drivers/staging/vc04_services/interface/vchiq_arm/vchiq_2835_arm.c

   1 /**
   2  * Copyright (c) 2010-2012 Broadcom. All rights reserved.
   3  *
   4  * Redistribution and use in source and binary forms, with or without
   5  * modification, are permitted provided that the following conditions
   6  * are met:
   7  * 1. Redistributions of source code must retain the above copyright
   8  *    notice, this list of conditions, and the following disclaimer,
   9  *    without modification.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 3. The names of the above-listed copyright holders may not be used
  14  *    to endorse or promote products derived from this software without
  15  *    specific prior written permission.
  16  *
  17  * ALTERNATIVELY, this software may be distributed under the terms of the
  18  * GNU General Public License ("GPL") version 2, as published by the Free
  19  * Software Foundation.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS
  22  * IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
  23  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  24  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
  25  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
  26  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
  27  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
  28  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
  29  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
  30  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
  31  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <linux/kernel.h>
  35 #include <linux/types.h>
  36 #include <linux/errno.h>
  37 #include <linux/interrupt.h>
  38 #include <linux/pagemap.h>
  39 #include <linux/dma-mapping.h>
  40 #include <linux/version.h>
  41 #include <linux/io.h>
  42 #include <linux/platform_device.h>
  43 #include <linux/uaccess.h>
  44 #include <linux/of.h>
  45 #include <asm/pgtable.h>
  46 #include <soc/bcm2835/raspberrypi-firmware.h>
  47
  48 #define dmac_map_area                   __glue(_CACHE,_dma_map_area)
  49 #define dmac_unmap_area                 __glue(_CACHE,_dma_unmap_area)
  50
  51 extern void dmac_map_area(const void *, size_t, int);
  52 extern void dmac_unmap_area(const void *, size_t, int);
  53
  54 #define TOTAL_SLOTS (VCHIQ_SLOT_ZERO_SLOTS + 2 * 32)
  55
  56 #define VCHIQ_ARM_ADDRESS(x) ((void *)((char *)x + g_virt_to_bus_offset))
  57
  58 #include "vchiq_arm.h"
  59 #include "vchiq_2835.h"
  60 #include "vchiq_connected.h"
  61 #include "vchiq_killable.h"
  62
  63 #define MAX_FRAGMENTS (VCHIQ_NUM_CURRENT_BULKS * 2)
  64
  65 #define BELL0   0x00
  66 #define BELL2   0x08
  67
  68 typedef struct vchiq_2835_state_struct {
  69    int inited;
  70    VCHIQ_ARM_STATE_T arm_state;
  71 } VCHIQ_2835_ARM_STATE_T;
  72
  73 static void __iomem *g_regs;
  74 static unsigned int g_cache_line_size = sizeof(CACHE_LINE_SIZE);
  75 static unsigned int g_fragments_size;
  76 static char *g_fragments_base;
  77 static char *g_free_fragments;
  78 static struct semaphore g_free_fragments_sema;
  79 static unsigned long g_virt_to_bus_offset;
  80
  81 extern int vchiq_arm_log_level;
  82
  83 static DEFINE_SEMAPHORE(g_free_fragments_mutex);
  84
  85 static irqreturn_t
  86 vchiq_doorbell_irq(int irq, void *dev_id);
  87
  88 static int
  89 create_pagelist(char __user *buf, size_t count, unsigned short type,
  90                 struct task_struct *task, PAGELIST_T ** ppagelist);
  91
  92 static void
  93 free_pagelist(PAGELIST_T *pagelist, int actual);
  94
  95 int vchiq_platform_init(struct platform_device *pdev, VCHIQ_STATE_T *state)
  96 {
  97         struct device *dev = &pdev->dev;
  98         struct rpi_firmware *fw = platform_get_drvdata(pdev);
  99         VCHIQ_SLOT_ZERO_T *vchiq_slot_zero;
 100         struct resource *res;
 101         void *slot_mem;
 102         dma_addr_t slot_phys;
 103         u32 channelbase;
 104         int slot_mem_size, frag_mem_size;
 105         int err, irq, i;
 106
 107         g_virt_to_bus_offset = virt_to_dma(dev, (void *)0);
 108
 109         (void)of_property_read_u32(dev->of_node, "cache-line-size",
 110                                    &g_cache_line_size);
 111         g_fragments_size = 2 * g_cache_line_size;
 112
 113         /* Allocate space for the channels in coherent memory */
 114         slot_mem_size = PAGE_ALIGN(TOTAL_SLOTS * VCHIQ_SLOT_SIZE);
 115         frag_mem_size = PAGE_ALIGN(g_fragments_size * MAX_FRAGMENTS);
 116
 117         slot_mem = dmam_alloc_coherent(dev, slot_mem_size + frag_mem_size,
 118                                        &slot_phys, GFP_KERNEL);
 119         if (!slot_mem) {
 120                 dev_err(dev, "could not allocate DMA memory\n");
 121                 return -ENOMEM;
 122         }
 123
 124         WARN_ON(((int)slot_mem & (PAGE_SIZE - 1)) != 0);
 125
 126         vchiq_slot_zero = vchiq_init_slots(slot_mem, slot_mem_size);
 127         if (!vchiq_slot_zero)
 128                 return -EINVAL;
 129
 130         vchiq_slot_zero->platform_data[VCHIQ_PLATFORM_FRAGMENTS_OFFSET_IDX] =
 131                 (int)slot_phys + slot_mem_size;
 132         vchiq_slot_zero->platform_data[VCHIQ_PLATFORM_FRAGMENTS_COUNT_IDX] =
 133                 MAX_FRAGMENTS;
 134
 135         g_fragments_base = (char *)slot_mem + slot_mem_size;
 136         slot_mem_size += frag_mem_size;
 137
 138         g_free_fragments = g_fragments_base;
 139         for (i = 0; i < (MAX_FRAGMENTS - 1); i++) {
 140                 *(char **)&g_fragments_base[i*g_fragments_size] =
 141                         &g_fragments_base[(i + 1)*g_fragments_size];
 142         }
 143         *(char **)&g_fragments_base[i * g_fragments_size] = NULL;
 144         sema_init(&g_free_fragments_sema, MAX_FRAGMENTS);
 145
 146         if (vchiq_init_state(state, vchiq_slot_zero, 0) != VCHIQ_SUCCESS)
 147                 return -EINVAL;
 148
 149         res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
 150         g_regs = devm_ioremap_resource(&pdev->dev, res);
 151         if (IS_ERR(g_regs))
 152                 return PTR_ERR(g_regs);
 153
 154         irq = platform_get_irq(pdev, 0);
 155         if (irq <= 0) {
 156                 dev_err(dev, "failed to get IRQ\n");
 157                 return irq;
 158         }
 159
 160         err = devm_request_irq(dev, irq, vchiq_doorbell_irq, IRQF_IRQPOLL,
 161                                "VCHIQ doorbell", state);
 162         if (err) {
 163                 dev_err(dev, "failed to register irq=%d\n", irq);
 164                 return err;
 165         }
 166
 167         /* Send the base address of the slots to VideoCore */
 168         channelbase = slot_phys;
 169         err = rpi_firmware_property(fw, RPI_FIRMWARE_VCHIQ_INIT,
 170                                     &channelbase, sizeof(channelbase));
 171         if (err || channelbase) {
 172                 dev_err(dev, "failed to set channelbase\n");
 173                 return err ? : -ENXIO;
 174         }
 175
 176         vchiq_log_info(vchiq_arm_log_level,
 177                 "vchiq_init - done (slots %x, phys %pad)",
 178                 (unsigned int)vchiq_slot_zero, &slot_phys);
 179
 180         vchiq_call_connected_callbacks();
 181
 182    return 0;
 183 }
 184
 185 VCHIQ_STATUS_T
 186 vchiq_platform_init_state(VCHIQ_STATE_T *state)
 187 {
 188    VCHIQ_STATUS_T status = VCHIQ_SUCCESS;
 189    state->platform_state = kzalloc(sizeof(VCHIQ_2835_ARM_STATE_T), GFP_KERNEL);
 190    ((VCHIQ_2835_ARM_STATE_T*)state->platform_state)->inited = 1;
 191    status = vchiq_arm_init_state(state, &((VCHIQ_2835_ARM_STATE_T*)state->platform_state)->arm_state);
 192    if(status != VCHIQ_SUCCESS)
 193    {
 194       ((VCHIQ_2835_ARM_STATE_T*)state->platform_state)->inited = 0;
 195    }
 196    return status;
 197 }
 198
 199 VCHIQ_ARM_STATE_T*
 200 vchiq_platform_get_arm_state(VCHIQ_STATE_T *state)
 201 {
 202    if(!((VCHIQ_2835_ARM_STATE_T*)state->platform_state)->inited)
 203    {
 204       BUG();
 205    }
 206    return &((VCHIQ_2835_ARM_STATE_T*)state->platform_state)->arm_state;
 207 }
 208
 209 void
 210 remote_event_signal(REMOTE_EVENT_T *event)
 211 {
 212         wmb();
 213
 214         event->fired = 1;
 215
 216         dsb();         /* data barrier operation */
 217
 218         if (event->armed)
 219                 writel(0, g_regs + BELL2); /* trigger vc interrupt */
 220 }
 221
 222 int
 223 vchiq_copy_from_user(void *dst, const void *src, int size)
 224 {
 225         if ((uint32_t)src < TASK_SIZE) {
 226                 return copy_from_user(dst, src, size);
 227         } else {
 228                 memcpy(dst, src, size);
 229                 return 0;
 230         }
 231 }
 232
 233 VCHIQ_STATUS_T
 234 vchiq_prepare_bulk_data(VCHIQ_BULK_T *bulk, VCHI_MEM_HANDLE_T memhandle,
 235         void *offset, int size, int dir)
 236 {
 237         PAGELIST_T *pagelist;
 238         int ret;
 239
 240         WARN_ON(memhandle != VCHI_MEM_HANDLE_INVALID);
 241
 242         ret = create_pagelist((char __user *)offset, size,
 243                         (dir == VCHIQ_BULK_RECEIVE)
 244                         ? PAGELIST_READ
 245                         : PAGELIST_WRITE,
 246                         current,
 247                         &pagelist);
 248         if (ret != 0)
 249                 return VCHIQ_ERROR;
 250
 251         bulk->handle = memhandle;
 252         bulk->data = VCHIQ_ARM_ADDRESS(pagelist);
 253
 254         /* Store the pagelist address in remote_data, which isn't used by the
 255            slave. */
 256         bulk->remote_data = pagelist;
 257
 258         return VCHIQ_SUCCESS;
 259 }
 260
 261 void
 262 vchiq_complete_bulk(VCHIQ_BULK_T *bulk)
 263 {
 264         if (bulk && bulk->remote_data && bulk->actual)
 265                 free_pagelist((PAGELIST_T *)bulk->remote_data, bulk->actual);
 266 }
 267
 268 void
 269 vchiq_transfer_bulk(VCHIQ_BULK_T *bulk)
 270 {
 271         /*
 272          * This should only be called on the master (VideoCore) side, but
 273          * provide an implementation to avoid the need for ifdefery.
 274          */
 275         BUG();
 276 }
 277
 278 void
 279 vchiq_dump_platform_state(void *dump_context)
 280 {
 281         char buf[80];
 282         int len;
 283         len = snprintf(buf, sizeof(buf),
 284                 "  Platform: 2835 (VC master)");
 285         vchiq_dump(dump_context, buf, len + 1);
 286 }
 287
 288 VCHIQ_STATUS_T
 289 vchiq_platform_suspend(VCHIQ_STATE_T *state)
 290 {
 291    return VCHIQ_ERROR;
 292 }
 293
 294 VCHIQ_STATUS_T
 295 vchiq_platform_resume(VCHIQ_STATE_T *state)
 296 {
 297    return VCHIQ_SUCCESS;
 298 }
 299
 300 void
 301 vchiq_platform_paused(VCHIQ_STATE_T *state)
 302 {
 303 }
 304
 305 void
 306 vchiq_platform_resumed(VCHIQ_STATE_T *state)
 307 {
 308 }
 309
 310 int
 311 vchiq_platform_videocore_wanted(VCHIQ_STATE_T* state)
 312 {
 313    return 1; // autosuspend not supported - videocore always wanted
 314 }
 315
 316 int
 317 vchiq_platform_use_suspend_timer(void)
 318 {
 319    return 0;
 320 }
 321 void
 322 vchiq_dump_platform_use_state(VCHIQ_STATE_T *state)
 323 {
 324         vchiq_log_info(vchiq_arm_log_level, "Suspend timer not in use");
 325 }
 326 void
 327 vchiq_platform_handle_timeout(VCHIQ_STATE_T *state)
 328 {
 329         (void)state;
 330 }
 331 /*
 332  * Local functions
 333  */
 334
 335 static irqreturn_t
 336 vchiq_doorbell_irq(int irq, void *dev_id)
 337 {
 338         VCHIQ_STATE_T *state = dev_id;
 339         irqreturn_t ret = IRQ_NONE;
 340         unsigned int status;
 341
 342         /* Read (and clear) the doorbell */
 343         status = readl(g_regs + BELL0);
 344
 345         if (status & 0x4) {  /* Was the doorbell rung? */
 346                 remote_event_pollall(state);
 347                 ret = IRQ_HANDLED;
 348         }
 349
 350         return ret;
 351 }
 352
 353 /* There is a potential problem with partial cache lines (pages?)
 354 ** at the ends of the block when reading. If the CPU accessed anything in
 355 ** the same line (page?) then it may have pulled old data into the cache,
 356 ** obscuring the new data underneath. We can solve this by transferring the
 357 ** partial cache lines separately, and allowing the ARM to copy into the
 358 ** cached area.
 359
 360 ** N.B. This implementation plays slightly fast and loose with the Linux
 361 ** driver programming rules, e.g. its use of dmac_map_area instead of
 362 ** dma_map_single, but it isn't a multi-platform driver and it benefits
 363 ** from increased speed as a result.
 364 */
 365
 366 static int
 367 create_pagelist(char __user *buf, size_t count, unsigned short type,
 368         struct task_struct *task, PAGELIST_T ** ppagelist)
 369 {
 370         PAGELIST_T *pagelist;
 371         struct page **pages;
 372         unsigned long *addrs;
 373         unsigned int num_pages, offset, i;
 374         char *addr, *base_addr, *next_addr;
 375         int run, addridx, actual_pages;
 376         unsigned long *need_release;
 377
 378         offset = (unsigned int)buf & (PAGE_SIZE - 1);
 379         num_pages = (count + offset + PAGE_SIZE - 1) / PAGE_SIZE;
 380
 381         *ppagelist = NULL;
 382
 383         /* Allocate enough storage to hold the page pointers and the page
 384         ** list
 385         */
 386         pagelist = kmalloc(sizeof(PAGELIST_T) +
 387                            (num_pages * sizeof(unsigned long)) +
 388                            sizeof(unsigned long) +
 389                            (num_pages * sizeof(pages[0])),
 390                            GFP_KERNEL);
 391
 392         vchiq_log_trace(vchiq_arm_log_level,
 393                 "create_pagelist - %x", (unsigned int)pagelist);
 394         if (!pagelist)
 395                 return -ENOMEM;
 396
 397         addrs = pagelist->addrs;
 398         need_release = (unsigned long *)(addrs + num_pages);
 399         pages = (struct page **)(addrs + num_pages + 1);
 400
 401         if (is_vmalloc_addr(buf)) {
 402                 int dir = (type == PAGELIST_WRITE) ?
 403                         DMA_TO_DEVICE : DMA_FROM_DEVICE;
 404                 unsigned long length = count;
 405                 unsigned int off = offset;
 406
 407                 for (actual_pages = 0; actual_pages < num_pages;
 408                      actual_pages++) {
 409                         struct page *pg = vmalloc_to_page(buf + (actual_pages *
 410                                                                  PAGE_SIZE));
 411                         size_t bytes = PAGE_SIZE - off;
 412
 413                         if (bytes > length)
 414                                 bytes = length;
 415                         pages[actual_pages] = pg;
 416                         dmac_map_area(page_address(pg) + off, bytes, dir);
 417                         length -= bytes;
 418                         off = 0;
 419                 }
 420                 *need_release = 0; /* do not try and release vmalloc pages */
 421         } else {
 422                 down_read(&task->mm->mmap_sem);
 423                 actual_pages = get_user_pages(task, task->mm,
 424                                           (unsigned long)buf & ~(PAGE_SIZE - 1),
 425                                           num_pages,
 426                                           (type == PAGELIST_READ) /*Write */ ,
 427                                           0 /*Force */ ,
 428                                           pages,
 429                                           NULL /*vmas */);
 430                 up_read(&task->mm->mmap_sem);
 431
 432                 if (actual_pages != num_pages) {
 433                         vchiq_log_info(vchiq_arm_log_level,
 434                                        "create_pagelist - only %d/%d pages locked",
 435                                        actual_pages,
 436                                        num_pages);
 437
 438                         /* This is probably due to the process being killed */
 439                         while (actual_pages > 0)
 440                         {
 441                                 actual_pages--;
 442                                 page_cache_release(pages[actual_pages]);
 443                         }
 444                         kfree(pagelist);
 445                         if (actual_pages == 0)
 446                                 actual_pages = -ENOMEM;
 447                         return actual_pages;
 448                 }
 449                 *need_release = 1; /* release user pages */
 450         }
 451
 452         pagelist->length = count;
 453         pagelist->type = type;
 454         pagelist->offset = offset;
 455
 456         /* Group the pages into runs of contiguous pages */
 457
 458         base_addr = VCHIQ_ARM_ADDRESS(page_address(pages[0]));
 459         next_addr = base_addr + PAGE_SIZE;
 460         addridx = 0;
 461         run = 0;
 462
 463         for (i = 1; i < num_pages; i++) {
 464                 addr = VCHIQ_ARM_ADDRESS(page_address(pages[i]));
 465                 if ((addr == next_addr) && (run < (PAGE_SIZE - 1))) {
 466                         next_addr += PAGE_SIZE;
 467                         run++;
 468                 } else {
 469                         addrs[addridx] = (unsigned long)base_addr + run;
 470                         addridx++;
 471                         base_addr = addr;
 472                         next_addr = addr + PAGE_SIZE;
 473                         run = 0;
 474                 }
 475         }
 476
 477         addrs[addridx] = (unsigned long)base_addr + run;
 478         addridx++;
 479
 480         /* Partial cache lines (fragments) require special measures */
 481         if ((type == PAGELIST_READ) &&
 482                 ((pagelist->offset & (g_cache_line_size - 1)) ||
 483                 ((pagelist->offset + pagelist->length) &
 484                 (g_cache_line_size - 1)))) {
 485                 char *fragments;
 486
 487                 if (down_interruptible(&g_free_fragments_sema) != 0) {
 488                         kfree(pagelist);
 489                         return -EINTR;
 490                 }
 491
 492                 WARN_ON(g_free_fragments == NULL);
 493
 494                 down(&g_free_fragments_mutex);
 495                 fragments = g_free_fragments;
 496                 WARN_ON(fragments == NULL);
 497                 g_free_fragments = *(char **) g_free_fragments;
 498                 up(&g_free_fragments_mutex);
 499                 pagelist->type = PAGELIST_READ_WITH_FRAGMENTS +
 500                         (fragments - g_fragments_base) / g_fragments_size;
 501         }
 502
 503         dmac_flush_range(pagelist, addrs + num_pages);
 504
 505         *ppagelist = pagelist;
 506
 507         return 0;
 508 }
 509
 510 static void
 511 free_pagelist(PAGELIST_T *pagelist, int actual)
 512 {
 513         unsigned long *need_release;
 514         struct page **pages;
 515         unsigned int num_pages, i;
 516
 517         vchiq_log_trace(vchiq_arm_log_level,
 518                 "free_pagelist - %x, %d", (unsigned int)pagelist, actual);
 519
 520         num_pages =
 521                 (pagelist->length + pagelist->offset + PAGE_SIZE - 1) /
 522                 PAGE_SIZE;
 523
 524         need_release = (unsigned long *)(pagelist->addrs + num_pages);
 525         pages = (struct page **)(pagelist->addrs + num_pages + 1);
 526
 527         /* Deal with any partial cache lines (fragments) */
 528         if (pagelist->type >= PAGELIST_READ_WITH_FRAGMENTS) {
 529                 char *fragments = g_fragments_base +
 530                         (pagelist->type - PAGELIST_READ_WITH_FRAGMENTS) *
 531                         g_fragments_size;
 532                 int head_bytes, tail_bytes;
 533                 head_bytes = (g_cache_line_size - pagelist->offset) &
 534                         (g_cache_line_size - 1);
 535                 tail_bytes = (pagelist->offset + actual) &
 536                         (g_cache_line_size - 1);
 537
 538                 if ((actual >= 0) && (head_bytes != 0)) {
 539                         if (head_bytes > actual)
 540                                 head_bytes = actual;
 541
 542                         memcpy((char *)page_address(pages[0]) +
 543                                 pagelist->offset,
 544                                 fragments,
 545                                 head_bytes);
 546                 }
 547                 if ((actual >= 0) && (head_bytes < actual) &&
 548                         (tail_bytes != 0)) {
 549                         memcpy((char *)page_address(pages[num_pages - 1]) +
 550                                 ((pagelist->offset + actual) &
 551                                 (PAGE_SIZE - 1) & ~(g_cache_line_size - 1)),
 552                                 fragments + g_cache_line_size,
 553                                 tail_bytes);
 554                 }
 555
 556                 down(&g_free_fragments_mutex);
 557                 *(char **)fragments = g_free_fragments;
 558                 g_free_fragments = fragments;
 559                 up(&g_free_fragments_mutex);
 560                 up(&g_free_fragments_sema);
 561         }
 562
 563         if (*need_release) {
 564                 unsigned int length = pagelist->length;
 565                 unsigned int offset = pagelist->offset;
 566
 567                 for (i = 0; i < num_pages; i++) {
 568                         struct page *pg = pages[i];
 569
 570                         if (pagelist->type != PAGELIST_WRITE) {
 571                                 unsigned int bytes = PAGE_SIZE - offset;
 572
 573                                 if (bytes > length)
 574                                         bytes = length;
 575                                 dmac_unmap_area(page_address(pg) + offset,
 576                                                 bytes, DMA_FROM_DEVICE);
 577                                 length -= bytes;
 578                                 offset = 0;
 579                                 set_page_dirty(pg);
 580                         }
 581                         page_cache_release(pg);
 582                 }
 583         }
 584
 585         kfree(pagelist);
 586 }