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 #include <drm/drmP.h>
  29 #include "radeon.h"
  30 #include "radeon_asic.h"
  31 #include "atom.h"
  32 #include "rs690d.h"
  33
  34 int rs690_mc_wait_for_idle(struct radeon_device *rdev)
  35 {
  36         unsigned i;
  37         uint32_t tmp;
  38
  39         for (i = 0; i < rdev->usec_timeout; i++) {
  40                 /* read MC_STATUS */
  41                 tmp = RREG32_MC(R_000090_MC_SYSTEM_STATUS);
  42                 if (G_000090_MC_SYSTEM_IDLE(tmp))
  43                         return 0;
  44                 udelay(1);
  45         }
  46         return -1;
  47 }
  48
  49 static void rs690_gpu_init(struct radeon_device *rdev)
  50 {
  51         /* FIXME: is this correct ? */
  52         r420_pipes_init(rdev);
  53         if (rs690_mc_wait_for_idle(rdev)) {
  54                 printk(KERN_WARNING "Failed to wait MC idle while "
  55                        "programming pipes. Bad things might happen.\n");
  56         }
  57 }
  58
  59 union igp_info {
  60         struct _ATOM_INTEGRATED_SYSTEM_INFO info;
  61         struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 info_v2;
  62 };
  63
  64 void rs690_pm_info(struct radeon_device *rdev)
  65 {
  66         int index = GetIndexIntoMasterTable(DATA, IntegratedSystemInfo);
  67         union igp_info *info;
  68         uint16_t data_offset;
  69         uint8_t frev, crev;
  70         fixed20_12 tmp;
  71
  72         if (atom_parse_data_header(rdev->mode_info.atom_context, index, NULL,
  73                                    &frev, &crev, &data_offset)) {
  74                 info = (union igp_info *)((uintptr_t)rdev->mode_info.atom_context->bios + data_offset);
  75
  76                 /* Get various system informations from bios */
  77                 switch (crev) {
  78                 case 1:
  79                         tmp.full = dfixed_const(100);
  80                         rdev->pm.igp_sideport_mclk.full = dfixed_const(le32_to_cpu(info->info.ulBootUpMemoryClock));
  81                         rdev->pm.igp_sideport_mclk.full = dfixed_div(rdev->pm.igp_sideport_mclk, tmp);
  82                         if (le16_to_cpu(info->info.usK8MemoryClock))
  83                                 rdev->pm.igp_system_mclk.full = dfixed_const(le16_to_cpu(info->info.usK8MemoryClock));
  84                         else if (rdev->clock.default_mclk) {
  85                                 rdev->pm.igp_system_mclk.full = dfixed_const(rdev->clock.default_mclk);
  86                                 rdev->pm.igp_system_mclk.full = dfixed_div(rdev->pm.igp_system_mclk, tmp);
  87                         } else
  88                                 rdev->pm.igp_system_mclk.full = dfixed_const(400);
  89                         rdev->pm.igp_ht_link_clk.full = dfixed_const(le16_to_cpu(info->info.usFSBClock));
  90                         rdev->pm.igp_ht_link_width.full = dfixed_const(info->info.ucHTLinkWidth);
  91                         break;
  92                 case 2:
  93                         tmp.full = dfixed_const(100);
  94                         rdev->pm.igp_sideport_mclk.full = dfixed_const(le32_to_cpu(info->info_v2.ulBootUpSidePortClock));
  95                         rdev->pm.igp_sideport_mclk.full = dfixed_div(rdev->pm.igp_sideport_mclk, tmp);
  96                         if (le32_to_cpu(info->info_v2.ulBootUpUMAClock))
  97                                 rdev->pm.igp_system_mclk.full = dfixed_const(le32_to_cpu(info->info_v2.ulBootUpUMAClock));
  98                         else if (rdev->clock.default_mclk)
  99                                 rdev->pm.igp_system_mclk.full = dfixed_const(rdev->clock.default_mclk);
 100                         else
 101                                 rdev->pm.igp_system_mclk.full = dfixed_const(66700);
 102                         rdev->pm.igp_system_mclk.full = dfixed_div(rdev->pm.igp_system_mclk, tmp);
 103                         rdev->pm.igp_ht_link_clk.full = dfixed_const(le32_to_cpu(info->info_v2.ulHTLinkFreq));
 104                         rdev->pm.igp_ht_link_clk.full = dfixed_div(rdev->pm.igp_ht_link_clk, tmp);
 105                         rdev->pm.igp_ht_link_width.full = dfixed_const(le16_to_cpu(info->info_v2.usMinHTLinkWidth));
 106                         break;
 107                 default:
 108                         /* We assume the slower possible clock ie worst case */
 109                         rdev->pm.igp_sideport_mclk.full = dfixed_const(200);
 110                         rdev->pm.igp_system_mclk.full = dfixed_const(200);
 111                         rdev->pm.igp_ht_link_clk.full = dfixed_const(1000);
 112                         rdev->pm.igp_ht_link_width.full = dfixed_const(8);
 113                         DRM_ERROR("No integrated system info for your GPU, using safe default\n");
 114                         break;
 115                 }
 116         } else {
 117                 /* We assume the slower possible clock ie worst case */
 118                 rdev->pm.igp_sideport_mclk.full = dfixed_const(200);
 119                 rdev->pm.igp_system_mclk.full = dfixed_const(200);
 120                 rdev->pm.igp_ht_link_clk.full = dfixed_const(1000);
 121                 rdev->pm.igp_ht_link_width.full = dfixed_const(8);
 122                 DRM_ERROR("No integrated system info for your GPU, using safe default\n");
 123         }
 124         /* Compute various bandwidth */
 125         /* k8_bandwidth = (memory_clk / 2) * 2 * 8 * 0.5 = memory_clk * 4  */
 126         tmp.full = dfixed_const(4);
 127         rdev->pm.k8_bandwidth.full = dfixed_mul(rdev->pm.igp_system_mclk, tmp);
 128         /* ht_bandwidth = ht_clk * 2 * ht_width / 8 * 0.8
 129          *              = ht_clk * ht_width / 5
 130          */
 131         tmp.full = dfixed_const(5);
 132         rdev->pm.ht_bandwidth.full = dfixed_mul(rdev->pm.igp_ht_link_clk,
 133                                                 rdev->pm.igp_ht_link_width);
 134         rdev->pm.ht_bandwidth.full = dfixed_div(rdev->pm.ht_bandwidth, tmp);
 135         if (tmp.full < rdev->pm.max_bandwidth.full) {
 136                 /* HT link is a limiting factor */
 137                 rdev->pm.max_bandwidth.full = tmp.full;
 138         }
 139         /* sideport_bandwidth = (sideport_clk / 2) * 2 * 2 * 0.7
 140          *                    = (sideport_clk * 14) / 10
 141          */
 142         tmp.full = dfixed_const(14);
 143         rdev->pm.sideport_bandwidth.full = dfixed_mul(rdev->pm.igp_sideport_mclk, tmp);
 144         tmp.full = dfixed_const(10);
 145         rdev->pm.sideport_bandwidth.full = dfixed_div(rdev->pm.sideport_bandwidth, tmp);
 146 }
 147
 148 static void rs690_mc_init(struct radeon_device *rdev)
 149 {
 150         u64 base;
 151         uint32_t h_addr, l_addr;
 152         unsigned long long k8_addr;
 153
 154         rs400_gart_adjust_size(rdev);
 155         rdev->mc.vram_is_ddr = true;
 156         rdev->mc.vram_width = 128;
 157         rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
 158         rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
 159         rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
 160         rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
 161         rdev->mc.visible_vram_size = rdev->mc.aper_size;
 162         base = RREG32_MC(R_000100_MCCFG_FB_LOCATION);
 163         base = G_000100_MC_FB_START(base) << 16;
 164         rdev->mc.igp_sideport_enabled = radeon_atombios_sideport_present(rdev);
 165         /* Some boards seem to be configured for 128MB of sideport memory,
 166          * but really only have 64MB.  Just skip the sideport and use
 167          * UMA memory.
 168          */
 169         if (rdev->mc.igp_sideport_enabled &&
 170             (rdev->mc.real_vram_size == (384 * 1024 * 1024))) {
 171                 base += 128 * 1024 * 1024;
 172                 rdev->mc.real_vram_size -= 128 * 1024 * 1024;
 173                 rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
 174         }
 175
 176         /* Use K8 direct mapping for fast fb access. */
 177         rdev->fastfb_working = false;
 178         h_addr = G_00005F_K8_ADDR_EXT(RREG32_MC(R_00005F_MC_MISC_UMA_CNTL));
 179         l_addr = RREG32_MC(R_00001E_K8_FB_LOCATION);
 180         k8_addr = ((unsigned long long)h_addr) << 32 | l_addr;
 181 #if defined(CONFIG_X86_32) && !defined(CONFIG_X86_PAE)
 182         if (k8_addr + rdev->mc.visible_vram_size < 0x100000000ULL)
 183 #endif
 184         {
 185                 /* FastFB shall be used with UMA memory. Here it is simply disabled when sideport
 186                  * memory is present.
 187                  */
 188                 if (rdev->mc.igp_sideport_enabled == false && radeon_fastfb == 1) {
 189                         DRM_INFO("Direct mapping: aper base at 0x%llx, replaced by direct mapping base 0x%llx.\n",
 190                                         (unsigned long long)rdev->mc.aper_base, k8_addr);
 191                         rdev->mc.aper_base = (resource_size_t)k8_addr;
 192                         rdev->fastfb_working = true;
 193                 }
 194         }
 195
 196         rs690_pm_info(rdev);
 197         radeon_vram_location(rdev, &rdev->mc, base);
 198         rdev->mc.gtt_base_align = rdev->mc.gtt_size - 1;
 199         radeon_gtt_location(rdev, &rdev->mc);
 200         radeon_update_bandwidth_info(rdev);
 201 }
 202
 203 void rs690_line_buffer_adjust(struct radeon_device *rdev,
 204                               struct drm_display_mode *mode1,
 205                               struct drm_display_mode *mode2)
 206 {
 207         u32 tmp;
 208
 209         /*
 210          * Line Buffer Setup
 211          * There is a single line buffer shared by both display controllers.
 212          * R_006520_DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
 213          * the display controllers.  The paritioning can either be done
 214          * manually or via one of four preset allocations specified in bits 1:0:
 215          *  0 - line buffer is divided in half and shared between crtc
 216          *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
 217          *  2 - D1 gets the whole buffer
 218          *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
 219          * Setting bit 2 of R_006520_DC_LB_MEMORY_SPLIT controls switches to manual
 220          * allocation mode. In manual allocation mode, D1 always starts at 0,
 221          * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
 222          */
 223         tmp = RREG32(R_006520_DC_LB_MEMORY_SPLIT) & C_006520_DC_LB_MEMORY_SPLIT;
 224         tmp &= ~C_006520_DC_LB_MEMORY_SPLIT_MODE;
 225         /* auto */
 226         if (mode1 && mode2) {
 227                 if (mode1->hdisplay > mode2->hdisplay) {
 228                         if (mode1->hdisplay > 2560)
 229                                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
 230                         else
 231                                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
 232                 } else if (mode2->hdisplay > mode1->hdisplay) {
 233                         if (mode2->hdisplay > 2560)
 234                                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
 235                         else
 236                                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
 237                 } else
 238                         tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
 239         } else if (mode1) {
 240                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_ONLY;
 241         } else if (mode2) {
 242                 tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
 243         }
 244         WREG32(R_006520_DC_LB_MEMORY_SPLIT, tmp);
 245 }
 246
 247 struct rs690_watermark {
 248         u32        lb_request_fifo_depth;
 249         fixed20_12 num_line_pair;
 250         fixed20_12 estimated_width;
 251         fixed20_12 worst_case_latency;
 252         fixed20_12 consumption_rate;
 253         fixed20_12 active_time;
 254         fixed20_12 dbpp;
 255         fixed20_12 priority_mark_max;
 256         fixed20_12 priority_mark;
 257         fixed20_12 sclk;
 258 };
 259
 260 static void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
 261                                          struct radeon_crtc *crtc,
 262                                          struct rs690_watermark *wm,
 263                                          bool low)
 264 {
 265         struct drm_display_mode *mode = &crtc->base.mode;
 266         fixed20_12 a, b, c;
 267         fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
 268         fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
 269         fixed20_12 sclk, core_bandwidth, max_bandwidth;
 270         u32 selected_sclk;
 271
 272         bzero(wm, sizeof(*wm)); /* avoid gcc warning */
 273         if (!crtc->base.enabled) {
 274                 /* FIXME: wouldn't it better to set priority mark to maximum */
 275                 wm->lb_request_fifo_depth = 4;
 276                 return;
 277         }
 278
 279         if (((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880)) &&
 280             (rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled)
 281                 selected_sclk = radeon_dpm_get_sclk(rdev, low);
 282         else
 283                 selected_sclk = rdev->pm.current_sclk;
 284
 285         /* sclk in Mhz */
 286         a.full = dfixed_const(100);
 287         sclk.full = dfixed_const(selected_sclk);
 288         sclk.full = dfixed_div(sclk, a);
 289
 290         /* core_bandwidth = sclk(Mhz) * 16 */
 291         a.full = dfixed_const(16);
 292         core_bandwidth.full = dfixed_div(rdev->pm.sclk, a);
 293
 294         if (crtc->vsc.full > dfixed_const(2))
 295                 wm->num_line_pair.full = dfixed_const(2);
 296         else
 297                 wm->num_line_pair.full = dfixed_const(1);
 298
 299         b.full = dfixed_const(mode->crtc_hdisplay);
 300         c.full = dfixed_const(256);
 301         a.full = dfixed_div(b, c);
 302         request_fifo_depth.full = dfixed_mul(a, wm->num_line_pair);
 303         request_fifo_depth.full = dfixed_ceil(request_fifo_depth);
 304         if (a.full < dfixed_const(4)) {
 305                 wm->lb_request_fifo_depth = 4;
 306         } else {
 307                 wm->lb_request_fifo_depth = dfixed_trunc(request_fifo_depth);
 308         }
 309
 310         /* Determine consumption rate
 311          *  pclk = pixel clock period(ns) = 1000 / (mode.clock / 1000)
 312          *  vtaps = number of vertical taps,
 313          *  vsc = vertical scaling ratio, defined as source/destination
 314          *  hsc = horizontal scaling ration, defined as source/destination
 315          */
 316         a.full = dfixed_const(mode->clock);
 317         b.full = dfixed_const(1000);
 318         a.full = dfixed_div(a, b);
 319         pclk.full = dfixed_div(b, a);
 320         if (crtc->rmx_type != RMX_OFF) {
 321                 b.full = dfixed_const(2);
 322                 if (crtc->vsc.full > b.full)
 323                         b.full = crtc->vsc.full;
 324                 b.full = dfixed_mul(b, crtc->hsc);
 325                 c.full = dfixed_const(2);
 326                 b.full = dfixed_div(b, c);
 327                 consumption_time.full = dfixed_div(pclk, b);
 328         } else {
 329                 consumption_time.full = pclk.full;
 330         }
 331         a.full = dfixed_const(1);
 332         wm->consumption_rate.full = dfixed_div(a, consumption_time);
 333
 334
 335         /* Determine line time
 336          *  LineTime = total time for one line of displayhtotal
 337          *  LineTime = total number of horizontal pixels
 338          *  pclk = pixel clock period(ns)
 339          */
 340         a.full = dfixed_const(crtc->base.mode.crtc_htotal);
 341         line_time.full = dfixed_mul(a, pclk);
 342
 343         /* Determine active time
 344          *  ActiveTime = time of active region of display within one line,
 345          *  hactive = total number of horizontal active pixels
 346          *  htotal = total number of horizontal pixels
 347          */
 348         a.full = dfixed_const(crtc->base.mode.crtc_htotal);
 349         b.full = dfixed_const(crtc->base.mode.crtc_hdisplay);
 350         wm->active_time.full = dfixed_mul(line_time, b);
 351         wm->active_time.full = dfixed_div(wm->active_time, a);
 352
 353         /* Maximun bandwidth is the minimun bandwidth of all component */
 354         max_bandwidth = core_bandwidth;
 355         if (rdev->mc.igp_sideport_enabled) {
 356                 if (max_bandwidth.full > rdev->pm.sideport_bandwidth.full &&
 357                         rdev->pm.sideport_bandwidth.full)
 358                         max_bandwidth = rdev->pm.sideport_bandwidth;
 359                 read_delay_latency.full = dfixed_const(370 * 800);
 360                 a.full = dfixed_const(1000);
 361                 b.full = dfixed_div(rdev->pm.igp_sideport_mclk, a);
 362                 read_delay_latency.full = dfixed_div(read_delay_latency, b);
 363                 read_delay_latency.full = dfixed_mul(read_delay_latency, a);
 364         } else {
 365                 if (max_bandwidth.full > rdev->pm.k8_bandwidth.full &&
 366                         rdev->pm.k8_bandwidth.full)
 367                         max_bandwidth = rdev->pm.k8_bandwidth;
 368                 if (max_bandwidth.full > rdev->pm.ht_bandwidth.full &&
 369                         rdev->pm.ht_bandwidth.full)
 370                         max_bandwidth = rdev->pm.ht_bandwidth;
 371                 read_delay_latency.full = dfixed_const(5000);
 372         }
 373
 374         /* sclk = system clocks(ns) = 1000 / max_bandwidth / 16 */
 375         a.full = dfixed_const(16);
 376         sclk.full = dfixed_mul(max_bandwidth, a);
 377         a.full = dfixed_const(1000);
 378         sclk.full = dfixed_div(a, sclk);
 379         /* Determine chunk time
 380          * ChunkTime = the time it takes the DCP to send one chunk of data
 381          * to the LB which consists of pipeline delay and inter chunk gap
 382          * sclk = system clock(ns)
 383          */
 384         a.full = dfixed_const(256 * 13);
 385         chunk_time.full = dfixed_mul(sclk, a);
 386         a.full = dfixed_const(10);
 387         chunk_time.full = dfixed_div(chunk_time, a);
 388
 389         /* Determine the worst case latency
 390          * NumLinePair = Number of line pairs to request(1=2 lines, 2=4 lines)
 391          * WorstCaseLatency = worst case time from urgent to when the MC starts
 392          *                    to return data
 393          * READ_DELAY_IDLE_MAX = constant of 1us
 394          * ChunkTime = time it takes the DCP to send one chunk of data to the LB
 395          *             which consists of pipeline delay and inter chunk gap
 396          */
 397         if (dfixed_trunc(wm->num_line_pair) > 1) {
 398                 a.full = dfixed_const(3);
 399                 wm->worst_case_latency.full = dfixed_mul(a, chunk_time);
 400                 wm->worst_case_latency.full += read_delay_latency.full;
 401         } else {
 402                 a.full = dfixed_const(2);
 403                 wm->worst_case_latency.full = dfixed_mul(a, chunk_time);
 404                 wm->worst_case_latency.full += read_delay_latency.full;
 405         }
 406
 407         /* Determine the tolerable latency
 408          * TolerableLatency = Any given request has only 1 line time
 409          *                    for the data to be returned
 410          * LBRequestFifoDepth = Number of chunk requests the LB can
 411          *                      put into the request FIFO for a display
 412          *  LineTime = total time for one line of display
 413          *  ChunkTime = the time it takes the DCP to send one chunk
 414          *              of data to the LB which consists of
 415          *  pipeline delay and inter chunk gap
 416          */
 417         if ((2+wm->lb_request_fifo_depth) >= dfixed_trunc(request_fifo_depth)) {
 418                 tolerable_latency.full = line_time.full;
 419         } else {
 420                 tolerable_latency.full = dfixed_const(wm->lb_request_fifo_depth - 2);
 421                 tolerable_latency.full = request_fifo_depth.full - tolerable_latency.full;
 422                 tolerable_latency.full = dfixed_mul(tolerable_latency, chunk_time);
 423                 tolerable_latency.full = line_time.full - tolerable_latency.full;
 424         }
 425         /* We assume worst case 32bits (4 bytes) */
 426         wm->dbpp.full = dfixed_const(4 * 8);
 427
 428         /* Determine the maximum priority mark
 429          *  width = viewport width in pixels
 430          */
 431         a.full = dfixed_const(16);
 432         wm->priority_mark_max.full = dfixed_const(crtc->base.mode.crtc_hdisplay);
 433         wm->priority_mark_max.full = dfixed_div(wm->priority_mark_max, a);
 434         wm->priority_mark_max.full = dfixed_ceil(wm->priority_mark_max);
 435
 436         /* Determine estimated width */
 437         estimated_width.full = tolerable_latency.full - wm->worst_case_latency.full;
 438         estimated_width.full = dfixed_div(estimated_width, consumption_time);
 439         if (dfixed_trunc(estimated_width) > crtc->base.mode.crtc_hdisplay) {
 440                 wm->priority_mark.full = dfixed_const(10);
 441         } else {
 442                 a.full = dfixed_const(16);
 443                 wm->priority_mark.full = dfixed_div(estimated_width, a);
 444                 wm->priority_mark.full = dfixed_ceil(wm->priority_mark);
 445                 wm->priority_mark.full = wm->priority_mark_max.full - wm->priority_mark.full;
 446         }
 447 }
 448
 449 static void rs690_compute_mode_priority(struct radeon_device *rdev,
 450                                         struct rs690_watermark *wm0,
 451                                         struct rs690_watermark *wm1,
 452                                         struct drm_display_mode *mode0,
 453                                         struct drm_display_mode *mode1,
 454                                         u32 *d1mode_priority_a_cnt,
 455                                         u32 *d2mode_priority_a_cnt)
 456 {
 457         fixed20_12 priority_mark02, priority_mark12, fill_rate;
 458         fixed20_12 a, b;
 459
 460         *d1mode_priority_a_cnt = S_006548_D1MODE_PRIORITY_A_OFF(1);
 461         *d2mode_priority_a_cnt = S_006548_D1MODE_PRIORITY_A_OFF(1);
 462
 463         if (mode0 && mode1) {
 464                 if (dfixed_trunc(wm0->dbpp) > 64)
 465                         a.full = dfixed_mul(wm0->dbpp, wm0->num_line_pair);
 466                 else
 467                         a.full = wm0->num_line_pair.full;
 468                 if (dfixed_trunc(wm1->dbpp) > 64)
 469                         b.full = dfixed_mul(wm1->dbpp, wm1->num_line_pair);
 470                 else
 471                         b.full = wm1->num_line_pair.full;
 472                 a.full += b.full;
 473                 fill_rate.full = dfixed_div(wm0->sclk, a);
 474                 if (wm0->consumption_rate.full > fill_rate.full) {
 475                         b.full = wm0->consumption_rate.full - fill_rate.full;
 476                         b.full = dfixed_mul(b, wm0->active_time);
 477                         a.full = dfixed_mul(wm0->worst_case_latency,
 478                                                 wm0->consumption_rate);
 479                         a.full = a.full + b.full;
 480                         b.full = dfixed_const(16 * 1000);
 481                         priority_mark02.full = dfixed_div(a, b);
 482                 } else {
 483                         a.full = dfixed_mul(wm0->worst_case_latency,
 484                                                 wm0->consumption_rate);
 485                         b.full = dfixed_const(16 * 1000);
 486                         priority_mark02.full = dfixed_div(a, b);
 487                 }
 488                 if (wm1->consumption_rate.full > fill_rate.full) {
 489                         b.full = wm1->consumption_rate.full - fill_rate.full;
 490                         b.full = dfixed_mul(b, wm1->active_time);
 491                         a.full = dfixed_mul(wm1->worst_case_latency,
 492                                                 wm1->consumption_rate);
 493                         a.full = a.full + b.full;
 494                         b.full = dfixed_const(16 * 1000);
 495                         priority_mark12.full = dfixed_div(a, b);
 496                 } else {
 497                         a.full = dfixed_mul(wm1->worst_case_latency,
 498                                                 wm1->consumption_rate);
 499                         b.full = dfixed_const(16 * 1000);
 500                         priority_mark12.full = dfixed_div(a, b);
 501                 }
 502                 if (wm0->priority_mark.full > priority_mark02.full)
 503                         priority_mark02.full = wm0->priority_mark.full;
 504                 if (wm0->priority_mark_max.full > priority_mark02.full)
 505                         priority_mark02.full = wm0->priority_mark_max.full;
 506                 if (wm1->priority_mark.full > priority_mark12.full)
 507                         priority_mark12.full = wm1->priority_mark.full;
 508                 if (wm1->priority_mark_max.full > priority_mark12.full)
 509                         priority_mark12.full = wm1->priority_mark_max.full;
 510                 *d1mode_priority_a_cnt = dfixed_trunc(priority_mark02);
 511                 *d2mode_priority_a_cnt = dfixed_trunc(priority_mark12);
 512                 if (rdev->disp_priority == 2) {
 513                         *d1mode_priority_a_cnt |= S_006548_D1MODE_PRIORITY_A_ALWAYS_ON(1);
 514                         *d2mode_priority_a_cnt |= S_006D48_D2MODE_PRIORITY_A_ALWAYS_ON(1);
 515                 }
 516         } else if (mode0) {
 517                 if (dfixed_trunc(wm0->dbpp) > 64)
 518                         a.full = dfixed_mul(wm0->dbpp, wm0->num_line_pair);
 519                 else
 520                         a.full = wm0->num_line_pair.full;
 521                 fill_rate.full = dfixed_div(wm0->sclk, a);
 522                 if (wm0->consumption_rate.full > fill_rate.full) {
 523                         b.full = wm0->consumption_rate.full - fill_rate.full;
 524                         b.full = dfixed_mul(b, wm0->active_time);
 525                         a.full = dfixed_mul(wm0->worst_case_latency,
 526                                                 wm0->consumption_rate);
 527                         a.full = a.full + b.full;
 528                         b.full = dfixed_const(16 * 1000);
 529                         priority_mark02.full = dfixed_div(a, b);
 530                 } else {
 531                         a.full = dfixed_mul(wm0->worst_case_latency,
 532                                                 wm0->consumption_rate);
 533                         b.full = dfixed_const(16 * 1000);
 534                         priority_mark02.full = dfixed_div(a, b);
 535                 }
 536                 if (wm0->priority_mark.full > priority_mark02.full)
 537                         priority_mark02.full = wm0->priority_mark.full;
 538                 if (wm0->priority_mark_max.full > priority_mark02.full)
 539                         priority_mark02.full = wm0->priority_mark_max.full;
 540                 *d1mode_priority_a_cnt = dfixed_trunc(priority_mark02);
 541                 if (rdev->disp_priority == 2)
 542                         *d1mode_priority_a_cnt |= S_006548_D1MODE_PRIORITY_A_ALWAYS_ON(1);
 543         } else if (mode1) {
 544                 if (dfixed_trunc(wm1->dbpp) > 64)
 545                         a.full = dfixed_mul(wm1->dbpp, wm1->num_line_pair);
 546                 else
 547                         a.full = wm1->num_line_pair.full;
 548                 fill_rate.full = dfixed_div(wm1->sclk, a);
 549                 if (wm1->consumption_rate.full > fill_rate.full) {
 550                         b.full = wm1->consumption_rate.full - fill_rate.full;
 551                         b.full = dfixed_mul(b, wm1->active_time);
 552                         a.full = dfixed_mul(wm1->worst_case_latency,
 553                                                 wm1->consumption_rate);
 554                         a.full = a.full + b.full;
 555                         b.full = dfixed_const(16 * 1000);
 556                         priority_mark12.full = dfixed_div(a, b);
 557                 } else {
 558                         a.full = dfixed_mul(wm1->worst_case_latency,
 559                                                 wm1->consumption_rate);
 560                         b.full = dfixed_const(16 * 1000);
 561                         priority_mark12.full = dfixed_div(a, b);
 562                 }
 563                 if (wm1->priority_mark.full > priority_mark12.full)
 564                         priority_mark12.full = wm1->priority_mark.full;
 565                 if (wm1->priority_mark_max.full > priority_mark12.full)
 566                         priority_mark12.full = wm1->priority_mark_max.full;
 567                 *d2mode_priority_a_cnt = dfixed_trunc(priority_mark12);
 568                 if (rdev->disp_priority == 2)
 569                         *d2mode_priority_a_cnt |= S_006D48_D2MODE_PRIORITY_A_ALWAYS_ON(1);
 570         }
 571 }
 572
 573 void rs690_bandwidth_update(struct radeon_device *rdev)
 574 {
 575         struct drm_display_mode *mode0 = NULL;
 576         struct drm_display_mode *mode1 = NULL;
 577         struct rs690_watermark wm0_high, wm0_low;
 578         struct rs690_watermark wm1_high, wm1_low;
 579         u32 tmp;
 580         u32 d1mode_priority_a_cnt, d1mode_priority_b_cnt;
 581         u32 d2mode_priority_a_cnt, d2mode_priority_b_cnt;
 582
 583         if (!rdev->mode_info.mode_config_initialized)
 584                 return;
 585
 586         radeon_update_display_priority(rdev);
 587
 588         if (rdev->mode_info.crtcs[0]->base.enabled)
 589                 mode0 = &rdev->mode_info.crtcs[0]->base.mode;
 590         if (rdev->mode_info.crtcs[1]->base.enabled)
 591                 mode1 = &rdev->mode_info.crtcs[1]->base.mode;
 592         /*
 593          * Set display0/1 priority up in the memory controller for
 594          * modes if the user specifies HIGH for displaypriority
 595          * option.
 596          */
 597         if ((rdev->disp_priority == 2) &&
 598             ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))) {
 599                 tmp = RREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER);
 600                 tmp &= C_000104_MC_DISP0R_INIT_LAT;
 601                 tmp &= C_000104_MC_DISP1R_INIT_LAT;
 602                 if (mode0)
 603                         tmp |= S_000104_MC_DISP0R_INIT_LAT(1);
 604                 if (mode1)
 605                         tmp |= S_000104_MC_DISP1R_INIT_LAT(1);
 606                 WREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER, tmp);
 607         }
 608         rs690_line_buffer_adjust(rdev, mode0, mode1);
 609
 610         if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
 611                 WREG32(R_006C9C_DCP_CONTROL, 0);
 612         if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
 613                 WREG32(R_006C9C_DCP_CONTROL, 2);
 614
 615         rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0_high, false);
 616         rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1_high, false);
 617
 618         rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0_low, true);
 619         rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1_low, true);
 620
 621         tmp = (wm0_high.lb_request_fifo_depth - 1);
 622         tmp |= (wm1_high.lb_request_fifo_depth - 1) << 16;
 623         WREG32(R_006D58_LB_MAX_REQ_OUTSTANDING, tmp);
 624
 625         rs690_compute_mode_priority(rdev,
 626                                     &wm0_high, &wm1_high,
 627                                     mode0, mode1,
 628                                     &d1mode_priority_a_cnt, &d2mode_priority_a_cnt);
 629         rs690_compute_mode_priority(rdev,
 630                                     &wm0_low, &wm1_low,
 631                                     mode0, mode1,
 632                                     &d1mode_priority_b_cnt, &d2mode_priority_b_cnt);
 633
 634         WREG32(R_006548_D1MODE_PRIORITY_A_CNT, d1mode_priority_a_cnt);
 635         WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, d1mode_priority_b_cnt);
 636         WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, d2mode_priority_a_cnt);
 637         WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, d2mode_priority_b_cnt);
 638 }
 639
 640 uint32_t rs690_mc_rreg(struct radeon_device *rdev, uint32_t reg)
 641 {
 642         uint32_t r;
 643
 644         spin_lock(&rdev->mc_idx_lock);
 645         WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg));
 646         r = RREG32(R_00007C_MC_DATA);
 647         WREG32(R_000078_MC_INDEX, ~C_000078_MC_IND_ADDR);
 648         spin_unlock(&rdev->mc_idx_lock);
 649         return r;
 650 }
 651
 652 void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
 653 {
 654         spin_lock(&rdev->mc_idx_lock);
 655         WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg) |
 656                 S_000078_MC_IND_WR_EN(1));
 657         WREG32(R_00007C_MC_DATA, v);
 658         WREG32(R_000078_MC_INDEX, 0x7F);
 659         spin_unlock(&rdev->mc_idx_lock);
 660 }
 661
 662 static void rs690_mc_program(struct radeon_device *rdev)
 663 {
 664         struct rv515_mc_save save;
 665
 666         /* Stops all mc clients */
 667         rv515_mc_stop(rdev, &save);
 668
 669         /* Wait for mc idle */
 670         if (rs690_mc_wait_for_idle(rdev))
 671                 dev_warn(rdev->dev, "Wait MC idle timeout before updating MC.\n");
 672         /* Program MC, should be a 32bits limited address space */
 673         WREG32_MC(R_000100_MCCFG_FB_LOCATION,
 674                         S_000100_MC_FB_START(rdev->mc.vram_start >> 16) |
 675                         S_000100_MC_FB_TOP(rdev->mc.vram_end >> 16));
 676         WREG32(R_000134_HDP_FB_LOCATION,
 677                 S_000134_HDP_FB_START(rdev->mc.vram_start >> 16));
 678
 679         rv515_mc_resume(rdev, &save);
 680 }
 681
 682 static int rs690_startup(struct radeon_device *rdev)
 683 {
 684         int r;
 685
 686         rs690_mc_program(rdev);
 687         /* Resume clock */
 688         rv515_clock_startup(rdev);
 689         /* Initialize GPU configuration (# pipes, ...) */
 690         rs690_gpu_init(rdev);
 691         /* Initialize GART (initialize after TTM so we can allocate
 692          * memory through TTM but finalize after TTM) */
 693         r = rs400_gart_enable(rdev);
 694         if (r)
 695                 return r;
 696
 697         /* allocate wb buffer */
 698         r = radeon_wb_init(rdev);
 699         if (r)
 700                 return r;
 701
 702         r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
 703         if (r) {
 704                 dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
 705                 return r;
 706         }
 707
 708         /* Enable IRQ */
 709         if (!rdev->irq.installed) {
 710                 r = radeon_irq_kms_init(rdev);
 711                 if (r)
 712                         return r;
 713         }
 714
 715         rs600_irq_set(rdev);
 716         rdev->config.r300.hdp_cntl = RREG32(RADEON_HOST_PATH_CNTL);
 717         /* 1M ring buffer */
 718         r = r100_cp_init(rdev, 1024 * 1024);
 719         if (r) {
 720                 dev_err(rdev->dev, "failed initializing CP (%d).\n", r);
 721                 return r;
 722         }
 723
 724         r = radeon_ib_pool_init(rdev);
 725         if (r) {
 726                 dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
 727                 return r;
 728         }
 729
 730         r = r600_audio_init(rdev);
 731         if (r) {
 732                 dev_err(rdev->dev, "failed initializing audio\n");
 733                 return r;
 734         }
 735
 736         return 0;
 737 }
 738
 739 int rs690_resume(struct radeon_device *rdev)
 740 {
 741         int r;
 742
 743         /* Make sur GART are not working */
 744         rs400_gart_disable(rdev);
 745         /* Resume clock before doing reset */
 746         rv515_clock_startup(rdev);
 747         /* Reset gpu before posting otherwise ATOM will enter infinite loop */
 748         if (radeon_asic_reset(rdev)) {
 749                 dev_warn(rdev->dev, "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
 750                         RREG32(R_000E40_RBBM_STATUS),
 751                         RREG32(R_0007C0_CP_STAT));
 752         }
 753         /* post */
 754         atom_asic_init(rdev->mode_info.atom_context);
 755         /* Resume clock after posting */
 756         rv515_clock_startup(rdev);
 757         /* Initialize surface registers */
 758         radeon_surface_init(rdev);
 759
 760         rdev->accel_working = true;
 761         r = rs690_startup(rdev);
 762         if (r) {
 763                 rdev->accel_working = false;
 764         }
 765         return r;
 766 }
 767
 768 int rs690_suspend(struct radeon_device *rdev)
 769 {
 770         radeon_pm_suspend(rdev);
 771         r600_audio_fini(rdev);
 772         r100_cp_disable(rdev);
 773         radeon_wb_disable(rdev);
 774         rs600_irq_disable(rdev);
 775         rs400_gart_disable(rdev);
 776         return 0;
 777 }
 778
 779 void rs690_fini(struct radeon_device *rdev)
 780 {
 781         radeon_pm_fini(rdev);
 782         r600_audio_fini(rdev);
 783         r100_cp_fini(rdev);
 784         radeon_wb_fini(rdev);
 785         radeon_ib_pool_fini(rdev);
 786         radeon_gem_fini(rdev);
 787         rs400_gart_fini(rdev);
 788         radeon_irq_kms_fini(rdev);
 789         radeon_fence_driver_fini(rdev);
 790         radeon_bo_fini(rdev);
 791         radeon_atombios_fini(rdev);
 792         kfree(rdev->bios);
 793         rdev->bios = NULL;
 794 }
 795
 796 int rs690_init(struct radeon_device *rdev)
 797 {
 798         int r;
 799
 800         /* Disable VGA */
 801         rv515_vga_render_disable(rdev);
 802         /* Initialize scratch registers */
 803         radeon_scratch_init(rdev);
 804         /* Initialize surface registers */
 805         radeon_surface_init(rdev);
 806         /* restore some register to sane defaults */
 807         r100_restore_sanity(rdev);
 808         /* TODO: disable VGA need to use VGA request */
 809         /* BIOS*/
 810         if (!radeon_get_bios(rdev)) {
 811                 if (ASIC_IS_AVIVO(rdev))
 812                         return -EINVAL;
 813         }
 814         if (rdev->is_atom_bios) {
 815                 r = radeon_atombios_init(rdev);
 816                 if (r)
 817                         return r;
 818         } else {
 819                 dev_err(rdev->dev, "Expecting atombios for RV515 GPU\n");
 820                 return -EINVAL;
 821         }
 822         /* Reset gpu before posting otherwise ATOM will enter infinite loop */
 823         if (radeon_asic_reset(rdev)) {
 824                 dev_warn(rdev->dev,
 825                         "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
 826                         RREG32(R_000E40_RBBM_STATUS),
 827                         RREG32(R_0007C0_CP_STAT));
 828         }
 829         /* check if cards are posted or not */
 830         if (radeon_boot_test_post_card(rdev) == false)
 831                 return -EINVAL;
 832
 833         /* Initialize clocks */
 834         radeon_get_clock_info(rdev->ddev);
 835         /* initialize memory controller */
 836         rs690_mc_init(rdev);
 837         rv515_debugfs(rdev);
 838         /* Fence driver */
 839         r = radeon_fence_driver_init(rdev);
 840         if (r)
 841                 return r;
 842         /* Memory manager */
 843         r = radeon_bo_init(rdev);
 844         if (r)
 845                 return r;
 846         r = rs400_gart_init(rdev);
 847         if (r)
 848                 return r;
 849         rs600_set_safe_registers(rdev);
 850
 851         /* Initialize power management */
 852         radeon_pm_init(rdev);
 853
 854         rdev->accel_working = true;
 855         r = rs690_startup(rdev);
 856         if (r) {
 857                 /* Somethings want wront with the accel init stop accel */
 858                 dev_err(rdev->dev, "Disabling GPU acceleration\n");
 859                 r100_cp_fini(rdev);
 860                 radeon_wb_fini(rdev);
 861                 radeon_ib_pool_fini(rdev);
 862                 rs400_gart_fini(rdev);
 863                 radeon_irq_kms_fini(rdev);
 864                 rdev->accel_working = false;
 865         }
 866         return 0;
 867 }