target/linux/ps3/patches-2.6.28/0010-ps3vram-ng.patch

   1 From d7ddc1aaee1ff6dd6a73bd3663b6c390800e0500 Mon Sep 17 00:00:00 2001
   2 From: Geert Uytterhoeven <Geert.Uytterhoeven@sonycom.com>
   3 Date: Wed, 25 Feb 2009 18:32:10 +0100
   4 Subject: [PATCH] ps3/block: Replace mtd/ps3vram by block/ps3vram
   5
   6 Convert the PS3 Video RAM Storage Driver from a MTD driver to a plain block
   7 device driver, as requested by Arnd Bergmann.
   8
   9 The ps3vram driver exposes unused video RAM on the PS3 as a block device
  10 suitable for storage or swap.  Fast data transfer is achieved using a local
  11 cache in system RAM and DMA transfers via the GPU.
  12
  13 The new driver is ca. 50% faster for reading, and ca. 10% for writing.
  14
  15 Signed-off-by: Geert Uytterhoeven <Geert.Uytterhoeven@sonycom.com>
  16 Acked-by: Jens Axboe <axboe@kernel.dk>
  17 Acked-by: Geoff Levand <geoffrey.levand@am.sony.com>
  18 Cc: Arnd Bergmann <arnd@arndb.de>
  19 Cc: David Woodhouse <David.Woodhouse@intel.com>
  20 Cc: Vivien Chappelier <vivien.chappelier@free.fr>
  21 Cc: Jim Paris <jim@jtan.com>
  22
  23 Notes:
  24
  25 The MTD-based PS3 Video RAM Storage Driver was integrated into the mainline
  26 kernel in 2.6.29-rc1.
  27
  28 Ideally, we think it would be best if the existing MTD-based ps3vram driver
  29 would be replaced by the new block-based ps3vram driver before 2.6.29 is
  30 released. This would relieve the burden of supporting two different swap space
  31 schemes on PS3 (swap on /dev/mtdblock0 vs. /dev/ps3vram) from the distro
  32 maintainer's shoulders, as in that case there would never have been a stable
  33 kernel version containing the MTD-based ps3vram driver.
  34
  35 Changes since previous submission (Wed, 4 Mar 2009 14:57:20 +0100 (CET)):
  36   - Use blk_queue_make_request() to get rid of the thread
  37   - Add a mutex (cfr. the old driver), as ps3vram_make_request() may be called
  38     concurrently
  39     TO DO (after 2.6.29): use a spinlock and a list to queue requests while the
  40     driver is busy
  41   - Remove the old MTD-based ps3vram driver and rename ps3vram-ng to ps3vram
  42   - Make PS3_VRAM depend on FB_PS3=y and m for now
  43     ps3vram relies on ps3fb being initialized first. The easiest way to do this
  44     is by making ps3vram modular, and ps3fb builtin
  45   - Remove the dependency on ps3fb_videomemory.size
  46     The loop to reduce ddr_size until it succeeds does the right thing anyway,
  47     and a few MiB of DDR RAM are reserved by the hypervisor
  48   - lv1 return codes can be int
  49   - Correct a few debug annotations
  50 ---
  51  arch/powerpc/platforms/ps3/Kconfig |    7 +
  52  drivers/block/Makefile             |    1 +
  53  drivers/block/ps3vram.c            |  865 ++++++++++++++++++++++++++++++++++++
  54  create mode 100644 drivers/block/ps3vram.c
  55  delete mode 100644 drivers/mtd/devices/ps3vram.c
  56
  57 diff --git a/arch/powerpc/platforms/ps3/Kconfig b/arch/powerpc/platforms/ps3/Kconfig
  58 index 920cf7a..740ef56 100644
  59 --- a/arch/powerpc/platforms/ps3/Kconfig
  60 +++ b/arch/powerpc/platforms/ps3/Kconfig
  61 @@ -128,6 +128,13 @@ config PS3_FLASH
  62           be disabled on the kernel command line using "ps3flash=off", to
  63           not allocate this fixed buffer.
  64
  65 +config PS3_VRAM
  66 +       tristate "PS3 Video RAM Storage Driver"
  67 +       depends on FB_PS3=y && BLOCK && m
  68 +       help
  69 +         This driver allows you to use excess PS3 video RAM as volatile
  70 +         storage or system swap.
  71 +
  72  config PS3_LPM
  73         tristate "PS3 Logical Performance Monitor support"
  74         depends on PPC_PS3
  75 diff --git a/drivers/block/Makefile b/drivers/block/Makefile
  76 index 204332b..87e120e 100644
  77 --- a/drivers/block/Makefile
  78 +++ b/drivers/block/Makefile
  79 @@ -9,6 +9,7 @@ obj-$(CONFIG_MAC_FLOPPY)        += swim3.o
  80  obj-$(CONFIG_BLK_DEV_FD)       += floppy.o
  81  obj-$(CONFIG_AMIGA_FLOPPY)     += amiflop.o
  82  obj-$(CONFIG_PS3_DISK)         += ps3disk.o
  83 +obj-$(CONFIG_PS3_VRAM)         += ps3vram.o
  84  obj-$(CONFIG_ATARI_FLOPPY)     += ataflop.o
  85  obj-$(CONFIG_AMIGA_Z2RAM)      += z2ram.o
  86  obj-$(CONFIG_BLK_DEV_RAM)      += brd.o
  87 diff --git a/drivers/block/ps3vram.c b/drivers/block/ps3vram.c
  88 new file mode 100644
  89 index 0000000..393ed67
  90 --- /dev/null
  91 +++ b/drivers/block/ps3vram.c
  92 @@ -0,0 +1,865 @@
  93 +/*
  94 + * ps3vram - Use extra PS3 video ram as MTD block device.
  95 + *
  96 + * Copyright 2009 Sony Corporation
  97 + *
  98 + * Based on the MTD ps3vram driver, which is
  99 + * Copyright (c) 2007-2008 Jim Paris <jim@jtan.com>
 100 + * Added support RSX DMA Vivien Chappelier <vivien.chappelier@free.fr>
 101 + */
 102 +
 103 +#include <linux/blkdev.h>
 104 +#include <linux/delay.h>
 105 +#include <linux/proc_fs.h>
 106 +#include <linux/seq_file.h>
 107 +
 108 +#include <asm/firmware.h>
 109 +#include <asm/lv1call.h>
 110 +#include <asm/ps3.h>
 111 +
 112 +
 113 +#define DEVICE_NAME            "ps3vram"
 114 +
 115 +
 116 +#define XDR_BUF_SIZE (2 * 1024 * 1024) /* XDR buffer (must be 1MiB aligned) */
 117 +#define XDR_IOIF 0x0c000000
 118 +
 119 +#define FIFO_BASE XDR_IOIF
 120 +#define FIFO_SIZE (64 * 1024)
 121 +
 122 +#define DMA_PAGE_SIZE (4 * 1024)
 123 +
 124 +#define CACHE_PAGE_SIZE (256 * 1024)
 125 +#define CACHE_PAGE_COUNT ((XDR_BUF_SIZE - FIFO_SIZE) / CACHE_PAGE_SIZE)
 126 +
 127 +#define CACHE_OFFSET CACHE_PAGE_SIZE
 128 +#define FIFO_OFFSET 0
 129 +
 130 +#define CTRL_PUT 0x10
 131 +#define CTRL_GET 0x11
 132 +#define CTRL_TOP 0x15
 133 +
 134 +#define UPLOAD_SUBCH   1
 135 +#define DOWNLOAD_SUBCH 2
 136 +
 137 +#define NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN   0x0000030c
 138 +#define NV_MEMORY_TO_MEMORY_FORMAT_NOTIFY      0x00000104
 139 +
 140 +#define L1GPU_CONTEXT_ATTRIBUTE_FB_BLIT 0x601
 141 +
 142 +#define CACHE_PAGE_PRESENT 1
 143 +#define CACHE_PAGE_DIRTY   2
 144 +
 145 +struct ps3vram_tag {
 146 +       unsigned int address;
 147 +       unsigned int flags;
 148 +};
 149 +
 150 +struct ps3vram_cache {
 151 +       unsigned int page_count;
 152 +       unsigned int page_size;
 153 +       struct ps3vram_tag *tags;
 154 +       unsigned int hit;
 155 +       unsigned int miss;
 156 +};
 157 +
 158 +struct ps3vram_priv {
 159 +       struct request_queue *queue;
 160 +       struct gendisk *gendisk;
 161 +
 162 +       u64 size;
 163 +
 164 +       u64 memory_handle;
 165 +       u64 context_handle;
 166 +       u32 *ctrl;
 167 +       u32 *reports;
 168 +       u8 __iomem *ddr_base;
 169 +       u8 *xdr_buf;
 170 +
 171 +       u32 *fifo_base;
 172 +       u32 *fifo_ptr;
 173 +
 174 +       struct ps3vram_cache cache;
 175 +
 176 +       /* Used to serialize cache/DMA operations */
 177 +       struct mutex lock;
 178 +};
 179 +
 180 +
 181 +static int ps3vram_major;
 182 +
 183 +
 184 +static struct block_device_operations ps3vram_fops = {
 185 +       .owner          = THIS_MODULE,
 186 +};
 187 +
 188 +
 189 +#define DMA_NOTIFIER_HANDLE_BASE 0x66604200 /* first DMA notifier handle */
 190 +#define DMA_NOTIFIER_OFFSET_BASE 0x1000     /* first DMA notifier offset */
 191 +#define DMA_NOTIFIER_SIZE        0x40
 192 +#define NOTIFIER 7     /* notifier used for completion report */
 193 +
 194 +static char *size = "256M";
 195 +module_param(size, charp, 0);
 196 +MODULE_PARM_DESC(size, "memory size");
 197 +
 198 +static u32 *ps3vram_get_notifier(u32 *reports, int notifier)
 199 +{
 200 +       return (void *)reports + DMA_NOTIFIER_OFFSET_BASE +
 201 +              DMA_NOTIFIER_SIZE * notifier;
 202 +}
 203 +
 204 +static void ps3vram_notifier_reset(struct ps3_system_bus_device *dev)
 205 +{
 206 +       struct ps3vram_priv *priv = dev->core.driver_data;
 207 +       u32 *notify = ps3vram_get_notifier(priv->reports, NOTIFIER);
 208 +       int i;
 209 +
 210 +       for (i = 0; i < 4; i++)
 211 +               notify[i] = 0xffffffff;
 212 +}
 213 +
 214 +static int ps3vram_notifier_wait(struct ps3_system_bus_device *dev,
 215 +                                unsigned int timeout_ms)
 216 +{
 217 +       struct ps3vram_priv *priv = dev->core.driver_data;
 218 +       u32 *notify = ps3vram_get_notifier(priv->reports, NOTIFIER);
 219 +       unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
 220 +
 221 +       do {
 222 +               if (!notify[3])
 223 +                       return 0;
 224 +               msleep(1);
 225 +       } while (time_before(jiffies, timeout));
 226 +
 227 +       return -ETIMEDOUT;
 228 +}
 229 +
 230 +static void ps3vram_init_ring(struct ps3_system_bus_device *dev)
 231 +{
 232 +       struct ps3vram_priv *priv = dev->core.driver_data;
 233 +
 234 +       priv->ctrl[CTRL_PUT] = FIFO_BASE + FIFO_OFFSET;
 235 +       priv->ctrl[CTRL_GET] = FIFO_BASE + FIFO_OFFSET;
 236 +}
 237 +
 238 +static int ps3vram_wait_ring(struct ps3_system_bus_device *dev,
 239 +                            unsigned int timeout_ms)
 240 +{
 241 +       struct ps3vram_priv *priv = dev->core.driver_data;
 242 +       unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
 243 +
 244 +       do {
 245 +               if (priv->ctrl[CTRL_PUT] == priv->ctrl[CTRL_GET])
 246 +                       return 0;
 247 +               msleep(1);
 248 +       } while (time_before(jiffies, timeout));
 249 +
 250 +       dev_warn(&dev->core, "FIFO timeout (%08x/%08x/%08x)\n",
 251 +                priv->ctrl[CTRL_PUT], priv->ctrl[CTRL_GET],
 252 +                priv->ctrl[CTRL_TOP]);
 253 +
 254 +       return -ETIMEDOUT;
 255 +}
 256 +
 257 +static void ps3vram_out_ring(struct ps3vram_priv *priv, u32 data)
 258 +{
 259 +       *(priv->fifo_ptr)++ = data;
 260 +}
 261 +
 262 +static void ps3vram_begin_ring(struct ps3vram_priv *priv, u32 chan, u32 tag,
 263 +                              u32 size)
 264 +{
 265 +       ps3vram_out_ring(priv, (size << 18) | (chan << 13) | tag);
 266 +}
 267 +
 268 +static void ps3vram_rewind_ring(struct ps3_system_bus_device *dev)
 269 +{
 270 +       struct ps3vram_priv *priv = dev->core.driver_data;
 271 +       int status;
 272 +
 273 +       ps3vram_out_ring(priv, 0x20000000 | (FIFO_BASE + FIFO_OFFSET));
 274 +
 275 +       priv->ctrl[CTRL_PUT] = FIFO_BASE + FIFO_OFFSET;
 276 +
 277 +       /* asking the HV for a blit will kick the FIFO */
 278 +       status = lv1_gpu_context_attribute(priv->context_handle,
 279 +                                          L1GPU_CONTEXT_ATTRIBUTE_FB_BLIT, 0,
 280 +                                          0, 0, 0);
 281 +       if (status)
 282 +               dev_err(&dev->core,
 283 +                       "%s: lv1_gpu_context_attribute failed %d\n", __func__,
 284 +                       status);
 285 +
 286 +       priv->fifo_ptr = priv->fifo_base;
 287 +}
 288 +
 289 +static void ps3vram_fire_ring(struct ps3_system_bus_device *dev)
 290 +{
 291 +       struct ps3vram_priv *priv = dev->core.driver_data;
 292 +       int status;
 293 +
 294 +       mutex_lock(&ps3_gpu_mutex);
 295 +
 296 +       priv->ctrl[CTRL_PUT] = FIFO_BASE + FIFO_OFFSET +
 297 +                              (priv->fifo_ptr - priv->fifo_base) * sizeof(u32);
 298 +
 299 +       /* asking the HV for a blit will kick the FIFO */
 300 +       status = lv1_gpu_context_attribute(priv->context_handle,
 301 +                                          L1GPU_CONTEXT_ATTRIBUTE_FB_BLIT, 0,
 302 +                                          0, 0, 0);
 303 +       if (status)
 304 +               dev_err(&dev->core,
 305 +                       "%s: lv1_gpu_context_attribute failed %d\n", __func__,
 306 +                       status);
 307 +
 308 +       if ((priv->fifo_ptr - priv->fifo_base) * sizeof(u32) >
 309 +           FIFO_SIZE - 1024) {
 310 +               dev_dbg(&dev->core, "FIFO full, rewinding\n");
 311 +               ps3vram_wait_ring(dev, 200);
 312 +               ps3vram_rewind_ring(dev);
 313 +       }
 314 +
 315 +       mutex_unlock(&ps3_gpu_mutex);
 316 +}
 317 +
 318 +static void ps3vram_bind(struct ps3_system_bus_device *dev)
 319 +{
 320 +       struct ps3vram_priv *priv = dev->core.driver_data;
 321 +
 322 +       ps3vram_begin_ring(priv, UPLOAD_SUBCH, 0, 1);
 323 +       ps3vram_out_ring(priv, 0x31337303);
 324 +       ps3vram_begin_ring(priv, UPLOAD_SUBCH, 0x180, 3);
 325 +       ps3vram_out_ring(priv, DMA_NOTIFIER_HANDLE_BASE + NOTIFIER);
 326 +       ps3vram_out_ring(priv, 0xfeed0001);     /* DMA system RAM instance */
 327 +       ps3vram_out_ring(priv, 0xfeed0000);     /* DMA video RAM instance */
 328 +
 329 +       ps3vram_begin_ring(priv, DOWNLOAD_SUBCH, 0, 1);
 330 +       ps3vram_out_ring(priv, 0x3137c0de);
 331 +       ps3vram_begin_ring(priv, DOWNLOAD_SUBCH, 0x180, 3);
 332 +       ps3vram_out_ring(priv, DMA_NOTIFIER_HANDLE_BASE + NOTIFIER);
 333 +       ps3vram_out_ring(priv, 0xfeed0000);     /* DMA video RAM instance */
 334 +       ps3vram_out_ring(priv, 0xfeed0001);     /* DMA system RAM instance */
 335 +
 336 +       ps3vram_fire_ring(dev);
 337 +}
 338 +
 339 +static int ps3vram_upload(struct ps3_system_bus_device *dev,
 340 +                         unsigned int src_offset, unsigned int dst_offset,
 341 +                         int len, int count)
 342 +{
 343 +       struct ps3vram_priv *priv = dev->core.driver_data;
 344 +
 345 +       ps3vram_begin_ring(priv, UPLOAD_SUBCH,
 346 +                          NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
 347 +       ps3vram_out_ring(priv, XDR_IOIF + src_offset);
 348 +       ps3vram_out_ring(priv, dst_offset);
 349 +       ps3vram_out_ring(priv, len);
 350 +       ps3vram_out_ring(priv, len);
 351 +       ps3vram_out_ring(priv, len);
 352 +       ps3vram_out_ring(priv, count);
 353 +       ps3vram_out_ring(priv, (1 << 8) | 1);
 354 +       ps3vram_out_ring(priv, 0);
 355 +
 356 +       ps3vram_notifier_reset(dev);
 357 +       ps3vram_begin_ring(priv, UPLOAD_SUBCH,
 358 +                          NV_MEMORY_TO_MEMORY_FORMAT_NOTIFY, 1);
 359 +       ps3vram_out_ring(priv, 0);
 360 +       ps3vram_begin_ring(priv, UPLOAD_SUBCH, 0x100, 1);
 361 +       ps3vram_out_ring(priv, 0);
 362 +       ps3vram_fire_ring(dev);
 363 +       if (ps3vram_notifier_wait(dev, 200) < 0) {
 364 +               dev_warn(&dev->core, "%s: Notifier timeout\n", __func__);
 365 +               return -1;
 366 +       }
 367 +
 368 +       return 0;
 369 +}
 370 +
 371 +static int ps3vram_download(struct ps3_system_bus_device *dev,
 372 +                           unsigned int src_offset, unsigned int dst_offset,
 373 +                           int len, int count)
 374 +{
 375 +       struct ps3vram_priv *priv = dev->core.driver_data;
 376 +
 377 +       ps3vram_begin_ring(priv, DOWNLOAD_SUBCH,
 378 +                          NV_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 8);
 379 +       ps3vram_out_ring(priv, src_offset);
 380 +       ps3vram_out_ring(priv, XDR_IOIF + dst_offset);
 381 +       ps3vram_out_ring(priv, len);
 382 +       ps3vram_out_ring(priv, len);
 383 +       ps3vram_out_ring(priv, len);
 384 +       ps3vram_out_ring(priv, count);
 385 +       ps3vram_out_ring(priv, (1 << 8) | 1);
 386 +       ps3vram_out_ring(priv, 0);
 387 +
 388 +       ps3vram_notifier_reset(dev);
 389 +       ps3vram_begin_ring(priv, DOWNLOAD_SUBCH,
 390 +                          NV_MEMORY_TO_MEMORY_FORMAT_NOTIFY, 1);
 391 +       ps3vram_out_ring(priv, 0);
 392 +       ps3vram_begin_ring(priv, DOWNLOAD_SUBCH, 0x100, 1);
 393 +       ps3vram_out_ring(priv, 0);
 394 +       ps3vram_fire_ring(dev);
 395 +       if (ps3vram_notifier_wait(dev, 200) < 0) {
 396 +               dev_warn(&dev->core, "%s: Notifier timeout\n", __func__);
 397 +               return -1;
 398 +       }
 399 +
 400 +       return 0;
 401 +}
 402 +
 403 +static void ps3vram_cache_evict(struct ps3_system_bus_device *dev, int entry)
 404 +{
 405 +       struct ps3vram_priv *priv = dev->core.driver_data;
 406 +       struct ps3vram_cache *cache = &priv->cache;
 407 +
 408 +       if (!(cache->tags[entry].flags & CACHE_PAGE_DIRTY))
 409 +               return;
 410 +
 411 +       dev_dbg(&dev->core, "Flushing %d: 0x%08x\n", entry,
 412 +               cache->tags[entry].address);
 413 +       if (ps3vram_upload(dev, CACHE_OFFSET + entry * cache->page_size,
 414 +                          cache->tags[entry].address, DMA_PAGE_SIZE,
 415 +                          cache->page_size / DMA_PAGE_SIZE) < 0) {
 416 +               dev_err(&dev->core,
 417 +                       "Failed to upload from 0x%x to " "0x%x size 0x%x\n",
 418 +                       entry * cache->page_size, cache->tags[entry].address,
 419 +                       cache->page_size);
 420 +       }
 421 +       cache->tags[entry].flags &= ~CACHE_PAGE_DIRTY;
 422 +}
 423 +
 424 +static void ps3vram_cache_load(struct ps3_system_bus_device *dev, int entry,
 425 +                              unsigned int address)
 426 +{
 427 +       struct ps3vram_priv *priv = dev->core.driver_data;
 428 +       struct ps3vram_cache *cache = &priv->cache;
 429 +
 430 +       dev_dbg(&dev->core, "Fetching %d: 0x%08x\n", entry, address);
 431 +       if (ps3vram_download(dev, address,
 432 +                            CACHE_OFFSET + entry * cache->page_size,
 433 +                            DMA_PAGE_SIZE,
 434 +                            cache->page_size / DMA_PAGE_SIZE) < 0) {
 435 +               dev_err(&dev->core,
 436 +                       "Failed to download from 0x%x to 0x%x size 0x%x\n",
 437 +                       address, entry * cache->page_size, cache->page_size);
 438 +       }
 439 +
 440 +       cache->tags[entry].address = address;
 441 +       cache->tags[entry].flags |= CACHE_PAGE_PRESENT;
 442 +}
 443 +
 444 +
 445 +static void ps3vram_cache_flush(struct ps3_system_bus_device *dev)
 446 +{
 447 +       struct ps3vram_priv *priv = dev->core.driver_data;
 448 +       struct ps3vram_cache *cache = &priv->cache;
 449 +       int i;
 450 +
 451 +       dev_dbg(&dev->core, "FLUSH\n");
 452 +       for (i = 0; i < cache->page_count; i++) {
 453 +               ps3vram_cache_evict(dev, i);
 454 +               cache->tags[i].flags = 0;
 455 +       }
 456 +}
 457 +
 458 +static unsigned int ps3vram_cache_match(struct ps3_system_bus_device *dev,
 459 +                                       loff_t address)
 460 +{
 461 +       struct ps3vram_priv *priv = dev->core.driver_data;
 462 +       struct ps3vram_cache *cache = &priv->cache;
 463 +       unsigned int base;
 464 +       unsigned int offset;
 465 +       int i;
 466 +       static int counter;
 467 +
 468 +       offset = (unsigned int) (address & (cache->page_size - 1));
 469 +       base = (unsigned int) (address - offset);
 470 +
 471 +       /* fully associative check */
 472 +       for (i = 0; i < cache->page_count; i++) {
 473 +               if ((cache->tags[i].flags & CACHE_PAGE_PRESENT) &&
 474 +                   cache->tags[i].address == base) {
 475 +                       cache->hit++;
 476 +                       dev_dbg(&dev->core, "Found entry %d: 0x%08x\n", i,
 477 +                               cache->tags[i].address);
 478 +                       return i;
 479 +               }
 480 +       }
 481 +
 482 +       /* choose a random entry */
 483 +       i = (jiffies + (counter++)) % cache->page_count;
 484 +       dev_dbg(&dev->core, "Using entry %d\n", i);
 485 +
 486 +       ps3vram_cache_evict(dev, i);
 487 +       ps3vram_cache_load(dev, i, base);
 488 +
 489 +       cache->miss++;
 490 +       return i;
 491 +}
 492 +
 493 +static int ps3vram_cache_init(struct ps3_system_bus_device *dev)
 494 +{
 495 +       struct ps3vram_priv *priv = dev->core.driver_data;
 496 +
 497 +       priv->cache.page_count = CACHE_PAGE_COUNT;
 498 +       priv->cache.page_size = CACHE_PAGE_SIZE;
 499 +       priv->cache.tags = kzalloc(sizeof(struct ps3vram_tag) *
 500 +                                  CACHE_PAGE_COUNT, GFP_KERNEL);
 501 +       if (priv->cache.tags == NULL) {
 502 +               dev_err(&dev->core, "Could not allocate cache tags\n");
 503 +               return -ENOMEM;
 504 +       }
 505 +
 506 +       dev_info(&dev->core, "Created ram cache: %d entries, %d KiB each\n",
 507 +               CACHE_PAGE_COUNT, CACHE_PAGE_SIZE / 1024);
 508 +
 509 +       return 0;
 510 +}
 511 +
 512 +static void ps3vram_cache_cleanup(struct ps3_system_bus_device *dev)
 513 +{
 514 +       struct ps3vram_priv *priv = dev->core.driver_data;
 515 +
 516 +       ps3vram_cache_flush(dev);
 517 +       kfree(priv->cache.tags);
 518 +}
 519 +
 520 +static int ps3vram_read(struct ps3_system_bus_device *dev, loff_t from,
 521 +                       size_t len, size_t *retlen, u_char *buf)
 522 +{
 523 +       struct ps3vram_priv *priv = dev->core.driver_data;
 524 +       unsigned int cached, count;
 525 +
 526 +       dev_dbg(&dev->core, "%s: from=0x%08x len=0x%zx\n", __func__,
 527 +               (unsigned int)from, len);
 528 +
 529 +       if (from >= priv->size)
 530 +               return -EIO;
 531 +
 532 +       if (len > priv->size - from)
 533 +               len = priv->size - from;
 534 +
 535 +       /* Copy from vram to buf */
 536 +       count = len;
 537 +       while (count) {
 538 +               unsigned int offset, avail;
 539 +               unsigned int entry;
 540 +
 541 +               offset = (unsigned int) (from & (priv->cache.page_size - 1));
 542 +               avail  = priv->cache.page_size - offset;
 543 +
 544 +               mutex_lock(&priv->lock);
 545 +
 546 +               entry = ps3vram_cache_match(dev, from);
 547 +               cached = CACHE_OFFSET + entry * priv->cache.page_size + offset;
 548 +
 549 +               dev_dbg(&dev->core, "%s: from=%08x cached=%08x offset=%08x "
 550 +                       "avail=%08x count=%08x\n", __func__,
 551 +                       (unsigned int)from, cached, offset, avail, count);
 552 +
 553 +               if (avail > count)
 554 +                       avail = count;
 555 +               memcpy(buf, priv->xdr_buf + cached, avail);
 556 +
 557 +               mutex_unlock(&priv->lock);
 558 +
 559 +               buf += avail;
 560 +               count -= avail;
 561 +               from += avail;
 562 +       }
 563 +
 564 +       *retlen = len;
 565 +       return 0;
 566 +}
 567 +
 568 +static int ps3vram_write(struct ps3_system_bus_device *dev, loff_t to,
 569 +                        size_t len, size_t *retlen, const u_char *buf)
 570 +{
 571 +       struct ps3vram_priv *priv = dev->core.driver_data;
 572 +       unsigned int cached, count;
 573 +
 574 +       if (to >= priv->size)
 575 +               return -EIO;
 576 +
 577 +       if (len > priv->size - to)
 578 +               len = priv->size - to;
 579 +
 580 +       /* Copy from buf to vram */
 581 +       count = len;
 582 +       while (count) {
 583 +               unsigned int offset, avail;
 584 +               unsigned int entry;
 585 +
 586 +               offset = (unsigned int) (to & (priv->cache.page_size - 1));
 587 +               avail  = priv->cache.page_size - offset;
 588 +
 589 +               mutex_lock(&priv->lock);
 590 +
 591 +               entry = ps3vram_cache_match(dev, to);
 592 +               cached = CACHE_OFFSET + entry * priv->cache.page_size + offset;
 593 +
 594 +               dev_dbg(&dev->core, "%s: to=%08x cached=%08x offset=%08x "
 595 +                       "avail=%08x count=%08x\n", __func__, (unsigned int)to,
 596 +                       cached, offset, avail, count);
 597 +
 598 +               if (avail > count)
 599 +                       avail = count;
 600 +               memcpy(priv->xdr_buf + cached, buf, avail);
 601 +
 602 +               priv->cache.tags[entry].flags |= CACHE_PAGE_DIRTY;
 603 +
 604 +               mutex_unlock(&priv->lock);
 605 +
 606 +               buf += avail;
 607 +               count -= avail;
 608 +               to += avail;
 609 +       }
 610 +
 611 +       *retlen = len;
 612 +       return 0;
 613 +}
 614 +
 615 +static int ps3vram_proc_show(struct seq_file *m, void *v)
 616 +{
 617 +       struct ps3vram_priv *priv = m->private;
 618 +
 619 +       seq_printf(m, "hit:%u\nmiss:%u\n", priv->cache.hit, priv->cache.miss);
 620 +       return 0;
 621 +}
 622 +
 623 +static int ps3vram_proc_open(struct inode *inode, struct file *file)
 624 +{
 625 +       return single_open(file, ps3vram_proc_show, PDE(inode)->data);
 626 +}
 627 +
 628 +static const struct file_operations ps3vram_proc_fops = {
 629 +       .owner          = THIS_MODULE,
 630 +       .open           = ps3vram_proc_open,
 631 +       .read           = seq_read,
 632 +       .llseek         = seq_lseek,
 633 +       .release        = single_release,
 634 +};
 635 +
 636 +static void __devinit ps3vram_proc_init(struct ps3_system_bus_device *dev)
 637 +{
 638 +       struct ps3vram_priv *priv = dev->core.driver_data;
 639 +       struct proc_dir_entry *pde;
 640 +
 641 +       pde = proc_create(DEVICE_NAME, 0444, NULL, &ps3vram_proc_fops);
 642 +       if (!pde) {
 643 +               dev_warn(&dev->core, "failed to create /proc entry\n");
 644 +               return;
 645 +       }
 646 +
 647 +       pde->owner = THIS_MODULE;
 648 +       pde->data = priv;
 649 +}
 650 +
 651 +static int ps3vram_make_request(struct request_queue *q, struct bio *bio)
 652 +{
 653 +       struct ps3_system_bus_device *dev = q->queuedata;
 654 +       int write = bio_data_dir(bio) == WRITE;
 655 +       const char *op = write ? "write" : "read";
 656 +       loff_t offset = bio->bi_sector << 9;
 657 +       int error = 0;
 658 +       struct bio_vec *bvec;
 659 +       unsigned int i;
 660 +
 661 +       dev_dbg(&dev->core, "%s\n", __func__);
 662 +
 663 +       bio_for_each_segment(bvec, bio, i) {
 664 +               /* PS3 is ppc64, so we don't handle highmem */
 665 +               char *ptr = page_address(bvec->bv_page) + bvec->bv_offset;
 666 +               size_t len = bvec->bv_len, retlen;
 667 +
 668 +               dev_dbg(&dev->core, "    %s %zu bytes at offset %llu\n", op,
 669 +                       len, offset);
 670 +               if (write)
 671 +                       error = ps3vram_write(dev, offset, len, &retlen, ptr);
 672 +               else
 673 +                       error = ps3vram_read(dev, offset, len, &retlen, ptr);
 674 +
 675 +               if (error) {
 676 +                       dev_err(&dev->core, "%s failed\n", op);
 677 +                       goto out;
 678 +               }
 679 +
 680 +               if (retlen != len) {
 681 +                       dev_err(&dev->core, "Short %s\n", op);
 682 +                       goto out;
 683 +               }
 684 +
 685 +               offset += len;
 686 +       }
 687 +
 688 +       dev_dbg(&dev->core, "%s completed\n", op);
 689 +
 690 +out:
 691 +       bio_endio(bio, error);
 692 +       return 0;
 693 +}
 694 +
 695 +static int __devinit ps3vram_probe(struct ps3_system_bus_device *dev)
 696 +{
 697 +       struct ps3vram_priv *priv;
 698 +       int error, status;
 699 +       struct request_queue *queue;
 700 +       struct gendisk *gendisk;
 701 +       u64 ddr_lpar, ctrl_lpar, info_lpar, reports_lpar, ddr_size,
 702 +           reports_size;
 703 +       char *rest;
 704 +
 705 +       priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 706 +       if (!priv) {
 707 +               error = -ENOMEM;
 708 +               goto fail;
 709 +       }
 710 +
 711 +       mutex_init(&priv->lock);
 712 +       dev->core.driver_data = priv;
 713 +
 714 +       priv = dev->core.driver_data;
 715 +
 716 +       /* Allocate XDR buffer (1MiB aligned) */
 717 +       priv->xdr_buf = (void *)__get_free_pages(GFP_KERNEL,
 718 +               get_order(XDR_BUF_SIZE));
 719 +       if (priv->xdr_buf == NULL) {
 720 +               dev_err(&dev->core, "Could not allocate XDR buffer\n");
 721 +               error = -ENOMEM;
 722 +               goto fail_free_priv;
 723 +       }
 724 +
 725 +       /* Put FIFO at begginning of XDR buffer */
 726 +       priv->fifo_base = (u32 *) (priv->xdr_buf + FIFO_OFFSET);
 727 +       priv->fifo_ptr = priv->fifo_base;
 728 +
 729 +       /* XXX: Need to open GPU, in case ps3fb or snd_ps3 aren't loaded */
 730 +       if (ps3_open_hv_device(dev)) {
 731 +               dev_err(&dev->core, "ps3_open_hv_device failed\n");
 732 +               error = -EAGAIN;
 733 +               goto out_close_gpu;
 734 +       }
 735 +
 736 +       /* Request memory */
 737 +       status = -1;
 738 +       ddr_size = ALIGN(memparse(size, &rest), 1024*1024);
 739 +       if (!ddr_size) {
 740 +               dev_err(&dev->core, "Specified size is too small\n");
 741 +               error = -EINVAL;
 742 +               goto out_close_gpu;
 743 +       }
 744 +
 745 +       while (ddr_size > 0) {
 746 +               status = lv1_gpu_memory_allocate(ddr_size, 0, 0, 0, 0,
 747 +                                                &priv->memory_handle,
 748 +                                                &ddr_lpar);
 749 +               if (!status)
 750 +                       break;
 751 +               ddr_size -= 1024*1024;
 752 +       }
 753 +       if (status) {
 754 +               dev_err(&dev->core, "lv1_gpu_memory_allocate failed %d\n",
 755 +                       status);
 756 +               error = -ENOMEM;
 757 +               goto out_free_xdr_buf;
 758 +       }
 759 +
 760 +       /* Request context */
 761 +       status = lv1_gpu_context_allocate(priv->memory_handle, 0,
 762 +                                         &priv->context_handle, &ctrl_lpar,
 763 +                                         &info_lpar, &reports_lpar,
 764 +                                         &reports_size);
 765 +       if (status) {
 766 +               dev_err(&dev->core, "lv1_gpu_context_allocate failed %d\n",
 767 +                       status);
 768 +               error = -ENOMEM;
 769 +               goto out_free_memory;
 770 +       }
 771 +
 772 +       /* Map XDR buffer to RSX */
 773 +       status = lv1_gpu_context_iomap(priv->context_handle, XDR_IOIF,
 774 +                                      ps3_mm_phys_to_lpar(__pa(priv->xdr_buf)),
 775 +                                      XDR_BUF_SIZE, 0);
 776 +       if (status) {
 777 +               dev_err(&dev->core, "lv1_gpu_context_iomap failed %d\n",
 778 +                       status);
 779 +               error = -ENOMEM;
 780 +               goto out_free_context;
 781 +       }
 782 +
 783 +       priv->ddr_base = ioremap_flags(ddr_lpar, ddr_size, _PAGE_NO_CACHE);
 784 +
 785 +       if (!priv->ddr_base) {
 786 +               dev_err(&dev->core, "ioremap DDR failed\n");
 787 +               error = -ENOMEM;
 788 +               goto out_free_context;
 789 +       }
 790 +
 791 +       priv->ctrl = ioremap(ctrl_lpar, 64 * 1024);
 792 +       if (!priv->ctrl) {
 793 +               dev_err(&dev->core, "ioremap CTRL failed\n");
 794 +               error = -ENOMEM;
 795 +               goto out_unmap_vram;
 796 +       }
 797 +
 798 +       priv->reports = ioremap(reports_lpar, reports_size);
 799 +       if (!priv->reports) {
 800 +               dev_err(&dev->core, "ioremap REPORTS failed\n");
 801 +               error = -ENOMEM;
 802 +               goto out_unmap_ctrl;
 803 +       }
 804 +
 805 +       mutex_lock(&ps3_gpu_mutex);
 806 +       ps3vram_init_ring(dev);
 807 +       mutex_unlock(&ps3_gpu_mutex);
 808 +
 809 +       priv->size = ddr_size;
 810 +
 811 +       ps3vram_bind(dev);
 812 +
 813 +       mutex_lock(&ps3_gpu_mutex);
 814 +       error = ps3vram_wait_ring(dev, 100);
 815 +       mutex_unlock(&ps3_gpu_mutex);
 816 +       if (error < 0) {
 817 +               dev_err(&dev->core, "Failed to initialize channels\n");
 818 +               error = -ETIMEDOUT;
 819 +               goto out_unmap_reports;
 820 +       }
 821 +
 822 +       ps3vram_cache_init(dev);
 823 +       ps3vram_proc_init(dev);
 824 +
 825 +       queue = blk_alloc_queue(GFP_KERNEL);
 826 +       if (!queue) {
 827 +               dev_err(&dev->core, "blk_alloc_queue failed\n");
 828 +               error = -ENOMEM;
 829 +               goto out_cache_cleanup;
 830 +       }
 831 +
 832 +       priv->queue = queue;
 833 +       queue->queuedata = dev;
 834 +       blk_queue_make_request(queue, ps3vram_make_request);
 835 +       blk_queue_max_phys_segments(queue, MAX_PHYS_SEGMENTS);
 836 +       blk_queue_max_hw_segments(queue, MAX_HW_SEGMENTS);
 837 +       blk_queue_max_segment_size(queue, MAX_SEGMENT_SIZE);
 838 +       blk_queue_max_sectors(queue, SAFE_MAX_SECTORS);
 839 +
 840 +       gendisk = alloc_disk(1);
 841 +       if (!gendisk) {
 842 +               dev_err(&dev->core, "alloc_disk failed\n");
 843 +               error = -ENOMEM;
 844 +               goto fail_cleanup_queue;
 845 +       }
 846 +
 847 +       priv->gendisk = gendisk;
 848 +       gendisk->major = ps3vram_major;
 849 +       gendisk->first_minor = 0;
 850 +       gendisk->fops = &ps3vram_fops;
 851 +       gendisk->queue = queue;
 852 +       gendisk->private_data = dev;
 853 +       gendisk->driverfs_dev = &dev->core;
 854 +       strlcpy(gendisk->disk_name, DEVICE_NAME, sizeof(gendisk->disk_name));
 855 +       set_capacity(gendisk, priv->size >> 9);
 856 +
 857 +       dev_info(&dev->core, "%s: Using %lu MiB of GPU memory\n",
 858 +                gendisk->disk_name, get_capacity(gendisk) >> 11);
 859 +
 860 +       add_disk(gendisk);
 861 +       return 0;
 862 +
 863 +fail_cleanup_queue:
 864 +       blk_cleanup_queue(queue);
 865 +out_cache_cleanup:
 866 +       remove_proc_entry(DEVICE_NAME, NULL);
 867 +       ps3vram_cache_cleanup(dev);
 868 +out_unmap_reports:
 869 +       iounmap(priv->reports);
 870 +out_unmap_ctrl:
 871 +       iounmap(priv->ctrl);
 872 +out_unmap_vram:
 873 +       iounmap(priv->ddr_base);
 874 +out_free_context:
 875 +       lv1_gpu_context_free(priv->context_handle);
 876 +out_free_memory:
 877 +       lv1_gpu_memory_free(priv->memory_handle);
 878 +out_close_gpu:
 879 +       ps3_close_hv_device(dev);
 880 +out_free_xdr_buf:
 881 +       free_pages((unsigned long) priv->xdr_buf, get_order(XDR_BUF_SIZE));
 882 +fail_free_priv:
 883 +       kfree(priv);
 884 +       dev->core.driver_data = NULL;
 885 +fail:
 886 +       return error;
 887 +}
 888 +
 889 +static int ps3vram_remove(struct ps3_system_bus_device *dev)
 890 +{
 891 +       struct ps3vram_priv *priv = dev->core.driver_data;
 892 +
 893 +       del_gendisk(priv->gendisk);
 894 +       put_disk(priv->gendisk);
 895 +       blk_cleanup_queue(priv->queue);
 896 +       remove_proc_entry(DEVICE_NAME, NULL);
 897 +       ps3vram_cache_cleanup(dev);
 898 +       iounmap(priv->reports);
 899 +       iounmap(priv->ctrl);
 900 +       iounmap(priv->ddr_base);
 901 +       lv1_gpu_context_free(priv->context_handle);
 902 +       lv1_gpu_memory_free(priv->memory_handle);
 903 +       ps3_close_hv_device(dev);
 904 +       free_pages((unsigned long) priv->xdr_buf, get_order(XDR_BUF_SIZE));
 905 +       kfree(priv);
 906 +       dev->core.driver_data = NULL;
 907 +       return 0;
 908 +}
 909 +
 910 +static struct ps3_system_bus_driver ps3vram = {
 911 +       .match_id       = PS3_MATCH_ID_GPU,
 912 +       .match_sub_id   = PS3_MATCH_SUB_ID_GPU_RAMDISK,
 913 +       .core.name      = DEVICE_NAME,
 914 +       .core.owner     = THIS_MODULE,
 915 +       .probe          = ps3vram_probe,
 916 +       .remove         = ps3vram_remove,
 917 +       .shutdown       = ps3vram_remove,
 918 +};
 919 +
 920 +
 921 +static int __init ps3vram_init(void)
 922 +{
 923 +       int error;
 924 +
 925 +       if (!firmware_has_feature(FW_FEATURE_PS3_LV1))
 926 +               return -ENODEV;
 927 +
 928 +       error = register_blkdev(0, DEVICE_NAME);
 929 +       if (error <= 0) {
 930 +               pr_err("%s: register_blkdev failed %d\n", DEVICE_NAME, error);
 931 +               return error;
 932 +       }
 933 +       ps3vram_major = error;
 934 +
 935 +       pr_info("%s: registered block device major %d\n", DEVICE_NAME,
 936 +               ps3vram_major);
 937 +
 938 +       error = ps3_system_bus_driver_register(&ps3vram);
 939 +       if (error)
 940 +               unregister_blkdev(ps3vram_major, DEVICE_NAME);
 941 +
 942 +       return error;
 943 +}
 944 +
 945 +static void __exit ps3vram_exit(void)
 946 +{
 947 +       ps3_system_bus_driver_unregister(&ps3vram);
 948 +       unregister_blkdev(ps3vram_major, DEVICE_NAME);
 949 +}
 950 +
 951 +module_init(ps3vram_init);
 952 +module_exit(ps3vram_exit);
 953 +
 954 +MODULE_LICENSE("GPL");
 955 +MODULE_DESCRIPTION("PS3 Video RAM Storage Driver");
 956 +MODULE_AUTHOR("Sony Corporation");
 957 +MODULE_ALIAS(PS3_MODULE_ALIAS_GPU_RAMDISK);