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