sys/dev/drm/radeon/rs690.c

   1 /*
   2  * Copyright 2008 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  * Copyright 2009 Jerome Glisse.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors: Dave Airlie
  25  *          Alex Deucher
  26  *          Jerome Glisse
  27  *
  28  * $FreeBSD: head/sys/dev/drm2/radeon/rs690.c 254885 2013-08-25 19:37:15Z dumbbell $
  29  */
  30
  31 #include <drm/drmP.h>
  32 #include "radeon.h"
  33 #include "radeon_asic.h"
  34 #include "atom.h"
  35 #include "rs690d.h"
  36
  37 int rs690_mc_wait_for_idle(struct radeon_device *rdev)
  38 {
  39         unsigned i;
  40         uint32_t tmp;
  41
  42         for (i = 0; i < rdev->usec_timeout; i++) {
  43                 /* read MC_STATUS */
  44                 tmp = RREG32_MC(R_000090_MC_SYSTEM_STATUS);
  45                 if (G_000090_MC_SYSTEM_IDLE(tmp))
  46                         return 0;
  47                 DRM_UDELAY(1);
  48         }
  49         return -1;
  50 }
  51
  52 static void rs690_gpu_init(struct radeon_device *rdev)
  53 {
  54         /* FIXME: is this correct ? */
  55         r420_pipes_init(rdev);
  56         if (rs690_mc_wait_for_idle(rdev)) {
  57                 DRM_ERROR("Failed to wait MC idle while "
  58                        "programming pipes. Bad things might happen.\n");
  59         }
  60 }
  61
  62 union igp_info {
  63         struct _ATOM_INTEGRATED_SYSTEM_INFO info;
  64         struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 info_v2;
  65 };
  66
  67 void rs690_pm_info(struct radeon_device *rdev)
  68 {
  69         int index = GetIndexIntoMasterTable(DATA, IntegratedSystemInfo);
  70         union igp_info *info;
  71         uint16_t data_offset;
  72         uint8_t frev, crev;
  73         fixed20_12 tmp;
  74
  75         if (atom_parse_data_header(rdev->mode_info.atom_context, index, NULL,
  76                                    &frev, &crev, &data_offset)) {
  77                 info = (union igp_info *)((uintptr_t)rdev->mode_info.atom_context->bios + data_offset);
  78
  79                 /* Get various system informations from bios */
  80                 switch (crev) {
  81                 case 1:
  82                         tmp.full = dfixed_const(100);
  83                         rdev->pm.igp_sideport_mclk.full = dfixed_const(le32_to_cpu(info->info.ulBootUpMemoryClock));
  84                         rdev->pm.igp_sideport_mclk.full = dfixed_div(rdev->pm.igp_sideport_mclk, tmp);
  85                         if (le16_to_cpu(info->info.usK8MemoryClock))
  86                                 rdev->pm.igp_system_mclk.full = dfixed_const(le16_to_cpu(info->info.usK8MemoryClock));
  87                         else if (rdev->clock.default_mclk) {
  88                                 rdev->pm.igp_system_mclk.full = dfixed_const(rdev->clock.default_mclk);
  89                                 rdev->pm.igp_system_mclk.full = dfixed_div(rdev->pm.igp_system_mclk, tmp);
  90                         } else
  91                                 rdev->pm.igp_system_mclk.full = dfixed_const(400);
  92                         rdev->pm.igp_ht_link_clk.full = dfixed_const(le16_to_cpu(info->info.usFSBClock));
  93                         rdev->pm.igp_ht_link_width.full = dfixed_const(info->info.ucHTLinkWidth);
  94                         break;
  95                 case 2:
  96                         tmp.full = dfixed_const(100);
  97                         rdev->pm.igp_sideport_mclk.full = dfixed_const(le32_to_cpu(info->info_v2.ulBootUpSidePortClock));
  98                         rdev->pm.igp_sideport_mclk.full = dfixed_div(rdev->pm.igp_sideport_mclk, tmp);
  99                         if (le32_to_cpu(info->info_v2.ulBootUpUMAClock))
 100                                 rdev->pm.igp_system_mclk.full = dfixed_const(le32_to_cpu(info->info_v2.ulBootUpUMAClock));
 101                         else if (rdev->clock.default_mclk)
 102                                 rdev->pm.igp_system_mclk.full = dfixed_const(rdev->clock.default_mclk);
 103                         else
 104                                 rdev->pm.igp_system_mclk.full = dfixed_const(66700);
 105                         rdev->pm.igp_system_mclk.full = dfixed_div(rdev->pm.igp_system_mclk, tmp);
 106                         rdev->pm.igp_ht_link_clk.full = dfixed_const(le32_to_cpu(info->info_v2.ulHTLinkFreq));
 107                         rdev->pm.igp_ht_link_clk.full = dfixed_div(rdev->pm.igp_ht_link_clk, tmp);
 108                         rdev->pm.igp_ht_link_width.full = dfixed_const(le16_to_cpu(info->info_v2.usMinHTLinkWidth));
 109                         break;
 110                 default:
 111                         /* We assume the slower possible clock ie worst case */
 112                         rdev->pm.igp_sideport_mclk.full = dfixed_const(200);
 113                         rdev->pm.igp_system_mclk.full = dfixed_const(200);
 114                         rdev->pm.igp_ht_link_clk.full = dfixed_const(1000);
 115                         rdev->pm.igp_ht_link_width.full = dfixed_const(8);
 116                         DRM_ERROR("No integrated system info for your GPU, using safe default\n");
 117                         break;
 118                 }
 119         } else {
 120                 /* We assume the slower possible clock ie worst case */
 121                 rdev->pm.igp_sideport_mclk.full = dfixed_const(200);
 122                 rdev->pm.igp_system_mclk.full = dfixed_const(200);
 123                 rdev->pm.igp_ht_link_clk.full = dfixed_const(1000);
 124                 rdev->pm.igp_ht_link_width.full = dfixed_const(8);
 125                 DRM_ERROR("No integrated system info for your GPU, using safe default\n");
 126         }
 127         /* Compute various bandwidth */
 128         /* k8_bandwidth = (memory_clk / 2) * 2 * 8 * 0.5 = memory_clk * 4  */
 129         tmp.full = dfixed_const(4);
 130         rdev->pm.k8_bandwidth.full = dfixed_mul(rdev->pm.igp_system_mclk, tmp);
 131         /* ht_bandwidth = ht_clk * 2 * ht_width / 8 * 0.8
 132          *              = ht_clk * ht_width / 5
 133          */
 134         tmp.full = dfixed_const(5);
 135         rdev->pm.ht_bandwidth.full = dfixed_mul(rdev->pm.igp_ht_link_clk,
 136                                                 rdev->pm.igp_ht_link_width);
 137         rdev->pm.ht_bandwidth.full = dfixed_div(rdev->pm.ht_bandwidth, tmp);
 138         if (tmp.full < rdev->pm.max_bandwidth.full) {
 139                 /* HT link is a limiting factor */
 140                 rdev->pm.max_bandwidth.full = tmp.full;
 141         }
 142         /* sideport_bandwidth = (sideport_clk / 2) * 2 * 2 * 0.7
 143          *                    = (sideport_clk * 14) / 10
 144          */
 145         tmp.full = dfixed_const(14);
 146         rdev->pm.sideport_bandwidth.full = dfixed_mul(rdev->pm.igp_sideport_mclk, tmp);
 147         tmp.full = dfixed_const(10);
 148         rdev->pm.sideport_bandwidth.full = dfixed_div(rdev->pm.sideport_bandwidth, tmp);
 149 }
 150
 151 static void rs690_mc_init(struct radeon_device *rdev)
 152 {
 153         u64 base;
 154
 155         rs400_gart_adjust_size(rdev);
 156         rdev->mc.vram_is_ddr = true;
 157         rdev->mc.vram_width = 128;
 158         rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
 159         rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
 160         rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0);
 161         rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0);
 162         rdev->mc.visible_vram_size = rdev->mc.aper_size;
 163         base = RREG32_MC(R_000100_MCCFG_FB_LOCATION);
 164         base = G_000100_MC_FB_START(base) << 16;
 165         rdev->mc.igp_sideport_enabled = radeon_atombios_sideport_present(rdev);
 166         rs690_pm_info(rdev);
 167         radeon_vram_location(rdev, &rdev->mc, base);
 168         rdev->mc.gtt_base_align = rdev->mc.gtt_size - 1;
 169         radeon_gtt_location(rdev, &rdev->mc);
 170         radeon_update_bandwidth_info(rdev);
 171 }
 172
 173 void rs690_line_buffer_adjust(struct radeon_device *rdev,
 174                               struct drm_display_mode *mode1,
 175                               struct drm_display_mode *mode2)
 176 {
 177         u32 tmp;
 178
 179         /*
 180          * Line Buffer Setup
 181          * There is a single line buffer shared by both display controllers.
 182          * R_006520_DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
 183          * the display controllers.  The paritioning can either be done
 184          * manually or via one of four preset allocations specified in bits 1:0:
 185          *  0 - line buffer is divided in half and shared between crtc
 186          *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
 187          *  2 - D1 gets the whole buffer
 188          *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
 189          * Setting bit 2 of R_006520_DC_LB_MEMORY_SPLIT controls switches to manual
 190          * allocation mode. In manual allocation mode, D1 always starts at 0,
 191          * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
 192          */
 193         tmp = RREG32(R_006520_DC_LB_MEMORY_SPLIT) & C_006520_DC_LB_MEMORY_SPLIT;
 194         tmp &= ~C_006520_DC_LB_MEMORY_SPLIT_MODE;
 195         /* auto */
 196         if (mode1 && mode2) {
 197                 if (mode1->hdisplay > mode2->hdisplay) {
 198                         if (mode1->hdisplay > 2560)
 199                                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
 200                         else
 201                                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
 202                 } else if (mode2->hdisplay > mode1->hdisplay) {
 203                         if (mode2->hdisplay > 2560)
 204                                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
 205                         else
 206                                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
 207                 } else
 208                         tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
 209         } else if (mode1) {
 210                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_ONLY;
 211         } else if (mode2) {
 212                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
 213         }
 214         WREG32(R_006520_DC_LB_MEMORY_SPLIT, tmp);
 215 }
 216
 217 struct rs690_watermark {
 218         u32        lb_request_fifo_depth;
 219         fixed20_12 num_line_pair;
 220         fixed20_12 estimated_width;
 221         fixed20_12 worst_case_latency;
 222         fixed20_12 consumption_rate;
 223         fixed20_12 active_time;
 224         fixed20_12 dbpp;
 225         fixed20_12 priority_mark_max;
 226         fixed20_12 priority_mark;
 227         fixed20_12 sclk;
 228 };
 229
 230 static void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
 231                                   struct radeon_crtc *crtc,
 232                                   struct rs690_watermark *wm)
 233 {
 234         struct drm_display_mode *mode = &crtc->base.mode;
 235         fixed20_12 a, b, c;
 236         fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
 237         fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
 238
 239         if (!crtc->base.enabled) {
 240                 /* FIXME: wouldn't it better to set priority mark to maximum */
 241                 wm->lb_request_fifo_depth = 4;
 242                 return;
 243         }
 244
 245         if (crtc->vsc.full > dfixed_const(2))
 246                 wm->num_line_pair.full = dfixed_const(2);
 247         else
 248                 wm->num_line_pair.full = dfixed_const(1);
 249
 250         b.full = dfixed_const(mode->crtc_hdisplay);
 251         c.full = dfixed_const(256);
 252         a.full = dfixed_div(b, c);
 253         request_fifo_depth.full = dfixed_mul(a, wm->num_line_pair);
 254         request_fifo_depth.full = dfixed_ceil(request_fifo_depth);
 255         if (a.full < dfixed_const(4)) {
 256                 wm->lb_request_fifo_depth = 4;
 257         } else {
 258                 wm->lb_request_fifo_depth = dfixed_trunc(request_fifo_depth);
 259         }
 260
 261         /* Determine consumption rate
 262          *  pclk = pixel clock period(ns) = 1000 / (mode.clock / 1000)
 263          *  vtaps = number of vertical taps,
 264          *  vsc = vertical scaling ratio, defined as source/destination
 265          *  hsc = horizontal scaling ration, defined as source/destination
 266          */
 267         a.full = dfixed_const(mode->clock);
 268         b.full = dfixed_const(1000);
 269         a.full = dfixed_div(a, b);
 270         pclk.full = dfixed_div(b, a);
 271         if (crtc->rmx_type != RMX_OFF) {
 272                 b.full = dfixed_const(2);
 273                 if (crtc->vsc.full > b.full)
 274                         b.full = crtc->vsc.full;
 275                 b.full = dfixed_mul(b, crtc->hsc);
 276                 c.full = dfixed_const(2);
 277                 b.full = dfixed_div(b, c);
 278                 consumption_time.full = dfixed_div(pclk, b);
 279         } else {
 280                 consumption_time.full = pclk.full;
 281         }
 282         a.full = dfixed_const(1);
 283         wm->consumption_rate.full = dfixed_div(a, consumption_time);
 284
 285
 286         /* Determine line time
 287          *  LineTime = total time for one line of displayhtotal
 288          *  LineTime = total number of horizontal pixels
 289          *  pclk = pixel clock period(ns)
 290          */
 291         a.full = dfixed_const(crtc->base.mode.crtc_htotal);
 292         line_time.full = dfixed_mul(a, pclk);
 293
 294         /* Determine active time
 295          *  ActiveTime = time of active region of display within one line,
 296          *  hactive = total number of horizontal active pixels
 297          *  htotal = total number of horizontal pixels
 298          */
 299         a.full = dfixed_const(crtc->base.mode.crtc_htotal);
 300         b.full = dfixed_const(crtc->base.mode.crtc_hdisplay);
 301         wm->active_time.full = dfixed_mul(line_time, b);
 302         wm->active_time.full = dfixed_div(wm->active_time, a);
 303
 304         /* Maximun bandwidth is the minimun bandwidth of all component */
 305         rdev->pm.max_bandwidth = rdev->pm.core_bandwidth;
 306         if (rdev->mc.igp_sideport_enabled) {
 307                 if (rdev->pm.max_bandwidth.full > rdev->pm.sideport_bandwidth.full &&
 308                         rdev->pm.sideport_bandwidth.full)
 309                         rdev->pm.max_bandwidth = rdev->pm.sideport_bandwidth;
 310 #ifdef DUMBBELL_WIP
 311                 read_delay_latency.full = dfixed_const(370 * 800 * 1000);
 312 #endif /* DUMBBELL_WIP */
 313                 read_delay_latency.full = UINT_MAX;
 314                 read_delay_latency.full = dfixed_div(read_delay_latency,
 315                         rdev->pm.igp_sideport_mclk);
 316         } else {
 317                 if (rdev->pm.max_bandwidth.full > rdev->pm.k8_bandwidth.full &&
 318                         rdev->pm.k8_bandwidth.full)
 319                         rdev->pm.max_bandwidth = rdev->pm.k8_bandwidth;
 320                 if (rdev->pm.max_bandwidth.full > rdev->pm.ht_bandwidth.full &&
 321                         rdev->pm.ht_bandwidth.full)
 322                         rdev->pm.max_bandwidth = rdev->pm.ht_bandwidth;
 323                 read_delay_latency.full = dfixed_const(5000);
 324         }
 325
 326         /* sclk = system clocks(ns) = 1000 / max_bandwidth / 16 */
 327         a.full = dfixed_const(16);
 328         rdev->pm.sclk.full = dfixed_mul(rdev->pm.max_bandwidth, a);
 329         a.full = dfixed_const(1000);
 330         rdev->pm.sclk.full = dfixed_div(a, rdev->pm.sclk);
 331         /* Determine chunk time
 332          * ChunkTime = the time it takes the DCP to send one chunk of data
 333          * to the LB which consists of pipeline delay and inter chunk gap
 334          * sclk = system clock(ns)
 335          */
 336         a.full = dfixed_const(256 * 13);
 337         chunk_time.full = dfixed_mul(rdev->pm.sclk, a);
 338         a.full = dfixed_const(10);
 339         chunk_time.full = dfixed_div(chunk_time, a);
 340
 341         /* Determine the worst case latency
 342          * NumLinePair = Number of line pairs to request(1=2 lines, 2=4 lines)
 343          * WorstCaseLatency = worst case time from urgent to when the MC starts
 344          *                    to return data
 345          * READ_DELAY_IDLE_MAX = constant of 1us
 346          * ChunkTime = time it takes the DCP to send one chunk of data to the LB
 347          *             which consists of pipeline delay and inter chunk gap
 348          */
 349         if (dfixed_trunc(wm->num_line_pair) > 1) {
 350                 a.full = dfixed_const(3);
 351                 wm->worst_case_latency.full = dfixed_mul(a, chunk_time);
 352                 wm->worst_case_latency.full += read_delay_latency.full;
 353         } else {
 354                 a.full = dfixed_const(2);
 355                 wm->worst_case_latency.full = dfixed_mul(a, chunk_time);
 356                 wm->worst_case_latency.full += read_delay_latency.full;
 357         }
 358
 359         /* Determine the tolerable latency
 360          * TolerableLatency = Any given request has only 1 line time
 361          *                    for the data to be returned
 362          * LBRequestFifoDepth = Number of chunk requests the LB can
 363          *                      put into the request FIFO for a display
 364          *  LineTime = total time for one line of display
 365          *  ChunkTime = the time it takes the DCP to send one chunk
 366          *              of data to the LB which consists of
 367          *  pipeline delay and inter chunk gap
 368          */
 369         if ((2+wm->lb_request_fifo_depth) >= dfixed_trunc(request_fifo_depth)) {
 370                 tolerable_latency.full = line_time.full;
 371         } else {
 372                 tolerable_latency.full = dfixed_const(wm->lb_request_fifo_depth - 2);
 373                 tolerable_latency.full = request_fifo_depth.full - tolerable_latency.full;
 374                 tolerable_latency.full = dfixed_mul(tolerable_latency, chunk_time);
 375                 tolerable_latency.full = line_time.full - tolerable_latency.full;
 376         }
 377         /* We assume worst case 32bits (4 bytes) */
 378         wm->dbpp.full = dfixed_const(4 * 8);
 379
 380         /* Determine the maximum priority mark
 381          *  width = viewport width in pixels
 382          */
 383         a.full = dfixed_const(16);
 384         wm->priority_mark_max.full = dfixed_const(crtc->base.mode.crtc_hdisplay);
 385         wm->priority_mark_max.full = dfixed_div(wm->priority_mark_max, a);
 386         wm->priority_mark_max.full = dfixed_ceil(wm->priority_mark_max);
 387
 388         /* Determine estimated width */
 389         estimated_width.full = tolerable_latency.full - wm->worst_case_latency.full;
 390         estimated_width.full = dfixed_div(estimated_width, consumption_time);
 391         if (dfixed_trunc(estimated_width) > crtc->base.mode.crtc_hdisplay) {
 392                 wm->priority_mark.full = dfixed_const(10);
 393         } else {
 394                 a.full = dfixed_const(16);
 395                 wm->priority_mark.full = dfixed_div(estimated_width, a);
 396                 wm->priority_mark.full = dfixed_ceil(wm->priority_mark);
 397                 wm->priority_mark.full = wm->priority_mark_max.full - wm->priority_mark.full;
 398         }
 399 }
 400
 401 void rs690_bandwidth_update(struct radeon_device *rdev)
 402 {
 403         struct drm_display_mode *mode0 = NULL;
 404         struct drm_display_mode *mode1 = NULL;
 405         struct rs690_watermark wm0;
 406         struct rs690_watermark wm1;
 407         u32 tmp;
 408         u32 d1mode_priority_a_cnt = S_006548_D1MODE_PRIORITY_A_OFF(1);
 409         u32 d2mode_priority_a_cnt = S_006548_D1MODE_PRIORITY_A_OFF(1);
 410         fixed20_12 priority_mark02, priority_mark12, fill_rate;
 411         fixed20_12 a, b;
 412
 413         radeon_update_display_priority(rdev);
 414
 415         if (rdev->mode_info.crtcs[0]->base.enabled)
 416                 mode0 = &rdev->mode_info.crtcs[0]->base.mode;
 417         if (rdev->mode_info.crtcs[1]->base.enabled)
 418                 mode1 = &rdev->mode_info.crtcs[1]->base.mode;
 419         /*
 420          * Set display0/1 priority up in the memory controller for
 421          * modes if the user specifies HIGH for displaypriority
 422          * option.
 423          */
 424         if ((rdev->disp_priority == 2) &&
 425             ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))) {
 426                 tmp = RREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER);
 427                 tmp &= C_000104_MC_DISP0R_INIT_LAT;
 428                 tmp &= C_000104_MC_DISP1R_INIT_LAT;
 429                 if (mode0)
 430                         tmp |= S_000104_MC_DISP0R_INIT_LAT(1);
 431                 if (mode1)
 432                         tmp |= S_000104_MC_DISP1R_INIT_LAT(1);
 433                 WREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER, tmp);
 434         }
 435         rs690_line_buffer_adjust(rdev, mode0, mode1);
 436
 437         if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
 438                 WREG32(R_006C9C_DCP_CONTROL, 0);
 439         if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
 440                 WREG32(R_006C9C_DCP_CONTROL, 2);
 441
 442         rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0);
 443         rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1);
 444
 445         tmp = (wm0.lb_request_fifo_depth - 1);
 446         tmp |= (wm1.lb_request_fifo_depth - 1) << 16;
 447         WREG32(R_006D58_LB_MAX_REQ_OUTSTANDING, tmp);
 448
 449         if (mode0 && mode1) {
 450                 if (dfixed_trunc(wm0.dbpp) > 64)
 451                         a.full = dfixed_mul(wm0.dbpp, wm0.num_line_pair);
 452                 else
 453                         a.full = wm0.num_line_pair.full;
 454                 if (dfixed_trunc(wm1.dbpp) > 64)
 455                         b.full = dfixed_mul(wm1.dbpp, wm1.num_line_pair);
 456                 else
 457                         b.full = wm1.num_line_pair.full;
 458                 a.full += b.full;
 459                 fill_rate.full = dfixed_div(wm0.sclk, a);
 460                 if (wm0.consumption_rate.full > fill_rate.full) {
 461                         b.full = wm0.consumption_rate.full - fill_rate.full;
 462                         b.full = dfixed_mul(b, wm0.active_time);
 463                         a.full = dfixed_mul(wm0.worst_case_latency,
 464                                                 wm0.consumption_rate);
 465                         a.full = a.full + b.full;
 466                         b.full = dfixed_const(16 * 1000);
 467                         priority_mark02.full = dfixed_div(a, b);
 468                 } else {
 469                         a.full = dfixed_mul(wm0.worst_case_latency,
 470                                                 wm0.consumption_rate);
 471                         b.full = dfixed_const(16 * 1000);
 472                         priority_mark02.full = dfixed_div(a, b);
 473                 }
 474                 if (wm1.consumption_rate.full > fill_rate.full) {
 475                         b.full = wm1.consumption_rate.full - fill_rate.full;
 476                         b.full = dfixed_mul(b, wm1.active_time);
 477                         a.full = dfixed_mul(wm1.worst_case_latency,
 478                                                 wm1.consumption_rate);
 479                         a.full = a.full + b.full;
 480                         b.full = dfixed_const(16 * 1000);
 481                         priority_mark12.full = dfixed_div(a, b);
 482                 } else {
 483                         a.full = dfixed_mul(wm1.worst_case_latency,
 484                                                 wm1.consumption_rate);
 485                         b.full = dfixed_const(16 * 1000);
 486                         priority_mark12.full = dfixed_div(a, b);
 487                 }
 488                 if (wm0.priority_mark.full > priority_mark02.full)
 489                         priority_mark02.full = wm0.priority_mark.full;
 490                 if (dfixed_trunc(priority_mark02) < 0)
 491                         priority_mark02.full = 0;
 492                 if (wm0.priority_mark_max.full > priority_mark02.full)
 493                         priority_mark02.full = wm0.priority_mark_max.full;
 494                 if (wm1.priority_mark.full > priority_mark12.full)
 495                         priority_mark12.full = wm1.priority_mark.full;
 496                 if (dfixed_trunc(priority_mark12) < 0)
 497                         priority_mark12.full = 0;
 498                 if (wm1.priority_mark_max.full > priority_mark12.full)
 499                         priority_mark12.full = wm1.priority_mark_max.full;
 500                 d1mode_priority_a_cnt = dfixed_trunc(priority_mark02);
 501                 d2mode_priority_a_cnt = dfixed_trunc(priority_mark12);
 502                 if (rdev->disp_priority == 2) {
 503                         d1mode_priority_a_cnt |= S_006548_D1MODE_PRIORITY_A_ALWAYS_ON(1);
 504                         d2mode_priority_a_cnt |= S_006D48_D2MODE_PRIORITY_A_ALWAYS_ON(1);
 505                 }
 506         } else if (mode0) {
 507                 if (dfixed_trunc(wm0.dbpp) > 64)
 508                         a.full = dfixed_mul(wm0.dbpp, wm0.num_line_pair);
 509                 else
 510                         a.full = wm0.num_line_pair.full;
 511                 fill_rate.full = dfixed_div(wm0.sclk, a);
 512                 if (wm0.consumption_rate.full > fill_rate.full) {
 513                         b.full = wm0.consumption_rate.full - fill_rate.full;
 514                         b.full = dfixed_mul(b, wm0.active_time);
 515                         a.full = dfixed_mul(wm0.worst_case_latency,
 516                                                 wm0.consumption_rate);
 517                         a.full = a.full + b.full;
 518                         b.full = dfixed_const(16 * 1000);
 519                         priority_mark02.full = dfixed_div(a, b);
 520                 } else {
 521                         a.full = dfixed_mul(wm0.worst_case_latency,
 522                                                 wm0.consumption_rate);
 523                         b.full = dfixed_const(16 * 1000);
 524                         priority_mark02.full = dfixed_div(a, b);
 525                 }
 526                 if (wm0.priority_mark.full > priority_mark02.full)
 527                         priority_mark02.full = wm0.priority_mark.full;
 528                 if (dfixed_trunc(priority_mark02) < 0)
 529                         priority_mark02.full = 0;
 530                 if (wm0.priority_mark_max.full > priority_mark02.full)
 531                         priority_mark02.full = wm0.priority_mark_max.full;
 532                 d1mode_priority_a_cnt = dfixed_trunc(priority_mark02);
 533                 if (rdev->disp_priority == 2)
 534                         d1mode_priority_a_cnt |= S_006548_D1MODE_PRIORITY_A_ALWAYS_ON(1);
 535         } else if (mode1) {
 536                 if (dfixed_trunc(wm1.dbpp) > 64)
 537                         a.full = dfixed_mul(wm1.dbpp, wm1.num_line_pair);
 538                 else
 539                         a.full = wm1.num_line_pair.full;
 540                 fill_rate.full = dfixed_div(wm1.sclk, a);
 541                 if (wm1.consumption_rate.full > fill_rate.full) {
 542                         b.full = wm1.consumption_rate.full - fill_rate.full;
 543                         b.full = dfixed_mul(b, wm1.active_time);
 544                         a.full = dfixed_mul(wm1.worst_case_latency,
 545                                                 wm1.consumption_rate);
 546                         a.full = a.full + b.full;
 547                         b.full = dfixed_const(16 * 1000);
 548                         priority_mark12.full = dfixed_div(a, b);
 549                 } else {
 550                         a.full = dfixed_mul(wm1.worst_case_latency,
 551                                                 wm1.consumption_rate);
 552                         b.full = dfixed_const(16 * 1000);
 553                         priority_mark12.full = dfixed_div(a, b);
 554                 }
 555                 if (wm1.priority_mark.full > priority_mark12.full)
 556                         priority_mark12.full = wm1.priority_mark.full;
 557                 if (dfixed_trunc(priority_mark12) < 0)
 558                         priority_mark12.full = 0;
 559                 if (wm1.priority_mark_max.full > priority_mark12.full)
 560                         priority_mark12.full = wm1.priority_mark_max.full;
 561                 d2mode_priority_a_cnt = dfixed_trunc(priority_mark12);
 562                 if (rdev->disp_priority == 2)
 563                         d2mode_priority_a_cnt |= S_006D48_D2MODE_PRIORITY_A_ALWAYS_ON(1);
 564         }
 565
 566         WREG32(R_006548_D1MODE_PRIORITY_A_CNT, d1mode_priority_a_cnt);
 567         WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, d1mode_priority_a_cnt);
 568         WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, d2mode_priority_a_cnt);
 569         WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, d2mode_priority_a_cnt);
 570 }
 571
 572 uint32_t rs690_mc_rreg(struct radeon_device *rdev, uint32_t reg)
 573 {
 574         uint32_t r;
 575
 576         WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg));
 577         r = RREG32(R_00007C_MC_DATA);
 578         WREG32(R_000078_MC_INDEX, ~C_000078_MC_IND_ADDR);
 579         return r;
 580 }
 581
 582 void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
 583 {
 584         WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg) |
 585                 S_000078_MC_IND_WR_EN(1));
 586         WREG32(R_00007C_MC_DATA, v);
 587         WREG32(R_000078_MC_INDEX, 0x7F);
 588 }
 589
 590 static void rs690_mc_program(struct radeon_device *rdev)
 591 {
 592         struct rv515_mc_save save;
 593
 594         /* Stops all mc clients */
 595         rv515_mc_stop(rdev, &save);
 596
 597         /* Wait for mc idle */
 598         if (rs690_mc_wait_for_idle(rdev))
 599                 dev_warn(rdev->dev, "Wait MC idle timeout before updating MC.\n");
 600         /* Program MC, should be a 32bits limited address space */
 601         WREG32_MC(R_000100_MCCFG_FB_LOCATION,
 602                         S_000100_MC_FB_START(rdev->mc.vram_start >> 16) |
 603                         S_000100_MC_FB_TOP(rdev->mc.vram_end >> 16));
 604         WREG32(R_000134_HDP_FB_LOCATION,
 605                 S_000134_HDP_FB_START(rdev->mc.vram_start >> 16));
 606
 607         rv515_mc_resume(rdev, &save);
 608 }
 609
 610 static int rs690_startup(struct radeon_device *rdev)
 611 {
 612         int r;
 613
 614         rs690_mc_program(rdev);
 615         /* Resume clock */
 616         rv515_clock_startup(rdev);
 617         /* Initialize GPU configuration (# pipes, ...) */
 618         rs690_gpu_init(rdev);
 619         /* Initialize GART (initialize after TTM so we can allocate
 620          * memory through TTM but finalize after TTM) */
 621         r = rs400_gart_enable(rdev);
 622         if (r)
 623                 return r;
 624
 625         /* allocate wb buffer */
 626         r = radeon_wb_init(rdev);
 627         if (r)
 628                 return r;
 629
 630         r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
 631         if (r) {
 632                 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
 633                 return r;
 634         }
 635
 636         /* Enable IRQ */
 637         rs600_irq_set(rdev);
 638         rdev->config.r300.hdp_cntl = RREG32(RADEON_HOST_PATH_CNTL);
 639         /* 1M ring buffer */
 640         r = r100_cp_init(rdev, 1024 * 1024);
 641         if (r) {
 642                 dev_err(rdev->dev, "failed initializing CP (%d).\n", r);
 643                 return r;
 644         }
 645
 646         r = radeon_ib_pool_init(rdev);
 647         if (r) {
 648                 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
 649                 return r;
 650         }
 651
 652         r = r600_audio_init(rdev);
 653         if (r) {
 654                 dev_err(rdev->dev, "failed initializing audio\n");
 655                 return r;
 656         }
 657
 658         return 0;
 659 }
 660
 661 int rs690_resume(struct radeon_device *rdev)
 662 {
 663         int r;
 664
 665         /* Make sur GART are not working */
 666         rs400_gart_disable(rdev);
 667         /* Resume clock before doing reset */
 668         rv515_clock_startup(rdev);
 669         /* Reset gpu before posting otherwise ATOM will enter infinite loop */
 670         if (radeon_asic_reset(rdev)) {
 671                 dev_warn(rdev->dev, "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
 672                         RREG32(R_000E40_RBBM_STATUS),
 673                         RREG32(R_0007C0_CP_STAT));
 674         }
 675         /* post */
 676         atom_asic_init(rdev->mode_info.atom_context);
 677         /* Resume clock after posting */
 678         rv515_clock_startup(rdev);
 679         /* Initialize surface registers */
 680         radeon_surface_init(rdev);
 681
 682         rdev->accel_working = true;
 683         r = rs690_startup(rdev);
 684         if (r) {
 685                 rdev->accel_working = false;
 686         }
 687         return r;
 688 }
 689
 690 int rs690_suspend(struct radeon_device *rdev)
 691 {
 692         r600_audio_fini(rdev);
 693         r100_cp_disable(rdev);
 694         radeon_wb_disable(rdev);
 695         rs600_irq_disable(rdev);
 696         rs400_gart_disable(rdev);
 697         return 0;
 698 }
 699
 700 void rs690_fini(struct radeon_device *rdev)
 701 {
 702         r600_audio_fini(rdev);
 703         r100_cp_fini(rdev);
 704         radeon_wb_fini(rdev);
 705         radeon_ib_pool_fini(rdev);
 706         radeon_gem_fini(rdev);
 707         rs400_gart_fini(rdev);
 708         radeon_irq_kms_fini(rdev);
 709         radeon_fence_driver_fini(rdev);
 710         radeon_bo_fini(rdev);
 711         radeon_atombios_fini(rdev);
 712         drm_free(rdev->bios, DRM_MEM_DRIVER);
 713         rdev->bios = NULL;
 714 }
 715
 716 int rs690_init(struct radeon_device *rdev)
 717 {
 718         int r;
 719
 720         /* Disable VGA */
 721         rv515_vga_render_disable(rdev);
 722         /* Initialize scratch registers */
 723         radeon_scratch_init(rdev);
 724         /* Initialize surface registers */
 725         radeon_surface_init(rdev);
 726         /* restore some register to sane defaults */
 727         r100_restore_sanity(rdev);
 728         /* TODO: disable VGA need to use VGA request */
 729         /* BIOS*/
 730         if (!radeon_get_bios(rdev)) {
 731                 if (ASIC_IS_AVIVO(rdev))
 732                         return -EINVAL;
 733         }
 734         if (rdev->is_atom_bios) {
 735                 r = radeon_atombios_init(rdev);
 736                 if (r)
 737                         return r;
 738         } else {
 739                 dev_err(rdev->dev, "Expecting atombios for RV515 GPU\n");
 740                 return -EINVAL;
 741         }
 742         /* Reset gpu before posting otherwise ATOM will enter infinite loop */
 743         if (radeon_asic_reset(rdev)) {
 744                 dev_warn(rdev->dev,
 745                         "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
 746                         RREG32(R_000E40_RBBM_STATUS),
 747                         RREG32(R_0007C0_CP_STAT));
 748         }
 749         /* check if cards are posted or not */
 750         if (radeon_boot_test_post_card(rdev) == false)
 751                 return -EINVAL;
 752
 753         /* Initialize clocks */
 754         radeon_get_clock_info(rdev->ddev);
 755         /* initialize memory controller */
 756         rs690_mc_init(rdev);
 757         rv515_debugfs(rdev);
 758         /* Fence driver */
 759         r = radeon_fence_driver_init(rdev);
 760         if (r)
 761                 return r;
 762         r = radeon_irq_kms_init(rdev);
 763         if (r)
 764                 return r;
 765         /* Memory manager */
 766         r = radeon_bo_init(rdev);
 767         if (r)
 768                 return r;
 769         r = rs400_gart_init(rdev);
 770         if (r)
 771                 return r;
 772         rs600_set_safe_registers(rdev);
 773
 774         rdev->accel_working = true;
 775         r = rs690_startup(rdev);
 776         if (r) {
 777                 /* Somethings want wront with the accel init stop accel */
 778                 dev_err(rdev->dev, "Disabling GPU acceleration\n");
 779                 r100_cp_fini(rdev);
 780                 radeon_wb_fini(rdev);
 781                 radeon_ib_pool_fini(rdev);
 782                 rs400_gart_fini(rdev);
 783                 radeon_irq_kms_fini(rdev);
 784                 rdev->accel_working = false;
 785         }
 786         return 0;
 787 }