xref: /openbmc/linux/drivers/gpu/drm/radeon/rs690.c (revision 95e9fd10)
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 "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 *)(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 void rs690_mc_init(struct radeon_device *rdev)
149 {
150 	u64 base;
151 
152 	rs400_gart_adjust_size(rdev);
153 	rdev->mc.vram_is_ddr = true;
154 	rdev->mc.vram_width = 128;
155 	rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
156 	rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
157 	rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
158 	rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
159 	rdev->mc.visible_vram_size = rdev->mc.aper_size;
160 	base = RREG32_MC(R_000100_MCCFG_FB_LOCATION);
161 	base = G_000100_MC_FB_START(base) << 16;
162 	rdev->mc.igp_sideport_enabled = radeon_atombios_sideport_present(rdev);
163 	rs690_pm_info(rdev);
164 	radeon_vram_location(rdev, &rdev->mc, base);
165 	rdev->mc.gtt_base_align = rdev->mc.gtt_size - 1;
166 	radeon_gtt_location(rdev, &rdev->mc);
167 	radeon_update_bandwidth_info(rdev);
168 }
169 
170 void rs690_line_buffer_adjust(struct radeon_device *rdev,
171 			      struct drm_display_mode *mode1,
172 			      struct drm_display_mode *mode2)
173 {
174 	u32 tmp;
175 
176 	/*
177 	 * Line Buffer Setup
178 	 * There is a single line buffer shared by both display controllers.
179 	 * R_006520_DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
180 	 * the display controllers.  The paritioning can either be done
181 	 * manually or via one of four preset allocations specified in bits 1:0:
182 	 *  0 - line buffer is divided in half and shared between crtc
183 	 *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
184 	 *  2 - D1 gets the whole buffer
185 	 *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
186 	 * Setting bit 2 of R_006520_DC_LB_MEMORY_SPLIT controls switches to manual
187 	 * allocation mode. In manual allocation mode, D1 always starts at 0,
188 	 * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
189 	 */
190 	tmp = RREG32(R_006520_DC_LB_MEMORY_SPLIT) & C_006520_DC_LB_MEMORY_SPLIT;
191 	tmp &= ~C_006520_DC_LB_MEMORY_SPLIT_MODE;
192 	/* auto */
193 	if (mode1 && mode2) {
194 		if (mode1->hdisplay > mode2->hdisplay) {
195 			if (mode1->hdisplay > 2560)
196 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
197 			else
198 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
199 		} else if (mode2->hdisplay > mode1->hdisplay) {
200 			if (mode2->hdisplay > 2560)
201 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
202 			else
203 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
204 		} else
205 			tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
206 	} else if (mode1) {
207 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_ONLY;
208 	} else if (mode2) {
209 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
210 	}
211 	WREG32(R_006520_DC_LB_MEMORY_SPLIT, tmp);
212 }
213 
214 struct rs690_watermark {
215 	u32        lb_request_fifo_depth;
216 	fixed20_12 num_line_pair;
217 	fixed20_12 estimated_width;
218 	fixed20_12 worst_case_latency;
219 	fixed20_12 consumption_rate;
220 	fixed20_12 active_time;
221 	fixed20_12 dbpp;
222 	fixed20_12 priority_mark_max;
223 	fixed20_12 priority_mark;
224 	fixed20_12 sclk;
225 };
226 
227 void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
228 				  struct radeon_crtc *crtc,
229 				  struct rs690_watermark *wm)
230 {
231 	struct drm_display_mode *mode = &crtc->base.mode;
232 	fixed20_12 a, b, c;
233 	fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
234 	fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
235 
236 	if (!crtc->base.enabled) {
237 		/* FIXME: wouldn't it better to set priority mark to maximum */
238 		wm->lb_request_fifo_depth = 4;
239 		return;
240 	}
241 
242 	if (crtc->vsc.full > dfixed_const(2))
243 		wm->num_line_pair.full = dfixed_const(2);
244 	else
245 		wm->num_line_pair.full = dfixed_const(1);
246 
247 	b.full = dfixed_const(mode->crtc_hdisplay);
248 	c.full = dfixed_const(256);
249 	a.full = dfixed_div(b, c);
250 	request_fifo_depth.full = dfixed_mul(a, wm->num_line_pair);
251 	request_fifo_depth.full = dfixed_ceil(request_fifo_depth);
252 	if (a.full < dfixed_const(4)) {
253 		wm->lb_request_fifo_depth = 4;
254 	} else {
255 		wm->lb_request_fifo_depth = dfixed_trunc(request_fifo_depth);
256 	}
257 
258 	/* Determine consumption rate
259 	 *  pclk = pixel clock period(ns) = 1000 / (mode.clock / 1000)
260 	 *  vtaps = number of vertical taps,
261 	 *  vsc = vertical scaling ratio, defined as source/destination
262 	 *  hsc = horizontal scaling ration, defined as source/destination
263 	 */
264 	a.full = dfixed_const(mode->clock);
265 	b.full = dfixed_const(1000);
266 	a.full = dfixed_div(a, b);
267 	pclk.full = dfixed_div(b, a);
268 	if (crtc->rmx_type != RMX_OFF) {
269 		b.full = dfixed_const(2);
270 		if (crtc->vsc.full > b.full)
271 			b.full = crtc->vsc.full;
272 		b.full = dfixed_mul(b, crtc->hsc);
273 		c.full = dfixed_const(2);
274 		b.full = dfixed_div(b, c);
275 		consumption_time.full = dfixed_div(pclk, b);
276 	} else {
277 		consumption_time.full = pclk.full;
278 	}
279 	a.full = dfixed_const(1);
280 	wm->consumption_rate.full = dfixed_div(a, consumption_time);
281 
282 
283 	/* Determine line time
284 	 *  LineTime = total time for one line of displayhtotal
285 	 *  LineTime = total number of horizontal pixels
286 	 *  pclk = pixel clock period(ns)
287 	 */
288 	a.full = dfixed_const(crtc->base.mode.crtc_htotal);
289 	line_time.full = dfixed_mul(a, pclk);
290 
291 	/* Determine active time
292 	 *  ActiveTime = time of active region of display within one line,
293 	 *  hactive = total number of horizontal active pixels
294 	 *  htotal = total number of horizontal pixels
295 	 */
296 	a.full = dfixed_const(crtc->base.mode.crtc_htotal);
297 	b.full = dfixed_const(crtc->base.mode.crtc_hdisplay);
298 	wm->active_time.full = dfixed_mul(line_time, b);
299 	wm->active_time.full = dfixed_div(wm->active_time, a);
300 
301 	/* Maximun bandwidth is the minimun bandwidth of all component */
302 	rdev->pm.max_bandwidth = rdev->pm.core_bandwidth;
303 	if (rdev->mc.igp_sideport_enabled) {
304 		if (rdev->pm.max_bandwidth.full > rdev->pm.sideport_bandwidth.full &&
305 			rdev->pm.sideport_bandwidth.full)
306 			rdev->pm.max_bandwidth = rdev->pm.sideport_bandwidth;
307 		read_delay_latency.full = dfixed_const(370 * 800 * 1000);
308 		read_delay_latency.full = dfixed_div(read_delay_latency,
309 			rdev->pm.igp_sideport_mclk);
310 	} else {
311 		if (rdev->pm.max_bandwidth.full > rdev->pm.k8_bandwidth.full &&
312 			rdev->pm.k8_bandwidth.full)
313 			rdev->pm.max_bandwidth = rdev->pm.k8_bandwidth;
314 		if (rdev->pm.max_bandwidth.full > rdev->pm.ht_bandwidth.full &&
315 			rdev->pm.ht_bandwidth.full)
316 			rdev->pm.max_bandwidth = rdev->pm.ht_bandwidth;
317 		read_delay_latency.full = dfixed_const(5000);
318 	}
319 
320 	/* sclk = system clocks(ns) = 1000 / max_bandwidth / 16 */
321 	a.full = dfixed_const(16);
322 	rdev->pm.sclk.full = dfixed_mul(rdev->pm.max_bandwidth, a);
323 	a.full = dfixed_const(1000);
324 	rdev->pm.sclk.full = dfixed_div(a, rdev->pm.sclk);
325 	/* Determine chunk time
326 	 * ChunkTime = the time it takes the DCP to send one chunk of data
327 	 * to the LB which consists of pipeline delay and inter chunk gap
328 	 * sclk = system clock(ns)
329 	 */
330 	a.full = dfixed_const(256 * 13);
331 	chunk_time.full = dfixed_mul(rdev->pm.sclk, a);
332 	a.full = dfixed_const(10);
333 	chunk_time.full = dfixed_div(chunk_time, a);
334 
335 	/* Determine the worst case latency
336 	 * NumLinePair = Number of line pairs to request(1=2 lines, 2=4 lines)
337 	 * WorstCaseLatency = worst case time from urgent to when the MC starts
338 	 *                    to return data
339 	 * READ_DELAY_IDLE_MAX = constant of 1us
340 	 * ChunkTime = time it takes the DCP to send one chunk of data to the LB
341 	 *             which consists of pipeline delay and inter chunk gap
342 	 */
343 	if (dfixed_trunc(wm->num_line_pair) > 1) {
344 		a.full = dfixed_const(3);
345 		wm->worst_case_latency.full = dfixed_mul(a, chunk_time);
346 		wm->worst_case_latency.full += read_delay_latency.full;
347 	} else {
348 		a.full = dfixed_const(2);
349 		wm->worst_case_latency.full = dfixed_mul(a, chunk_time);
350 		wm->worst_case_latency.full += read_delay_latency.full;
351 	}
352 
353 	/* Determine the tolerable latency
354 	 * TolerableLatency = Any given request has only 1 line time
355 	 *                    for the data to be returned
356 	 * LBRequestFifoDepth = Number of chunk requests the LB can
357 	 *                      put into the request FIFO for a display
358 	 *  LineTime = total time for one line of display
359 	 *  ChunkTime = the time it takes the DCP to send one chunk
360 	 *              of data to the LB which consists of
361 	 *  pipeline delay and inter chunk gap
362 	 */
363 	if ((2+wm->lb_request_fifo_depth) >= dfixed_trunc(request_fifo_depth)) {
364 		tolerable_latency.full = line_time.full;
365 	} else {
366 		tolerable_latency.full = dfixed_const(wm->lb_request_fifo_depth - 2);
367 		tolerable_latency.full = request_fifo_depth.full - tolerable_latency.full;
368 		tolerable_latency.full = dfixed_mul(tolerable_latency, chunk_time);
369 		tolerable_latency.full = line_time.full - tolerable_latency.full;
370 	}
371 	/* We assume worst case 32bits (4 bytes) */
372 	wm->dbpp.full = dfixed_const(4 * 8);
373 
374 	/* Determine the maximum priority mark
375 	 *  width = viewport width in pixels
376 	 */
377 	a.full = dfixed_const(16);
378 	wm->priority_mark_max.full = dfixed_const(crtc->base.mode.crtc_hdisplay);
379 	wm->priority_mark_max.full = dfixed_div(wm->priority_mark_max, a);
380 	wm->priority_mark_max.full = dfixed_ceil(wm->priority_mark_max);
381 
382 	/* Determine estimated width */
383 	estimated_width.full = tolerable_latency.full - wm->worst_case_latency.full;
384 	estimated_width.full = dfixed_div(estimated_width, consumption_time);
385 	if (dfixed_trunc(estimated_width) > crtc->base.mode.crtc_hdisplay) {
386 		wm->priority_mark.full = dfixed_const(10);
387 	} else {
388 		a.full = dfixed_const(16);
389 		wm->priority_mark.full = dfixed_div(estimated_width, a);
390 		wm->priority_mark.full = dfixed_ceil(wm->priority_mark);
391 		wm->priority_mark.full = wm->priority_mark_max.full - wm->priority_mark.full;
392 	}
393 }
394 
395 void rs690_bandwidth_update(struct radeon_device *rdev)
396 {
397 	struct drm_display_mode *mode0 = NULL;
398 	struct drm_display_mode *mode1 = NULL;
399 	struct rs690_watermark wm0;
400 	struct rs690_watermark wm1;
401 	u32 tmp;
402 	u32 d1mode_priority_a_cnt = S_006548_D1MODE_PRIORITY_A_OFF(1);
403 	u32 d2mode_priority_a_cnt = S_006548_D1MODE_PRIORITY_A_OFF(1);
404 	fixed20_12 priority_mark02, priority_mark12, fill_rate;
405 	fixed20_12 a, b;
406 
407 	radeon_update_display_priority(rdev);
408 
409 	if (rdev->mode_info.crtcs[0]->base.enabled)
410 		mode0 = &rdev->mode_info.crtcs[0]->base.mode;
411 	if (rdev->mode_info.crtcs[1]->base.enabled)
412 		mode1 = &rdev->mode_info.crtcs[1]->base.mode;
413 	/*
414 	 * Set display0/1 priority up in the memory controller for
415 	 * modes if the user specifies HIGH for displaypriority
416 	 * option.
417 	 */
418 	if ((rdev->disp_priority == 2) &&
419 	    ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))) {
420 		tmp = RREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER);
421 		tmp &= C_000104_MC_DISP0R_INIT_LAT;
422 		tmp &= C_000104_MC_DISP1R_INIT_LAT;
423 		if (mode0)
424 			tmp |= S_000104_MC_DISP0R_INIT_LAT(1);
425 		if (mode1)
426 			tmp |= S_000104_MC_DISP1R_INIT_LAT(1);
427 		WREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER, tmp);
428 	}
429 	rs690_line_buffer_adjust(rdev, mode0, mode1);
430 
431 	if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
432 		WREG32(R_006C9C_DCP_CONTROL, 0);
433 	if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
434 		WREG32(R_006C9C_DCP_CONTROL, 2);
435 
436 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0);
437 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1);
438 
439 	tmp = (wm0.lb_request_fifo_depth - 1);
440 	tmp |= (wm1.lb_request_fifo_depth - 1) << 16;
441 	WREG32(R_006D58_LB_MAX_REQ_OUTSTANDING, tmp);
442 
443 	if (mode0 && mode1) {
444 		if (dfixed_trunc(wm0.dbpp) > 64)
445 			a.full = dfixed_mul(wm0.dbpp, wm0.num_line_pair);
446 		else
447 			a.full = wm0.num_line_pair.full;
448 		if (dfixed_trunc(wm1.dbpp) > 64)
449 			b.full = dfixed_mul(wm1.dbpp, wm1.num_line_pair);
450 		else
451 			b.full = wm1.num_line_pair.full;
452 		a.full += b.full;
453 		fill_rate.full = dfixed_div(wm0.sclk, a);
454 		if (wm0.consumption_rate.full > fill_rate.full) {
455 			b.full = wm0.consumption_rate.full - fill_rate.full;
456 			b.full = dfixed_mul(b, wm0.active_time);
457 			a.full = dfixed_mul(wm0.worst_case_latency,
458 						wm0.consumption_rate);
459 			a.full = a.full + b.full;
460 			b.full = dfixed_const(16 * 1000);
461 			priority_mark02.full = dfixed_div(a, b);
462 		} else {
463 			a.full = dfixed_mul(wm0.worst_case_latency,
464 						wm0.consumption_rate);
465 			b.full = dfixed_const(16 * 1000);
466 			priority_mark02.full = dfixed_div(a, b);
467 		}
468 		if (wm1.consumption_rate.full > fill_rate.full) {
469 			b.full = wm1.consumption_rate.full - fill_rate.full;
470 			b.full = dfixed_mul(b, wm1.active_time);
471 			a.full = dfixed_mul(wm1.worst_case_latency,
472 						wm1.consumption_rate);
473 			a.full = a.full + b.full;
474 			b.full = dfixed_const(16 * 1000);
475 			priority_mark12.full = dfixed_div(a, b);
476 		} else {
477 			a.full = dfixed_mul(wm1.worst_case_latency,
478 						wm1.consumption_rate);
479 			b.full = dfixed_const(16 * 1000);
480 			priority_mark12.full = dfixed_div(a, b);
481 		}
482 		if (wm0.priority_mark.full > priority_mark02.full)
483 			priority_mark02.full = wm0.priority_mark.full;
484 		if (dfixed_trunc(priority_mark02) < 0)
485 			priority_mark02.full = 0;
486 		if (wm0.priority_mark_max.full > priority_mark02.full)
487 			priority_mark02.full = wm0.priority_mark_max.full;
488 		if (wm1.priority_mark.full > priority_mark12.full)
489 			priority_mark12.full = wm1.priority_mark.full;
490 		if (dfixed_trunc(priority_mark12) < 0)
491 			priority_mark12.full = 0;
492 		if (wm1.priority_mark_max.full > priority_mark12.full)
493 			priority_mark12.full = wm1.priority_mark_max.full;
494 		d1mode_priority_a_cnt = dfixed_trunc(priority_mark02);
495 		d2mode_priority_a_cnt = dfixed_trunc(priority_mark12);
496 		if (rdev->disp_priority == 2) {
497 			d1mode_priority_a_cnt |= S_006548_D1MODE_PRIORITY_A_ALWAYS_ON(1);
498 			d2mode_priority_a_cnt |= S_006D48_D2MODE_PRIORITY_A_ALWAYS_ON(1);
499 		}
500 	} else if (mode0) {
501 		if (dfixed_trunc(wm0.dbpp) > 64)
502 			a.full = dfixed_mul(wm0.dbpp, wm0.num_line_pair);
503 		else
504 			a.full = wm0.num_line_pair.full;
505 		fill_rate.full = dfixed_div(wm0.sclk, a);
506 		if (wm0.consumption_rate.full > fill_rate.full) {
507 			b.full = wm0.consumption_rate.full - fill_rate.full;
508 			b.full = dfixed_mul(b, wm0.active_time);
509 			a.full = dfixed_mul(wm0.worst_case_latency,
510 						wm0.consumption_rate);
511 			a.full = a.full + b.full;
512 			b.full = dfixed_const(16 * 1000);
513 			priority_mark02.full = dfixed_div(a, b);
514 		} else {
515 			a.full = dfixed_mul(wm0.worst_case_latency,
516 						wm0.consumption_rate);
517 			b.full = dfixed_const(16 * 1000);
518 			priority_mark02.full = dfixed_div(a, b);
519 		}
520 		if (wm0.priority_mark.full > priority_mark02.full)
521 			priority_mark02.full = wm0.priority_mark.full;
522 		if (dfixed_trunc(priority_mark02) < 0)
523 			priority_mark02.full = 0;
524 		if (wm0.priority_mark_max.full > priority_mark02.full)
525 			priority_mark02.full = wm0.priority_mark_max.full;
526 		d1mode_priority_a_cnt = dfixed_trunc(priority_mark02);
527 		if (rdev->disp_priority == 2)
528 			d1mode_priority_a_cnt |= S_006548_D1MODE_PRIORITY_A_ALWAYS_ON(1);
529 	} else if (mode1) {
530 		if (dfixed_trunc(wm1.dbpp) > 64)
531 			a.full = dfixed_mul(wm1.dbpp, wm1.num_line_pair);
532 		else
533 			a.full = wm1.num_line_pair.full;
534 		fill_rate.full = dfixed_div(wm1.sclk, a);
535 		if (wm1.consumption_rate.full > fill_rate.full) {
536 			b.full = wm1.consumption_rate.full - fill_rate.full;
537 			b.full = dfixed_mul(b, wm1.active_time);
538 			a.full = dfixed_mul(wm1.worst_case_latency,
539 						wm1.consumption_rate);
540 			a.full = a.full + b.full;
541 			b.full = dfixed_const(16 * 1000);
542 			priority_mark12.full = dfixed_div(a, b);
543 		} else {
544 			a.full = dfixed_mul(wm1.worst_case_latency,
545 						wm1.consumption_rate);
546 			b.full = dfixed_const(16 * 1000);
547 			priority_mark12.full = dfixed_div(a, b);
548 		}
549 		if (wm1.priority_mark.full > priority_mark12.full)
550 			priority_mark12.full = wm1.priority_mark.full;
551 		if (dfixed_trunc(priority_mark12) < 0)
552 			priority_mark12.full = 0;
553 		if (wm1.priority_mark_max.full > priority_mark12.full)
554 			priority_mark12.full = wm1.priority_mark_max.full;
555 		d2mode_priority_a_cnt = dfixed_trunc(priority_mark12);
556 		if (rdev->disp_priority == 2)
557 			d2mode_priority_a_cnt |= S_006D48_D2MODE_PRIORITY_A_ALWAYS_ON(1);
558 	}
559 
560 	WREG32(R_006548_D1MODE_PRIORITY_A_CNT, d1mode_priority_a_cnt);
561 	WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, d1mode_priority_a_cnt);
562 	WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, d2mode_priority_a_cnt);
563 	WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, d2mode_priority_a_cnt);
564 }
565 
566 uint32_t rs690_mc_rreg(struct radeon_device *rdev, uint32_t reg)
567 {
568 	uint32_t r;
569 
570 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg));
571 	r = RREG32(R_00007C_MC_DATA);
572 	WREG32(R_000078_MC_INDEX, ~C_000078_MC_IND_ADDR);
573 	return r;
574 }
575 
576 void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
577 {
578 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg) |
579 		S_000078_MC_IND_WR_EN(1));
580 	WREG32(R_00007C_MC_DATA, v);
581 	WREG32(R_000078_MC_INDEX, 0x7F);
582 }
583 
584 void rs690_mc_program(struct radeon_device *rdev)
585 {
586 	struct rv515_mc_save save;
587 
588 	/* Stops all mc clients */
589 	rv515_mc_stop(rdev, &save);
590 
591 	/* Wait for mc idle */
592 	if (rs690_mc_wait_for_idle(rdev))
593 		dev_warn(rdev->dev, "Wait MC idle timeout before updating MC.\n");
594 	/* Program MC, should be a 32bits limited address space */
595 	WREG32_MC(R_000100_MCCFG_FB_LOCATION,
596 			S_000100_MC_FB_START(rdev->mc.vram_start >> 16) |
597 			S_000100_MC_FB_TOP(rdev->mc.vram_end >> 16));
598 	WREG32(R_000134_HDP_FB_LOCATION,
599 		S_000134_HDP_FB_START(rdev->mc.vram_start >> 16));
600 
601 	rv515_mc_resume(rdev, &save);
602 }
603 
604 static int rs690_startup(struct radeon_device *rdev)
605 {
606 	int r;
607 
608 	rs690_mc_program(rdev);
609 	/* Resume clock */
610 	rv515_clock_startup(rdev);
611 	/* Initialize GPU configuration (# pipes, ...) */
612 	rs690_gpu_init(rdev);
613 	/* Initialize GART (initialize after TTM so we can allocate
614 	 * memory through TTM but finalize after TTM) */
615 	r = rs400_gart_enable(rdev);
616 	if (r)
617 		return r;
618 
619 	/* allocate wb buffer */
620 	r = radeon_wb_init(rdev);
621 	if (r)
622 		return r;
623 
624 	r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
625 	if (r) {
626 		dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
627 		return r;
628 	}
629 
630 	/* Enable IRQ */
631 	rs600_irq_set(rdev);
632 	rdev->config.r300.hdp_cntl = RREG32(RADEON_HOST_PATH_CNTL);
633 	/* 1M ring buffer */
634 	r = r100_cp_init(rdev, 1024 * 1024);
635 	if (r) {
636 		dev_err(rdev->dev, "failed initializing CP (%d).\n", r);
637 		return r;
638 	}
639 
640 	r = radeon_ib_pool_init(rdev);
641 	if (r) {
642 		dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
643 		return r;
644 	}
645 
646 	r = r600_audio_init(rdev);
647 	if (r) {
648 		dev_err(rdev->dev, "failed initializing audio\n");
649 		return r;
650 	}
651 
652 	return 0;
653 }
654 
655 int rs690_resume(struct radeon_device *rdev)
656 {
657 	int r;
658 
659 	/* Make sur GART are not working */
660 	rs400_gart_disable(rdev);
661 	/* Resume clock before doing reset */
662 	rv515_clock_startup(rdev);
663 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
664 	if (radeon_asic_reset(rdev)) {
665 		dev_warn(rdev->dev, "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
666 			RREG32(R_000E40_RBBM_STATUS),
667 			RREG32(R_0007C0_CP_STAT));
668 	}
669 	/* post */
670 	atom_asic_init(rdev->mode_info.atom_context);
671 	/* Resume clock after posting */
672 	rv515_clock_startup(rdev);
673 	/* Initialize surface registers */
674 	radeon_surface_init(rdev);
675 
676 	rdev->accel_working = true;
677 	r = rs690_startup(rdev);
678 	if (r) {
679 		rdev->accel_working = false;
680 	}
681 	return r;
682 }
683 
684 int rs690_suspend(struct radeon_device *rdev)
685 {
686 	r600_audio_fini(rdev);
687 	r100_cp_disable(rdev);
688 	radeon_wb_disable(rdev);
689 	rs600_irq_disable(rdev);
690 	rs400_gart_disable(rdev);
691 	return 0;
692 }
693 
694 void rs690_fini(struct radeon_device *rdev)
695 {
696 	r600_audio_fini(rdev);
697 	r100_cp_fini(rdev);
698 	radeon_wb_fini(rdev);
699 	radeon_ib_pool_fini(rdev);
700 	radeon_gem_fini(rdev);
701 	rs400_gart_fini(rdev);
702 	radeon_irq_kms_fini(rdev);
703 	radeon_fence_driver_fini(rdev);
704 	radeon_bo_fini(rdev);
705 	radeon_atombios_fini(rdev);
706 	kfree(rdev->bios);
707 	rdev->bios = NULL;
708 }
709 
710 int rs690_init(struct radeon_device *rdev)
711 {
712 	int r;
713 
714 	/* Disable VGA */
715 	rv515_vga_render_disable(rdev);
716 	/* Initialize scratch registers */
717 	radeon_scratch_init(rdev);
718 	/* Initialize surface registers */
719 	radeon_surface_init(rdev);
720 	/* restore some register to sane defaults */
721 	r100_restore_sanity(rdev);
722 	/* TODO: disable VGA need to use VGA request */
723 	/* BIOS*/
724 	if (!radeon_get_bios(rdev)) {
725 		if (ASIC_IS_AVIVO(rdev))
726 			return -EINVAL;
727 	}
728 	if (rdev->is_atom_bios) {
729 		r = radeon_atombios_init(rdev);
730 		if (r)
731 			return r;
732 	} else {
733 		dev_err(rdev->dev, "Expecting atombios for RV515 GPU\n");
734 		return -EINVAL;
735 	}
736 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
737 	if (radeon_asic_reset(rdev)) {
738 		dev_warn(rdev->dev,
739 			"GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
740 			RREG32(R_000E40_RBBM_STATUS),
741 			RREG32(R_0007C0_CP_STAT));
742 	}
743 	/* check if cards are posted or not */
744 	if (radeon_boot_test_post_card(rdev) == false)
745 		return -EINVAL;
746 
747 	/* Initialize clocks */
748 	radeon_get_clock_info(rdev->ddev);
749 	/* initialize memory controller */
750 	rs690_mc_init(rdev);
751 	rv515_debugfs(rdev);
752 	/* Fence driver */
753 	r = radeon_fence_driver_init(rdev);
754 	if (r)
755 		return r;
756 	r = radeon_irq_kms_init(rdev);
757 	if (r)
758 		return r;
759 	/* Memory manager */
760 	r = radeon_bo_init(rdev);
761 	if (r)
762 		return r;
763 	r = rs400_gart_init(rdev);
764 	if (r)
765 		return r;
766 	rs600_set_safe_registers(rdev);
767 
768 	rdev->accel_working = true;
769 	r = rs690_startup(rdev);
770 	if (r) {
771 		/* Somethings want wront with the accel init stop accel */
772 		dev_err(rdev->dev, "Disabling GPU acceleration\n");
773 		r100_cp_fini(rdev);
774 		radeon_wb_fini(rdev);
775 		radeon_ib_pool_fini(rdev);
776 		rs400_gart_fini(rdev);
777 		radeon_irq_kms_fini(rdev);
778 		rdev->accel_working = false;
779 	}
780 	return 0;
781 }
782