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

   1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
   2 /* Copyright (c) 2010-2012 Broadcom. All rights reserved. */
   3
   4 #include <linux/kernel.h>
   5 #include <linux/types.h>
   6 #include <linux/errno.h>
   7 #include <linux/interrupt.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/dma-mapping.h>
  10 #include <linux/io.h>
  11 #include <linux/platform_device.h>
  12 #include <linux/uaccess.h>
  13 #include <linux/mm.h>
  14 #include <linux/of.h>
  15 #include <soc/bcm2835/raspberrypi-firmware.h>
  16
  17 #define TOTAL_SLOTS (VCHIQ_SLOT_ZERO_SLOTS + 2 * 32)
  18
  19 #include "vchiq_arm.h"
  20 #include "vchiq_connected.h"
  21 #include "vchiq_pagelist.h"
  22
  23 #define MAX_FRAGMENTS (VCHIQ_NUM_CURRENT_BULKS * 2)
  24
  25 #define VCHIQ_PLATFORM_FRAGMENTS_OFFSET_IDX 0
  26 #define VCHIQ_PLATFORM_FRAGMENTS_COUNT_IDX  1
  27
  28 #define BELL0   0x00
  29 #define BELL2   0x08
  30
  31 struct vchiq_2835_state {
  32         int inited;
  33         struct vchiq_arm_state arm_state;
  34 };
  35
  36 struct vchiq_pagelist_info {
  37         struct pagelist *pagelist;
  38         size_t pagelist_buffer_size;
  39         dma_addr_t dma_addr;
  40         enum dma_data_direction dma_dir;
  41         unsigned int num_pages;
  42         unsigned int pages_need_release;
  43         struct page **pages;
  44         struct scatterlist *scatterlist;
  45         unsigned int scatterlist_mapped;
  46 };
  47
  48 static void __iomem *g_regs;
  49 /* This value is the size of the L2 cache lines as understood by the
  50  * VPU firmware, which determines the required alignment of the
  51  * offsets/sizes in pagelists.
  52  *
  53  * Modern VPU firmware looks for a DT "cache-line-size" property in
  54  * the VCHIQ node and will overwrite it with the actual L2 cache size,
  55  * which the kernel must then respect.  That property was rejected
  56  * upstream, so we have to use the VPU firmware's compatibility value
  57  * of 32.
  58  */
  59 static unsigned int g_cache_line_size = 32;
  60 static unsigned int g_fragments_size;
  61 static char *g_fragments_base;
  62 static char *g_free_fragments;
  63 static struct semaphore g_free_fragments_sema;
  64 static struct device *g_dev;
  65
  66 static DEFINE_SEMAPHORE(g_free_fragments_mutex);
  67
  68 static irqreturn_t
  69 vchiq_doorbell_irq(int irq, void *dev_id);
  70
  71 static struct vchiq_pagelist_info *
  72 create_pagelist(char __user *buf, size_t count, unsigned short type);
  73
  74 static void
  75 free_pagelist(struct vchiq_pagelist_info *pagelistinfo,
  76               int actual);
  77
  78 int vchiq_platform_init(struct platform_device *pdev, struct vchiq_state *state)
  79 {
  80         struct device *dev = &pdev->dev;
  81         struct vchiq_drvdata *drvdata = platform_get_drvdata(pdev);
  82         struct rpi_firmware *fw = drvdata->fw;
  83         struct vchiq_slot_zero *vchiq_slot_zero;
  84         struct resource *res;
  85         void *slot_mem;
  86         dma_addr_t slot_phys;
  87         u32 channelbase;
  88         int slot_mem_size, frag_mem_size;
  89         int err, irq, i;
  90
  91         /*
  92          * VCHI messages between the CPU and firmware use
  93          * 32-bit bus addresses.
  94          */
  95         err = dma_set_mask_and_coherent(dev, DMA_BIT_MASK(32));
  96
  97         if (err < 0)
  98                 return err;
  99
 100         g_cache_line_size = drvdata->cache_line_size;
 101         g_fragments_size = 2 * g_cache_line_size;
 102
 103         /* Allocate space for the channels in coherent memory */
 104         slot_mem_size = PAGE_ALIGN(TOTAL_SLOTS * VCHIQ_SLOT_SIZE);
 105         frag_mem_size = PAGE_ALIGN(g_fragments_size * MAX_FRAGMENTS);
 106
 107         slot_mem = dmam_alloc_coherent(dev, slot_mem_size + frag_mem_size,
 108                                        &slot_phys, GFP_KERNEL);
 109         if (!slot_mem) {
 110                 dev_err(dev, "could not allocate DMA memory\n");
 111                 return -ENOMEM;
 112         }
 113
 114         WARN_ON(((unsigned long)slot_mem & (PAGE_SIZE - 1)) != 0);
 115
 116         vchiq_slot_zero = vchiq_init_slots(slot_mem, slot_mem_size);
 117         if (!vchiq_slot_zero)
 118                 return -EINVAL;
 119
 120         vchiq_slot_zero->platform_data[VCHIQ_PLATFORM_FRAGMENTS_OFFSET_IDX] =
 121                 (int)slot_phys + slot_mem_size;
 122         vchiq_slot_zero->platform_data[VCHIQ_PLATFORM_FRAGMENTS_COUNT_IDX] =
 123                 MAX_FRAGMENTS;
 124
 125         g_fragments_base = (char *)slot_mem + slot_mem_size;
 126
 127         g_free_fragments = g_fragments_base;
 128         for (i = 0; i < (MAX_FRAGMENTS - 1); i++) {
 129                 *(char **)&g_fragments_base[i*g_fragments_size] =
 130                         &g_fragments_base[(i + 1)*g_fragments_size];
 131         }
 132         *(char **)&g_fragments_base[i * g_fragments_size] = NULL;
 133         sema_init(&g_free_fragments_sema, MAX_FRAGMENTS);
 134
 135         if (vchiq_init_state(state, vchiq_slot_zero) != VCHIQ_SUCCESS)
 136                 return -EINVAL;
 137
 138         res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
 139         g_regs = devm_ioremap_resource(&pdev->dev, res);
 140         if (IS_ERR(g_regs))
 141                 return PTR_ERR(g_regs);
 142
 143         irq = platform_get_irq(pdev, 0);
 144         if (irq <= 0) {
 145                 dev_err(dev, "failed to get IRQ\n");
 146                 return irq;
 147         }
 148
 149         err = devm_request_irq(dev, irq, vchiq_doorbell_irq, IRQF_IRQPOLL,
 150                                "VCHIQ doorbell", state);
 151         if (err) {
 152                 dev_err(dev, "failed to register irq=%d\n", irq);
 153                 return err;
 154         }
 155
 156         /* Send the base address of the slots to VideoCore */
 157         channelbase = slot_phys;
 158         err = rpi_firmware_property(fw, RPI_FIRMWARE_VCHIQ_INIT,
 159                                     &channelbase, sizeof(channelbase));
 160         if (err || channelbase) {
 161                 dev_err(dev, "failed to set channelbase\n");
 162                 return err ? : -ENXIO;
 163         }
 164
 165         g_dev = dev;
 166         vchiq_log_info(vchiq_arm_log_level,
 167                 "vchiq_init - done (slots %pK, phys %pad)",
 168                 vchiq_slot_zero, &slot_phys);
 169
 170         vchiq_call_connected_callbacks();
 171
 172         return 0;
 173 }
 174
 175 VCHIQ_STATUS_T
 176 vchiq_platform_init_state(struct vchiq_state *state)
 177 {
 178         VCHIQ_STATUS_T status = VCHIQ_SUCCESS;
 179         struct vchiq_2835_state *platform_state;
 180
 181         state->platform_state = kzalloc(sizeof(*platform_state), GFP_KERNEL);
 182         if (!state->platform_state)
 183                 return VCHIQ_ERROR;
 184
 185         platform_state = (struct vchiq_2835_state *)state->platform_state;
 186
 187         platform_state->inited = 1;
 188         status = vchiq_arm_init_state(state, &platform_state->arm_state);
 189
 190         if (status != VCHIQ_SUCCESS)
 191                 platform_state->inited = 0;
 192
 193         return status;
 194 }
 195
 196 struct vchiq_arm_state*
 197 vchiq_platform_get_arm_state(struct vchiq_state *state)
 198 {
 199         struct vchiq_2835_state *platform_state;
 200
 201         platform_state   = (struct vchiq_2835_state *)state->platform_state;
 202
 203         WARN_ON_ONCE(!platform_state->inited);
 204
 205         return &platform_state->arm_state;
 206 }
 207
 208 void
 209 remote_event_signal(struct remote_event *event)
 210 {
 211         wmb();
 212
 213         event->fired = 1;
 214
 215         dsb(sy);         /* data barrier operation */
 216
 217         if (event->armed)
 218                 writel(0, g_regs + BELL2); /* trigger vc interrupt */
 219 }
 220
 221 VCHIQ_STATUS_T
 222 vchiq_prepare_bulk_data(struct vchiq_bulk *bulk, void *offset, int size,
 223                         int dir)
 224 {
 225         struct vchiq_pagelist_info *pagelistinfo;
 226
 227         pagelistinfo = create_pagelist((char __user *)offset, size,
 228                                        (dir == VCHIQ_BULK_RECEIVE)
 229                                        ? PAGELIST_READ
 230                                        : PAGELIST_WRITE);
 231
 232         if (!pagelistinfo)
 233                 return VCHIQ_ERROR;
 234
 235         bulk->data = (void *)(unsigned long)pagelistinfo->dma_addr;
 236
 237         /*
 238          * Store the pagelistinfo address in remote_data,
 239          * which isn't used by the slave.
 240          */
 241         bulk->remote_data = pagelistinfo;
 242
 243         return VCHIQ_SUCCESS;
 244 }
 245
 246 void
 247 vchiq_complete_bulk(struct vchiq_bulk *bulk)
 248 {
 249         if (bulk && bulk->remote_data && bulk->actual)
 250                 free_pagelist((struct vchiq_pagelist_info *)bulk->remote_data,
 251                               bulk->actual);
 252 }
 253
 254 void
 255 vchiq_dump_platform_state(void *dump_context)
 256 {
 257         char buf[80];
 258         int len;
 259
 260         len = snprintf(buf, sizeof(buf),
 261                 "  Platform: 2835 (VC master)");
 262         vchiq_dump(dump_context, buf, len + 1);
 263 }
 264
 265 VCHIQ_STATUS_T
 266 vchiq_platform_suspend(struct vchiq_state *state)
 267 {
 268         return VCHIQ_ERROR;
 269 }
 270
 271 VCHIQ_STATUS_T
 272 vchiq_platform_resume(struct vchiq_state *state)
 273 {
 274         return VCHIQ_SUCCESS;
 275 }
 276
 277 void
 278 vchiq_platform_paused(struct vchiq_state *state)
 279 {
 280 }
 281
 282 void
 283 vchiq_platform_resumed(struct vchiq_state *state)
 284 {
 285 }
 286
 287 int
 288 vchiq_platform_videocore_wanted(struct vchiq_state *state)
 289 {
 290         return 1; // autosuspend not supported - videocore always wanted
 291 }
 292
 293 int
 294 vchiq_platform_use_suspend_timer(void)
 295 {
 296         return 0;
 297 }
 298 void
 299 vchiq_dump_platform_use_state(struct vchiq_state *state)
 300 {
 301         vchiq_log_info(vchiq_arm_log_level, "Suspend timer not in use");
 302 }
 303 void
 304 vchiq_platform_handle_timeout(struct vchiq_state *state)
 305 {
 306         (void)state;
 307 }
 308 /*
 309  * Local functions
 310  */
 311
 312 static irqreturn_t
 313 vchiq_doorbell_irq(int irq, void *dev_id)
 314 {
 315         struct vchiq_state *state = dev_id;
 316         irqreturn_t ret = IRQ_NONE;
 317         unsigned int status;
 318
 319         /* Read (and clear) the doorbell */
 320         status = readl(g_regs + BELL0);
 321
 322         if (status & 0x4) {  /* Was the doorbell rung? */
 323                 remote_event_pollall(state);
 324                 ret = IRQ_HANDLED;
 325         }
 326
 327         return ret;
 328 }
 329
 330 static void
 331 cleanup_pagelistinfo(struct vchiq_pagelist_info *pagelistinfo)
 332 {
 333         if (pagelistinfo->scatterlist_mapped) {
 334                 dma_unmap_sg(g_dev, pagelistinfo->scatterlist,
 335                              pagelistinfo->num_pages, pagelistinfo->dma_dir);
 336         }
 337
 338         if (pagelistinfo->pages_need_release) {
 339                 unsigned int i;
 340
 341                 for (i = 0; i < pagelistinfo->num_pages; i++)
 342                         put_page(pagelistinfo->pages[i]);
 343         }
 344
 345         dma_free_coherent(g_dev, pagelistinfo->pagelist_buffer_size,
 346                           pagelistinfo->pagelist, pagelistinfo->dma_addr);
 347 }
 348
 349 /* There is a potential problem with partial cache lines (pages?)
 350  * at the ends of the block when reading. If the CPU accessed anything in
 351  * the same line (page?) then it may have pulled old data into the cache,
 352  * obscuring the new data underneath. We can solve this by transferring the
 353  * partial cache lines separately, and allowing the ARM to copy into the
 354  * cached area.
 355  */
 356
 357 static struct vchiq_pagelist_info *
 358 create_pagelist(char __user *buf, size_t count, unsigned short type)
 359 {
 360         struct pagelist *pagelist;
 361         struct vchiq_pagelist_info *pagelistinfo;
 362         struct page **pages;
 363         u32 *addrs;
 364         unsigned int num_pages, offset, i, k;
 365         int actual_pages;
 366         size_t pagelist_size;
 367         struct scatterlist *scatterlist, *sg;
 368         int dma_buffers;
 369         dma_addr_t dma_addr;
 370
 371         if (count >= INT_MAX - PAGE_SIZE)
 372                 return NULL;
 373
 374         offset = ((unsigned int)(unsigned long)buf & (PAGE_SIZE - 1));
 375         num_pages = DIV_ROUND_UP(count + offset, PAGE_SIZE);
 376
 377         if (num_pages > (SIZE_MAX - sizeof(struct pagelist) -
 378                          sizeof(struct vchiq_pagelist_info)) /
 379                         (sizeof(u32) + sizeof(pages[0]) +
 380                          sizeof(struct scatterlist)))
 381                 return NULL;
 382
 383         pagelist_size = sizeof(struct pagelist) +
 384                         (num_pages * sizeof(u32)) +
 385                         (num_pages * sizeof(pages[0]) +
 386                         (num_pages * sizeof(struct scatterlist))) +
 387                         sizeof(struct vchiq_pagelist_info);
 388
 389         /* Allocate enough storage to hold the page pointers and the page
 390          * list
 391          */
 392         pagelist = dma_alloc_coherent(g_dev, pagelist_size, &dma_addr,
 393                                       GFP_KERNEL);
 394
 395         vchiq_log_trace(vchiq_arm_log_level, "%s - %pK", __func__, pagelist);
 396
 397         if (!pagelist)
 398                 return NULL;
 399
 400         addrs           = pagelist->addrs;
 401         pages           = (struct page **)(addrs + num_pages);
 402         scatterlist     = (struct scatterlist *)(pages + num_pages);
 403         pagelistinfo    = (struct vchiq_pagelist_info *)
 404                           (scatterlist + num_pages);
 405
 406         pagelist->length = count;
 407         pagelist->type = type;
 408         pagelist->offset = offset;
 409
 410         /* Populate the fields of the pagelistinfo structure */
 411         pagelistinfo->pagelist = pagelist;
 412         pagelistinfo->pagelist_buffer_size = pagelist_size;
 413         pagelistinfo->dma_addr = dma_addr;
 414         pagelistinfo->dma_dir =  (type == PAGELIST_WRITE) ?
 415                                   DMA_TO_DEVICE : DMA_FROM_DEVICE;
 416         pagelistinfo->num_pages = num_pages;
 417         pagelistinfo->pages_need_release = 0;
 418         pagelistinfo->pages = pages;
 419         pagelistinfo->scatterlist = scatterlist;
 420         pagelistinfo->scatterlist_mapped = 0;
 421
 422         if (is_vmalloc_addr(buf)) {
 423                 unsigned long length = count;
 424                 unsigned int off = offset;
 425
 426                 for (actual_pages = 0; actual_pages < num_pages;
 427                      actual_pages++) {
 428                         struct page *pg = vmalloc_to_page(buf + (actual_pages *
 429                                                                  PAGE_SIZE));
 430                         size_t bytes = PAGE_SIZE - off;
 431
 432                         if (!pg) {
 433                                 cleanup_pagelistinfo(pagelistinfo);
 434                                 return NULL;
 435                         }
 436
 437                         if (bytes > length)
 438                                 bytes = length;
 439                         pages[actual_pages] = pg;
 440                         length -= bytes;
 441                         off = 0;
 442                 }
 443                 /* do not try and release vmalloc pages */
 444         } else {
 445                 actual_pages = get_user_pages_fast(
 446                                           (unsigned long)buf & PAGE_MASK,
 447                                           num_pages,
 448                                           type == PAGELIST_READ,
 449                                           pages);
 450
 451                 if (actual_pages != num_pages) {
 452                         vchiq_log_info(vchiq_arm_log_level,
 453                                        "%s - only %d/%d pages locked",
 454                                        __func__, actual_pages, num_pages);
 455
 456                         /* This is probably due to the process being killed */
 457                         while (actual_pages > 0) {
 458                                 actual_pages--;
 459                                 put_page(pages[actual_pages]);
 460                         }
 461                         cleanup_pagelistinfo(pagelistinfo);
 462                         return NULL;
 463                 }
 464                  /* release user pages */
 465                 pagelistinfo->pages_need_release = 1;
 466         }
 467
 468         /*
 469          * Initialize the scatterlist so that the magic cookie
 470          *  is filled if debugging is enabled
 471          */
 472         sg_init_table(scatterlist, num_pages);
 473         /* Now set the pages for each scatterlist */
 474         for (i = 0; i < num_pages; i++) {
 475                 unsigned int len = PAGE_SIZE - offset;
 476
 477                 if (len > count)
 478                         len = count;
 479                 sg_set_page(scatterlist + i, pages[i], len, offset);
 480                 offset = 0;
 481                 count -= len;
 482         }
 483
 484         dma_buffers = dma_map_sg(g_dev,
 485                                  scatterlist,
 486                                  num_pages,
 487                                  pagelistinfo->dma_dir);
 488
 489         if (dma_buffers == 0) {
 490                 cleanup_pagelistinfo(pagelistinfo);
 491                 return NULL;
 492         }
 493
 494         pagelistinfo->scatterlist_mapped = 1;
 495
 496         /* Combine adjacent blocks for performance */
 497         k = 0;
 498         for_each_sg(scatterlist, sg, dma_buffers, i) {
 499                 u32 len = sg_dma_len(sg);
 500                 u32 addr = sg_dma_address(sg);
 501
 502                 /* Note: addrs is the address + page_count - 1
 503                  * The firmware expects blocks after the first to be page-
 504                  * aligned and a multiple of the page size
 505                  */
 506                 WARN_ON(len == 0);
 507                 WARN_ON(i && (i != (dma_buffers - 1)) && (len & ~PAGE_MASK));
 508                 WARN_ON(i && (addr & ~PAGE_MASK));
 509                 if (k > 0 &&
 510                     ((addrs[k - 1] & PAGE_MASK) +
 511                      (((addrs[k - 1] & ~PAGE_MASK) + 1) << PAGE_SHIFT))
 512                     == (addr & PAGE_MASK))
 513                         addrs[k - 1] += ((len + PAGE_SIZE - 1) >> PAGE_SHIFT);
 514                 else
 515                         addrs[k++] = (addr & PAGE_MASK) |
 516                                 (((len + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1);
 517         }
 518
 519         /* Partial cache lines (fragments) require special measures */
 520         if ((type == PAGELIST_READ) &&
 521                 ((pagelist->offset & (g_cache_line_size - 1)) ||
 522                 ((pagelist->offset + pagelist->length) &
 523                 (g_cache_line_size - 1)))) {
 524                 char *fragments;
 525
 526                 if (down_killable(&g_free_fragments_sema)) {
 527                         cleanup_pagelistinfo(pagelistinfo);
 528                         return NULL;
 529                 }
 530
 531                 WARN_ON(g_free_fragments == NULL);
 532
 533                 down(&g_free_fragments_mutex);
 534                 fragments = g_free_fragments;
 535                 WARN_ON(fragments == NULL);
 536                 g_free_fragments = *(char **) g_free_fragments;
 537                 up(&g_free_fragments_mutex);
 538                 pagelist->type = PAGELIST_READ_WITH_FRAGMENTS +
 539                         (fragments - g_fragments_base) / g_fragments_size;
 540         }
 541
 542         return pagelistinfo;
 543 }
 544
 545 static void
 546 free_pagelist(struct vchiq_pagelist_info *pagelistinfo,
 547               int actual)
 548 {
 549         struct pagelist *pagelist = pagelistinfo->pagelist;
 550         struct page **pages = pagelistinfo->pages;
 551         unsigned int num_pages = pagelistinfo->num_pages;
 552
 553         vchiq_log_trace(vchiq_arm_log_level, "%s - %pK, %d",
 554                         __func__, pagelistinfo->pagelist, actual);
 555
 556         /*
 557          * NOTE: dma_unmap_sg must be called before the
 558          * cpu can touch any of the data/pages.
 559          */
 560         dma_unmap_sg(g_dev, pagelistinfo->scatterlist,
 561                      pagelistinfo->num_pages, pagelistinfo->dma_dir);
 562         pagelistinfo->scatterlist_mapped = 0;
 563
 564         /* Deal with any partial cache lines (fragments) */
 565         if (pagelist->type >= PAGELIST_READ_WITH_FRAGMENTS) {
 566                 char *fragments = g_fragments_base +
 567                         (pagelist->type - PAGELIST_READ_WITH_FRAGMENTS) *
 568                         g_fragments_size;
 569                 int head_bytes, tail_bytes;
 570
 571                 head_bytes = (g_cache_line_size - pagelist->offset) &
 572                         (g_cache_line_size - 1);
 573                 tail_bytes = (pagelist->offset + actual) &
 574                         (g_cache_line_size - 1);
 575
 576                 if ((actual >= 0) && (head_bytes != 0)) {
 577                         if (head_bytes > actual)
 578                                 head_bytes = actual;
 579
 580                         memcpy((char *)kmap(pages[0]) +
 581                                 pagelist->offset,
 582                                 fragments,
 583                                 head_bytes);
 584                         kunmap(pages[0]);
 585                 }
 586                 if ((actual >= 0) && (head_bytes < actual) &&
 587                         (tail_bytes != 0)) {
 588                         memcpy((char *)kmap(pages[num_pages - 1]) +
 589                                 ((pagelist->offset + actual) &
 590                                 (PAGE_SIZE - 1) & ~(g_cache_line_size - 1)),
 591                                 fragments + g_cache_line_size,
 592                                 tail_bytes);
 593                         kunmap(pages[num_pages - 1]);
 594                 }
 595
 596                 down(&g_free_fragments_mutex);
 597                 *(char **)fragments = g_free_fragments;
 598                 g_free_fragments = fragments;
 599                 up(&g_free_fragments_mutex);
 600                 up(&g_free_fragments_sema);
 601         }
 602
 603         /* Need to mark all the pages dirty. */
 604         if (pagelist->type != PAGELIST_WRITE &&
 605             pagelistinfo->pages_need_release) {
 606                 unsigned int i;
 607
 608                 for (i = 0; i < num_pages; i++)
 609                         set_page_dirty(pages[i]);
 610         }
 611
 612         cleanup_pagelistinfo(pagelistinfo);
 613 }