drivers/gpu/drm/radeon/si.c

   1 /*
   2  * Copyright 2011 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Authors: Alex Deucher
  23  */
  24 #include <linux/firmware.h>
  25 #include <linux/platform_device.h>
  26 #include <linux/slab.h>
  27 #include <linux/module.h>
  28 #include <drm/drmP.h>
  29 #include "radeon.h"
  30 #include "radeon_asic.h"
  31 #include <drm/radeon_drm.h>
  32 #include "sid.h"
  33 #include "atom.h"
  34 #include "si_blit_shaders.h"
  35
  36 #define SI_PFP_UCODE_SIZE 2144
  37 #define SI_PM4_UCODE_SIZE 2144
  38 #define SI_CE_UCODE_SIZE 2144
  39 #define SI_RLC_UCODE_SIZE 2048
  40 #define SI_MC_UCODE_SIZE 7769
  41
  42 MODULE_FIRMWARE("radeon/TAHITI_pfp.bin");
  43 MODULE_FIRMWARE("radeon/TAHITI_me.bin");
  44 MODULE_FIRMWARE("radeon/TAHITI_ce.bin");
  45 MODULE_FIRMWARE("radeon/TAHITI_mc.bin");
  46 MODULE_FIRMWARE("radeon/TAHITI_rlc.bin");
  47 MODULE_FIRMWARE("radeon/PITCAIRN_pfp.bin");
  48 MODULE_FIRMWARE("radeon/PITCAIRN_me.bin");
  49 MODULE_FIRMWARE("radeon/PITCAIRN_ce.bin");
  50 MODULE_FIRMWARE("radeon/PITCAIRN_mc.bin");
  51 MODULE_FIRMWARE("radeon/PITCAIRN_rlc.bin");
  52 MODULE_FIRMWARE("radeon/VERDE_pfp.bin");
  53 MODULE_FIRMWARE("radeon/VERDE_me.bin");
  54 MODULE_FIRMWARE("radeon/VERDE_ce.bin");
  55 MODULE_FIRMWARE("radeon/VERDE_mc.bin");
  56 MODULE_FIRMWARE("radeon/VERDE_rlc.bin");
  57
  58 extern int r600_ih_ring_alloc(struct radeon_device *rdev);
  59 extern void r600_ih_ring_fini(struct radeon_device *rdev);
  60 extern void evergreen_fix_pci_max_read_req_size(struct radeon_device *rdev);
  61 extern void evergreen_mc_stop(struct radeon_device *rdev, struct evergreen_mc_save *save);
  62 extern void evergreen_mc_resume(struct radeon_device *rdev, struct evergreen_mc_save *save);
  63 extern u32 evergreen_get_number_of_dram_channels(struct radeon_device *rdev);
  64 extern void evergreen_print_gpu_status_regs(struct radeon_device *rdev);
  65 extern bool evergreen_is_display_hung(struct radeon_device *rdev);
  66
  67 /* get temperature in millidegrees */
  68 int si_get_temp(struct radeon_device *rdev)
  69 {
  70         u32 temp;
  71         int actual_temp = 0;
  72
  73         temp = (RREG32(CG_MULT_THERMAL_STATUS) & CTF_TEMP_MASK) >>
  74                 CTF_TEMP_SHIFT;
  75
  76         if (temp & 0x200)
  77                 actual_temp = 255;
  78         else
  79                 actual_temp = temp & 0x1ff;
  80
  81         actual_temp = (actual_temp * 1000);
  82
  83         return actual_temp;
  84 }
  85
  86 #define TAHITI_IO_MC_REGS_SIZE 36
  87
  88 static const u32 tahiti_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
  89         {0x0000006f, 0x03044000},
  90         {0x00000070, 0x0480c018},
  91         {0x00000071, 0x00000040},
  92         {0x00000072, 0x01000000},
  93         {0x00000074, 0x000000ff},
  94         {0x00000075, 0x00143400},
  95         {0x00000076, 0x08ec0800},
  96         {0x00000077, 0x040000cc},
  97         {0x00000079, 0x00000000},
  98         {0x0000007a, 0x21000409},
  99         {0x0000007c, 0x00000000},
 100         {0x0000007d, 0xe8000000},
 101         {0x0000007e, 0x044408a8},
 102         {0x0000007f, 0x00000003},
 103         {0x00000080, 0x00000000},
 104         {0x00000081, 0x01000000},
 105         {0x00000082, 0x02000000},
 106         {0x00000083, 0x00000000},
 107         {0x00000084, 0xe3f3e4f4},
 108         {0x00000085, 0x00052024},
 109         {0x00000087, 0x00000000},
 110         {0x00000088, 0x66036603},
 111         {0x00000089, 0x01000000},
 112         {0x0000008b, 0x1c0a0000},
 113         {0x0000008c, 0xff010000},
 114         {0x0000008e, 0xffffefff},
 115         {0x0000008f, 0xfff3efff},
 116         {0x00000090, 0xfff3efbf},
 117         {0x00000094, 0x00101101},
 118         {0x00000095, 0x00000fff},
 119         {0x00000096, 0x00116fff},
 120         {0x00000097, 0x60010000},
 121         {0x00000098, 0x10010000},
 122         {0x00000099, 0x00006000},
 123         {0x0000009a, 0x00001000},
 124         {0x0000009f, 0x00a77400}
 125 };
 126
 127 static const u32 pitcairn_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
 128         {0x0000006f, 0x03044000},
 129         {0x00000070, 0x0480c018},
 130         {0x00000071, 0x00000040},
 131         {0x00000072, 0x01000000},
 132         {0x00000074, 0x000000ff},
 133         {0x00000075, 0x00143400},
 134         {0x00000076, 0x08ec0800},
 135         {0x00000077, 0x040000cc},
 136         {0x00000079, 0x00000000},
 137         {0x0000007a, 0x21000409},
 138         {0x0000007c, 0x00000000},
 139         {0x0000007d, 0xe8000000},
 140         {0x0000007e, 0x044408a8},
 141         {0x0000007f, 0x00000003},
 142         {0x00000080, 0x00000000},
 143         {0x00000081, 0x01000000},
 144         {0x00000082, 0x02000000},
 145         {0x00000083, 0x00000000},
 146         {0x00000084, 0xe3f3e4f4},
 147         {0x00000085, 0x00052024},
 148         {0x00000087, 0x00000000},
 149         {0x00000088, 0x66036603},
 150         {0x00000089, 0x01000000},
 151         {0x0000008b, 0x1c0a0000},
 152         {0x0000008c, 0xff010000},
 153         {0x0000008e, 0xffffefff},
 154         {0x0000008f, 0xfff3efff},
 155         {0x00000090, 0xfff3efbf},
 156         {0x00000094, 0x00101101},
 157         {0x00000095, 0x00000fff},
 158         {0x00000096, 0x00116fff},
 159         {0x00000097, 0x60010000},
 160         {0x00000098, 0x10010000},
 161         {0x00000099, 0x00006000},
 162         {0x0000009a, 0x00001000},
 163         {0x0000009f, 0x00a47400}
 164 };
 165
 166 static const u32 verde_io_mc_regs[TAHITI_IO_MC_REGS_SIZE][2] = {
 167         {0x0000006f, 0x03044000},
 168         {0x00000070, 0x0480c018},
 169         {0x00000071, 0x00000040},
 170         {0x00000072, 0x01000000},
 171         {0x00000074, 0x000000ff},
 172         {0x00000075, 0x00143400},
 173         {0x00000076, 0x08ec0800},
 174         {0x00000077, 0x040000cc},
 175         {0x00000079, 0x00000000},
 176         {0x0000007a, 0x21000409},
 177         {0x0000007c, 0x00000000},
 178         {0x0000007d, 0xe8000000},
 179         {0x0000007e, 0x044408a8},
 180         {0x0000007f, 0x00000003},
 181         {0x00000080, 0x00000000},
 182         {0x00000081, 0x01000000},
 183         {0x00000082, 0x02000000},
 184         {0x00000083, 0x00000000},
 185         {0x00000084, 0xe3f3e4f4},
 186         {0x00000085, 0x00052024},
 187         {0x00000087, 0x00000000},
 188         {0x00000088, 0x66036603},
 189         {0x00000089, 0x01000000},
 190         {0x0000008b, 0x1c0a0000},
 191         {0x0000008c, 0xff010000},
 192         {0x0000008e, 0xffffefff},
 193         {0x0000008f, 0xfff3efff},
 194         {0x00000090, 0xfff3efbf},
 195         {0x00000094, 0x00101101},
 196         {0x00000095, 0x00000fff},
 197         {0x00000096, 0x00116fff},
 198         {0x00000097, 0x60010000},
 199         {0x00000098, 0x10010000},
 200         {0x00000099, 0x00006000},
 201         {0x0000009a, 0x00001000},
 202         {0x0000009f, 0x00a37400}
 203 };
 204
 205 /* ucode loading */
 206 static int si_mc_load_microcode(struct radeon_device *rdev)
 207 {
 208         const __be32 *fw_data;
 209         u32 running, blackout = 0;
 210         u32 *io_mc_regs;
 211         int i, ucode_size, regs_size;
 212
 213         if (!rdev->mc_fw)
 214                 return -EINVAL;
 215
 216         switch (rdev->family) {
 217         case CHIP_TAHITI:
 218                 io_mc_regs = (u32 *)&tahiti_io_mc_regs;
 219                 ucode_size = SI_MC_UCODE_SIZE;
 220                 regs_size = TAHITI_IO_MC_REGS_SIZE;
 221                 break;
 222         case CHIP_PITCAIRN:
 223                 io_mc_regs = (u32 *)&pitcairn_io_mc_regs;
 224                 ucode_size = SI_MC_UCODE_SIZE;
 225                 regs_size = TAHITI_IO_MC_REGS_SIZE;
 226                 break;
 227         case CHIP_VERDE:
 228         default:
 229                 io_mc_regs = (u32 *)&verde_io_mc_regs;
 230                 ucode_size = SI_MC_UCODE_SIZE;
 231                 regs_size = TAHITI_IO_MC_REGS_SIZE;
 232                 break;
 233         }
 234
 235         running = RREG32(MC_SEQ_SUP_CNTL) & RUN_MASK;
 236
 237         if (running == 0) {
 238                 if (running) {
 239                         blackout = RREG32(MC_SHARED_BLACKOUT_CNTL);
 240                         WREG32(MC_SHARED_BLACKOUT_CNTL, blackout | 1);
 241                 }
 242
 243                 /* reset the engine and set to writable */
 244                 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
 245                 WREG32(MC_SEQ_SUP_CNTL, 0x00000010);
 246
 247                 /* load mc io regs */
 248                 for (i = 0; i < regs_size; i++) {
 249                         WREG32(MC_SEQ_IO_DEBUG_INDEX, io_mc_regs[(i << 1)]);
 250                         WREG32(MC_SEQ_IO_DEBUG_DATA, io_mc_regs[(i << 1) + 1]);
 251                 }
 252                 /* load the MC ucode */
 253                 fw_data = (const __be32 *)rdev->mc_fw->data;
 254                 for (i = 0; i < ucode_size; i++)
 255                         WREG32(MC_SEQ_SUP_PGM, be32_to_cpup(fw_data++));
 256
 257                 /* put the engine back into the active state */
 258                 WREG32(MC_SEQ_SUP_CNTL, 0x00000008);
 259                 WREG32(MC_SEQ_SUP_CNTL, 0x00000004);
 260                 WREG32(MC_SEQ_SUP_CNTL, 0x00000001);
 261
 262                 /* wait for training to complete */
 263                 for (i = 0; i < rdev->usec_timeout; i++) {
 264                         if (RREG32(MC_SEQ_TRAIN_WAKEUP_CNTL) & TRAIN_DONE_D0)
 265                                 break;
 266                         udelay(1);
 267                 }
 268                 for (i = 0; i < rdev->usec_timeout; i++) {
 269                         if (RREG32(MC_SEQ_TRAIN_WAKEUP_CNTL) & TRAIN_DONE_D1)
 270                                 break;
 271                         udelay(1);
 272                 }
 273
 274                 if (running)
 275                         WREG32(MC_SHARED_BLACKOUT_CNTL, blackout);
 276         }
 277
 278         return 0;
 279 }
 280
 281 static int si_init_microcode(struct radeon_device *rdev)
 282 {
 283         struct platform_device *pdev;
 284         const char *chip_name;
 285         const char *rlc_chip_name;
 286         size_t pfp_req_size, me_req_size, ce_req_size, rlc_req_size, mc_req_size;
 287         char fw_name[30];
 288         int err;
 289
 290         DRM_DEBUG("\n");
 291
 292         pdev = platform_device_register_simple("radeon_cp", 0, NULL, 0);
 293         err = IS_ERR(pdev);
 294         if (err) {
 295                 printk(KERN_ERR "radeon_cp: Failed to register firmware\n");
 296                 return -EINVAL;
 297         }
 298
 299         switch (rdev->family) {
 300         case CHIP_TAHITI:
 301                 chip_name = "TAHITI";
 302                 rlc_chip_name = "TAHITI";
 303                 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
 304                 me_req_size = SI_PM4_UCODE_SIZE * 4;
 305                 ce_req_size = SI_CE_UCODE_SIZE * 4;
 306                 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
 307                 mc_req_size = SI_MC_UCODE_SIZE * 4;
 308                 break;
 309         case CHIP_PITCAIRN:
 310                 chip_name = "PITCAIRN";
 311                 rlc_chip_name = "PITCAIRN";
 312                 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
 313                 me_req_size = SI_PM4_UCODE_SIZE * 4;
 314                 ce_req_size = SI_CE_UCODE_SIZE * 4;
 315                 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
 316                 mc_req_size = SI_MC_UCODE_SIZE * 4;
 317                 break;
 318         case CHIP_VERDE:
 319                 chip_name = "VERDE";
 320                 rlc_chip_name = "VERDE";
 321                 pfp_req_size = SI_PFP_UCODE_SIZE * 4;
 322                 me_req_size = SI_PM4_UCODE_SIZE * 4;
 323                 ce_req_size = SI_CE_UCODE_SIZE * 4;
 324                 rlc_req_size = SI_RLC_UCODE_SIZE * 4;
 325                 mc_req_size = SI_MC_UCODE_SIZE * 4;
 326                 break;
 327         default: BUG();
 328         }
 329
 330         DRM_INFO("Loading %s Microcode\n", chip_name);
 331
 332         snprintf(fw_name, sizeof(fw_name), "radeon/%s_pfp.bin", chip_name);
 333         err = request_firmware(&rdev->pfp_fw, fw_name, &pdev->dev);
 334         if (err)
 335                 goto out;
 336         if (rdev->pfp_fw->size != pfp_req_size) {
 337                 printk(KERN_ERR
 338                        "si_cp: Bogus length %zu in firmware \"%s\"\n",
 339                        rdev->pfp_fw->size, fw_name);
 340                 err = -EINVAL;
 341                 goto out;
 342         }
 343
 344         snprintf(fw_name, sizeof(fw_name), "radeon/%s_me.bin", chip_name);
 345         err = request_firmware(&rdev->me_fw, fw_name, &pdev->dev);
 346         if (err)
 347                 goto out;
 348         if (rdev->me_fw->size != me_req_size) {
 349                 printk(KERN_ERR
 350                        "si_cp: Bogus length %zu in firmware \"%s\"\n",
 351                        rdev->me_fw->size, fw_name);
 352                 err = -EINVAL;
 353         }
 354
 355         snprintf(fw_name, sizeof(fw_name), "radeon/%s_ce.bin", chip_name);
 356         err = request_firmware(&rdev->ce_fw, fw_name, &pdev->dev);
 357         if (err)
 358                 goto out;
 359         if (rdev->ce_fw->size != ce_req_size) {
 360                 printk(KERN_ERR
 361                        "si_cp: Bogus length %zu in firmware \"%s\"\n",
 362                        rdev->ce_fw->size, fw_name);
 363                 err = -EINVAL;
 364         }
 365
 366         snprintf(fw_name, sizeof(fw_name), "radeon/%s_rlc.bin", rlc_chip_name);
 367         err = request_firmware(&rdev->rlc_fw, fw_name, &pdev->dev);
 368         if (err)
 369                 goto out;
 370         if (rdev->rlc_fw->size != rlc_req_size) {
 371                 printk(KERN_ERR
 372                        "si_rlc: Bogus length %zu in firmware \"%s\"\n",
 373                        rdev->rlc_fw->size, fw_name);
 374                 err = -EINVAL;
 375         }
 376
 377         snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc.bin", chip_name);
 378         err = request_firmware(&rdev->mc_fw, fw_name, &pdev->dev);
 379         if (err)
 380                 goto out;
 381         if (rdev->mc_fw->size != mc_req_size) {
 382                 printk(KERN_ERR
 383                        "si_mc: Bogus length %zu in firmware \"%s\"\n",
 384                        rdev->mc_fw->size, fw_name);
 385                 err = -EINVAL;
 386         }
 387
 388 out:
 389         platform_device_unregister(pdev);
 390
 391         if (err) {
 392                 if (err != -EINVAL)
 393                         printk(KERN_ERR
 394                                "si_cp: Failed to load firmware \"%s\"\n",
 395                                fw_name);
 396                 release_firmware(rdev->pfp_fw);
 397                 rdev->pfp_fw = NULL;
 398                 release_firmware(rdev->me_fw);
 399                 rdev->me_fw = NULL;
 400                 release_firmware(rdev->ce_fw);
 401                 rdev->ce_fw = NULL;
 402                 release_firmware(rdev->rlc_fw);
 403                 rdev->rlc_fw = NULL;
 404                 release_firmware(rdev->mc_fw);
 405                 rdev->mc_fw = NULL;
 406         }
 407         return err;
 408 }
 409
 410 /* watermark setup */
 411 static u32 dce6_line_buffer_adjust(struct radeon_device *rdev,
 412                                    struct radeon_crtc *radeon_crtc,
 413                                    struct drm_display_mode *mode,
 414                                    struct drm_display_mode *other_mode)
 415 {
 416         u32 tmp;
 417         /*
 418          * Line Buffer Setup
 419          * There are 3 line buffers, each one shared by 2 display controllers.
 420          * DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
 421          * the display controllers.  The paritioning is done via one of four
 422          * preset allocations specified in bits 21:20:
 423          *  0 - half lb
 424          *  2 - whole lb, other crtc must be disabled
 425          */
 426         /* this can get tricky if we have two large displays on a paired group
 427          * of crtcs.  Ideally for multiple large displays we'd assign them to
 428          * non-linked crtcs for maximum line buffer allocation.
 429          */
 430         if (radeon_crtc->base.enabled && mode) {
 431                 if (other_mode)
 432                         tmp = 0; /* 1/2 */
 433                 else
 434                         tmp = 2; /* whole */
 435         } else
 436                 tmp = 0;
 437
 438         WREG32(DC_LB_MEMORY_SPLIT + radeon_crtc->crtc_offset,
 439                DC_LB_MEMORY_CONFIG(tmp));
 440
 441         if (radeon_crtc->base.enabled && mode) {
 442                 switch (tmp) {
 443                 case 0:
 444                 default:
 445                         return 4096 * 2;
 446                 case 2:
 447                         return 8192 * 2;
 448                 }
 449         }
 450
 451         /* controller not enabled, so no lb used */
 452         return 0;
 453 }
 454
 455 static u32 si_get_number_of_dram_channels(struct radeon_device *rdev)
 456 {
 457         u32 tmp = RREG32(MC_SHARED_CHMAP);
 458
 459         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
 460         case 0:
 461         default:
 462                 return 1;
 463         case 1:
 464                 return 2;
 465         case 2:
 466                 return 4;
 467         case 3:
 468                 return 8;
 469         case 4:
 470                 return 3;
 471         case 5:
 472                 return 6;
 473         case 6:
 474                 return 10;
 475         case 7:
 476                 return 12;
 477         case 8:
 478                 return 16;
 479         }
 480 }
 481
 482 struct dce6_wm_params {
 483         u32 dram_channels; /* number of dram channels */
 484         u32 yclk;          /* bandwidth per dram data pin in kHz */
 485         u32 sclk;          /* engine clock in kHz */
 486         u32 disp_clk;      /* display clock in kHz */
 487         u32 src_width;     /* viewport width */
 488         u32 active_time;   /* active display time in ns */
 489         u32 blank_time;    /* blank time in ns */
 490         bool interlaced;    /* mode is interlaced */
 491         fixed20_12 vsc;    /* vertical scale ratio */
 492         u32 num_heads;     /* number of active crtcs */
 493         u32 bytes_per_pixel; /* bytes per pixel display + overlay */
 494         u32 lb_size;       /* line buffer allocated to pipe */
 495         u32 vtaps;         /* vertical scaler taps */
 496 };
 497
 498 static u32 dce6_dram_bandwidth(struct dce6_wm_params *wm)
 499 {
 500         /* Calculate raw DRAM Bandwidth */
 501         fixed20_12 dram_efficiency; /* 0.7 */
 502         fixed20_12 yclk, dram_channels, bandwidth;
 503         fixed20_12 a;
 504
 505         a.full = dfixed_const(1000);
 506         yclk.full = dfixed_const(wm->yclk);
 507         yclk.full = dfixed_div(yclk, a);
 508         dram_channels.full = dfixed_const(wm->dram_channels * 4);
 509         a.full = dfixed_const(10);
 510         dram_efficiency.full = dfixed_const(7);
 511         dram_efficiency.full = dfixed_div(dram_efficiency, a);
 512         bandwidth.full = dfixed_mul(dram_channels, yclk);
 513         bandwidth.full = dfixed_mul(bandwidth, dram_efficiency);
 514
 515         return dfixed_trunc(bandwidth);
 516 }
 517
 518 static u32 dce6_dram_bandwidth_for_display(struct dce6_wm_params *wm)
 519 {
 520         /* Calculate DRAM Bandwidth and the part allocated to display. */
 521         fixed20_12 disp_dram_allocation; /* 0.3 to 0.7 */
 522         fixed20_12 yclk, dram_channels, bandwidth;
 523         fixed20_12 a;
 524
 525         a.full = dfixed_const(1000);
 526         yclk.full = dfixed_const(wm->yclk);
 527         yclk.full = dfixed_div(yclk, a);
 528         dram_channels.full = dfixed_const(wm->dram_channels * 4);
 529         a.full = dfixed_const(10);
 530         disp_dram_allocation.full = dfixed_const(3); /* XXX worse case value 0.3 */
 531         disp_dram_allocation.full = dfixed_div(disp_dram_allocation, a);
 532         bandwidth.full = dfixed_mul(dram_channels, yclk);
 533         bandwidth.full = dfixed_mul(bandwidth, disp_dram_allocation);
 534
 535         return dfixed_trunc(bandwidth);
 536 }
 537
 538 static u32 dce6_data_return_bandwidth(struct dce6_wm_params *wm)
 539 {
 540         /* Calculate the display Data return Bandwidth */
 541         fixed20_12 return_efficiency; /* 0.8 */
 542         fixed20_12 sclk, bandwidth;
 543         fixed20_12 a;
 544
 545         a.full = dfixed_const(1000);
 546         sclk.full = dfixed_const(wm->sclk);
 547         sclk.full = dfixed_div(sclk, a);
 548         a.full = dfixed_const(10);
 549         return_efficiency.full = dfixed_const(8);
 550         return_efficiency.full = dfixed_div(return_efficiency, a);
 551         a.full = dfixed_const(32);
 552         bandwidth.full = dfixed_mul(a, sclk);
 553         bandwidth.full = dfixed_mul(bandwidth, return_efficiency);
 554
 555         return dfixed_trunc(bandwidth);
 556 }
 557
 558 static u32 dce6_get_dmif_bytes_per_request(struct dce6_wm_params *wm)
 559 {
 560         return 32;
 561 }
 562
 563 static u32 dce6_dmif_request_bandwidth(struct dce6_wm_params *wm)
 564 {
 565         /* Calculate the DMIF Request Bandwidth */
 566         fixed20_12 disp_clk_request_efficiency; /* 0.8 */
 567         fixed20_12 disp_clk, sclk, bandwidth;
 568         fixed20_12 a, b1, b2;
 569         u32 min_bandwidth;
 570
 571         a.full = dfixed_const(1000);
 572         disp_clk.full = dfixed_const(wm->disp_clk);
 573         disp_clk.full = dfixed_div(disp_clk, a);
 574         a.full = dfixed_const(dce6_get_dmif_bytes_per_request(wm) / 2);
 575         b1.full = dfixed_mul(a, disp_clk);
 576
 577         a.full = dfixed_const(1000);
 578         sclk.full = dfixed_const(wm->sclk);
 579         sclk.full = dfixed_div(sclk, a);
 580         a.full = dfixed_const(dce6_get_dmif_bytes_per_request(wm));
 581         b2.full = dfixed_mul(a, sclk);
 582
 583         a.full = dfixed_const(10);
 584         disp_clk_request_efficiency.full = dfixed_const(8);
 585         disp_clk_request_efficiency.full = dfixed_div(disp_clk_request_efficiency, a);
 586
 587         min_bandwidth = min(dfixed_trunc(b1), dfixed_trunc(b2));
 588
 589         a.full = dfixed_const(min_bandwidth);
 590         bandwidth.full = dfixed_mul(a, disp_clk_request_efficiency);
 591
 592         return dfixed_trunc(bandwidth);
 593 }
 594
 595 static u32 dce6_available_bandwidth(struct dce6_wm_params *wm)
 596 {
 597         /* Calculate the Available bandwidth. Display can use this temporarily but not in average. */
 598         u32 dram_bandwidth = dce6_dram_bandwidth(wm);
 599         u32 data_return_bandwidth = dce6_data_return_bandwidth(wm);
 600         u32 dmif_req_bandwidth = dce6_dmif_request_bandwidth(wm);
 601
 602         return min(dram_bandwidth, min(data_return_bandwidth, dmif_req_bandwidth));
 603 }
 604
 605 static u32 dce6_average_bandwidth(struct dce6_wm_params *wm)
 606 {
 607         /* Calculate the display mode Average Bandwidth
 608          * DisplayMode should contain the source and destination dimensions,
 609          * timing, etc.
 610          */
 611         fixed20_12 bpp;
 612         fixed20_12 line_time;
 613         fixed20_12 src_width;
 614         fixed20_12 bandwidth;
 615         fixed20_12 a;
 616
 617         a.full = dfixed_const(1000);
 618         line_time.full = dfixed_const(wm->active_time + wm->blank_time);
 619         line_time.full = dfixed_div(line_time, a);
 620         bpp.full = dfixed_const(wm->bytes_per_pixel);
 621         src_width.full = dfixed_const(wm->src_width);
 622         bandwidth.full = dfixed_mul(src_width, bpp);
 623         bandwidth.full = dfixed_mul(bandwidth, wm->vsc);
 624         bandwidth.full = dfixed_div(bandwidth, line_time);
 625
 626         return dfixed_trunc(bandwidth);
 627 }
 628
 629 static u32 dce6_latency_watermark(struct dce6_wm_params *wm)
 630 {
 631         /* First calcualte the latency in ns */
 632         u32 mc_latency = 2000; /* 2000 ns. */
 633         u32 available_bandwidth = dce6_available_bandwidth(wm);
 634         u32 worst_chunk_return_time = (512 * 8 * 1000) / available_bandwidth;
 635         u32 cursor_line_pair_return_time = (128 * 4 * 1000) / available_bandwidth;
 636         u32 dc_latency = 40000000 / wm->disp_clk; /* dc pipe latency */
 637         u32 other_heads_data_return_time = ((wm->num_heads + 1) * worst_chunk_return_time) +
 638                 (wm->num_heads * cursor_line_pair_return_time);
 639         u32 latency = mc_latency + other_heads_data_return_time + dc_latency;
 640         u32 max_src_lines_per_dst_line, lb_fill_bw, line_fill_time;
 641         u32 tmp, dmif_size = 12288;
 642         fixed20_12 a, b, c;
 643
 644         if (wm->num_heads == 0)
 645                 return 0;
 646
 647         a.full = dfixed_const(2);
 648         b.full = dfixed_const(1);
 649         if ((wm->vsc.full > a.full) ||
 650             ((wm->vsc.full > b.full) && (wm->vtaps >= 3)) ||
 651             (wm->vtaps >= 5) ||
 652             ((wm->vsc.full >= a.full) && wm->interlaced))
 653                 max_src_lines_per_dst_line = 4;
 654         else
 655                 max_src_lines_per_dst_line = 2;
 656
 657         a.full = dfixed_const(available_bandwidth);
 658         b.full = dfixed_const(wm->num_heads);
 659         a.full = dfixed_div(a, b);
 660
 661         b.full = dfixed_const(mc_latency + 512);
 662         c.full = dfixed_const(wm->disp_clk);
 663         b.full = dfixed_div(b, c);
 664
 665         c.full = dfixed_const(dmif_size);
 666         b.full = dfixed_div(c, b);
 667
 668         tmp = min(dfixed_trunc(a), dfixed_trunc(b));
 669
 670         b.full = dfixed_const(1000);
 671         c.full = dfixed_const(wm->disp_clk);
 672         b.full = dfixed_div(c, b);
 673         c.full = dfixed_const(wm->bytes_per_pixel);
 674         b.full = dfixed_mul(b, c);
 675
 676         lb_fill_bw = min(tmp, dfixed_trunc(b));
 677
 678         a.full = dfixed_const(max_src_lines_per_dst_line * wm->src_width * wm->bytes_per_pixel);
 679         b.full = dfixed_const(1000);
 680         c.full = dfixed_const(lb_fill_bw);
 681         b.full = dfixed_div(c, b);
 682         a.full = dfixed_div(a, b);
 683         line_fill_time = dfixed_trunc(a);
 684
 685         if (line_fill_time < wm->active_time)
 686                 return latency;
 687         else
 688                 return latency + (line_fill_time - wm->active_time);
 689
 690 }
 691
 692 static bool dce6_average_bandwidth_vs_dram_bandwidth_for_display(struct dce6_wm_params *wm)
 693 {
 694         if (dce6_average_bandwidth(wm) <=
 695             (dce6_dram_bandwidth_for_display(wm) / wm->num_heads))
 696                 return true;
 697         else
 698                 return false;
 699 };
 700
 701 static bool dce6_average_bandwidth_vs_available_bandwidth(struct dce6_wm_params *wm)
 702 {
 703         if (dce6_average_bandwidth(wm) <=
 704             (dce6_available_bandwidth(wm) / wm->num_heads))
 705                 return true;
 706         else
 707                 return false;
 708 };
 709
 710 static bool dce6_check_latency_hiding(struct dce6_wm_params *wm)
 711 {
 712         u32 lb_partitions = wm->lb_size / wm->src_width;
 713         u32 line_time = wm->active_time + wm->blank_time;
 714         u32 latency_tolerant_lines;
 715         u32 latency_hiding;
 716         fixed20_12 a;
 717
 718         a.full = dfixed_const(1);
 719         if (wm->vsc.full > a.full)
 720                 latency_tolerant_lines = 1;
 721         else {
 722                 if (lb_partitions <= (wm->vtaps + 1))
 723                         latency_tolerant_lines = 1;
 724                 else
 725                         latency_tolerant_lines = 2;
 726         }
 727
 728         latency_hiding = (latency_tolerant_lines * line_time + wm->blank_time);
 729
 730         if (dce6_latency_watermark(wm) <= latency_hiding)
 731                 return true;
 732         else
 733                 return false;
 734 }
 735
 736 static void dce6_program_watermarks(struct radeon_device *rdev,
 737                                          struct radeon_crtc *radeon_crtc,
 738                                          u32 lb_size, u32 num_heads)
 739 {
 740         struct drm_display_mode *mode = &radeon_crtc->base.mode;
 741         struct dce6_wm_params wm;
 742         u32 pixel_period;
 743         u32 line_time = 0;
 744         u32 latency_watermark_a = 0, latency_watermark_b = 0;
 745         u32 priority_a_mark = 0, priority_b_mark = 0;
 746         u32 priority_a_cnt = PRIORITY_OFF;
 747         u32 priority_b_cnt = PRIORITY_OFF;
 748         u32 tmp, arb_control3;
 749         fixed20_12 a, b, c;
 750
 751         if (radeon_crtc->base.enabled && num_heads && mode) {
 752                 pixel_period = 1000000 / (u32)mode->clock;
 753                 line_time = min((u32)mode->crtc_htotal * pixel_period, (u32)65535);
 754                 priority_a_cnt = 0;
 755                 priority_b_cnt = 0;
 756
 757                 wm.yclk = rdev->pm.current_mclk * 10;
 758                 wm.sclk = rdev->pm.current_sclk * 10;
 759                 wm.disp_clk = mode->clock;
 760                 wm.src_width = mode->crtc_hdisplay;
 761                 wm.active_time = mode->crtc_hdisplay * pixel_period;
 762                 wm.blank_time = line_time - wm.active_time;
 763                 wm.interlaced = false;
 764                 if (mode->flags & DRM_MODE_FLAG_INTERLACE)
 765                         wm.interlaced = true;
 766                 wm.vsc = radeon_crtc->vsc;
 767                 wm.vtaps = 1;
 768                 if (radeon_crtc->rmx_type != RMX_OFF)
 769                         wm.vtaps = 2;
 770                 wm.bytes_per_pixel = 4; /* XXX: get this from fb config */
 771                 wm.lb_size = lb_size;
 772                 if (rdev->family == CHIP_ARUBA)
 773                         wm.dram_channels = evergreen_get_number_of_dram_channels(rdev);
 774                 else
 775                         wm.dram_channels = si_get_number_of_dram_channels(rdev);
 776                 wm.num_heads = num_heads;
 777
 778                 /* set for high clocks */
 779                 latency_watermark_a = min(dce6_latency_watermark(&wm), (u32)65535);
 780                 /* set for low clocks */
 781                 /* wm.yclk = low clk; wm.sclk = low clk */
 782                 latency_watermark_b = min(dce6_latency_watermark(&wm), (u32)65535);
 783
 784                 /* possibly force display priority to high */
 785                 /* should really do this at mode validation time... */
 786                 if (!dce6_average_bandwidth_vs_dram_bandwidth_for_display(&wm) ||
 787                     !dce6_average_bandwidth_vs_available_bandwidth(&wm) ||
 788                     !dce6_check_latency_hiding(&wm) ||
 789                     (rdev->disp_priority == 2)) {
 790                         DRM_DEBUG_KMS("force priority to high\n");
 791                         priority_a_cnt |= PRIORITY_ALWAYS_ON;
 792                         priority_b_cnt |= PRIORITY_ALWAYS_ON;
 793                 }
 794
 795                 a.full = dfixed_const(1000);
 796                 b.full = dfixed_const(mode->clock);
 797                 b.full = dfixed_div(b, a);
 798                 c.full = dfixed_const(latency_watermark_a);
 799                 c.full = dfixed_mul(c, b);
 800                 c.full = dfixed_mul(c, radeon_crtc->hsc);
 801                 c.full = dfixed_div(c, a);
 802                 a.full = dfixed_const(16);
 803                 c.full = dfixed_div(c, a);
 804                 priority_a_mark = dfixed_trunc(c);
 805                 priority_a_cnt |= priority_a_mark & PRIORITY_MARK_MASK;
 806
 807                 a.full = dfixed_const(1000);
 808                 b.full = dfixed_const(mode->clock);
 809                 b.full = dfixed_div(b, a);
 810                 c.full = dfixed_const(latency_watermark_b);
 811                 c.full = dfixed_mul(c, b);
 812                 c.full = dfixed_mul(c, radeon_crtc->hsc);
 813                 c.full = dfixed_div(c, a);
 814                 a.full = dfixed_const(16);
 815                 c.full = dfixed_div(c, a);
 816                 priority_b_mark = dfixed_trunc(c);
 817                 priority_b_cnt |= priority_b_mark & PRIORITY_MARK_MASK;
 818         }
 819
 820         /* select wm A */
 821         arb_control3 = RREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset);
 822         tmp = arb_control3;
 823         tmp &= ~LATENCY_WATERMARK_MASK(3);
 824         tmp |= LATENCY_WATERMARK_MASK(1);
 825         WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, tmp);
 826         WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
 827                (LATENCY_LOW_WATERMARK(latency_watermark_a) |
 828                 LATENCY_HIGH_WATERMARK(line_time)));
 829         /* select wm B */
 830         tmp = RREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset);
 831         tmp &= ~LATENCY_WATERMARK_MASK(3);
 832         tmp |= LATENCY_WATERMARK_MASK(2);
 833         WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, tmp);
 834         WREG32(DPG_PIPE_LATENCY_CONTROL + radeon_crtc->crtc_offset,
 835                (LATENCY_LOW_WATERMARK(latency_watermark_b) |
 836                 LATENCY_HIGH_WATERMARK(line_time)));
 837         /* restore original selection */
 838         WREG32(DPG_PIPE_ARBITRATION_CONTROL3 + radeon_crtc->crtc_offset, arb_control3);
 839
 840         /* write the priority marks */
 841         WREG32(PRIORITY_A_CNT + radeon_crtc->crtc_offset, priority_a_cnt);
 842         WREG32(PRIORITY_B_CNT + radeon_crtc->crtc_offset, priority_b_cnt);
 843
 844 }
 845
 846 void dce6_bandwidth_update(struct radeon_device *rdev)
 847 {
 848         struct drm_display_mode *mode0 = NULL;
 849         struct drm_display_mode *mode1 = NULL;
 850         u32 num_heads = 0, lb_size;
 851         int i;
 852
 853         radeon_update_display_priority(rdev);
 854
 855         for (i = 0; i < rdev->num_crtc; i++) {
 856                 if (rdev->mode_info.crtcs[i]->base.enabled)
 857                         num_heads++;
 858         }
 859         for (i = 0; i < rdev->num_crtc; i += 2) {
 860                 mode0 = &rdev->mode_info.crtcs[i]->base.mode;
 861                 mode1 = &rdev->mode_info.crtcs[i+1]->base.mode;
 862                 lb_size = dce6_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i], mode0, mode1);
 863                 dce6_program_watermarks(rdev, rdev->mode_info.crtcs[i], lb_size, num_heads);
 864                 lb_size = dce6_line_buffer_adjust(rdev, rdev->mode_info.crtcs[i+1], mode1, mode0);
 865                 dce6_program_watermarks(rdev, rdev->mode_info.crtcs[i+1], lb_size, num_heads);
 866         }
 867 }
 868
 869 /*
 870  * Core functions
 871  */
 872 static void si_tiling_mode_table_init(struct radeon_device *rdev)
 873 {
 874         const u32 num_tile_mode_states = 32;
 875         u32 reg_offset, gb_tile_moden, split_equal_to_row_size;
 876
 877         switch (rdev->config.si.mem_row_size_in_kb) {
 878         case 1:
 879                 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_1KB;
 880                 break;
 881         case 2:
 882         default:
 883                 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_2KB;
 884                 break;
 885         case 4:
 886                 split_equal_to_row_size = ADDR_SURF_TILE_SPLIT_4KB;
 887                 break;
 888         }
 889
 890         if ((rdev->family == CHIP_TAHITI) ||
 891             (rdev->family == CHIP_PITCAIRN)) {
 892                 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
 893                         switch (reg_offset) {
 894                         case 0:  /* non-AA compressed depth or any compressed stencil */
 895                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
 896                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
 897                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 898                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
 899                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
 900                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
 901                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
 902                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
 903                                 break;
 904                         case 1:  /* 2xAA/4xAA compressed depth only */
 905                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
 906                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
 907                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 908                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
 909                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
 910                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
 911                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
 912                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
 913                                 break;
 914                         case 2:  /* 8xAA compressed depth only */
 915                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
 916                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
 917                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 918                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
 919                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
 920                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
 921                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
 922                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
 923                                 break;
 924                         case 3:  /* 2xAA/4xAA compressed depth with stencil (for depth buffer) */
 925                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
 926                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
 927                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 928                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
 929                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
 930                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
 931                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
 932                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
 933                                 break;
 934                         case 4:  /* Maps w/ a dimension less than the 2D macro-tile dimensions (for mipmapped depth textures) */
 935                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
 936                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
 937                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 938                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
 939                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
 940                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
 941                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
 942                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
 943                                 break;
 944                         case 5:  /* Uncompressed 16bpp depth - and stencil buffer allocated with it */
 945                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
 946                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
 947                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 948                                                  TILE_SPLIT(split_equal_to_row_size) |
 949                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
 950                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
 951                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
 952                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
 953                                 break;
 954                         case 6:  /* Uncompressed 32bpp depth - and stencil buffer allocated with it */
 955                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
 956                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
 957                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 958                                                  TILE_SPLIT(split_equal_to_row_size) |
 959                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
 960                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
 961                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
 962                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
 963                                 break;
 964                         case 7:  /* Uncompressed 8bpp stencil without depth (drivers typically do not use) */
 965                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
 966                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
 967                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 968                                                  TILE_SPLIT(split_equal_to_row_size) |
 969                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
 970                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
 971                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
 972                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
 973                                 break;
 974                         case 8:  /* 1D and 1D Array Surfaces */
 975                                 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
 976                                                  MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
 977                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 978                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
 979                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
 980                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
 981                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
 982                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
 983                                 break;
 984                         case 9:  /* Displayable maps. */
 985                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
 986                                                  MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
 987                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 988                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
 989                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
 990                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
 991                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
 992                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
 993                                 break;
 994                         case 10:  /* Display 8bpp. */
 995                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
 996                                                  MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
 997                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
 998                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
 999                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1000                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1001                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1002                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1003                                 break;
1004                         case 11:  /* Display 16bpp. */
1005                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1006                                                  MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1007                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1008                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1009                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1010                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1011                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1012                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1013                                 break;
1014                         case 12:  /* Display 32bpp. */
1015                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1016                                                  MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1017                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1018                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
1019                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1020                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1021                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1022                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
1023                                 break;
1024                         case 13:  /* Thin. */
1025                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1026                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1027                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1028                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
1029                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1030                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1031                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1032                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1033                                 break;
1034                         case 14:  /* Thin 8 bpp. */
1035                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1036                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1037                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1038                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1039                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1040                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1041                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1042                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
1043                                 break;
1044                         case 15:  /* Thin 16 bpp. */
1045                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1046                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1047                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1048                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1049                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1050                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1051                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1052                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
1053                                 break;
1054                         case 16:  /* Thin 32 bpp. */
1055                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1056                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1057                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1058                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
1059                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1060                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1061                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1062                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
1063                                 break;
1064                         case 17:  /* Thin 64 bpp. */
1065                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1066                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1067                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1068                                                  TILE_SPLIT(split_equal_to_row_size) |
1069                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1070                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1071                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1072                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
1073                                 break;
1074                         case 21:  /* 8 bpp PRT. */
1075                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1076                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1077                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1078                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1079                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1080                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
1081                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1082                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1083                                 break;
1084                         case 22:  /* 16 bpp PRT */
1085                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1086                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1087                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1088                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1089                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1090                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1091                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1092                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
1093                                 break;
1094                         case 23:  /* 32 bpp PRT */
1095                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1096                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1097                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1098                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1099                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1100                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1101                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1102                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1103                                 break;
1104                         case 24:  /* 64 bpp PRT */
1105                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1106                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1107                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1108                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
1109                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1110                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1111                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1112                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1113                                 break;
1114                         case 25:  /* 128 bpp PRT */
1115                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1116                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1117                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1118                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
1119                                                  NUM_BANKS(ADDR_SURF_8_BANK) |
1120                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1121                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1122                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
1123                                 break;
1124                         default:
1125                                 gb_tile_moden = 0;
1126                                 break;
1127                         }
1128                         WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
1129                 }
1130         } else if (rdev->family == CHIP_VERDE) {
1131                 for (reg_offset = 0; reg_offset < num_tile_mode_states; reg_offset++) {
1132                         switch (reg_offset) {
1133                         case 0:  /* non-AA compressed depth or any compressed stencil */
1134                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1135                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1136                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1137                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
1138                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1139                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1140                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1141                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
1142                                 break;
1143                         case 1:  /* 2xAA/4xAA compressed depth only */
1144                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1145                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1146                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1147                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
1148                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1149                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1150                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1151                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
1152                                 break;
1153                         case 2:  /* 8xAA compressed depth only */
1154                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1155                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1156                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1157                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1158                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1159                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1160                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1161                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
1162                                 break;
1163                         case 3:  /* 2xAA/4xAA compressed depth with stencil (for depth buffer) */
1164                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1165                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1166                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1167                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_128B) |
1168                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1169                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1170                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1171                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
1172                                 break;
1173                         case 4:  /* Maps w/ a dimension less than the 2D macro-tile dimensions (for mipmapped depth textures) */
1174                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1175                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1176                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1177                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
1178                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1179                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1180                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1181                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1182                                 break;
1183                         case 5:  /* Uncompressed 16bpp depth - and stencil buffer allocated with it */
1184                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1185                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1186                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1187                                                  TILE_SPLIT(split_equal_to_row_size) |
1188                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1189                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1190                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1191                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1192                                 break;
1193                         case 6:  /* Uncompressed 32bpp depth - and stencil buffer allocated with it */
1194                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1195                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1196                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1197                                                  TILE_SPLIT(split_equal_to_row_size) |
1198                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1199                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1200                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1201                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1202                                 break;
1203                         case 7:  /* Uncompressed 8bpp stencil without depth (drivers typically do not use) */
1204                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1205                                                  MICRO_TILE_MODE(ADDR_SURF_DEPTH_MICRO_TILING) |
1206                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1207                                                  TILE_SPLIT(split_equal_to_row_size) |
1208                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1209                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1210                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1211                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
1212                                 break;
1213                         case 8:  /* 1D and 1D Array Surfaces */
1214                                 gb_tile_moden = (ARRAY_MODE(ARRAY_LINEAR_ALIGNED) |
1215                                                  MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1216                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1217                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
1218                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1219                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1220                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1221                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1222                                 break;
1223                         case 9:  /* Displayable maps. */
1224                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1225                                                  MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1226                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1227                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
1228                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1229                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1230                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1231                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1232                                 break;
1233                         case 10:  /* Display 8bpp. */
1234                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1235                                                  MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1236                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1237                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1238                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1239                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1240                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1241                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
1242                                 break;
1243                         case 11:  /* Display 16bpp. */
1244                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1245                                                  MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1246                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1247                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1248                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1249                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1250                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1251                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1252                                 break;
1253                         case 12:  /* Display 32bpp. */
1254                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1255                                                  MICRO_TILE_MODE(ADDR_SURF_DISPLAY_MICRO_TILING) |
1256                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1257                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
1258                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1259                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1260                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1261                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1262                                 break;
1263                         case 13:  /* Thin. */
1264                                 gb_tile_moden = (ARRAY_MODE(ARRAY_1D_TILED_THIN1) |
1265                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1266                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1267                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_64B) |
1268                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1269                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1270                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1271                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1272                                 break;
1273                         case 14:  /* Thin 8 bpp. */
1274                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1275                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1276                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1277                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1278                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1279                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1280                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1281                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1282                                 break;
1283                         case 15:  /* Thin 16 bpp. */
1284                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1285                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1286                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1287                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1288                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1289                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1290                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1291                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1292                                 break;
1293                         case 16:  /* Thin 32 bpp. */
1294                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1295                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1296                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1297                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
1298                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1299                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1300                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1301                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1302                                 break;
1303                         case 17:  /* Thin 64 bpp. */
1304                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1305                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1306                                                  PIPE_CONFIG(ADDR_SURF_P4_8x16) |
1307                                                  TILE_SPLIT(split_equal_to_row_size) |
1308                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1309                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1310                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1311                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1312                                 break;
1313                         case 21:  /* 8 bpp PRT. */
1314                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1315                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1316                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1317                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1318                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1319                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_2) |
1320                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1321                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1322                                 break;
1323                         case 22:  /* 16 bpp PRT */
1324                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1325                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1326                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1327                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1328                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1329                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1330                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_4) |
1331                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_4));
1332                                 break;
1333                         case 23:  /* 32 bpp PRT */
1334                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1335                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1336                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1337                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_256B) |
1338                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1339                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1340                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_2) |
1341                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1342                                 break;
1343                         case 24:  /* 64 bpp PRT */
1344                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1345                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1346                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1347                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_512B) |
1348                                                  NUM_BANKS(ADDR_SURF_16_BANK) |
1349                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1350                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1351                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_2));
1352                                 break;
1353                         case 25:  /* 128 bpp PRT */
1354                                 gb_tile_moden = (ARRAY_MODE(ARRAY_2D_TILED_THIN1) |
1355                                                  MICRO_TILE_MODE(ADDR_SURF_THIN_MICRO_TILING) |
1356                                                  PIPE_CONFIG(ADDR_SURF_P8_32x32_8x16) |
1357                                                  TILE_SPLIT(ADDR_SURF_TILE_SPLIT_1KB) |
1358                                                  NUM_BANKS(ADDR_SURF_8_BANK) |
1359                                                  BANK_WIDTH(ADDR_SURF_BANK_WIDTH_1) |
1360                                                  BANK_HEIGHT(ADDR_SURF_BANK_HEIGHT_1) |
1361                                                  MACRO_TILE_ASPECT(ADDR_SURF_MACRO_ASPECT_1));
1362                                 break;
1363                         default:
1364                                 gb_tile_moden = 0;
1365                                 break;
1366                         }
1367                         WREG32(GB_TILE_MODE0 + (reg_offset * 4), gb_tile_moden);
1368                 }
1369         } else
1370                 DRM_ERROR("unknown asic: 0x%x\n", rdev->family);
1371 }
1372
1373 static void si_select_se_sh(struct radeon_device *rdev,
1374                             u32 se_num, u32 sh_num)
1375 {
1376         u32 data = INSTANCE_BROADCAST_WRITES;
1377
1378         if ((se_num == 0xffffffff) && (sh_num == 0xffffffff))
1379                 data = SH_BROADCAST_WRITES | SE_BROADCAST_WRITES;
1380         else if (se_num == 0xffffffff)
1381                 data |= SE_BROADCAST_WRITES | SH_INDEX(sh_num);
1382         else if (sh_num == 0xffffffff)
1383                 data |= SH_BROADCAST_WRITES | SE_INDEX(se_num);
1384         else
1385                 data |= SH_INDEX(sh_num) | SE_INDEX(se_num);
1386         WREG32(GRBM_GFX_INDEX, data);
1387 }
1388
1389 static u32 si_create_bitmask(u32 bit_width)
1390 {
1391         u32 i, mask = 0;
1392
1393         for (i = 0; i < bit_width; i++) {
1394                 mask <<= 1;
1395                 mask |= 1;
1396         }
1397         return mask;
1398 }
1399
1400 static u32 si_get_cu_enabled(struct radeon_device *rdev, u32 cu_per_sh)
1401 {
1402         u32 data, mask;
1403
1404         data = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
1405         if (data & 1)
1406                 data &= INACTIVE_CUS_MASK;
1407         else
1408                 data = 0;
1409         data |= RREG32(GC_USER_SHADER_ARRAY_CONFIG);
1410
1411         data >>= INACTIVE_CUS_SHIFT;
1412
1413         mask = si_create_bitmask(cu_per_sh);
1414
1415         return ~data & mask;
1416 }
1417
1418 static void si_setup_spi(struct radeon_device *rdev,
1419                          u32 se_num, u32 sh_per_se,
1420                          u32 cu_per_sh)
1421 {
1422         int i, j, k;
1423         u32 data, mask, active_cu;
1424
1425         for (i = 0; i < se_num; i++) {
1426                 for (j = 0; j < sh_per_se; j++) {
1427                         si_select_se_sh(rdev, i, j);
1428                         data = RREG32(SPI_STATIC_THREAD_MGMT_3);
1429                         active_cu = si_get_cu_enabled(rdev, cu_per_sh);
1430
1431                         mask = 1;
1432                         for (k = 0; k < 16; k++) {
1433                                 mask <<= k;
1434                                 if (active_cu & mask) {
1435                                         data &= ~mask;
1436                                         WREG32(SPI_STATIC_THREAD_MGMT_3, data);
1437                                         break;
1438                                 }
1439                         }
1440                 }
1441         }
1442         si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
1443 }
1444
1445 static u32 si_get_rb_disabled(struct radeon_device *rdev,
1446                               u32 max_rb_num, u32 se_num,
1447                               u32 sh_per_se)
1448 {
1449         u32 data, mask;
1450
1451         data = RREG32(CC_RB_BACKEND_DISABLE);
1452         if (data & 1)
1453                 data &= BACKEND_DISABLE_MASK;
1454         else
1455                 data = 0;
1456         data |= RREG32(GC_USER_RB_BACKEND_DISABLE);
1457
1458         data >>= BACKEND_DISABLE_SHIFT;
1459
1460         mask = si_create_bitmask(max_rb_num / se_num / sh_per_se);
1461
1462         return data & mask;
1463 }
1464
1465 static void si_setup_rb(struct radeon_device *rdev,
1466                         u32 se_num, u32 sh_per_se,
1467                         u32 max_rb_num)
1468 {
1469         int i, j;
1470         u32 data, mask;
1471         u32 disabled_rbs = 0;
1472         u32 enabled_rbs = 0;
1473
1474         for (i = 0; i < se_num; i++) {
1475                 for (j = 0; j < sh_per_se; j++) {
1476                         si_select_se_sh(rdev, i, j);
1477                         data = si_get_rb_disabled(rdev, max_rb_num, se_num, sh_per_se);
1478                         disabled_rbs |= data << ((i * sh_per_se + j) * TAHITI_RB_BITMAP_WIDTH_PER_SH);
1479                 }
1480         }
1481         si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
1482
1483         mask = 1;
1484         for (i = 0; i < max_rb_num; i++) {
1485                 if (!(disabled_rbs & mask))
1486                         enabled_rbs |= mask;
1487                 mask <<= 1;
1488         }
1489
1490         for (i = 0; i < se_num; i++) {
1491                 si_select_se_sh(rdev, i, 0xffffffff);
1492                 data = 0;
1493                 for (j = 0; j < sh_per_se; j++) {
1494                         switch (enabled_rbs & 3) {
1495                         case 1:
1496                                 data |= (RASTER_CONFIG_RB_MAP_0 << (i * sh_per_se + j) * 2);
1497                                 break;
1498                         case 2:
1499                                 data |= (RASTER_CONFIG_RB_MAP_3 << (i * sh_per_se + j) * 2);
1500                                 break;
1501                         case 3:
1502                         default:
1503                                 data |= (RASTER_CONFIG_RB_MAP_2 << (i * sh_per_se + j) * 2);
1504                                 break;
1505                         }
1506                         enabled_rbs >>= 2;
1507                 }
1508                 WREG32(PA_SC_RASTER_CONFIG, data);
1509         }
1510         si_select_se_sh(rdev, 0xffffffff, 0xffffffff);
1511 }
1512
1513 static void si_gpu_init(struct radeon_device *rdev)
1514 {
1515         u32 gb_addr_config = 0;
1516         u32 mc_shared_chmap, mc_arb_ramcfg;
1517         u32 sx_debug_1;
1518         u32 hdp_host_path_cntl;
1519         u32 tmp;
1520         int i, j;
1521
1522         switch (rdev->family) {
1523         case CHIP_TAHITI:
1524                 rdev->config.si.max_shader_engines = 2;
1525                 rdev->config.si.max_tile_pipes = 12;
1526                 rdev->config.si.max_cu_per_sh = 8;
1527                 rdev->config.si.max_sh_per_se = 2;
1528                 rdev->config.si.max_backends_per_se = 4;
1529                 rdev->config.si.max_texture_channel_caches = 12;
1530                 rdev->config.si.max_gprs = 256;
1531                 rdev->config.si.max_gs_threads = 32;
1532                 rdev->config.si.max_hw_contexts = 8;
1533
1534                 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
1535                 rdev->config.si.sc_prim_fifo_size_backend = 0x100;
1536                 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
1537                 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1538                 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
1539                 break;
1540         case CHIP_PITCAIRN:
1541                 rdev->config.si.max_shader_engines = 2;
1542                 rdev->config.si.max_tile_pipes = 8;
1543                 rdev->config.si.max_cu_per_sh = 5;
1544                 rdev->config.si.max_sh_per_se = 2;
1545                 rdev->config.si.max_backends_per_se = 4;
1546                 rdev->config.si.max_texture_channel_caches = 8;
1547                 rdev->config.si.max_gprs = 256;
1548                 rdev->config.si.max_gs_threads = 32;
1549                 rdev->config.si.max_hw_contexts = 8;
1550
1551                 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
1552                 rdev->config.si.sc_prim_fifo_size_backend = 0x100;
1553                 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
1554                 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1555                 gb_addr_config = TAHITI_GB_ADDR_CONFIG_GOLDEN;
1556                 break;
1557         case CHIP_VERDE:
1558         default:
1559                 rdev->config.si.max_shader_engines = 1;
1560                 rdev->config.si.max_tile_pipes = 4;
1561                 rdev->config.si.max_cu_per_sh = 2;
1562                 rdev->config.si.max_sh_per_se = 2;
1563                 rdev->config.si.max_backends_per_se = 4;
1564                 rdev->config.si.max_texture_channel_caches = 4;
1565                 rdev->config.si.max_gprs = 256;
1566                 rdev->config.si.max_gs_threads = 32;
1567                 rdev->config.si.max_hw_contexts = 8;
1568
1569                 rdev->config.si.sc_prim_fifo_size_frontend = 0x20;
1570                 rdev->config.si.sc_prim_fifo_size_backend = 0x40;
1571                 rdev->config.si.sc_hiz_tile_fifo_size = 0x30;
1572                 rdev->config.si.sc_earlyz_tile_fifo_size = 0x130;
1573                 gb_addr_config = VERDE_GB_ADDR_CONFIG_GOLDEN;
1574                 break;
1575         }
1576
1577         /* Initialize HDP */
1578         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
1579                 WREG32((0x2c14 + j), 0x00000000);
1580                 WREG32((0x2c18 + j), 0x00000000);
1581                 WREG32((0x2c1c + j), 0x00000000);
1582                 WREG32((0x2c20 + j), 0x00000000);
1583                 WREG32((0x2c24 + j), 0x00000000);
1584         }
1585
1586         WREG32(GRBM_CNTL, GRBM_READ_TIMEOUT(0xff));
1587
1588         evergreen_fix_pci_max_read_req_size(rdev);
1589
1590         WREG32(BIF_FB_EN, FB_READ_EN | FB_WRITE_EN);
1591
1592         mc_shared_chmap = RREG32(MC_SHARED_CHMAP);
1593         mc_arb_ramcfg = RREG32(MC_ARB_RAMCFG);
1594
1595         rdev->config.si.num_tile_pipes = rdev->config.si.max_tile_pipes;
1596         rdev->config.si.mem_max_burst_length_bytes = 256;
1597         tmp = (mc_arb_ramcfg & NOOFCOLS_MASK) >> NOOFCOLS_SHIFT;
1598         rdev->config.si.mem_row_size_in_kb = (4 * (1 << (8 + tmp))) / 1024;
1599         if (rdev->config.si.mem_row_size_in_kb > 4)
1600                 rdev->config.si.mem_row_size_in_kb = 4;
1601         /* XXX use MC settings? */
1602         rdev->config.si.shader_engine_tile_size = 32;
1603         rdev->config.si.num_gpus = 1;
1604         rdev->config.si.multi_gpu_tile_size = 64;
1605
1606         /* fix up row size */
1607         gb_addr_config &= ~ROW_SIZE_MASK;
1608         switch (rdev->config.si.mem_row_size_in_kb) {
1609         case 1:
1610         default:
1611                 gb_addr_config |= ROW_SIZE(0);
1612                 break;
1613         case 2:
1614                 gb_addr_config |= ROW_SIZE(1);
1615                 break;
1616         case 4:
1617                 gb_addr_config |= ROW_SIZE(2);
1618                 break;
1619         }
1620
1621         /* setup tiling info dword.  gb_addr_config is not adequate since it does
1622          * not have bank info, so create a custom tiling dword.
1623          * bits 3:0   num_pipes
1624          * bits 7:4   num_banks
1625          * bits 11:8  group_size
1626          * bits 15:12 row_size
1627          */
1628         rdev->config.si.tile_config = 0;
1629         switch (rdev->config.si.num_tile_pipes) {
1630         case 1:
1631                 rdev->config.si.tile_config |= (0 << 0);
1632                 break;
1633         case 2:
1634                 rdev->config.si.tile_config |= (1 << 0);
1635                 break;
1636         case 4:
1637                 rdev->config.si.tile_config |= (2 << 0);
1638                 break;
1639         case 8:
1640         default:
1641                 /* XXX what about 12? */
1642                 rdev->config.si.tile_config |= (3 << 0);
1643                 break;
1644         }
1645         switch ((mc_arb_ramcfg & NOOFBANK_MASK) >> NOOFBANK_SHIFT) {
1646         case 0: /* four banks */
1647                 rdev->config.si.tile_config |= 0 << 4;
1648                 break;
1649         case 1: /* eight banks */
1650                 rdev->config.si.tile_config |= 1 << 4;
1651                 break;
1652         case 2: /* sixteen banks */
1653         default:
1654                 rdev->config.si.tile_config |= 2 << 4;
1655                 break;
1656         }
1657         rdev->config.si.tile_config |=
1658                 ((gb_addr_config & PIPE_INTERLEAVE_SIZE_MASK) >> PIPE_INTERLEAVE_SIZE_SHIFT) << 8;
1659         rdev->config.si.tile_config |=
1660                 ((gb_addr_config & ROW_SIZE_MASK) >> ROW_SIZE_SHIFT) << 12;
1661
1662         WREG32(GB_ADDR_CONFIG, gb_addr_config);
1663         WREG32(DMIF_ADDR_CONFIG, gb_addr_config);
1664         WREG32(HDP_ADDR_CONFIG, gb_addr_config);
1665         WREG32(DMA_TILING_CONFIG + DMA0_REGISTER_OFFSET, gb_addr_config);
1666         WREG32(DMA_TILING_CONFIG + DMA1_REGISTER_OFFSET, gb_addr_config);
1667
1668         si_tiling_mode_table_init(rdev);
1669
1670         si_setup_rb(rdev, rdev->config.si.max_shader_engines,
1671                     rdev->config.si.max_sh_per_se,
1672                     rdev->config.si.max_backends_per_se);
1673
1674         si_setup_spi(rdev, rdev->config.si.max_shader_engines,
1675                      rdev->config.si.max_sh_per_se,
1676                      rdev->config.si.max_cu_per_sh);
1677
1678
1679         /* set HW defaults for 3D engine */
1680         WREG32(CP_QUEUE_THRESHOLDS, (ROQ_IB1_START(0x16) |
1681                                      ROQ_IB2_START(0x2b)));
1682         WREG32(CP_MEQ_THRESHOLDS, MEQ1_START(0x30) | MEQ2_START(0x60));
1683
1684         sx_debug_1 = RREG32(SX_DEBUG_1);
1685         WREG32(SX_DEBUG_1, sx_debug_1);
1686
1687         WREG32(SPI_CONFIG_CNTL_1, VTX_DONE_DELAY(4));
1688
1689         WREG32(PA_SC_FIFO_SIZE, (SC_FRONTEND_PRIM_FIFO_SIZE(rdev->config.si.sc_prim_fifo_size_frontend) |
1690                                  SC_BACKEND_PRIM_FIFO_SIZE(rdev->config.si.sc_prim_fifo_size_backend) |
1691                                  SC_HIZ_TILE_FIFO_SIZE(rdev->config.si.sc_hiz_tile_fifo_size) |
1692                                  SC_EARLYZ_TILE_FIFO_SIZE(rdev->config.si.sc_earlyz_tile_fifo_size)));
1693
1694         WREG32(VGT_NUM_INSTANCES, 1);
1695
1696         WREG32(CP_PERFMON_CNTL, 0);
1697
1698         WREG32(SQ_CONFIG, 0);
1699
1700         WREG32(PA_SC_FORCE_EOV_MAX_CNTS, (FORCE_EOV_MAX_CLK_CNT(4095) |
1701                                           FORCE_EOV_MAX_REZ_CNT(255)));
1702
1703         WREG32(VGT_CACHE_INVALIDATION, CACHE_INVALIDATION(VC_AND_TC) |
1704                AUTO_INVLD_EN(ES_AND_GS_AUTO));
1705
1706         WREG32(VGT_GS_VERTEX_REUSE, 16);
1707         WREG32(PA_SC_LINE_STIPPLE_STATE, 0);
1708
1709         WREG32(CB_PERFCOUNTER0_SELECT0, 0);
1710         WREG32(CB_PERFCOUNTER0_SELECT1, 0);
1711         WREG32(CB_PERFCOUNTER1_SELECT0, 0);
1712         WREG32(CB_PERFCOUNTER1_SELECT1, 0);
1713         WREG32(CB_PERFCOUNTER2_SELECT0, 0);
1714         WREG32(CB_PERFCOUNTER2_SELECT1, 0);
1715         WREG32(CB_PERFCOUNTER3_SELECT0, 0);
1716         WREG32(CB_PERFCOUNTER3_SELECT1, 0);
1717
1718         tmp = RREG32(HDP_MISC_CNTL);
1719         tmp |= HDP_FLUSH_INVALIDATE_CACHE;
1720         WREG32(HDP_MISC_CNTL, tmp);
1721
1722         hdp_host_path_cntl = RREG32(HDP_HOST_PATH_CNTL);
1723         WREG32(HDP_HOST_PATH_CNTL, hdp_host_path_cntl);
1724
1725         WREG32(PA_CL_ENHANCE, CLIP_VTX_REORDER_ENA | NUM_CLIP_SEQ(3));
1726
1727         udelay(50);
1728 }
1729
1730 /*
1731  * GPU scratch registers helpers function.
1732  */
1733 static void si_scratch_init(struct radeon_device *rdev)
1734 {
1735         int i;
1736
1737         rdev->scratch.num_reg = 7;
1738         rdev->scratch.reg_base = SCRATCH_REG0;
1739         for (i = 0; i < rdev->scratch.num_reg; i++) {
1740                 rdev->scratch.free[i] = true;
1741                 rdev->scratch.reg[i] = rdev->scratch.reg_base + (i * 4);
1742         }
1743 }
1744
1745 void si_fence_ring_emit(struct radeon_device *rdev,
1746                         struct radeon_fence *fence)
1747 {
1748         struct radeon_ring *ring = &rdev->ring[fence->ring];
1749         u64 addr = rdev->fence_drv[fence->ring].gpu_addr;
1750
1751         /* flush read cache over gart */
1752         radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
1753         radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
1754         radeon_ring_write(ring, 0);
1755         radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
1756         radeon_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
1757                           PACKET3_TC_ACTION_ENA |
1758                           PACKET3_SH_KCACHE_ACTION_ENA |
1759                           PACKET3_SH_ICACHE_ACTION_ENA);
1760         radeon_ring_write(ring, 0xFFFFFFFF);
1761         radeon_ring_write(ring, 0);
1762         radeon_ring_write(ring, 10); /* poll interval */
1763         /* EVENT_WRITE_EOP - flush caches, send int */
1764         radeon_ring_write(ring, PACKET3(PACKET3_EVENT_WRITE_EOP, 4));
1765         radeon_ring_write(ring, EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5));
1766         radeon_ring_write(ring, addr & 0xffffffff);
1767         radeon_ring_write(ring, (upper_32_bits(addr) & 0xff) | DATA_SEL(1) | INT_SEL(2));
1768         radeon_ring_write(ring, fence->seq);
1769         radeon_ring_write(ring, 0);
1770 }
1771
1772 /*
1773  * IB stuff
1774  */
1775 void si_ring_ib_execute(struct radeon_device *rdev, struct radeon_ib *ib)
1776 {
1777         struct radeon_ring *ring = &rdev->ring[ib->ring];
1778         u32 header;
1779
1780         if (ib->is_const_ib) {
1781                 /* set switch buffer packet before const IB */
1782                 radeon_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
1783                 radeon_ring_write(ring, 0);
1784
1785                 header = PACKET3(PACKET3_INDIRECT_BUFFER_CONST, 2);
1786         } else {
1787                 u32 next_rptr;
1788                 if (ring->rptr_save_reg) {
1789                         next_rptr = ring->wptr + 3 + 4 + 8;
1790                         radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
1791                         radeon_ring_write(ring, ((ring->rptr_save_reg -
1792                                                   PACKET3_SET_CONFIG_REG_START) >> 2));
1793                         radeon_ring_write(ring, next_rptr);
1794                 } else if (rdev->wb.enabled) {
1795                         next_rptr = ring->wptr + 5 + 4 + 8;
1796                         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
1797                         radeon_ring_write(ring, (1 << 8));
1798                         radeon_ring_write(ring, ring->next_rptr_gpu_addr & 0xfffffffc);
1799                         radeon_ring_write(ring, upper_32_bits(ring->next_rptr_gpu_addr) & 0xffffffff);
1800                         radeon_ring_write(ring, next_rptr);
1801                 }
1802
1803                 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
1804         }
1805
1806         radeon_ring_write(ring, header);
1807         radeon_ring_write(ring,
1808 #ifdef __BIG_ENDIAN
1809                           (2 << 0) |
1810 #endif
1811                           (ib->gpu_addr & 0xFFFFFFFC));
1812         radeon_ring_write(ring, upper_32_bits(ib->gpu_addr) & 0xFFFF);
1813         radeon_ring_write(ring, ib->length_dw |
1814                           (ib->vm ? (ib->vm->id << 24) : 0));
1815
1816         if (!ib->is_const_ib) {
1817                 /* flush read cache over gart for this vmid */
1818                 radeon_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1));
1819                 radeon_ring_write(ring, (CP_COHER_CNTL2 - PACKET3_SET_CONFIG_REG_START) >> 2);
1820                 radeon_ring_write(ring, ib->vm ? ib->vm->id : 0);
1821                 radeon_ring_write(ring, PACKET3(PACKET3_SURFACE_SYNC, 3));
1822                 radeon_ring_write(ring, PACKET3_TCL1_ACTION_ENA |
1823                                   PACKET3_TC_ACTION_ENA |
1824                                   PACKET3_SH_KCACHE_ACTION_ENA |
1825                                   PACKET3_SH_ICACHE_ACTION_ENA);
1826                 radeon_ring_write(ring, 0xFFFFFFFF);
1827                 radeon_ring_write(ring, 0);
1828                 radeon_ring_write(ring, 10); /* poll interval */
1829         }
1830 }
1831
1832 /*
1833  * CP.
1834  */
1835 static void si_cp_enable(struct radeon_device *rdev, bool enable)
1836 {
1837         if (enable)
1838                 WREG32(CP_ME_CNTL, 0);
1839         else {
1840                 radeon_ttm_set_active_vram_size(rdev, rdev->mc.visible_vram_size);
1841                 WREG32(CP_ME_CNTL, (CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT));
1842                 WREG32(SCRATCH_UMSK, 0);
1843                 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
1844                 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
1845                 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
1846         }
1847         udelay(50);
1848 }
1849
1850 static int si_cp_load_microcode(struct radeon_device *rdev)
1851 {
1852         const __be32 *fw_data;
1853         int i;
1854
1855         if (!rdev->me_fw || !rdev->pfp_fw)
1856                 return -EINVAL;
1857
1858         si_cp_enable(rdev, false);
1859
1860         /* PFP */
1861         fw_data = (const __be32 *)rdev->pfp_fw->data;
1862         WREG32(CP_PFP_UCODE_ADDR, 0);
1863         for (i = 0; i < SI_PFP_UCODE_SIZE; i++)
1864                 WREG32(CP_PFP_UCODE_DATA, be32_to_cpup(fw_data++));
1865         WREG32(CP_PFP_UCODE_ADDR, 0);
1866
1867         /* CE */
1868         fw_data = (const __be32 *)rdev->ce_fw->data;
1869         WREG32(CP_CE_UCODE_ADDR, 0);
1870         for (i = 0; i < SI_CE_UCODE_SIZE; i++)
1871                 WREG32(CP_CE_UCODE_DATA, be32_to_cpup(fw_data++));
1872         WREG32(CP_CE_UCODE_ADDR, 0);
1873
1874         /* ME */
1875         fw_data = (const __be32 *)rdev->me_fw->data;
1876         WREG32(CP_ME_RAM_WADDR, 0);
1877         for (i = 0; i < SI_PM4_UCODE_SIZE; i++)
1878                 WREG32(CP_ME_RAM_DATA, be32_to_cpup(fw_data++));
1879         WREG32(CP_ME_RAM_WADDR, 0);
1880
1881         WREG32(CP_PFP_UCODE_ADDR, 0);
1882         WREG32(CP_CE_UCODE_ADDR, 0);
1883         WREG32(CP_ME_RAM_WADDR, 0);
1884         WREG32(CP_ME_RAM_RADDR, 0);
1885         return 0;
1886 }
1887
1888 static int si_cp_start(struct radeon_device *rdev)
1889 {
1890         struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
1891         int r, i;
1892
1893         r = radeon_ring_lock(rdev, ring, 7 + 4);
1894         if (r) {
1895                 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
1896                 return r;
1897         }
1898         /* init the CP */
1899         radeon_ring_write(ring, PACKET3(PACKET3_ME_INITIALIZE, 5));
1900         radeon_ring_write(ring, 0x1);
1901         radeon_ring_write(ring, 0x0);
1902         radeon_ring_write(ring, rdev->config.si.max_hw_contexts - 1);
1903         radeon_ring_write(ring, PACKET3_ME_INITIALIZE_DEVICE_ID(1));
1904         radeon_ring_write(ring, 0);
1905         radeon_ring_write(ring, 0);
1906
1907         /* init the CE partitions */
1908         radeon_ring_write(ring, PACKET3(PACKET3_SET_BASE, 2));
1909         radeon_ring_write(ring, PACKET3_BASE_INDEX(CE_PARTITION_BASE));
1910         radeon_ring_write(ring, 0xc000);
1911         radeon_ring_write(ring, 0xe000);
1912         radeon_ring_unlock_commit(rdev, ring);
1913
1914         si_cp_enable(rdev, true);
1915
1916         r = radeon_ring_lock(rdev, ring, si_default_size + 10);
1917         if (r) {
1918                 DRM_ERROR("radeon: cp failed to lock ring (%d).\n", r);
1919                 return r;
1920         }
1921
1922         /* setup clear context state */
1923         radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
1924         radeon_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
1925
1926         for (i = 0; i < si_default_size; i++)
1927                 radeon_ring_write(ring, si_default_state[i]);
1928
1929         radeon_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
1930         radeon_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
1931
1932         /* set clear context state */
1933         radeon_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
1934         radeon_ring_write(ring, 0);
1935
1936         radeon_ring_write(ring, PACKET3(PACKET3_SET_CONTEXT_REG, 2));
1937         radeon_ring_write(ring, 0x00000316);
1938         radeon_ring_write(ring, 0x0000000e); /* VGT_VERTEX_REUSE_BLOCK_CNTL */
1939         radeon_ring_write(ring, 0x00000010); /* VGT_OUT_DEALLOC_CNTL */
1940
1941         radeon_ring_unlock_commit(rdev, ring);
1942
1943         for (i = RADEON_RING_TYPE_GFX_INDEX; i <= CAYMAN_RING_TYPE_CP2_INDEX; ++i) {
1944                 ring = &rdev->ring[i];
1945                 r = radeon_ring_lock(rdev, ring, 2);
1946
1947                 /* clear the compute context state */
1948                 radeon_ring_write(ring, PACKET3_COMPUTE(PACKET3_CLEAR_STATE, 0));
1949                 radeon_ring_write(ring, 0);
1950
1951                 radeon_ring_unlock_commit(rdev, ring);
1952         }
1953
1954         return 0;
1955 }
1956
1957 static void si_cp_fini(struct radeon_device *rdev)
1958 {
1959         struct radeon_ring *ring;
1960         si_cp_enable(rdev, false);
1961
1962         ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
1963         radeon_ring_fini(rdev, ring);
1964         radeon_scratch_free(rdev, ring->rptr_save_reg);
1965
1966         ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
1967         radeon_ring_fini(rdev, ring);
1968         radeon_scratch_free(rdev, ring->rptr_save_reg);
1969
1970         ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
1971         radeon_ring_fini(rdev, ring);
1972         radeon_scratch_free(rdev, ring->rptr_save_reg);
1973 }
1974
1975 static int si_cp_resume(struct radeon_device *rdev)
1976 {
1977         struct radeon_ring *ring;
1978         u32 tmp;
1979         u32 rb_bufsz;
1980         int r;
1981
1982         /* Reset cp; if cp is reset, then PA, SH, VGT also need to be reset */
1983         WREG32(GRBM_SOFT_RESET, (SOFT_RESET_CP |
1984                                  SOFT_RESET_PA |
1985                                  SOFT_RESET_VGT |
1986                                  SOFT_RESET_SPI |
1987                                  SOFT_RESET_SX));
1988         RREG32(GRBM_SOFT_RESET);
1989         mdelay(15);
1990         WREG32(GRBM_SOFT_RESET, 0);
1991         RREG32(GRBM_SOFT_RESET);
1992
1993         WREG32(CP_SEM_WAIT_TIMER, 0x0);
1994         WREG32(CP_SEM_INCOMPLETE_TIMER_CNTL, 0x0);
1995
1996         /* Set the write pointer delay */
1997         WREG32(CP_RB_WPTR_DELAY, 0);
1998
1999         WREG32(CP_DEBUG, 0);
2000         WREG32(SCRATCH_ADDR, ((rdev->wb.gpu_addr + RADEON_WB_SCRATCH_OFFSET) >> 8) & 0xFFFFFFFF);
2001
2002         /* ring 0 - compute and gfx */
2003         /* Set ring buffer size */
2004         ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
2005         rb_bufsz = drm_order(ring->ring_size / 8);
2006         tmp = (drm_order(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
2007 #ifdef __BIG_ENDIAN
2008         tmp |= BUF_SWAP_32BIT;
2009 #endif
2010         WREG32(CP_RB0_CNTL, tmp);
2011
2012         /* Initialize the ring buffer's read and write pointers */
2013         WREG32(CP_RB0_CNTL, tmp | RB_RPTR_WR_ENA);
2014         ring->wptr = 0;
2015         WREG32(CP_RB0_WPTR, ring->wptr);
2016
2017         /* set the wb address whether it's enabled or not */
2018         WREG32(CP_RB0_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFFFFFFFC);
2019         WREG32(CP_RB0_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP_RPTR_OFFSET) & 0xFF);
2020
2021         if (rdev->wb.enabled)
2022                 WREG32(SCRATCH_UMSK, 0xff);
2023         else {
2024                 tmp |= RB_NO_UPDATE;
2025                 WREG32(SCRATCH_UMSK, 0);
2026         }
2027
2028         mdelay(1);
2029         WREG32(CP_RB0_CNTL, tmp);
2030
2031         WREG32(CP_RB0_BASE, ring->gpu_addr >> 8);
2032
2033         ring->rptr = RREG32(CP_RB0_RPTR);
2034
2035         /* ring1  - compute only */
2036         /* Set ring buffer size */
2037         ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
2038         rb_bufsz = drm_order(ring->ring_size / 8);
2039         tmp = (drm_order(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
2040 #ifdef __BIG_ENDIAN
2041         tmp |= BUF_SWAP_32BIT;
2042 #endif
2043         WREG32(CP_RB1_CNTL, tmp);
2044
2045         /* Initialize the ring buffer's read and write pointers */
2046         WREG32(CP_RB1_CNTL, tmp | RB_RPTR_WR_ENA);
2047         ring->wptr = 0;
2048         WREG32(CP_RB1_WPTR, ring->wptr);
2049
2050         /* set the wb address whether it's enabled or not */
2051         WREG32(CP_RB1_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET) & 0xFFFFFFFC);
2052         WREG32(CP_RB1_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP1_RPTR_OFFSET) & 0xFF);
2053
2054         mdelay(1);
2055         WREG32(CP_RB1_CNTL, tmp);
2056
2057         WREG32(CP_RB1_BASE, ring->gpu_addr >> 8);
2058
2059         ring->rptr = RREG32(CP_RB1_RPTR);
2060
2061         /* ring2 - compute only */
2062         /* Set ring buffer size */
2063         ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
2064         rb_bufsz = drm_order(ring->ring_size / 8);
2065         tmp = (drm_order(RADEON_GPU_PAGE_SIZE/8) << 8) | rb_bufsz;
2066 #ifdef __BIG_ENDIAN
2067         tmp |= BUF_SWAP_32BIT;
2068 #endif
2069         WREG32(CP_RB2_CNTL, tmp);
2070
2071         /* Initialize the ring buffer's read and write pointers */
2072         WREG32(CP_RB2_CNTL, tmp | RB_RPTR_WR_ENA);
2073         ring->wptr = 0;
2074         WREG32(CP_RB2_WPTR, ring->wptr);
2075
2076         /* set the wb address whether it's enabled or not */
2077         WREG32(CP_RB2_RPTR_ADDR, (rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET) & 0xFFFFFFFC);
2078         WREG32(CP_RB2_RPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + RADEON_WB_CP2_RPTR_OFFSET) & 0xFF);
2079
2080         mdelay(1);
2081         WREG32(CP_RB2_CNTL, tmp);
2082
2083         WREG32(CP_RB2_BASE, ring->gpu_addr >> 8);
2084
2085         ring->rptr = RREG32(CP_RB2_RPTR);
2086
2087         /* start the rings */
2088         si_cp_start(rdev);
2089         rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = true;
2090         rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = true;
2091         rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = true;
2092         r = radeon_ring_test(rdev, RADEON_RING_TYPE_GFX_INDEX, &rdev->ring[RADEON_RING_TYPE_GFX_INDEX]);
2093         if (r) {
2094                 rdev->ring[RADEON_RING_TYPE_GFX_INDEX].ready = false;
2095                 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
2096                 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
2097                 return r;
2098         }
2099         r = radeon_ring_test(rdev, CAYMAN_RING_TYPE_CP1_INDEX, &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX]);
2100         if (r) {
2101                 rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX].ready = false;
2102         }
2103         r = radeon_ring_test(rdev, CAYMAN_RING_TYPE_CP2_INDEX, &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX]);
2104         if (r) {
2105                 rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX].ready = false;
2106         }
2107
2108         return 0;
2109 }
2110
2111 bool si_gpu_is_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
2112 {
2113         u32 srbm_status;
2114         u32 grbm_status, grbm_status2;
2115         u32 grbm_status_se0, grbm_status_se1;
2116
2117         srbm_status = RREG32(SRBM_STATUS);
2118         grbm_status = RREG32(GRBM_STATUS);
2119         grbm_status2 = RREG32(GRBM_STATUS2);
2120         grbm_status_se0 = RREG32(GRBM_STATUS_SE0);
2121         grbm_status_se1 = RREG32(GRBM_STATUS_SE1);
2122         if (!(grbm_status & GUI_ACTIVE)) {
2123                 radeon_ring_lockup_update(ring);
2124                 return false;
2125         }
2126         /* force CP activities */
2127         radeon_ring_force_activity(rdev, ring);
2128         return radeon_ring_test_lockup(rdev, ring);
2129 }
2130
2131 static u32 si_gpu_check_soft_reset(struct radeon_device *rdev)
2132 {
2133         u32 reset_mask = 0;
2134         u32 tmp;
2135
2136         /* GRBM_STATUS */
2137         tmp = RREG32(GRBM_STATUS);
2138         if (tmp & (PA_BUSY | SC_BUSY |
2139                    BCI_BUSY | SX_BUSY |
2140                    TA_BUSY | VGT_BUSY |
2141                    DB_BUSY | CB_BUSY |
2142                    GDS_BUSY | SPI_BUSY |
2143                    IA_BUSY | IA_BUSY_NO_DMA))
2144                 reset_mask |= RADEON_RESET_GFX;
2145
2146         if (tmp & (CF_RQ_PENDING | PF_RQ_PENDING |
2147                    CP_BUSY | CP_COHERENCY_BUSY))
2148                 reset_mask |= RADEON_RESET_CP;
2149
2150         if (tmp & GRBM_EE_BUSY)
2151                 reset_mask |= RADEON_RESET_GRBM | RADEON_RESET_GFX | RADEON_RESET_CP;
2152
2153         /* GRBM_STATUS2 */
2154         tmp = RREG32(GRBM_STATUS2);
2155         if (tmp & (RLC_RQ_PENDING | RLC_BUSY))
2156                 reset_mask |= RADEON_RESET_RLC;
2157
2158         /* DMA_STATUS_REG 0 */
2159         tmp = RREG32(DMA_STATUS_REG + DMA0_REGISTER_OFFSET);
2160         if (!(tmp & DMA_IDLE))
2161                 reset_mask |= RADEON_RESET_DMA;
2162
2163         /* DMA_STATUS_REG 1 */
2164         tmp = RREG32(DMA_STATUS_REG + DMA1_REGISTER_OFFSET);
2165         if (!(tmp & DMA_IDLE))
2166                 reset_mask |= RADEON_RESET_DMA1;
2167
2168         /* SRBM_STATUS2 */
2169         tmp = RREG32(SRBM_STATUS2);
2170         if (tmp & DMA_BUSY)
2171                 reset_mask |= RADEON_RESET_DMA;
2172
2173         if (tmp & DMA1_BUSY)
2174                 reset_mask |= RADEON_RESET_DMA1;
2175
2176         /* SRBM_STATUS */
2177         tmp = RREG32(SRBM_STATUS);
2178
2179         if (tmp & IH_BUSY)
2180                 reset_mask |= RADEON_RESET_IH;
2181
2182         if (tmp & SEM_BUSY)
2183                 reset_mask |= RADEON_RESET_SEM;
2184
2185         if (tmp & GRBM_RQ_PENDING)
2186                 reset_mask |= RADEON_RESET_GRBM;
2187
2188         if (tmp & VMC_BUSY)
2189                 reset_mask |= RADEON_RESET_VMC;
2190
2191         if (tmp & (MCB_BUSY | MCB_NON_DISPLAY_BUSY |
2192                    MCC_BUSY | MCD_BUSY))
2193                 reset_mask |= RADEON_RESET_MC;
2194
2195         if (evergreen_is_display_hung(rdev))
2196                 reset_mask |= RADEON_RESET_DISPLAY;
2197
2198         /* VM_L2_STATUS */
2199         tmp = RREG32(VM_L2_STATUS);
2200         if (tmp & L2_BUSY)
2201                 reset_mask |= RADEON_RESET_VMC;
2202
2203         return reset_mask;
2204 }
2205
2206 static void si_gpu_soft_reset(struct radeon_device *rdev, u32 reset_mask)
2207 {
2208         struct evergreen_mc_save save;
2209         u32 grbm_soft_reset = 0, srbm_soft_reset = 0;
2210         u32 tmp;
2211
2212         if (reset_mask == 0)
2213                 return;
2214
2215         dev_info(rdev->dev, "GPU softreset: 0x%08X\n", reset_mask);
2216
2217         evergreen_print_gpu_status_regs(rdev);
2218         dev_info(rdev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_ADDR   0x%08X\n",
2219                  RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR));
2220         dev_info(rdev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
2221                  RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS));
2222
2223         evergreen_mc_stop(rdev, &save);
2224         if (evergreen_mc_wait_for_idle(rdev)) {
2225                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
2226         }
2227
2228         /* Disable CP parsing/prefetching */
2229         WREG32(CP_ME_CNTL, CP_ME_HALT | CP_PFP_HALT | CP_CE_HALT);
2230
2231         if (reset_mask & RADEON_RESET_DMA) {
2232                 /* dma0 */
2233                 tmp = RREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET);
2234                 tmp &= ~DMA_RB_ENABLE;
2235                 WREG32(DMA_RB_CNTL + DMA0_REGISTER_OFFSET, tmp);
2236         }
2237         if (reset_mask & RADEON_RESET_DMA1) {
2238                 /* dma1 */
2239                 tmp = RREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET);
2240                 tmp &= ~DMA_RB_ENABLE;
2241                 WREG32(DMA_RB_CNTL + DMA1_REGISTER_OFFSET, tmp);
2242         }
2243
2244         if (reset_mask & (RADEON_RESET_GFX | RADEON_RESET_COMPUTE | RADEON_RESET_CP)) {
2245                 grbm_soft_reset = SOFT_RESET_CB |
2246                         SOFT_RESET_DB |
2247                         SOFT_RESET_GDS |
2248                         SOFT_RESET_PA |
2249                         SOFT_RESET_SC |
2250                         SOFT_RESET_BCI |
2251                         SOFT_RESET_SPI |
2252                         SOFT_RESET_SX |
2253                         SOFT_RESET_TC |
2254                         SOFT_RESET_TA |
2255                         SOFT_RESET_VGT |
2256                         SOFT_RESET_IA;
2257         }
2258
2259         if (reset_mask & RADEON_RESET_CP) {
2260                 grbm_soft_reset |= SOFT_RESET_CP | SOFT_RESET_VGT;
2261
2262                 srbm_soft_reset |= SOFT_RESET_GRBM;
2263         }
2264
2265         if (reset_mask & RADEON_RESET_DMA)
2266                 srbm_soft_reset |= SOFT_RESET_DMA;
2267
2268         if (reset_mask & RADEON_RESET_DMA1)
2269                 srbm_soft_reset |= SOFT_RESET_DMA1;
2270
2271         if (reset_mask & RADEON_RESET_DISPLAY)
2272                 srbm_soft_reset |= SOFT_RESET_DC;
2273
2274         if (reset_mask & RADEON_RESET_RLC)
2275                 grbm_soft_reset |= SOFT_RESET_RLC;
2276
2277         if (reset_mask & RADEON_RESET_SEM)
2278                 srbm_soft_reset |= SOFT_RESET_SEM;
2279
2280         if (reset_mask & RADEON_RESET_IH)
2281                 srbm_soft_reset |= SOFT_RESET_IH;
2282
2283         if (reset_mask & RADEON_RESET_GRBM)
2284                 srbm_soft_reset |= SOFT_RESET_GRBM;
2285
2286         if (reset_mask & RADEON_RESET_VMC)
2287                 srbm_soft_reset |= SOFT_RESET_VMC;
2288
2289         if (reset_mask & RADEON_RESET_MC)
2290                 srbm_soft_reset |= SOFT_RESET_MC;
2291
2292         if (grbm_soft_reset) {
2293                 tmp = RREG32(GRBM_SOFT_RESET);
2294                 tmp |= grbm_soft_reset;
2295                 dev_info(rdev->dev, "GRBM_SOFT_RESET=0x%08X\n", tmp);
2296                 WREG32(GRBM_SOFT_RESET, tmp);
2297                 tmp = RREG32(GRBM_SOFT_RESET);
2298
2299                 udelay(50);
2300
2301                 tmp &= ~grbm_soft_reset;
2302                 WREG32(GRBM_SOFT_RESET, tmp);
2303                 tmp = RREG32(GRBM_SOFT_RESET);
2304         }
2305
2306         if (srbm_soft_reset) {
2307                 tmp = RREG32(SRBM_SOFT_RESET);
2308                 tmp |= srbm_soft_reset;
2309                 dev_info(rdev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
2310                 WREG32(SRBM_SOFT_RESET, tmp);
2311                 tmp = RREG32(SRBM_SOFT_RESET);
2312
2313                 udelay(50);
2314
2315                 tmp &= ~srbm_soft_reset;
2316                 WREG32(SRBM_SOFT_RESET, tmp);
2317                 tmp = RREG32(SRBM_SOFT_RESET);
2318         }
2319
2320         /* Wait a little for things to settle down */
2321         udelay(50);
2322
2323         evergreen_mc_resume(rdev, &save);
2324         udelay(50);
2325
2326         evergreen_print_gpu_status_regs(rdev);
2327 }
2328
2329 int si_asic_reset(struct radeon_device *rdev)
2330 {
2331         u32 reset_mask;
2332
2333         reset_mask = si_gpu_check_soft_reset(rdev);
2334
2335         if (reset_mask)
2336                 r600_set_bios_scratch_engine_hung(rdev, true);
2337
2338         si_gpu_soft_reset(rdev, reset_mask);
2339
2340         reset_mask = si_gpu_check_soft_reset(rdev);
2341
2342         if (!reset_mask)
2343                 r600_set_bios_scratch_engine_hung(rdev, false);
2344
2345         return 0;
2346 }
2347
2348 /* MC */
2349 static void si_mc_program(struct radeon_device *rdev)
2350 {
2351         struct evergreen_mc_save save;
2352         u32 tmp;
2353         int i, j;
2354
2355         /* Initialize HDP */
2356         for (i = 0, j = 0; i < 32; i++, j += 0x18) {
2357                 WREG32((0x2c14 + j), 0x00000000);
2358                 WREG32((0x2c18 + j), 0x00000000);
2359                 WREG32((0x2c1c + j), 0x00000000);
2360                 WREG32((0x2c20 + j), 0x00000000);
2361                 WREG32((0x2c24 + j), 0x00000000);
2362         }
2363         WREG32(HDP_REG_COHERENCY_FLUSH_CNTL, 0);
2364
2365         evergreen_mc_stop(rdev, &save);
2366         if (radeon_mc_wait_for_idle(rdev)) {
2367                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
2368         }
2369         /* Lockout access through VGA aperture*/
2370         WREG32(VGA_HDP_CONTROL, VGA_MEMORY_DISABLE);
2371         /* Update configuration */
2372         WREG32(MC_VM_SYSTEM_APERTURE_LOW_ADDR,
2373                rdev->mc.vram_start >> 12);
2374         WREG32(MC_VM_SYSTEM_APERTURE_HIGH_ADDR,
2375                rdev->mc.vram_end >> 12);
2376         WREG32(MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
2377                rdev->vram_scratch.gpu_addr >> 12);
2378         tmp = ((rdev->mc.vram_end >> 24) & 0xFFFF) << 16;
2379         tmp |= ((rdev->mc.vram_start >> 24) & 0xFFFF);
2380         WREG32(MC_VM_FB_LOCATION, tmp);
2381         /* XXX double check these! */
2382         WREG32(HDP_NONSURFACE_BASE, (rdev->mc.vram_start >> 8));
2383         WREG32(HDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
2384         WREG32(HDP_NONSURFACE_SIZE, 0x3FFFFFFF);
2385         WREG32(MC_VM_AGP_BASE, 0);
2386         WREG32(MC_VM_AGP_TOP, 0x0FFFFFFF);
2387         WREG32(MC_VM_AGP_BOT, 0x0FFFFFFF);
2388         if (radeon_mc_wait_for_idle(rdev)) {
2389                 dev_warn(rdev->dev, "Wait for MC idle timedout !\n");
2390         }
2391         evergreen_mc_resume(rdev, &save);
2392         /* we need to own VRAM, so turn off the VGA renderer here
2393          * to stop it overwriting our objects */
2394         rv515_vga_render_disable(rdev);
2395 }
2396
2397 /* SI MC address space is 40 bits */
2398 static void si_vram_location(struct radeon_device *rdev,
2399                              struct radeon_mc *mc, u64 base)
2400 {
2401         mc->vram_start = base;
2402         if (mc->mc_vram_size > (0xFFFFFFFFFFULL - base + 1)) {
2403                 dev_warn(rdev->dev, "limiting VRAM to PCI aperture size\n");
2404                 mc->real_vram_size = mc->aper_size;
2405                 mc->mc_vram_size = mc->aper_size;
2406         }
2407         mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
2408         dev_info(rdev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
2409                         mc->mc_vram_size >> 20, mc->vram_start,
2410                         mc->vram_end, mc->real_vram_size >> 20);
2411 }
2412
2413 static void si_gtt_location(struct radeon_device *rdev, struct radeon_mc *mc)
2414 {
2415         u64 size_af, size_bf;
2416
2417         size_af = ((0xFFFFFFFFFFULL - mc->vram_end) + mc->gtt_base_align) & ~mc->gtt_base_align;
2418         size_bf = mc->vram_start & ~mc->gtt_base_align;
2419         if (size_bf > size_af) {
2420                 if (mc->gtt_size > size_bf) {
2421                         dev_warn(rdev->dev, "limiting GTT\n");
2422                         mc->gtt_size = size_bf;
2423                 }
2424                 mc->gtt_start = (mc->vram_start & ~mc->gtt_base_align) - mc->gtt_size;
2425         } else {
2426                 if (mc->gtt_size > size_af) {
2427                         dev_warn(rdev->dev, "limiting GTT\n");
2428                         mc->gtt_size = size_af;
2429                 }
2430                 mc->gtt_start = (mc->vram_end + 1 + mc->gtt_base_align) & ~mc->gtt_base_align;
2431         }
2432         mc->gtt_end = mc->gtt_start + mc->gtt_size - 1;
2433         dev_info(rdev->dev, "GTT: %lluM 0x%016llX - 0x%016llX\n",
2434                         mc->gtt_size >> 20, mc->gtt_start, mc->gtt_end);
2435 }
2436
2437 static void si_vram_gtt_location(struct radeon_device *rdev,
2438                                  struct radeon_mc *mc)
2439 {
2440         if (mc->mc_vram_size > 0xFFC0000000ULL) {
2441                 /* leave room for at least 1024M GTT */
2442                 dev_warn(rdev->dev, "limiting VRAM\n");
2443                 mc->real_vram_size = 0xFFC0000000ULL;
2444                 mc->mc_vram_size = 0xFFC0000000ULL;
2445         }
2446         si_vram_location(rdev, &rdev->mc, 0);
2447         rdev->mc.gtt_base_align = 0;
2448         si_gtt_location(rdev, mc);
2449 }
2450
2451 static int si_mc_init(struct radeon_device *rdev)
2452 {
2453         u32 tmp;
2454         int chansize, numchan;
2455
2456         /* Get VRAM informations */
2457         rdev->mc.vram_is_ddr = true;
2458         tmp = RREG32(MC_ARB_RAMCFG);
2459         if (tmp & CHANSIZE_OVERRIDE) {
2460                 chansize = 16;
2461         } else if (tmp & CHANSIZE_MASK) {
2462                 chansize = 64;
2463         } else {
2464                 chansize = 32;
2465         }
2466         tmp = RREG32(MC_SHARED_CHMAP);
2467         switch ((tmp & NOOFCHAN_MASK) >> NOOFCHAN_SHIFT) {
2468         case 0:
2469         default:
2470                 numchan = 1;
2471                 break;
2472         case 1:
2473                 numchan = 2;
2474                 break;
2475         case 2:
2476                 numchan = 4;
2477                 break;
2478         case 3:
2479                 numchan = 8;
2480                 break;
2481         case 4:
2482                 numchan = 3;
2483                 break;
2484         case 5:
2485                 numchan = 6;
2486                 break;
2487         case 6:
2488                 numchan = 10;
2489                 break;
2490         case 7:
2491                 numchan = 12;
2492                 break;
2493         case 8:
2494                 numchan = 16;
2495                 break;
2496         }
2497         rdev->mc.vram_width = numchan * chansize;
2498         /* Could aper size report 0 ? */
2499         rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
2500         rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
2501         /* size in MB on si */
2502         rdev->mc.mc_vram_size = RREG32(CONFIG_MEMSIZE) * 1024 * 1024;
2503         rdev->mc.real_vram_size = RREG32(CONFIG_MEMSIZE) * 1024 * 1024;
2504         rdev->mc.visible_vram_size = rdev->mc.aper_size;
2505         si_vram_gtt_location(rdev, &rdev->mc);
2506         radeon_update_bandwidth_info(rdev);
2507
2508         return 0;
2509 }
2510
2511 /*
2512  * GART
2513  */
2514 void si_pcie_gart_tlb_flush(struct radeon_device *rdev)
2515 {
2516         /* flush hdp cache */
2517         WREG32(HDP_MEM_COHERENCY_FLUSH_CNTL, 0x1);
2518
2519         /* bits 0-15 are the VM contexts0-15 */
2520         WREG32(VM_INVALIDATE_REQUEST, 1);
2521 }
2522
2523 static int si_pcie_gart_enable(struct radeon_device *rdev)
2524 {
2525         int r, i;
2526
2527         if (rdev->gart.robj == NULL) {
2528                 dev_err(rdev->dev, "No VRAM object for PCIE GART.\n");
2529                 return -EINVAL;
2530         }
2531         r = radeon_gart_table_vram_pin(rdev);
2532         if (r)
2533                 return r;
2534         radeon_gart_restore(rdev);
2535         /* Setup TLB control */
2536         WREG32(MC_VM_MX_L1_TLB_CNTL,
2537                (0xA << 7) |
2538                ENABLE_L1_TLB |
2539                SYSTEM_ACCESS_MODE_NOT_IN_SYS |
2540                ENABLE_ADVANCED_DRIVER_MODEL |
2541                SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
2542         /* Setup L2 cache */
2543         WREG32(VM_L2_CNTL, ENABLE_L2_CACHE |
2544                ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
2545                ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
2546                EFFECTIVE_L2_QUEUE_SIZE(7) |
2547                CONTEXT1_IDENTITY_ACCESS_MODE(1));
2548         WREG32(VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS | INVALIDATE_L2_CACHE);
2549         WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
2550                L2_CACHE_BIGK_FRAGMENT_SIZE(0));
2551         /* setup context0 */
2552         WREG32(VM_CONTEXT0_PAGE_TABLE_START_ADDR, rdev->mc.gtt_start >> 12);
2553         WREG32(VM_CONTEXT0_PAGE_TABLE_END_ADDR, rdev->mc.gtt_end >> 12);
2554         WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR, rdev->gart.table_addr >> 12);
2555         WREG32(VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
2556                         (u32)(rdev->dummy_page.addr >> 12));
2557         WREG32(VM_CONTEXT0_CNTL2, 0);
2558         WREG32(VM_CONTEXT0_CNTL, (ENABLE_CONTEXT | PAGE_TABLE_DEPTH(0) |
2559                                   RANGE_PROTECTION_FAULT_ENABLE_DEFAULT));
2560
2561         WREG32(0x15D4, 0);
2562         WREG32(0x15D8, 0);
2563         WREG32(0x15DC, 0);
2564
2565         /* empty context1-15 */
2566         /* set vm size, must be a multiple of 4 */
2567         WREG32(VM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
2568         WREG32(VM_CONTEXT1_PAGE_TABLE_END_ADDR, rdev->vm_manager.max_pfn);
2569         /* Assign the pt base to something valid for now; the pts used for
2570          * the VMs are determined by the application and setup and assigned
2571          * on the fly in the vm part of radeon_gart.c
2572          */
2573         for (i = 1; i < 16; i++) {
2574                 if (i < 8)
2575                         WREG32(VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (i << 2),
2576                                rdev->gart.table_addr >> 12);
2577                 else
2578                         WREG32(VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((i - 8) << 2),
2579                                rdev->gart.table_addr >> 12);
2580         }
2581
2582         /* enable context1-15 */
2583         WREG32(VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
2584                (u32)(rdev->dummy_page.addr >> 12));
2585         WREG32(VM_CONTEXT1_CNTL2, 4);
2586         WREG32(VM_CONTEXT1_CNTL, ENABLE_CONTEXT | PAGE_TABLE_DEPTH(1) |
2587                                 RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
2588                                 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT |
2589                                 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT |
2590                                 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT |
2591                                 PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT |
2592                                 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT |
2593                                 VALID_PROTECTION_FAULT_ENABLE_INTERRUPT |
2594                                 VALID_PROTECTION_FAULT_ENABLE_DEFAULT |
2595                                 READ_PROTECTION_FAULT_ENABLE_INTERRUPT |
2596                                 READ_PROTECTION_FAULT_ENABLE_DEFAULT |
2597                                 WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT |
2598                                 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT);
2599
2600         si_pcie_gart_tlb_flush(rdev);
2601         DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
2602                  (unsigned)(rdev->mc.gtt_size >> 20),
2603                  (unsigned long long)rdev->gart.table_addr);
2604         rdev->gart.ready = true;
2605         return 0;
2606 }
2607
2608 static void si_pcie_gart_disable(struct radeon_device *rdev)
2609 {
2610         /* Disable all tables */
2611         WREG32(VM_CONTEXT0_CNTL, 0);
2612         WREG32(VM_CONTEXT1_CNTL, 0);
2613         /* Setup TLB control */
2614         WREG32(MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE_NOT_IN_SYS |
2615                SYSTEM_APERTURE_UNMAPPED_ACCESS_PASS_THRU);
2616         /* Setup L2 cache */
2617         WREG32(VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE |
2618                ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE |
2619                EFFECTIVE_L2_QUEUE_SIZE(7) |
2620                CONTEXT1_IDENTITY_ACCESS_MODE(1));
2621         WREG32(VM_L2_CNTL2, 0);
2622         WREG32(VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY |
2623                L2_CACHE_BIGK_FRAGMENT_SIZE(0));
2624         radeon_gart_table_vram_unpin(rdev);
2625 }
2626
2627 static void si_pcie_gart_fini(struct radeon_device *rdev)
2628 {
2629         si_pcie_gart_disable(rdev);
2630         radeon_gart_table_vram_free(rdev);
2631         radeon_gart_fini(rdev);
2632 }
2633
2634 /* vm parser */
2635 static bool si_vm_reg_valid(u32 reg)
2636 {
2637         /* context regs are fine */
2638         if (reg >= 0x28000)
2639                 return true;
2640
2641         /* check config regs */
2642         switch (reg) {
2643         case GRBM_GFX_INDEX:
2644         case CP_STRMOUT_CNTL:
2645         case VGT_VTX_VECT_EJECT_REG:
2646         case VGT_CACHE_INVALIDATION:
2647         case VGT_ESGS_RING_SIZE:
2648         case VGT_GSVS_RING_SIZE:
2649         case VGT_GS_VERTEX_REUSE:
2650         case VGT_PRIMITIVE_TYPE:
2651         case VGT_INDEX_TYPE:
2652         case VGT_NUM_INDICES:
2653         case VGT_NUM_INSTANCES:
2654         case VGT_TF_RING_SIZE:
2655         case VGT_HS_OFFCHIP_PARAM:
2656         case VGT_TF_MEMORY_BASE:
2657         case PA_CL_ENHANCE:
2658         case PA_SU_LINE_STIPPLE_VALUE:
2659         case PA_SC_LINE_STIPPLE_STATE:
2660         case PA_SC_ENHANCE:
2661         case SQC_CACHES:
2662         case SPI_STATIC_THREAD_MGMT_1:
2663         case SPI_STATIC_THREAD_MGMT_2:
2664         case SPI_STATIC_THREAD_MGMT_3:
2665         case SPI_PS_MAX_WAVE_ID:
2666         case SPI_CONFIG_CNTL:
2667         case SPI_CONFIG_CNTL_1:
2668         case TA_CNTL_AUX:
2669                 return true;
2670         default:
2671                 DRM_ERROR("Invalid register 0x%x in CS\n", reg);
2672                 return false;
2673         }
2674 }
2675
2676 static int si_vm_packet3_ce_check(struct radeon_device *rdev,
2677                                   u32 *ib, struct radeon_cs_packet *pkt)
2678 {
2679         switch (pkt->opcode) {
2680         case PACKET3_NOP:
2681         case PACKET3_SET_BASE:
2682         case PACKET3_SET_CE_DE_COUNTERS:
2683         case PACKET3_LOAD_CONST_RAM:
2684         case PACKET3_WRITE_CONST_RAM:
2685         case PACKET3_WRITE_CONST_RAM_OFFSET:
2686         case PACKET3_DUMP_CONST_RAM:
2687         case PACKET3_INCREMENT_CE_COUNTER:
2688         case PACKET3_WAIT_ON_DE_COUNTER:
2689         case PACKET3_CE_WRITE:
2690                 break;
2691         default:
2692                 DRM_ERROR("Invalid CE packet3: 0x%x\n", pkt->opcode);
2693                 return -EINVAL;
2694         }
2695         return 0;
2696 }
2697
2698 static int si_vm_packet3_gfx_check(struct radeon_device *rdev,
2699                                    u32 *ib, struct radeon_cs_packet *pkt)
2700 {
2701         u32 idx = pkt->idx + 1;
2702         u32 idx_value = ib[idx];
2703         u32 start_reg, end_reg, reg, i;
2704         u32 command, info;
2705
2706         switch (pkt->opcode) {
2707         case PACKET3_NOP:
2708         case PACKET3_SET_BASE:
2709         case PACKET3_CLEAR_STATE:
2710         case PACKET3_INDEX_BUFFER_SIZE:
2711         case PACKET3_DISPATCH_DIRECT:
2712         case PACKET3_DISPATCH_INDIRECT:
2713         case PACKET3_ALLOC_GDS:
2714         case PACKET3_WRITE_GDS_RAM:
2715         case PACKET3_ATOMIC_GDS:
2716         case PACKET3_ATOMIC:
2717         case PACKET3_OCCLUSION_QUERY:
2718         case PACKET3_SET_PREDICATION:
2719         case PACKET3_COND_EXEC:
2720         case PACKET3_PRED_EXEC:
2721         case PACKET3_DRAW_INDIRECT:
2722         case PACKET3_DRAW_INDEX_INDIRECT:
2723         case PACKET3_INDEX_BASE:
2724         case PACKET3_DRAW_INDEX_2:
2725         case PACKET3_CONTEXT_CONTROL:
2726         case PACKET3_INDEX_TYPE:
2727         case PACKET3_DRAW_INDIRECT_MULTI:
2728         case PACKET3_DRAW_INDEX_AUTO:
2729         case PACKET3_DRAW_INDEX_IMMD:
2730         case PACKET3_NUM_INSTANCES:
2731         case PACKET3_DRAW_INDEX_MULTI_AUTO:
2732         case PACKET3_STRMOUT_BUFFER_UPDATE:
2733         case PACKET3_DRAW_INDEX_OFFSET_2:
2734         case PACKET3_DRAW_INDEX_MULTI_ELEMENT:
2735         case PACKET3_DRAW_INDEX_INDIRECT_MULTI:
2736         case PACKET3_MPEG_INDEX:
2737         case PACKET3_WAIT_REG_MEM:
2738         case PACKET3_MEM_WRITE:
2739         case PACKET3_PFP_SYNC_ME:
2740         case PACKET3_SURFACE_SYNC:
2741         case PACKET3_EVENT_WRITE:
2742         case PACKET3_EVENT_WRITE_EOP:
2743         case PACKET3_EVENT_WRITE_EOS:
2744         case PACKET3_SET_CONTEXT_REG:
2745         case PACKET3_SET_CONTEXT_REG_INDIRECT:
2746         case PACKET3_SET_SH_REG:
2747         case PACKET3_SET_SH_REG_OFFSET:
2748         case PACKET3_INCREMENT_DE_COUNTER:
2749         case PACKET3_WAIT_ON_CE_COUNTER:
2750         case PACKET3_WAIT_ON_AVAIL_BUFFER:
2751         case PACKET3_ME_WRITE:
2752                 break;
2753         case PACKET3_COPY_DATA:
2754                 if ((idx_value & 0xf00) == 0) {
2755                         reg = ib[idx + 3] * 4;
2756                         if (!si_vm_reg_valid(reg))
2757                                 return -EINVAL;
2758                 }
2759                 break;
2760         case PACKET3_WRITE_DATA:
2761                 if ((idx_value & 0xf00) == 0) {
2762                         start_reg = ib[idx + 1] * 4;
2763                         if (idx_value & 0x10000) {
2764                                 if (!si_vm_reg_valid(start_reg))
2765                                         return -EINVAL;
2766                         } else {
2767                                 for (i = 0; i < (pkt->count - 2); i++) {
2768                                         reg = start_reg + (4 * i);
2769                                         if (!si_vm_reg_valid(reg))
2770                                                 return -EINVAL;
2771                                 }
2772                         }
2773                 }
2774                 break;
2775         case PACKET3_COND_WRITE:
2776                 if (idx_value & 0x100) {
2777                         reg = ib[idx + 5] * 4;
2778                         if (!si_vm_reg_valid(reg))
2779                                 return -EINVAL;
2780                 }
2781                 break;
2782         case PACKET3_COPY_DW:
2783                 if (idx_value & 0x2) {
2784                         reg = ib[idx + 3] * 4;
2785                         if (!si_vm_reg_valid(reg))
2786                                 return -EINVAL;
2787                 }
2788                 break;
2789         case PACKET3_SET_CONFIG_REG:
2790                 start_reg = (idx_value << 2) + PACKET3_SET_CONFIG_REG_START;
2791                 end_reg = 4 * pkt->count + start_reg - 4;
2792                 if ((start_reg < PACKET3_SET_CONFIG_REG_START) ||
2793                     (start_reg >= PACKET3_SET_CONFIG_REG_END) ||
2794                     (end_reg >= PACKET3_SET_CONFIG_REG_END)) {
2795                         DRM_ERROR("bad PACKET3_SET_CONFIG_REG\n");
2796                         return -EINVAL;
2797                 }
2798                 for (i = 0; i < pkt->count; i++) {
2799                         reg = start_reg + (4 * i);
2800                         if (!si_vm_reg_valid(reg))
2801                                 return -EINVAL;
2802                 }
2803                 break;
2804         case PACKET3_CP_DMA:
2805                 command = ib[idx + 4];
2806                 info = ib[idx + 1];
2807                 if (command & PACKET3_CP_DMA_CMD_SAS) {
2808                         /* src address space is register */
2809                         if (((info & 0x60000000) >> 29) == 0) {
2810                                 start_reg = idx_value << 2;
2811                                 if (command & PACKET3_CP_DMA_CMD_SAIC) {
2812                                         reg = start_reg;
2813                                         if (!si_vm_reg_valid(reg)) {
2814                                                 DRM_ERROR("CP DMA Bad SRC register\n");
2815                                                 return -EINVAL;
2816                                         }
2817                                 } else {
2818                                         for (i = 0; i < (command & 0x1fffff); i++) {
2819                                                 reg = start_reg + (4 * i);
2820                                                 if (!si_vm_reg_valid(reg)) {
2821                                                         DRM_ERROR("CP DMA Bad SRC register\n");
2822                                                         return -EINVAL;
2823                                                 }
2824                                         }
2825                                 }
2826                         }
2827                 }
2828                 if (command & PACKET3_CP_DMA_CMD_DAS) {
2829                         /* dst address space is register */
2830                         if (((info & 0x00300000) >> 20) == 0) {
2831                                 start_reg = ib[idx + 2];
2832                                 if (command & PACKET3_CP_DMA_CMD_DAIC) {
2833                                         reg = start_reg;
2834                                         if (!si_vm_reg_valid(reg)) {
2835                                                 DRM_ERROR("CP DMA Bad DST register\n");
2836                                                 return -EINVAL;
2837                                         }
2838                                 } else {
2839                                         for (i = 0; i < (command & 0x1fffff); i++) {
2840                                                 reg = start_reg + (4 * i);
2841                                                 if (!si_vm_reg_valid(reg)) {
2842                                                         DRM_ERROR("CP DMA Bad DST register\n");
2843                                                         return -EINVAL;
2844                                                 }
2845                                         }
2846                                 }
2847                         }
2848                 }
2849                 break;
2850         default:
2851                 DRM_ERROR("Invalid GFX packet3: 0x%x\n", pkt->opcode);
2852                 return -EINVAL;
2853         }
2854         return 0;
2855 }
2856
2857 static int si_vm_packet3_compute_check(struct radeon_device *rdev,
2858                                        u32 *ib, struct radeon_cs_packet *pkt)
2859 {
2860         u32 idx = pkt->idx + 1;
2861         u32 idx_value = ib[idx];
2862         u32 start_reg, reg, i;
2863
2864         switch (pkt->opcode) {
2865         case PACKET3_NOP:
2866         case PACKET3_SET_BASE:
2867         case PACKET3_CLEAR_STATE:
2868         case PACKET3_DISPATCH_DIRECT:
2869         case PACKET3_DISPATCH_INDIRECT:
2870         case PACKET3_ALLOC_GDS:
2871         case PACKET3_WRITE_GDS_RAM:
2872         case PACKET3_ATOMIC_GDS:
2873         case PACKET3_ATOMIC:
2874         case PACKET3_OCCLUSION_QUERY:
2875         case PACKET3_SET_PREDICATION:
2876         case PACKET3_COND_EXEC:
2877         case PACKET3_PRED_EXEC:
2878         case PACKET3_CONTEXT_CONTROL:
2879         case PACKET3_STRMOUT_BUFFER_UPDATE:
2880         case PACKET3_WAIT_REG_MEM:
2881         case PACKET3_MEM_WRITE:
2882         case PACKET3_PFP_SYNC_ME:
2883         case PACKET3_SURFACE_SYNC:
2884         case PACKET3_EVENT_WRITE:
2885         case PACKET3_EVENT_WRITE_EOP:
2886         case PACKET3_EVENT_WRITE_EOS:
2887         case PACKET3_SET_CONTEXT_REG:
2888         case PACKET3_SET_CONTEXT_REG_INDIRECT:
2889         case PACKET3_SET_SH_REG:
2890         case PACKET3_SET_SH_REG_OFFSET:
2891         case PACKET3_INCREMENT_DE_COUNTER:
2892         case PACKET3_WAIT_ON_CE_COUNTER:
2893         case PACKET3_WAIT_ON_AVAIL_BUFFER:
2894         case PACKET3_ME_WRITE:
2895                 break;
2896         case PACKET3_COPY_DATA:
2897                 if ((idx_value & 0xf00) == 0) {
2898                         reg = ib[idx + 3] * 4;
2899                         if (!si_vm_reg_valid(reg))
2900                                 return -EINVAL;
2901                 }
2902                 break;
2903         case PACKET3_WRITE_DATA:
2904                 if ((idx_value & 0xf00) == 0) {
2905                         start_reg = ib[idx + 1] * 4;
2906                         if (idx_value & 0x10000) {
2907                                 if (!si_vm_reg_valid(start_reg))
2908                                         return -EINVAL;
2909                         } else {
2910                                 for (i = 0; i < (pkt->count - 2); i++) {
2911                                         reg = start_reg + (4 * i);
2912                                         if (!si_vm_reg_valid(reg))
2913                                                 return -EINVAL;
2914                                 }
2915                         }
2916                 }
2917                 break;
2918         case PACKET3_COND_WRITE:
2919                 if (idx_value & 0x100) {
2920                         reg = ib[idx + 5] * 4;
2921                         if (!si_vm_reg_valid(reg))
2922                                 return -EINVAL;
2923                 }
2924                 break;
2925         case PACKET3_COPY_DW:
2926                 if (idx_value & 0x2) {
2927                         reg = ib[idx + 3] * 4;
2928                         if (!si_vm_reg_valid(reg))
2929                                 return -EINVAL;
2930                 }
2931                 break;
2932         default:
2933                 DRM_ERROR("Invalid Compute packet3: 0x%x\n", pkt->opcode);
2934                 return -EINVAL;
2935         }
2936         return 0;
2937 }
2938
2939 int si_ib_parse(struct radeon_device *rdev, struct radeon_ib *ib)
2940 {
2941         int ret = 0;
2942         u32 idx = 0;
2943         struct radeon_cs_packet pkt;
2944
2945         do {
2946                 pkt.idx = idx;
2947                 pkt.type = RADEON_CP_PACKET_GET_TYPE(ib->ptr[idx]);
2948                 pkt.count = RADEON_CP_PACKET_GET_COUNT(ib->ptr[idx]);
2949                 pkt.one_reg_wr = 0;
2950                 switch (pkt.type) {
2951                 case RADEON_PACKET_TYPE0:
2952                         dev_err(rdev->dev, "Packet0 not allowed!\n");
2953                         ret = -EINVAL;
2954                         break;
2955                 case RADEON_PACKET_TYPE2:
2956                         idx += 1;
2957                         break;
2958                 case RADEON_PACKET_TYPE3:
2959                         pkt.opcode = RADEON_CP_PACKET3_GET_OPCODE(ib->ptr[idx]);
2960                         if (ib->is_const_ib)
2961                                 ret = si_vm_packet3_ce_check(rdev, ib->ptr, &pkt);
2962                         else {
2963                                 switch (ib->ring) {
2964                                 case RADEON_RING_TYPE_GFX_INDEX:
2965                                         ret = si_vm_packet3_gfx_check(rdev, ib->ptr, &pkt);
2966                                         break;
2967                                 case CAYMAN_RING_TYPE_CP1_INDEX:
2968                                 case CAYMAN_RING_TYPE_CP2_INDEX:
2969                                         ret = si_vm_packet3_compute_check(rdev, ib->ptr, &pkt);
2970                                         break;
2971                                 default:
2972                                         dev_err(rdev->dev, "Non-PM4 ring %d !\n", ib->ring);
2973                                         ret = -EINVAL;
2974                                         break;
2975                                 }
2976                         }
2977                         idx += pkt.count + 2;
2978                         break;
2979                 default:
2980                         dev_err(rdev->dev, "Unknown packet type %d !\n", pkt.type);
2981                         ret = -EINVAL;
2982                         break;
2983                 }
2984                 if (ret)
2985                         break;
2986         } while (idx < ib->length_dw);
2987
2988         return ret;
2989 }
2990
2991 /*
2992  * vm
2993  */
2994 int si_vm_init(struct radeon_device *rdev)
2995 {
2996         /* number of VMs */
2997         rdev->vm_manager.nvm = 16;
2998         /* base offset of vram pages */
2999         rdev->vm_manager.vram_base_offset = 0;
3000
3001         return 0;
3002 }
3003
3004 void si_vm_fini(struct radeon_device *rdev)
3005 {
3006 }
3007
3008 /**
3009  * si_vm_set_page - update the page tables using the CP
3010  *
3011  * @rdev: radeon_device pointer
3012  * @pe: addr of the page entry
3013  * @addr: dst addr to write into pe
3014  * @count: number of page entries to update
3015  * @incr: increase next addr by incr bytes
3016  * @flags: access flags
3017  *
3018  * Update the page tables using the CP (cayman-si).
3019  */
3020 void si_vm_set_page(struct radeon_device *rdev, uint64_t pe,
3021                     uint64_t addr, unsigned count,
3022                     uint32_t incr, uint32_t flags)
3023 {
3024         struct radeon_ring *ring = &rdev->ring[rdev->asic->vm.pt_ring_index];
3025         uint32_t r600_flags = cayman_vm_page_flags(rdev, flags);
3026         uint64_t value;
3027         unsigned ndw;
3028
3029         if (rdev->asic->vm.pt_ring_index == RADEON_RING_TYPE_GFX_INDEX) {
3030                 while (count) {
3031                         ndw = 2 + count * 2;
3032                         if (ndw > 0x3FFE)
3033                                 ndw = 0x3FFE;
3034
3035                         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, ndw));
3036                         radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
3037                                                  WRITE_DATA_DST_SEL(1)));
3038                         radeon_ring_write(ring, pe);
3039                         radeon_ring_write(ring, upper_32_bits(pe));
3040                         for (; ndw > 2; ndw -= 2, --count, pe += 8) {
3041                                 if (flags & RADEON_VM_PAGE_SYSTEM) {
3042                                         value = radeon_vm_map_gart(rdev, addr);
3043                                         value &= 0xFFFFFFFFFFFFF000ULL;
3044                                 } else if (flags & RADEON_VM_PAGE_VALID) {
3045                                         value = addr;
3046                                 } else {
3047                                         value = 0;
3048                                 }
3049                                 addr += incr;
3050                                 value |= r600_flags;
3051                                 radeon_ring_write(ring, value);
3052                                 radeon_ring_write(ring, upper_32_bits(value));
3053                         }
3054                 }
3055         } else {
3056                 /* DMA */
3057                 if (flags & RADEON_VM_PAGE_SYSTEM) {
3058                         while (count) {
3059                                 ndw = count * 2;
3060                                 if (ndw > 0xFFFFE)
3061                                         ndw = 0xFFFFE;
3062
3063                                 /* for non-physically contiguous pages (system) */
3064                                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_WRITE, 0, 0, 0, ndw));
3065                                 radeon_ring_write(ring, pe);
3066                                 radeon_ring_write(ring, upper_32_bits(pe) & 0xff);
3067                                 for (; ndw > 0; ndw -= 2, --count, pe += 8) {
3068                                         if (flags & RADEON_VM_PAGE_SYSTEM) {
3069                                                 value = radeon_vm_map_gart(rdev, addr);
3070                                                 value &= 0xFFFFFFFFFFFFF000ULL;
3071                                         } else if (flags & RADEON_VM_PAGE_VALID) {
3072                                                 value = addr;
3073                                         } else {
3074                                                 value = 0;
3075                                         }
3076                                         addr += incr;
3077                                         value |= r600_flags;
3078                                         radeon_ring_write(ring, value);
3079                                         radeon_ring_write(ring, upper_32_bits(value));
3080                                 }
3081                         }
3082                 } else {
3083                         while (count) {
3084                                 ndw = count * 2;
3085                                 if (ndw > 0xFFFFE)
3086                                         ndw = 0xFFFFE;
3087
3088                                 if (flags & RADEON_VM_PAGE_VALID)
3089                                         value = addr;
3090                                 else
3091                                         value = 0;
3092                                 /* for physically contiguous pages (vram) */
3093                                 radeon_ring_write(ring, DMA_PTE_PDE_PACKET(ndw));
3094                                 radeon_ring_write(ring, pe); /* dst addr */
3095                                 radeon_ring_write(ring, upper_32_bits(pe) & 0xff);
3096                                 radeon_ring_write(ring, r600_flags); /* mask */
3097                                 radeon_ring_write(ring, 0);
3098                                 radeon_ring_write(ring, value); /* value */
3099                                 radeon_ring_write(ring, upper_32_bits(value));
3100                                 radeon_ring_write(ring, incr); /* increment size */
3101                                 radeon_ring_write(ring, 0);
3102                                 pe += ndw * 4;
3103                                 addr += (ndw / 2) * incr;
3104                                 count -= ndw / 2;
3105                         }
3106                 }
3107         }
3108 }
3109
3110 void si_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
3111 {
3112         struct radeon_ring *ring = &rdev->ring[ridx];
3113
3114         if (vm == NULL)
3115                 return;
3116
3117         /* write new base address */
3118         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3119         radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
3120                                  WRITE_DATA_DST_SEL(0)));
3121
3122         if (vm->id < 8) {
3123                 radeon_ring_write(ring,
3124                                   (VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2);
3125         } else {
3126                 radeon_ring_write(ring,
3127                                   (VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm->id - 8) << 2)) >> 2);
3128         }
3129         radeon_ring_write(ring, 0);
3130         radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
3131
3132         /* flush hdp cache */
3133         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3134         radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
3135                                  WRITE_DATA_DST_SEL(0)));
3136         radeon_ring_write(ring, HDP_MEM_COHERENCY_FLUSH_CNTL >> 2);
3137         radeon_ring_write(ring, 0);
3138         radeon_ring_write(ring, 0x1);
3139
3140         /* bits 0-15 are the VM contexts0-15 */
3141         radeon_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3142         radeon_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
3143                                  WRITE_DATA_DST_SEL(0)));
3144         radeon_ring_write(ring, VM_INVALIDATE_REQUEST >> 2);
3145         radeon_ring_write(ring, 0);
3146         radeon_ring_write(ring, 1 << vm->id);
3147
3148         /* sync PFP to ME, otherwise we might get invalid PFP reads */
3149         radeon_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
3150         radeon_ring_write(ring, 0x0);
3151 }
3152
3153 void si_dma_vm_flush(struct radeon_device *rdev, int ridx, struct radeon_vm *vm)
3154 {
3155         struct radeon_ring *ring = &rdev->ring[ridx];
3156
3157         if (vm == NULL)
3158                 return;
3159
3160         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
3161         if (vm->id < 8) {
3162                 radeon_ring_write(ring, (0xf << 16) | ((VM_CONTEXT0_PAGE_TABLE_BASE_ADDR + (vm->id << 2)) >> 2));
3163         } else {
3164                 radeon_ring_write(ring, (0xf << 16) | ((VM_CONTEXT8_PAGE_TABLE_BASE_ADDR + ((vm->id - 8) << 2)) >> 2));
3165         }
3166         radeon_ring_write(ring, vm->pd_gpu_addr >> 12);
3167
3168         /* flush hdp cache */
3169         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
3170         radeon_ring_write(ring, (0xf << 16) | (HDP_MEM_COHERENCY_FLUSH_CNTL >> 2));
3171         radeon_ring_write(ring, 1);
3172
3173         /* bits 0-7 are the VM contexts0-7 */
3174         radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_SRBM_WRITE, 0, 0, 0, 0));
3175         radeon_ring_write(ring, (0xf << 16) | (VM_INVALIDATE_REQUEST >> 2));
3176         radeon_ring_write(ring, 1 << vm->id);
3177 }
3178
3179 /*
3180  * RLC
3181  */
3182 void si_rlc_fini(struct radeon_device *rdev)
3183 {
3184         int r;
3185
3186         /* save restore block */
3187         if (rdev->rlc.save_restore_obj) {
3188                 r = radeon_bo_reserve(rdev->rlc.save_restore_obj, false);
3189                 if (unlikely(r != 0))
3190                         dev_warn(rdev->dev, "(%d) reserve RLC sr bo failed\n", r);
3191                 radeon_bo_unpin(rdev->rlc.save_restore_obj);
3192                 radeon_bo_unreserve(rdev->rlc.save_restore_obj);
3193
3194                 radeon_bo_unref(&rdev->rlc.save_restore_obj);
3195                 rdev->rlc.save_restore_obj = NULL;
3196         }
3197
3198         /* clear state block */
3199         if (rdev->rlc.clear_state_obj) {
3200                 r = radeon_bo_reserve(rdev->rlc.clear_state_obj, false);
3201                 if (unlikely(r != 0))
3202                         dev_warn(rdev->dev, "(%d) reserve RLC c bo failed\n", r);
3203                 radeon_bo_unpin(rdev->rlc.clear_state_obj);
3204                 radeon_bo_unreserve(rdev->rlc.clear_state_obj);
3205
3206                 radeon_bo_unref(&rdev->rlc.clear_state_obj);
3207                 rdev->rlc.clear_state_obj = NULL;
3208         }
3209 }
3210
3211 int si_rlc_init(struct radeon_device *rdev)
3212 {
3213         int r;
3214
3215         /* save restore block */
3216         if (rdev->rlc.save_restore_obj == NULL) {
3217                 r = radeon_bo_create(rdev, RADEON_GPU_PAGE_SIZE, PAGE_SIZE, true,
3218                                      RADEON_GEM_DOMAIN_VRAM, NULL,
3219                                      &rdev->rlc.save_restore_obj);
3220                 if (r) {
3221                         dev_warn(rdev->dev, "(%d) create RLC sr bo failed\n", r);
3222                         return r;
3223                 }
3224         }
3225
3226         r = radeon_bo_reserve(rdev->rlc.save_restore_obj, false);
3227         if (unlikely(r != 0)) {
3228                 si_rlc_fini(rdev);
3229                 return r;
3230         }
3231         r = radeon_bo_pin(rdev->rlc.save_restore_obj, RADEON_GEM_DOMAIN_VRAM,
3232                           &rdev->rlc.save_restore_gpu_addr);
3233         radeon_bo_unreserve(rdev->rlc.save_restore_obj);
3234         if (r) {
3235                 dev_warn(rdev->dev, "(%d) pin RLC sr bo failed\n", r);
3236                 si_rlc_fini(rdev);
3237                 return r;
3238         }
3239
3240         /* clear state block */
3241         if (rdev->rlc.clear_state_obj == NULL) {
3242                 r = radeon_bo_create(rdev, RADEON_GPU_PAGE_SIZE, PAGE_SIZE, true,
3243                                      RADEON_GEM_DOMAIN_VRAM, NULL,
3244                                      &rdev->rlc.clear_state_obj);
3245                 if (r) {
3246                         dev_warn(rdev->dev, "(%d) create RLC c bo failed\n", r);
3247                         si_rlc_fini(rdev);
3248                         return r;
3249                 }
3250         }
3251         r = radeon_bo_reserve(rdev->rlc.clear_state_obj, false);
3252         if (unlikely(r != 0)) {
3253                 si_rlc_fini(rdev);
3254                 return r;
3255         }
3256         r = radeon_bo_pin(rdev->rlc.clear_state_obj, RADEON_GEM_DOMAIN_VRAM,
3257                           &rdev->rlc.clear_state_gpu_addr);
3258         radeon_bo_unreserve(rdev->rlc.clear_state_obj);
3259         if (r) {
3260                 dev_warn(rdev->dev, "(%d) pin RLC c bo failed\n", r);
3261                 si_rlc_fini(rdev);
3262                 return r;
3263         }
3264
3265         return 0;
3266 }
3267
3268 static void si_rlc_stop(struct radeon_device *rdev)
3269 {
3270         WREG32(RLC_CNTL, 0);
3271 }
3272
3273 static void si_rlc_start(struct radeon_device *rdev)
3274 {
3275         WREG32(RLC_CNTL, RLC_ENABLE);
3276 }
3277
3278 static int si_rlc_resume(struct radeon_device *rdev)
3279 {
3280         u32 i;
3281         const __be32 *fw_data;
3282
3283         if (!rdev->rlc_fw)
3284                 return -EINVAL;
3285
3286         si_rlc_stop(rdev);
3287
3288         WREG32(RLC_RL_BASE, 0);
3289         WREG32(RLC_RL_SIZE, 0);
3290         WREG32(RLC_LB_CNTL, 0);
3291         WREG32(RLC_LB_CNTR_MAX, 0xffffffff);
3292         WREG32(RLC_LB_CNTR_INIT, 0);
3293
3294         WREG32(RLC_SAVE_AND_RESTORE_BASE, rdev->rlc.save_restore_gpu_addr >> 8);
3295         WREG32(RLC_CLEAR_STATE_RESTORE_BASE, rdev->rlc.clear_state_gpu_addr >> 8);
3296
3297         WREG32(RLC_MC_CNTL, 0);
3298         WREG32(RLC_UCODE_CNTL, 0);
3299
3300         fw_data = (const __be32 *)rdev->rlc_fw->data;
3301         for (i = 0; i < SI_RLC_UCODE_SIZE; i++) {
3302                 WREG32(RLC_UCODE_ADDR, i);
3303                 WREG32(RLC_UCODE_DATA, be32_to_cpup(fw_data++));
3304         }
3305         WREG32(RLC_UCODE_ADDR, 0);
3306
3307         si_rlc_start(rdev);
3308
3309         return 0;
3310 }
3311
3312 static void si_enable_interrupts(struct radeon_device *rdev)
3313 {
3314         u32 ih_cntl = RREG32(IH_CNTL);
3315         u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
3316
3317         ih_cntl |= ENABLE_INTR;
3318         ih_rb_cntl |= IH_RB_ENABLE;
3319         WREG32(IH_CNTL, ih_cntl);
3320         WREG32(IH_RB_CNTL, ih_rb_cntl);
3321         rdev->ih.enabled = true;
3322 }
3323
3324 static void si_disable_interrupts(struct radeon_device *rdev)
3325 {
3326         u32 ih_rb_cntl = RREG32(IH_RB_CNTL);
3327         u32 ih_cntl = RREG32(IH_CNTL);
3328
3329         ih_rb_cntl &= ~IH_RB_ENABLE;
3330         ih_cntl &= ~ENABLE_INTR;
3331         WREG32(IH_RB_CNTL, ih_rb_cntl);
3332         WREG32(IH_CNTL, ih_cntl);
3333         /* set rptr, wptr to 0 */
3334         WREG32(IH_RB_RPTR, 0);
3335         WREG32(IH_RB_WPTR, 0);
3336         rdev->ih.enabled = false;
3337         rdev->ih.rptr = 0;
3338 }
3339
3340 static void si_disable_interrupt_state(struct radeon_device *rdev)
3341 {
3342         u32 tmp;
3343
3344         WREG32(CP_INT_CNTL_RING0, CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE);
3345         WREG32(CP_INT_CNTL_RING1, 0);
3346         WREG32(CP_INT_CNTL_RING2, 0);
3347         tmp = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
3348         WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, tmp);
3349         tmp = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
3350         WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, tmp);
3351         WREG32(GRBM_INT_CNTL, 0);
3352         WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
3353         WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
3354         if (rdev->num_crtc >= 4) {
3355                 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
3356                 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
3357         }
3358         if (rdev->num_crtc >= 6) {
3359                 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
3360                 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
3361         }
3362
3363         WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, 0);
3364         WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, 0);
3365         if (rdev->num_crtc >= 4) {
3366                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, 0);
3367                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, 0);
3368         }
3369         if (rdev->num_crtc >= 6) {
3370                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, 0);
3371                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, 0);
3372         }
3373
3374         WREG32(DACA_AUTODETECT_INT_CONTROL, 0);
3375
3376         tmp = RREG32(DC_HPD1_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3377         WREG32(DC_HPD1_INT_CONTROL, tmp);
3378         tmp = RREG32(DC_HPD2_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3379         WREG32(DC_HPD2_INT_CONTROL, tmp);
3380         tmp = RREG32(DC_HPD3_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3381         WREG32(DC_HPD3_INT_CONTROL, tmp);
3382         tmp = RREG32(DC_HPD4_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3383         WREG32(DC_HPD4_INT_CONTROL, tmp);
3384         tmp = RREG32(DC_HPD5_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3385         WREG32(DC_HPD5_INT_CONTROL, tmp);
3386         tmp = RREG32(DC_HPD6_INT_CONTROL) & DC_HPDx_INT_POLARITY;
3387         WREG32(DC_HPD6_INT_CONTROL, tmp);
3388
3389 }
3390
3391 static int si_irq_init(struct radeon_device *rdev)
3392 {
3393         int ret = 0;
3394         int rb_bufsz;
3395         u32 interrupt_cntl, ih_cntl, ih_rb_cntl;
3396
3397         /* allocate ring */
3398         ret = r600_ih_ring_alloc(rdev);
3399         if (ret)
3400                 return ret;
3401
3402         /* disable irqs */
3403         si_disable_interrupts(rdev);
3404
3405         /* init rlc */
3406         ret = si_rlc_resume(rdev);
3407         if (ret) {
3408                 r600_ih_ring_fini(rdev);
3409                 return ret;
3410         }
3411
3412         /* setup interrupt control */
3413         /* set dummy read address to ring address */
3414         WREG32(INTERRUPT_CNTL2, rdev->ih.gpu_addr >> 8);
3415         interrupt_cntl = RREG32(INTERRUPT_CNTL);
3416         /* IH_DUMMY_RD_OVERRIDE=0 - dummy read disabled with msi, enabled without msi
3417          * IH_DUMMY_RD_OVERRIDE=1 - dummy read controlled by IH_DUMMY_RD_EN
3418          */
3419         interrupt_cntl &= ~IH_DUMMY_RD_OVERRIDE;
3420         /* IH_REQ_NONSNOOP_EN=1 if ring is in non-cacheable memory, e.g., vram */
3421         interrupt_cntl &= ~IH_REQ_NONSNOOP_EN;
3422         WREG32(INTERRUPT_CNTL, interrupt_cntl);
3423
3424         WREG32(IH_RB_BASE, rdev->ih.gpu_addr >> 8);
3425         rb_bufsz = drm_order(rdev->ih.ring_size / 4);
3426
3427         ih_rb_cntl = (IH_WPTR_OVERFLOW_ENABLE |
3428                       IH_WPTR_OVERFLOW_CLEAR |
3429                       (rb_bufsz << 1));
3430
3431         if (rdev->wb.enabled)
3432                 ih_rb_cntl |= IH_WPTR_WRITEBACK_ENABLE;
3433
3434         /* set the writeback address whether it's enabled or not */
3435         WREG32(IH_RB_WPTR_ADDR_LO, (rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFFFFFFFC);
3436         WREG32(IH_RB_WPTR_ADDR_HI, upper_32_bits(rdev->wb.gpu_addr + R600_WB_IH_WPTR_OFFSET) & 0xFF);
3437
3438         WREG32(IH_RB_CNTL, ih_rb_cntl);
3439
3440         /* set rptr, wptr to 0 */
3441         WREG32(IH_RB_RPTR, 0);
3442         WREG32(IH_RB_WPTR, 0);
3443
3444         /* Default settings for IH_CNTL (disabled at first) */
3445         ih_cntl = MC_WRREQ_CREDIT(0x10) | MC_WR_CLEAN_CNT(0x10) | MC_VMID(0);
3446         /* RPTR_REARM only works if msi's are enabled */
3447         if (rdev->msi_enabled)
3448                 ih_cntl |= RPTR_REARM;
3449         WREG32(IH_CNTL, ih_cntl);
3450
3451         /* force the active interrupt state to all disabled */
3452         si_disable_interrupt_state(rdev);
3453
3454         pci_set_master(rdev->pdev);
3455
3456         /* enable irqs */
3457         si_enable_interrupts(rdev);
3458
3459         return ret;
3460 }
3461
3462 int si_irq_set(struct radeon_device *rdev)
3463 {
3464         u32 cp_int_cntl = CNTX_BUSY_INT_ENABLE | CNTX_EMPTY_INT_ENABLE;
3465         u32 cp_int_cntl1 = 0, cp_int_cntl2 = 0;
3466         u32 crtc1 = 0, crtc2 = 0, crtc3 = 0, crtc4 = 0, crtc5 = 0, crtc6 = 0;
3467         u32 hpd1, hpd2, hpd3, hpd4, hpd5, hpd6;
3468         u32 grbm_int_cntl = 0;
3469         u32 grph1 = 0, grph2 = 0, grph3 = 0, grph4 = 0, grph5 = 0, grph6 = 0;
3470         u32 dma_cntl, dma_cntl1;
3471
3472         if (!rdev->irq.installed) {
3473                 WARN(1, "Can't enable IRQ/MSI because no handler is installed\n");
3474                 return -EINVAL;
3475         }
3476         /* don't enable anything if the ih is disabled */
3477         if (!rdev->ih.enabled) {
3478                 si_disable_interrupts(rdev);
3479                 /* force the active interrupt state to all disabled */
3480                 si_disable_interrupt_state(rdev);
3481                 return 0;
3482         }
3483
3484         hpd1 = RREG32(DC_HPD1_INT_CONTROL) & ~DC_HPDx_INT_EN;
3485         hpd2 = RREG32(DC_HPD2_INT_CONTROL) & ~DC_HPDx_INT_EN;
3486         hpd3 = RREG32(DC_HPD3_INT_CONTROL) & ~DC_HPDx_INT_EN;
3487         hpd4 = RREG32(DC_HPD4_INT_CONTROL) & ~DC_HPDx_INT_EN;
3488         hpd5 = RREG32(DC_HPD5_INT_CONTROL) & ~DC_HPDx_INT_EN;
3489         hpd6 = RREG32(DC_HPD6_INT_CONTROL) & ~DC_HPDx_INT_EN;
3490
3491         dma_cntl = RREG32(DMA_CNTL + DMA0_REGISTER_OFFSET) & ~TRAP_ENABLE;
3492         dma_cntl1 = RREG32(DMA_CNTL + DMA1_REGISTER_OFFSET) & ~TRAP_ENABLE;
3493
3494         /* enable CP interrupts on all rings */
3495         if (atomic_read(&rdev->irq.ring_int[RADEON_RING_TYPE_GFX_INDEX])) {
3496                 DRM_DEBUG("si_irq_set: sw int gfx\n");
3497                 cp_int_cntl |= TIME_STAMP_INT_ENABLE;
3498         }
3499         if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP1_INDEX])) {
3500                 DRM_DEBUG("si_irq_set: sw int cp1\n");
3501                 cp_int_cntl1 |= TIME_STAMP_INT_ENABLE;
3502         }
3503         if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_CP2_INDEX])) {
3504                 DRM_DEBUG("si_irq_set: sw int cp2\n");
3505                 cp_int_cntl2 |= TIME_STAMP_INT_ENABLE;
3506         }
3507         if (atomic_read(&rdev->irq.ring_int[R600_RING_TYPE_DMA_INDEX])) {
3508                 DRM_DEBUG("si_irq_set: sw int dma\n");
3509                 dma_cntl |= TRAP_ENABLE;
3510         }
3511
3512         if (atomic_read(&rdev->irq.ring_int[CAYMAN_RING_TYPE_DMA1_INDEX])) {
3513                 DRM_DEBUG("si_irq_set: sw int dma1\n");
3514                 dma_cntl1 |= TRAP_ENABLE;
3515         }
3516         if (rdev->irq.crtc_vblank_int[0] ||
3517             atomic_read(&rdev->irq.pflip[0])) {
3518                 DRM_DEBUG("si_irq_set: vblank 0\n");
3519                 crtc1 |= VBLANK_INT_MASK;
3520         }
3521         if (rdev->irq.crtc_vblank_int[1] ||
3522             atomic_read(&rdev->irq.pflip[1])) {
3523                 DRM_DEBUG("si_irq_set: vblank 1\n");
3524                 crtc2 |= VBLANK_INT_MASK;
3525         }
3526         if (rdev->irq.crtc_vblank_int[2] ||
3527             atomic_read(&rdev->irq.pflip[2])) {
3528                 DRM_DEBUG("si_irq_set: vblank 2\n");
3529                 crtc3 |= VBLANK_INT_MASK;
3530         }
3531         if (rdev->irq.crtc_vblank_int[3] ||
3532             atomic_read(&rdev->irq.pflip[3])) {
3533                 DRM_DEBUG("si_irq_set: vblank 3\n");
3534                 crtc4 |= VBLANK_INT_MASK;
3535         }
3536         if (rdev->irq.crtc_vblank_int[4] ||
3537             atomic_read(&rdev->irq.pflip[4])) {
3538                 DRM_DEBUG("si_irq_set: vblank 4\n");
3539                 crtc5 |= VBLANK_INT_MASK;
3540         }
3541         if (rdev->irq.crtc_vblank_int[5] ||
3542             atomic_read(&rdev->irq.pflip[5])) {
3543                 DRM_DEBUG("si_irq_set: vblank 5\n");
3544                 crtc6 |= VBLANK_INT_MASK;
3545         }
3546         if (rdev->irq.hpd[0]) {
3547                 DRM_DEBUG("si_irq_set: hpd 1\n");
3548                 hpd1 |= DC_HPDx_INT_EN;
3549         }
3550         if (rdev->irq.hpd[1]) {
3551                 DRM_DEBUG("si_irq_set: hpd 2\n");
3552                 hpd2 |= DC_HPDx_INT_EN;
3553         }
3554         if (rdev->irq.hpd[2]) {
3555                 DRM_DEBUG("si_irq_set: hpd 3\n");
3556                 hpd3 |= DC_HPDx_INT_EN;
3557         }
3558         if (rdev->irq.hpd[3]) {
3559                 DRM_DEBUG("si_irq_set: hpd 4\n");
3560                 hpd4 |= DC_HPDx_INT_EN;
3561         }
3562         if (rdev->irq.hpd[4]) {
3563                 DRM_DEBUG("si_irq_set: hpd 5\n");
3564                 hpd5 |= DC_HPDx_INT_EN;
3565         }
3566         if (rdev->irq.hpd[5]) {
3567                 DRM_DEBUG("si_irq_set: hpd 6\n");
3568                 hpd6 |= DC_HPDx_INT_EN;
3569         }
3570
3571         WREG32(CP_INT_CNTL_RING0, cp_int_cntl);
3572         WREG32(CP_INT_CNTL_RING1, cp_int_cntl1);
3573         WREG32(CP_INT_CNTL_RING2, cp_int_cntl2);
3574
3575         WREG32(DMA_CNTL + DMA0_REGISTER_OFFSET, dma_cntl);
3576         WREG32(DMA_CNTL + DMA1_REGISTER_OFFSET, dma_cntl1);
3577
3578         WREG32(GRBM_INT_CNTL, grbm_int_cntl);
3579
3580         WREG32(INT_MASK + EVERGREEN_CRTC0_REGISTER_OFFSET, crtc1);
3581         WREG32(INT_MASK + EVERGREEN_CRTC1_REGISTER_OFFSET, crtc2);
3582         if (rdev->num_crtc >= 4) {
3583                 WREG32(INT_MASK + EVERGREEN_CRTC2_REGISTER_OFFSET, crtc3);
3584                 WREG32(INT_MASK + EVERGREEN_CRTC3_REGISTER_OFFSET, crtc4);
3585         }
3586         if (rdev->num_crtc >= 6) {
3587                 WREG32(INT_MASK + EVERGREEN_CRTC4_REGISTER_OFFSET, crtc5);
3588                 WREG32(INT_MASK + EVERGREEN_CRTC5_REGISTER_OFFSET, crtc6);
3589         }
3590
3591         WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC0_REGISTER_OFFSET, grph1);
3592         WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC1_REGISTER_OFFSET, grph2);
3593         if (rdev->num_crtc >= 4) {
3594                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC2_REGISTER_OFFSET, grph3);
3595                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC3_REGISTER_OFFSET, grph4);
3596         }
3597         if (rdev->num_crtc >= 6) {
3598                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC4_REGISTER_OFFSET, grph5);
3599                 WREG32(GRPH_INT_CONTROL + EVERGREEN_CRTC5_REGISTER_OFFSET, grph6);
3600         }
3601
3602         WREG32(DC_HPD1_INT_CONTROL, hpd1);
3603         WREG32(DC_HPD2_INT_CONTROL, hpd2);
3604         WREG32(DC_HPD3_INT_CONTROL, hpd3);
3605         WREG32(DC_HPD4_INT_CONTROL, hpd4);
3606         WREG32(DC_HPD5_INT_CONTROL, hpd5);
3607         WREG32(DC_HPD6_INT_CONTROL, hpd6);
3608
3609         return 0;
3610 }
3611
3612 static inline void si_irq_ack(struct radeon_device *rdev)
3613 {
3614         u32 tmp;
3615
3616         rdev->irq.stat_regs.evergreen.disp_int = RREG32(DISP_INTERRUPT_STATUS);
3617         rdev->irq.stat_regs.evergreen.disp_int_cont = RREG32(DISP_INTERRUPT_STATUS_CONTINUE);
3618         rdev->irq.stat_regs.evergreen.disp_int_cont2 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE2);
3619         rdev->irq.stat_regs.evergreen.disp_int_cont3 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE3);
3620         rdev->irq.stat_regs.evergreen.disp_int_cont4 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE4);
3621         rdev->irq.stat_regs.evergreen.disp_int_cont5 = RREG32(DISP_INTERRUPT_STATUS_CONTINUE5);
3622         rdev->irq.stat_regs.evergreen.d1grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET);
3623         rdev->irq.stat_regs.evergreen.d2grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET);
3624         if (rdev->num_crtc >= 4) {
3625                 rdev->irq.stat_regs.evergreen.d3grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET);
3626                 rdev->irq.stat_regs.evergreen.d4grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET);
3627         }
3628         if (rdev->num_crtc >= 6) {
3629                 rdev->irq.stat_regs.evergreen.d5grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET);
3630                 rdev->irq.stat_regs.evergreen.d6grph_int = RREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET);
3631         }
3632
3633         if (rdev->irq.stat_regs.evergreen.d1grph_int & GRPH_PFLIP_INT_OCCURRED)
3634                 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
3635         if (rdev->irq.stat_regs.evergreen.d2grph_int & GRPH_PFLIP_INT_OCCURRED)
3636                 WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
3637         if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT)
3638                 WREG32(VBLANK_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VBLANK_ACK);
3639         if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT)
3640                 WREG32(VLINE_STATUS + EVERGREEN_CRTC0_REGISTER_OFFSET, VLINE_ACK);
3641         if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT)
3642                 WREG32(VBLANK_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VBLANK_ACK);
3643         if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT)
3644                 WREG32(VLINE_STATUS + EVERGREEN_CRTC1_REGISTER_OFFSET, VLINE_ACK);
3645
3646         if (rdev->num_crtc >= 4) {
3647                 if (rdev->irq.stat_regs.evergreen.d3grph_int & GRPH_PFLIP_INT_OCCURRED)
3648                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
3649                 if (rdev->irq.stat_regs.evergreen.d4grph_int & GRPH_PFLIP_INT_OCCURRED)
3650                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
3651                 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT)
3652                         WREG32(VBLANK_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VBLANK_ACK);
3653                 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT)
3654                         WREG32(VLINE_STATUS + EVERGREEN_CRTC2_REGISTER_OFFSET, VLINE_ACK);
3655                 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT)
3656                         WREG32(VBLANK_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VBLANK_ACK);
3657                 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT)
3658                         WREG32(VLINE_STATUS + EVERGREEN_CRTC3_REGISTER_OFFSET, VLINE_ACK);
3659         }
3660
3661         if (rdev->num_crtc >= 6) {
3662                 if (rdev->irq.stat_regs.evergreen.d5grph_int & GRPH_PFLIP_INT_OCCURRED)
3663                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
3664                 if (rdev->irq.stat_regs.evergreen.d6grph_int & GRPH_PFLIP_INT_OCCURRED)
3665                         WREG32(GRPH_INT_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, GRPH_PFLIP_INT_CLEAR);
3666                 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT)
3667                         WREG32(VBLANK_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VBLANK_ACK);
3668                 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT)
3669                         WREG32(VLINE_STATUS + EVERGREEN_CRTC4_REGISTER_OFFSET, VLINE_ACK);
3670                 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT)
3671                         WREG32(VBLANK_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VBLANK_ACK);
3672                 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT)
3673                         WREG32(VLINE_STATUS + EVERGREEN_CRTC5_REGISTER_OFFSET, VLINE_ACK);
3674         }
3675
3676         if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
3677                 tmp = RREG32(DC_HPD1_INT_CONTROL);
3678                 tmp |= DC_HPDx_INT_ACK;
3679                 WREG32(DC_HPD1_INT_CONTROL, tmp);
3680         }
3681         if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
3682                 tmp = RREG32(DC_HPD2_INT_CONTROL);
3683                 tmp |= DC_HPDx_INT_ACK;
3684                 WREG32(DC_HPD2_INT_CONTROL, tmp);
3685         }
3686         if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
3687                 tmp = RREG32(DC_HPD3_INT_CONTROL);
3688                 tmp |= DC_HPDx_INT_ACK;
3689                 WREG32(DC_HPD3_INT_CONTROL, tmp);
3690         }
3691         if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
3692                 tmp = RREG32(DC_HPD4_INT_CONTROL);
3693                 tmp |= DC_HPDx_INT_ACK;
3694                 WREG32(DC_HPD4_INT_CONTROL, tmp);
3695         }
3696         if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
3697                 tmp = RREG32(DC_HPD5_INT_CONTROL);
3698                 tmp |= DC_HPDx_INT_ACK;
3699                 WREG32(DC_HPD5_INT_CONTROL, tmp);
3700         }
3701         if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
3702                 tmp = RREG32(DC_HPD5_INT_CONTROL);
3703                 tmp |= DC_HPDx_INT_ACK;
3704                 WREG32(DC_HPD6_INT_CONTROL, tmp);
3705         }
3706 }
3707
3708 static void si_irq_disable(struct radeon_device *rdev)
3709 {
3710         si_disable_interrupts(rdev);
3711         /* Wait and acknowledge irq */
3712         mdelay(1);
3713         si_irq_ack(rdev);
3714         si_disable_interrupt_state(rdev);
3715 }
3716
3717 static void si_irq_suspend(struct radeon_device *rdev)
3718 {
3719         si_irq_disable(rdev);
3720         si_rlc_stop(rdev);
3721 }
3722
3723 static void si_irq_fini(struct radeon_device *rdev)
3724 {
3725         si_irq_suspend(rdev);
3726         r600_ih_ring_fini(rdev);
3727 }
3728
3729 static inline u32 si_get_ih_wptr(struct radeon_device *rdev)
3730 {
3731         u32 wptr, tmp;
3732
3733         if (rdev->wb.enabled)
3734                 wptr = le32_to_cpu(rdev->wb.wb[R600_WB_IH_WPTR_OFFSET/4]);
3735         else
3736                 wptr = RREG32(IH_RB_WPTR);
3737
3738         if (wptr & RB_OVERFLOW) {
3739                 /* When a ring buffer overflow happen start parsing interrupt
3740                  * from the last not overwritten vector (wptr + 16). Hopefully
3741                  * this should allow us to catchup.
3742                  */
3743                 dev_warn(rdev->dev, "IH ring buffer overflow (0x%08X, %d, %d)\n",
3744                         wptr, rdev->ih.rptr, (wptr + 16) + rdev->ih.ptr_mask);
3745                 rdev->ih.rptr = (wptr + 16) & rdev->ih.ptr_mask;
3746                 tmp = RREG32(IH_RB_CNTL);
3747                 tmp |= IH_WPTR_OVERFLOW_CLEAR;
3748                 WREG32(IH_RB_CNTL, tmp);
3749         }
3750         return (wptr & rdev->ih.ptr_mask);
3751 }
3752
3753 /*        SI IV Ring
3754  * Each IV ring entry is 128 bits:
3755  * [7:0]    - interrupt source id
3756  * [31:8]   - reserved
3757  * [59:32]  - interrupt source data
3758  * [63:60]  - reserved
3759  * [71:64]  - RINGID
3760  * [79:72]  - VMID
3761  * [127:80] - reserved
3762  */
3763 int si_irq_process(struct radeon_device *rdev)
3764 {
3765         u32 wptr;
3766         u32 rptr;
3767         u32 src_id, src_data, ring_id;
3768         u32 ring_index;
3769         bool queue_hotplug = false;
3770
3771         if (!rdev->ih.enabled || rdev->shutdown)
3772                 return IRQ_NONE;
3773
3774         wptr = si_get_ih_wptr(rdev);
3775
3776 restart_ih:
3777         /* is somebody else already processing irqs? */
3778         if (atomic_xchg(&rdev->ih.lock, 1))
3779                 return IRQ_NONE;
3780
3781         rptr = rdev->ih.rptr;
3782         DRM_DEBUG("si_irq_process start: rptr %d, wptr %d\n", rptr, wptr);
3783
3784         /* Order reading of wptr vs. reading of IH ring data */
3785         rmb();
3786
3787         /* display interrupts */
3788         si_irq_ack(rdev);
3789
3790         while (rptr != wptr) {
3791                 /* wptr/rptr are in bytes! */
3792                 ring_index = rptr / 4;
3793                 src_id =  le32_to_cpu(rdev->ih.ring[ring_index]) & 0xff;
3794                 src_data = le32_to_cpu(rdev->ih.ring[ring_index + 1]) & 0xfffffff;
3795                 ring_id = le32_to_cpu(rdev->ih.ring[ring_index + 2]) & 0xff;
3796
3797                 switch (src_id) {
3798                 case 1: /* D1 vblank/vline */
3799                         switch (src_data) {
3800                         case 0: /* D1 vblank */
3801                                 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VBLANK_INTERRUPT) {
3802                                         if (rdev->irq.crtc_vblank_int[0]) {
3803                                                 drm_handle_vblank(rdev->ddev, 0);
3804                                                 rdev->pm.vblank_sync = true;
3805                                                 wake_up(&rdev->irq.vblank_queue);
3806                                         }
3807                                         if (atomic_read(&rdev->irq.pflip[0]))
3808                                                 radeon_crtc_handle_flip(rdev, 0);
3809                                         rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VBLANK_INTERRUPT;
3810                                         DRM_DEBUG("IH: D1 vblank\n");
3811                                 }
3812                                 break;
3813                         case 1: /* D1 vline */
3814                                 if (rdev->irq.stat_regs.evergreen.disp_int & LB_D1_VLINE_INTERRUPT) {
3815                                         rdev->irq.stat_regs.evergreen.disp_int &= ~LB_D1_VLINE_INTERRUPT;
3816                                         DRM_DEBUG("IH: D1 vline\n");
3817                                 }
3818                                 break;
3819                         default:
3820                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3821                                 break;
3822                         }
3823                         break;
3824                 case 2: /* D2 vblank/vline */
3825                         switch (src_data) {
3826                         case 0: /* D2 vblank */
3827                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VBLANK_INTERRUPT) {
3828                                         if (rdev->irq.crtc_vblank_int[1]) {
3829                                                 drm_handle_vblank(rdev->ddev, 1);
3830                                                 rdev->pm.vblank_sync = true;
3831                                                 wake_up(&rdev->irq.vblank_queue);
3832                                         }
3833                                         if (atomic_read(&rdev->irq.pflip[1]))
3834                                                 radeon_crtc_handle_flip(rdev, 1);
3835                                         rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VBLANK_INTERRUPT;
3836                                         DRM_DEBUG("IH: D2 vblank\n");
3837                                 }
3838                                 break;
3839                         case 1: /* D2 vline */
3840                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont & LB_D2_VLINE_INTERRUPT) {
3841                                         rdev->irq.stat_regs.evergreen.disp_int_cont &= ~LB_D2_VLINE_INTERRUPT;
3842                                         DRM_DEBUG("IH: D2 vline\n");
3843                                 }
3844                                 break;
3845                         default:
3846                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3847                                 break;
3848                         }
3849                         break;
3850                 case 3: /* D3 vblank/vline */
3851                         switch (src_data) {
3852                         case 0: /* D3 vblank */
3853                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VBLANK_INTERRUPT) {
3854                                         if (rdev->irq.crtc_vblank_int[2]) {
3855                                                 drm_handle_vblank(rdev->ddev, 2);
3856                                                 rdev->pm.vblank_sync = true;
3857                                                 wake_up(&rdev->irq.vblank_queue);
3858                                         }
3859                                         if (atomic_read(&rdev->irq.pflip[2]))
3860                                                 radeon_crtc_handle_flip(rdev, 2);
3861                                         rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VBLANK_INTERRUPT;
3862                                         DRM_DEBUG("IH: D3 vblank\n");
3863                                 }
3864                                 break;
3865                         case 1: /* D3 vline */
3866                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & LB_D3_VLINE_INTERRUPT) {
3867                                         rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~LB_D3_VLINE_INTERRUPT;
3868                                         DRM_DEBUG("IH: D3 vline\n");
3869                                 }
3870                                 break;
3871                         default:
3872                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3873                                 break;
3874                         }
3875                         break;
3876                 case 4: /* D4 vblank/vline */
3877                         switch (src_data) {
3878                         case 0: /* D4 vblank */
3879                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VBLANK_INTERRUPT) {
3880                                         if (rdev->irq.crtc_vblank_int[3]) {
3881                                                 drm_handle_vblank(rdev->ddev, 3);
3882                                                 rdev->pm.vblank_sync = true;
3883                                                 wake_up(&rdev->irq.vblank_queue);
3884                                         }
3885                                         if (atomic_read(&rdev->irq.pflip[3]))
3886                                                 radeon_crtc_handle_flip(rdev, 3);
3887                                         rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VBLANK_INTERRUPT;
3888                                         DRM_DEBUG("IH: D4 vblank\n");
3889                                 }
3890                                 break;
3891                         case 1: /* D4 vline */
3892                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & LB_D4_VLINE_INTERRUPT) {
3893                                         rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~LB_D4_VLINE_INTERRUPT;
3894                                         DRM_DEBUG("IH: D4 vline\n");
3895                                 }
3896                                 break;
3897                         default:
3898                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3899                                 break;
3900                         }
3901                         break;
3902                 case 5: /* D5 vblank/vline */
3903                         switch (src_data) {
3904                         case 0: /* D5 vblank */
3905                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VBLANK_INTERRUPT) {
3906                                         if (rdev->irq.crtc_vblank_int[4]) {
3907                                                 drm_handle_vblank(rdev->ddev, 4);
3908                                                 rdev->pm.vblank_sync = true;
3909                                                 wake_up(&rdev->irq.vblank_queue);
3910                                         }
3911                                         if (atomic_read(&rdev->irq.pflip[4]))
3912                                                 radeon_crtc_handle_flip(rdev, 4);
3913                                         rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VBLANK_INTERRUPT;
3914                                         DRM_DEBUG("IH: D5 vblank\n");
3915                                 }
3916                                 break;
3917                         case 1: /* D5 vline */
3918                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & LB_D5_VLINE_INTERRUPT) {
3919                                         rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~LB_D5_VLINE_INTERRUPT;
3920                                         DRM_DEBUG("IH: D5 vline\n");
3921                                 }
3922                                 break;
3923                         default:
3924                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3925                                 break;
3926                         }
3927                         break;
3928                 case 6: /* D6 vblank/vline */
3929                         switch (src_data) {
3930                         case 0: /* D6 vblank */
3931                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VBLANK_INTERRUPT) {
3932                                         if (rdev->irq.crtc_vblank_int[5]) {
3933                                                 drm_handle_vblank(rdev->ddev, 5);
3934                                                 rdev->pm.vblank_sync = true;
3935                                                 wake_up(&rdev->irq.vblank_queue);
3936                                         }
3937                                         if (atomic_read(&rdev->irq.pflip[5]))
3938                                                 radeon_crtc_handle_flip(rdev, 5);
3939                                         rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VBLANK_INTERRUPT;
3940                                         DRM_DEBUG("IH: D6 vblank\n");
3941                                 }
3942                                 break;
3943                         case 1: /* D6 vline */
3944                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & LB_D6_VLINE_INTERRUPT) {
3945                                         rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~LB_D6_VLINE_INTERRUPT;
3946                                         DRM_DEBUG("IH: D6 vline\n");
3947                                 }
3948                                 break;
3949                         default:
3950                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
3951                                 break;
3952                         }
3953                         break;
3954                 case 42: /* HPD hotplug */
3955                         switch (src_data) {
3956                         case 0:
3957                                 if (rdev->irq.stat_regs.evergreen.disp_int & DC_HPD1_INTERRUPT) {
3958                                         rdev->irq.stat_regs.evergreen.disp_int &= ~DC_HPD1_INTERRUPT;
3959                                         queue_hotplug = true;
3960                                         DRM_DEBUG("IH: HPD1\n");
3961                                 }
3962                                 break;
3963                         case 1:
3964                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont & DC_HPD2_INTERRUPT) {
3965                                         rdev->irq.stat_regs.evergreen.disp_int_cont &= ~DC_HPD2_INTERRUPT;
3966                                         queue_hotplug = true;
3967                                         DRM_DEBUG("IH: HPD2\n");
3968                                 }
3969                                 break;
3970                         case 2:
3971                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont2 & DC_HPD3_INTERRUPT) {
3972                                         rdev->irq.stat_regs.evergreen.disp_int_cont2 &= ~DC_HPD3_INTERRUPT;
3973                                         queue_hotplug = true;
3974                                         DRM_DEBUG("IH: HPD3\n");
3975                                 }
3976                                 break;
3977                         case 3:
3978                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont3 & DC_HPD4_INTERRUPT) {
3979                                         rdev->irq.stat_regs.evergreen.disp_int_cont3 &= ~DC_HPD4_INTERRUPT;
3980                                         queue_hotplug = true;
3981                                         DRM_DEBUG("IH: HPD4\n");
3982                                 }
3983                                 break;
3984                         case 4:
3985                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont4 & DC_HPD5_INTERRUPT) {
3986                                         rdev->irq.stat_regs.evergreen.disp_int_cont4 &= ~DC_HPD5_INTERRUPT;
3987                                         queue_hotplug = true;
3988                                         DRM_DEBUG("IH: HPD5\n");
3989                                 }
3990                                 break;
3991                         case 5:
3992                                 if (rdev->irq.stat_regs.evergreen.disp_int_cont5 & DC_HPD6_INTERRUPT) {
3993                                         rdev->irq.stat_regs.evergreen.disp_int_cont5 &= ~DC_HPD6_INTERRUPT;
3994                                         queue_hotplug = true;
3995                                         DRM_DEBUG("IH: HPD6\n");
3996                                 }
3997                                 break;
3998                         default:
3999                                 DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
4000                                 break;
4001                         }
4002                         break;
4003                 case 146:
4004                 case 147:
4005                         dev_err(rdev->dev, "GPU fault detected: %d 0x%08x\n", src_id, src_data);
4006                         dev_err(rdev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_ADDR   0x%08X\n",
4007                                 RREG32(VM_CONTEXT1_PROTECTION_FAULT_ADDR));
4008                         dev_err(rdev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
4009                                 RREG32(VM_CONTEXT1_PROTECTION_FAULT_STATUS));
4010                         /* reset addr and status */
4011                         WREG32_P(VM_CONTEXT1_CNTL2, 1, ~1);
4012                         break;
4013                 case 176: /* RINGID0 CP_INT */
4014                         radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
4015                         break;
4016                 case 177: /* RINGID1 CP_INT */
4017                         radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
4018                         break;
4019                 case 178: /* RINGID2 CP_INT */
4020                         radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
4021                         break;
4022                 case 181: /* CP EOP event */
4023                         DRM_DEBUG("IH: CP EOP\n");
4024                         switch (ring_id) {
4025                         case 0:
4026                                 radeon_fence_process(rdev, RADEON_RING_TYPE_GFX_INDEX);
4027                                 break;
4028                         case 1:
4029                                 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
4030                                 break;
4031                         case 2:
4032                                 radeon_fence_process(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
4033                                 break;
4034                         }
4035                         break;
4036                 case 224: /* DMA trap event */
4037                         DRM_DEBUG("IH: DMA trap\n");
4038                         radeon_fence_process(rdev, R600_RING_TYPE_DMA_INDEX);
4039                         break;
4040                 case 233: /* GUI IDLE */
4041                         DRM_DEBUG("IH: GUI idle\n");
4042                         break;
4043                 case 244: /* DMA trap event */
4044                         DRM_DEBUG("IH: DMA1 trap\n");
4045                         radeon_fence_process(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
4046                         break;
4047                 default:
4048                         DRM_DEBUG("Unhandled interrupt: %d %d\n", src_id, src_data);
4049                         break;
4050                 }
4051
4052                 /* wptr/rptr are in bytes! */
4053                 rptr += 16;
4054                 rptr &= rdev->ih.ptr_mask;
4055         }
4056         if (queue_hotplug)
4057                 schedule_work(&rdev->hotplug_work);
4058         rdev->ih.rptr = rptr;
4059         WREG32(IH_RB_RPTR, rdev->ih.rptr);
4060         atomic_set(&rdev->ih.lock, 0);
4061
4062         /* make sure wptr hasn't changed while processing */
4063         wptr = si_get_ih_wptr(rdev);
4064         if (wptr != rptr)
4065                 goto restart_ih;
4066
4067         return IRQ_HANDLED;
4068 }
4069
4070 /**
4071  * si_copy_dma - copy pages using the DMA engine
4072  *
4073  * @rdev: radeon_device pointer
4074  * @src_offset: src GPU address
4075  * @dst_offset: dst GPU address
4076  * @num_gpu_pages: number of GPU pages to xfer
4077  * @fence: radeon fence object
4078  *
4079  * Copy GPU paging using the DMA engine (SI).
4080  * Used by the radeon ttm implementation to move pages if
4081  * registered as the asic copy callback.
4082  */
4083 int si_copy_dma(struct radeon_device *rdev,
4084                 uint64_t src_offset, uint64_t dst_offset,
4085                 unsigned num_gpu_pages,
4086                 struct radeon_fence **fence)
4087 {
4088         struct radeon_semaphore *sem = NULL;
4089         int ring_index = rdev->asic->copy.dma_ring_index;
4090         struct radeon_ring *ring = &rdev->ring[ring_index];
4091         u32 size_in_bytes, cur_size_in_bytes;
4092         int i, num_loops;
4093         int r = 0;
4094
4095         r = radeon_semaphore_create(rdev, &sem);
4096         if (r) {
4097                 DRM_ERROR("radeon: moving bo (%d).\n", r);
4098                 return r;
4099         }
4100
4101         size_in_bytes = (num_gpu_pages << RADEON_GPU_PAGE_SHIFT);
4102         num_loops = DIV_ROUND_UP(size_in_bytes, 0xfffff);
4103         r = radeon_ring_lock(rdev, ring, num_loops * 5 + 11);
4104         if (r) {
4105                 DRM_ERROR("radeon: moving bo (%d).\n", r);
4106                 radeon_semaphore_free(rdev, &sem, NULL);
4107                 return r;
4108         }
4109
4110         if (radeon_fence_need_sync(*fence, ring->idx)) {
4111                 radeon_semaphore_sync_rings(rdev, sem, (*fence)->ring,
4112                                             ring->idx);
4113                 radeon_fence_note_sync(*fence, ring->idx);
4114         } else {
4115                 radeon_semaphore_free(rdev, &sem, NULL);
4116         }
4117
4118         for (i = 0; i < num_loops; i++) {
4119                 cur_size_in_bytes = size_in_bytes;
4120                 if (cur_size_in_bytes > 0xFFFFF)
4121                         cur_size_in_bytes = 0xFFFFF;
4122                 size_in_bytes -= cur_size_in_bytes;
4123                 radeon_ring_write(ring, DMA_PACKET(DMA_PACKET_COPY, 1, 0, 0, cur_size_in_bytes));
4124                 radeon_ring_write(ring, dst_offset & 0xffffffff);
4125                 radeon_ring_write(ring, src_offset & 0xffffffff);
4126                 radeon_ring_write(ring, upper_32_bits(dst_offset) & 0xff);
4127                 radeon_ring_write(ring, upper_32_bits(src_offset) & 0xff);
4128                 src_offset += cur_size_in_bytes;
4129                 dst_offset += cur_size_in_bytes;
4130         }
4131
4132         r = radeon_fence_emit(rdev, fence, ring->idx);
4133         if (r) {
4134                 radeon_ring_unlock_undo(rdev, ring);
4135                 return r;
4136         }
4137
4138         radeon_ring_unlock_commit(rdev, ring);
4139         radeon_semaphore_free(rdev, &sem, *fence);
4140
4141         return r;
4142 }
4143
4144 /*
4145  * startup/shutdown callbacks
4146  */
4147 static int si_startup(struct radeon_device *rdev)
4148 {
4149         struct radeon_ring *ring;
4150         int r;
4151
4152         if (!rdev->me_fw || !rdev->pfp_fw || !rdev->ce_fw ||
4153             !rdev->rlc_fw || !rdev->mc_fw) {
4154                 r = si_init_microcode(rdev);
4155                 if (r) {
4156                         DRM_ERROR("Failed to load firmware!\n");
4157                         return r;
4158                 }
4159         }
4160
4161         r = si_mc_load_microcode(rdev);
4162         if (r) {
4163                 DRM_ERROR("Failed to load MC firmware!\n");
4164                 return r;
4165         }
4166
4167         r = r600_vram_scratch_init(rdev);
4168         if (r)
4169                 return r;
4170
4171         si_mc_program(rdev);
4172         r = si_pcie_gart_enable(rdev);
4173         if (r)
4174                 return r;
4175         si_gpu_init(rdev);
4176
4177 #if 0
4178         r = evergreen_blit_init(rdev);
4179         if (r) {
4180                 r600_blit_fini(rdev);
4181                 rdev->asic->copy = NULL;
4182                 dev_warn(rdev->dev, "failed blitter (%d) falling back to memcpy\n", r);
4183         }
4184 #endif
4185         /* allocate rlc buffers */
4186         r = si_rlc_init(rdev);
4187         if (r) {
4188                 DRM_ERROR("Failed to init rlc BOs!\n");
4189                 return r;
4190         }
4191
4192         /* allocate wb buffer */
4193         r = radeon_wb_init(rdev);
4194         if (r)
4195                 return r;
4196
4197         r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
4198         if (r) {
4199                 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
4200                 return r;
4201         }
4202
4203         r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP1_INDEX);
4204         if (r) {
4205                 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
4206                 return r;
4207         }
4208
4209         r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_CP2_INDEX);
4210         if (r) {
4211                 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
4212                 return r;
4213         }
4214
4215         r = radeon_fence_driver_start_ring(rdev, R600_RING_TYPE_DMA_INDEX);
4216         if (r) {
4217                 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
4218                 return r;
4219         }
4220
4221         r = radeon_fence_driver_start_ring(rdev, CAYMAN_RING_TYPE_DMA1_INDEX);
4222         if (r) {
4223                 dev_err(rdev->dev, "failed initializing DMA fences (%d).\n", r);
4224                 return r;
4225         }
4226
4227         /* Enable IRQ */
4228         r = si_irq_init(rdev);
4229         if (r) {
4230                 DRM_ERROR("radeon: IH init failed (%d).\n", r);
4231                 radeon_irq_kms_fini(rdev);
4232                 return r;
4233         }
4234         si_irq_set(rdev);
4235
4236         ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
4237         r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP_RPTR_OFFSET,
4238                              CP_RB0_RPTR, CP_RB0_WPTR,
4239                              0, 0xfffff, RADEON_CP_PACKET2);
4240         if (r)
4241                 return r;
4242
4243         ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
4244         r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP1_RPTR_OFFSET,
4245                              CP_RB1_RPTR, CP_RB1_WPTR,
4246                              0, 0xfffff, RADEON_CP_PACKET2);
4247         if (r)
4248                 return r;
4249
4250         ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
4251         r = radeon_ring_init(rdev, ring, ring->ring_size, RADEON_WB_CP2_RPTR_OFFSET,
4252                              CP_RB2_RPTR, CP_RB2_WPTR,
4253                              0, 0xfffff, RADEON_CP_PACKET2);
4254         if (r)
4255                 return r;
4256
4257         ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
4258         r = radeon_ring_init(rdev, ring, ring->ring_size, R600_WB_DMA_RPTR_OFFSET,
4259                              DMA_RB_RPTR + DMA0_REGISTER_OFFSET,
4260                              DMA_RB_WPTR + DMA0_REGISTER_OFFSET,
4261                              2, 0x3fffc, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
4262         if (r)
4263                 return r;
4264
4265         ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
4266         r = radeon_ring_init(rdev, ring, ring->ring_size, CAYMAN_WB_DMA1_RPTR_OFFSET,
4267                              DMA_RB_RPTR + DMA1_REGISTER_OFFSET,
4268                              DMA_RB_WPTR + DMA1_REGISTER_OFFSET,
4269                              2, 0x3fffc, DMA_PACKET(DMA_PACKET_NOP, 0, 0, 0, 0));
4270         if (r)
4271                 return r;
4272
4273         r = si_cp_load_microcode(rdev);
4274         if (r)
4275                 return r;
4276         r = si_cp_resume(rdev);
4277         if (r)
4278                 return r;
4279
4280         r = cayman_dma_resume(rdev);
4281         if (r)
4282                 return r;
4283
4284         r = radeon_ib_pool_init(rdev);
4285         if (r) {
4286                 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
4287                 return r;
4288         }
4289
4290         r = radeon_vm_manager_init(rdev);
4291         if (r) {
4292                 dev_err(rdev->dev, "vm manager initialization failed (%d).\n", r);
4293                 return r;
4294         }
4295
4296         return 0;
4297 }
4298
4299 int si_resume(struct radeon_device *rdev)
4300 {
4301         int r;
4302
4303         /* Do not reset GPU before posting, on rv770 hw unlike on r500 hw,
4304          * posting will perform necessary task to bring back GPU into good
4305          * shape.
4306          */
4307         /* post card */
4308         atom_asic_init(rdev->mode_info.atom_context);
4309
4310         rdev->accel_working = true;
4311         r = si_startup(rdev);
4312         if (r) {
4313                 DRM_ERROR("si startup failed on resume\n");
4314                 rdev->accel_working = false;
4315                 return r;
4316         }
4317
4318         return r;
4319
4320 }
4321
4322 int si_suspend(struct radeon_device *rdev)
4323 {
4324         si_cp_enable(rdev, false);
4325         cayman_dma_stop(rdev);
4326         si_irq_suspend(rdev);
4327         radeon_wb_disable(rdev);
4328         si_pcie_gart_disable(rdev);
4329         return 0;
4330 }
4331
4332 /* Plan is to move initialization in that function and use
4333  * helper function so that radeon_device_init pretty much
4334  * do nothing more than calling asic specific function. This
4335  * should also allow to remove a bunch of callback function
4336  * like vram_info.
4337  */
4338 int si_init(struct radeon_device *rdev)
4339 {
4340         struct radeon_ring *ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
4341         int r;
4342
4343         /* Read BIOS */
4344         if (!radeon_get_bios(rdev)) {
4345                 if (ASIC_IS_AVIVO(rdev))
4346                         return -EINVAL;
4347         }
4348         /* Must be an ATOMBIOS */
4349         if (!rdev->is_atom_bios) {
4350                 dev_err(rdev->dev, "Expecting atombios for cayman GPU\n");
4351                 return -EINVAL;
4352         }
4353         r = radeon_atombios_init(rdev);
4354         if (r)
4355                 return r;
4356
4357         /* Post card if necessary */
4358         if (!radeon_card_posted(rdev)) {
4359                 if (!rdev->bios) {
4360                         dev_err(rdev->dev, "Card not posted and no BIOS - ignoring\n");
4361                         return -EINVAL;
4362                 }
4363                 DRM_INFO("GPU not posted. posting now...\n");
4364                 atom_asic_init(rdev->mode_info.atom_context);
4365         }
4366         /* Initialize scratch registers */
4367         si_scratch_init(rdev);
4368         /* Initialize surface registers */
4369         radeon_surface_init(rdev);
4370         /* Initialize clocks */
4371         radeon_get_clock_info(rdev->ddev);
4372
4373         /* Fence driver */
4374         r = radeon_fence_driver_init(rdev);
4375         if (r)
4376                 return r;
4377
4378         /* initialize memory controller */
4379         r = si_mc_init(rdev);
4380         if (r)
4381                 return r;
4382         /* Memory manager */
4383         r = radeon_bo_init(rdev);
4384         if (r)
4385                 return r;
4386
4387         r = radeon_irq_kms_init(rdev);
4388         if (r)
4389                 return r;
4390
4391         ring = &rdev->ring[RADEON_RING_TYPE_GFX_INDEX];
4392         ring->ring_obj = NULL;
4393         r600_ring_init(rdev, ring, 1024 * 1024);
4394
4395         ring = &rdev->ring[CAYMAN_RING_TYPE_CP1_INDEX];
4396         ring->ring_obj = NULL;
4397         r600_ring_init(rdev, ring, 1024 * 1024);
4398
4399         ring = &rdev->ring[CAYMAN_RING_TYPE_CP2_INDEX];
4400         ring->ring_obj = NULL;
4401         r600_ring_init(rdev, ring, 1024 * 1024);
4402
4403         ring = &rdev->ring[R600_RING_TYPE_DMA_INDEX];
4404         ring->ring_obj = NULL;
4405         r600_ring_init(rdev, ring, 64 * 1024);
4406
4407         ring = &rdev->ring[CAYMAN_RING_TYPE_DMA1_INDEX];
4408         ring->ring_obj = NULL;
4409         r600_ring_init(rdev, ring, 64 * 1024);
4410
4411         rdev->ih.ring_obj = NULL;
4412         r600_ih_ring_init(rdev, 64 * 1024);
4413
4414         r = r600_pcie_gart_init(rdev);
4415         if (r)
4416                 return r;
4417
4418         rdev->accel_working = true;
4419         r = si_startup(rdev);
4420         if (r) {
4421                 dev_err(rdev->dev, "disabling GPU acceleration\n");
4422                 si_cp_fini(rdev);
4423                 cayman_dma_fini(rdev);
4424                 si_irq_fini(rdev);
4425                 si_rlc_fini(rdev);
4426                 radeon_wb_fini(rdev);
4427                 radeon_ib_pool_fini(rdev);
4428                 radeon_vm_manager_fini(rdev);
4429                 radeon_irq_kms_fini(rdev);
4430                 si_pcie_gart_fini(rdev);
4431                 rdev->accel_working = false;
4432         }
4433
4434         /* Don't start up if the MC ucode is missing.
4435          * The default clocks and voltages before the MC ucode
4436          * is loaded are not suffient for advanced operations.
4437          */
4438         if (!rdev->mc_fw) {
4439                 DRM_ERROR("radeon: MC ucode required for NI+.\n");
4440                 return -EINVAL;
4441         }
4442
4443         return 0;
4444 }
4445
4446 void si_fini(struct radeon_device *rdev)
4447 {
4448 #if 0
4449         r600_blit_fini(rdev);
4450 #endif
4451         si_cp_fini(rdev);
4452         cayman_dma_fini(rdev);
4453         si_irq_fini(rdev);
4454         si_rlc_fini(rdev);
4455         radeon_wb_fini(rdev);
4456         radeon_vm_manager_fini(rdev);
4457         radeon_ib_pool_fini(rdev);
4458         radeon_irq_kms_fini(rdev);
4459         si_pcie_gart_fini(rdev);
4460         r600_vram_scratch_fini(rdev);
4461         radeon_gem_fini(rdev);
4462         radeon_fence_driver_fini(rdev);
4463         radeon_bo_fini(rdev);
4464         radeon_atombios_fini(rdev);
4465         kfree(rdev->bios);
4466         rdev->bios = NULL;
4467 }
4468
4469 /**
4470  * si_get_gpu_clock - return GPU clock counter snapshot
4471  *
4472  * @rdev: radeon_device pointer
4473  *
4474  * Fetches a GPU clock counter snapshot (SI).
4475  * Returns the 64 bit clock counter snapshot.
4476  */
4477 uint64_t si_get_gpu_clock(struct radeon_device *rdev)
4478 {
4479         uint64_t clock;
4480
4481         mutex_lock(&rdev->gpu_clock_mutex);
4482         WREG32(RLC_CAPTURE_GPU_CLOCK_COUNT, 1);
4483         clock = (uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_LSB) |
4484                 ((uint64_t)RREG32(RLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
4485         mutex_unlock(&rdev->gpu_clock_mutex);
4486         return clock;
4487 }