disable_device(dev);
3793  fail_free:
3794         kfree(ohci);
3795         pmac_ohci_off(dev);
3796  fail:
3797         return err;
3798 }
3799
3800 static void pci_remove(struct pci_dev *dev)
3801 {
3802         struct fw_ohci *ohci = pci_get_drvdata(dev);
3803
3804         /*
3805          * If the removal is happening from the suspend state, LPS won't be
3806          * enabled and host registers (eg., IntMaskClear) won't be accessible.
3807          */
3808         if (reg_read(ohci, OHCI1394_HCControlSet) & OHCI1394_HCControl_LPS) {
3809                 reg_write(ohci, OHCI1394_IntMaskClear, ~0);
3810                 flush_writes(ohci);
3811         }
3812         cancel_work_sync(&ohci->bus_reset_work);
3813         fw_core_remove_card(&ohci->card);
3814
3815         /*
3816          * FIXME: Fail all pending packets here, now that the upper
3817          * layers can't queue any more.
3818          */
3819
3820         software_reset(ohci);
3821         free_irq(dev->irq, ohci);
3822
3823         if (ohci->next_config_rom && ohci->next_config_rom != ohci->config_rom)
3824                 dma_free_coherent(ohci->card.device, CONFIG_ROM_SIZE,
3825                                   ohci->next_config_rom, ohci->next_config_rom_bus);
3826         if (ohci->config_rom)
3827                 dma_free_coherent(ohci->card.device, CONFIG_ROM_SIZE,
3828                                   ohci->config_rom, ohci->config_rom_bus);
3829         ar_context_release(&ohci->ar_request_ctx);
3830         ar_context_release(&ohci->ar_response_ctx);
3831         dma_free_coherent(ohci->card.device, PAGE_SIZE,
3832                           ohci->misc_buffer, ohci->misc_buffer_bus);
3833         context_release(&ohci->at_request_ctx);
3834         context_release(&ohci->at_response_ctx);
3835         kfree(ohci->it_context_list);
3836         kfree(ohci->ir_context_list);
3837         pci_disable_msi(dev);
3838         pci_iounmap(dev, ohci->registers);
3839         pci_release_region(dev, 0);
3840         pci_disable_device(dev);
3841         kfree(ohci);
3842         pmac_ohci_off(dev);
3843
3844         dev_notice(&dev->dev, "removed fw-ohci device\n");
3845 }
3846
3847 #ifdef CONFIG_PM
3848 static int pci_suspend(struct pci_dev *dev, pm_message_t state)
3849 {
3850         struct fw_ohci *ohci = pci_get_drvdata(dev);
3851         int err;
3852
3853         software_reset(ohci);
3854         err = pci_save_state(dev);
3855         if (err) {
3856                 ohci_err(ohci, "pci_save_state failed\n");
3857                 return err;
3858         }
3859         err = pci_set_power_state(dev, pci_choose_state(dev, state));
3860         if (err)
3861                 ohci_err(ohci, "pci_set_power_state failed with %d\n", err);
3862         pmac_ohci_off(dev);
3863
3864         return 0;
3865 }
3866
3867 static int pci_resume(struct pci_dev *dev)
3868 {
3869         struct fw_ohci *ohci = pci_get_drvdata(dev);
3870         int err;
3871
3872         pmac_ohci_on(dev);
3873         pci_set_power_state(dev, PCI_D0);
3874         pci_restore_state(dev);
3875         err = pci_enable_device(dev);
3876         if (err) {
3877                 ohci_err(ohci, "pci_enable_device failed\n");
3878                 return err;
3879         }
3880
3881         /* Some systems don't setup GUID register on resume from ram  */
3882         if (!reg_read(ohci, OHCI1394_GUIDLo) &&
3883                                         !reg_read(ohci, OHCI1394_GUIDHi)) {
3884                 reg_write(ohci, OHCI1394_GUIDLo, (u32)ohci->card.guid);
3885                 reg_write(ohci, OHCI1394_GUIDHi, (u32)(ohci->card.guid >> 32));
3886         }
3887
3888         err = ohci_enable(&ohci->card, NULL, 0);
3889         if (err)
3890                 return err;
3891
3892         ohci_resume_iso_dma(ohci);
3893
3894         return 0;
3895 }
3896 #endif
3897
3898 static const struct pci_device_id pci_table[] = {
3899         { PCI_DEVICE_CLASS(PCI_CLASS_SERIAL_FIREWIRE_OHCI, ~0) },
3900         { }
3901 };
3902
3903 MODULE_DEVICE_TABLE(pci, pci_table);
3904
3905 static struct pci_driver fw_ohci_pci_driver = {
3906         .name           = ohci_driver_name,
3907         .id_table       = pci_table,
3908         .probe          = pci_probe,
3909         .remove         = pci_remove,
3910 #ifdef CONFIG_PM
3911         .resume         = pci_resume,
3912         .suspend        = pci_suspend,
3913 #endif
3914 };
3915
3916 static int __init fw_ohci_init(void)
3917 {
3918         selfid_workqueue = alloc_workqueue(KBUILD_MODNAME, WQ_MEM_RECLAIM, 0);
3919         if (!selfid_workqueue)
3920                 return -ENOMEM;
3921
3922         return pci_register_driver(&fw_ohci_pci_driver);
3923 }
3924
3925 static void __exit fw_ohci_cleanup(void)
3926 {
3927         pci_unregister_driver(&fw_ohci_pci_driver);
3928         destroy_workqueue(selfid_workqueue);
3929 }
3930
3931 module_init(fw_ohci_init);
3932 module_exit(fw_ohci_cleanup);
3933
3934 MODULE_AUTHOR("Kristian Hoegsberg <krh@bitplanet.net>");
3935 MODULE_DESCRIPTION("Driver for PCI OHCI IEEE1394 controllers");
3936 MODULE_LICENSE("GPL");
3937
3938 /* Provide a module alias so root-on-sbp2 initrds don't break. */
3939 MODULE_ALIAS("ohci1394");