xref: /linux/drivers/gpu/drm/radeon/rs690.c (revision fad837c16cdd856c68ce2e1335ad0fe836ed8ecd)
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 "atom.h"
31 #include "rs690d.h"
32 
33 static int rs690_mc_wait_for_idle(struct radeon_device *rdev)
34 {
35 	unsigned i;
36 	uint32_t tmp;
37 
38 	for (i = 0; i < rdev->usec_timeout; i++) {
39 		/* read MC_STATUS */
40 		tmp = RREG32_MC(R_000090_MC_SYSTEM_STATUS);
41 		if (G_000090_MC_SYSTEM_IDLE(tmp))
42 			return 0;
43 		udelay(1);
44 	}
45 	return -1;
46 }
47 
48 static void rs690_gpu_init(struct radeon_device *rdev)
49 {
50 	/* FIXME: HDP same place on rs690 ? */
51 	r100_hdp_reset(rdev);
52 	/* FIXME: is this correct ? */
53 	r420_pipes_init(rdev);
54 	if (rs690_mc_wait_for_idle(rdev)) {
55 		printk(KERN_WARNING "Failed to wait MC idle while "
56 		       "programming pipes. Bad things might happen.\n");
57 	}
58 }
59 
60 void rs690_pm_info(struct radeon_device *rdev)
61 {
62 	int index = GetIndexIntoMasterTable(DATA, IntegratedSystemInfo);
63 	struct _ATOM_INTEGRATED_SYSTEM_INFO *info;
64 	struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 *info_v2;
65 	void *ptr;
66 	uint16_t data_offset;
67 	uint8_t frev, crev;
68 	fixed20_12 tmp;
69 
70 	atom_parse_data_header(rdev->mode_info.atom_context, index, NULL,
71 			       &frev, &crev, &data_offset);
72 	ptr = rdev->mode_info.atom_context->bios + data_offset;
73 	info = (struct _ATOM_INTEGRATED_SYSTEM_INFO *)ptr;
74 	info_v2 = (struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 *)ptr;
75 	/* Get various system informations from bios */
76 	switch (crev) {
77 	case 1:
78 		tmp.full = rfixed_const(100);
79 		rdev->pm.igp_sideport_mclk.full = rfixed_const(info->ulBootUpMemoryClock);
80 		rdev->pm.igp_sideport_mclk.full = rfixed_div(rdev->pm.igp_sideport_mclk, tmp);
81 		rdev->pm.igp_system_mclk.full = rfixed_const(le16_to_cpu(info->usK8MemoryClock));
82 		rdev->pm.igp_ht_link_clk.full = rfixed_const(le16_to_cpu(info->usFSBClock));
83 		rdev->pm.igp_ht_link_width.full = rfixed_const(info->ucHTLinkWidth);
84 		break;
85 	case 2:
86 		tmp.full = rfixed_const(100);
87 		rdev->pm.igp_sideport_mclk.full = rfixed_const(info_v2->ulBootUpSidePortClock);
88 		rdev->pm.igp_sideport_mclk.full = rfixed_div(rdev->pm.igp_sideport_mclk, tmp);
89 		rdev->pm.igp_system_mclk.full = rfixed_const(info_v2->ulBootUpUMAClock);
90 		rdev->pm.igp_system_mclk.full = rfixed_div(rdev->pm.igp_system_mclk, tmp);
91 		rdev->pm.igp_ht_link_clk.full = rfixed_const(info_v2->ulHTLinkFreq);
92 		rdev->pm.igp_ht_link_clk.full = rfixed_div(rdev->pm.igp_ht_link_clk, tmp);
93 		rdev->pm.igp_ht_link_width.full = rfixed_const(le16_to_cpu(info_v2->usMinHTLinkWidth));
94 		break;
95 	default:
96 		tmp.full = rfixed_const(100);
97 		/* We assume the slower possible clock ie worst case */
98 		/* DDR 333Mhz */
99 		rdev->pm.igp_sideport_mclk.full = rfixed_const(333);
100 		/* FIXME: system clock ? */
101 		rdev->pm.igp_system_mclk.full = rfixed_const(100);
102 		rdev->pm.igp_system_mclk.full = rfixed_div(rdev->pm.igp_system_mclk, tmp);
103 		rdev->pm.igp_ht_link_clk.full = rfixed_const(200);
104 		rdev->pm.igp_ht_link_width.full = rfixed_const(8);
105 		DRM_ERROR("No integrated system info for your GPU, using safe default\n");
106 		break;
107 	}
108 	/* Compute various bandwidth */
109 	/* k8_bandwidth = (memory_clk / 2) * 2 * 8 * 0.5 = memory_clk * 4  */
110 	tmp.full = rfixed_const(4);
111 	rdev->pm.k8_bandwidth.full = rfixed_mul(rdev->pm.igp_system_mclk, tmp);
112 	/* ht_bandwidth = ht_clk * 2 * ht_width / 8 * 0.8
113 	 *              = ht_clk * ht_width / 5
114 	 */
115 	tmp.full = rfixed_const(5);
116 	rdev->pm.ht_bandwidth.full = rfixed_mul(rdev->pm.igp_ht_link_clk,
117 						rdev->pm.igp_ht_link_width);
118 	rdev->pm.ht_bandwidth.full = rfixed_div(rdev->pm.ht_bandwidth, tmp);
119 	if (tmp.full < rdev->pm.max_bandwidth.full) {
120 		/* HT link is a limiting factor */
121 		rdev->pm.max_bandwidth.full = tmp.full;
122 	}
123 	/* sideport_bandwidth = (sideport_clk / 2) * 2 * 2 * 0.7
124 	 *                    = (sideport_clk * 14) / 10
125 	 */
126 	tmp.full = rfixed_const(14);
127 	rdev->pm.sideport_bandwidth.full = rfixed_mul(rdev->pm.igp_sideport_mclk, tmp);
128 	tmp.full = rfixed_const(10);
129 	rdev->pm.sideport_bandwidth.full = rfixed_div(rdev->pm.sideport_bandwidth, tmp);
130 }
131 
132 void rs690_mc_init(struct radeon_device *rdev)
133 {
134 	fixed20_12 a;
135 	u64 base;
136 
137 	rs400_gart_adjust_size(rdev);
138 	rdev->mc.vram_is_ddr = true;
139 	rdev->mc.vram_width = 128;
140 	rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
141 	rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
142 	rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0);
143 	rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0);
144 	rdev->mc.visible_vram_size = rdev->mc.aper_size;
145 	base = RREG32_MC(R_000100_MCCFG_FB_LOCATION);
146 	base = G_000100_MC_FB_START(base) << 16;
147 	rs690_pm_info(rdev);
148 	/* FIXME: we should enforce default clock in case GPU is not in
149 	 * default setup
150 	 */
151 	a.full = rfixed_const(100);
152 	rdev->pm.sclk.full = rfixed_const(rdev->clock.default_sclk);
153 	rdev->pm.sclk.full = rfixed_div(rdev->pm.sclk, a);
154 	a.full = rfixed_const(16);
155 	/* core_bandwidth = sclk(Mhz) * 16 */
156 	rdev->pm.core_bandwidth.full = rfixed_div(rdev->pm.sclk, a);
157 	rdev->mc.igp_sideport_enabled = radeon_atombios_sideport_present(rdev);
158 	radeon_vram_location(rdev, &rdev->mc, base);
159 	radeon_gtt_location(rdev, &rdev->mc);
160 }
161 
162 void rs690_line_buffer_adjust(struct radeon_device *rdev,
163 			      struct drm_display_mode *mode1,
164 			      struct drm_display_mode *mode2)
165 {
166 	u32 tmp;
167 
168 	/*
169 	 * Line Buffer Setup
170 	 * There is a single line buffer shared by both display controllers.
171 	 * R_006520_DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
172 	 * the display controllers.  The paritioning can either be done
173 	 * manually or via one of four preset allocations specified in bits 1:0:
174 	 *  0 - line buffer is divided in half and shared between crtc
175 	 *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
176 	 *  2 - D1 gets the whole buffer
177 	 *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
178 	 * Setting bit 2 of R_006520_DC_LB_MEMORY_SPLIT controls switches to manual
179 	 * allocation mode. In manual allocation mode, D1 always starts at 0,
180 	 * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
181 	 */
182 	tmp = RREG32(R_006520_DC_LB_MEMORY_SPLIT) & C_006520_DC_LB_MEMORY_SPLIT;
183 	tmp &= ~C_006520_DC_LB_MEMORY_SPLIT_MODE;
184 	/* auto */
185 	if (mode1 && mode2) {
186 		if (mode1->hdisplay > mode2->hdisplay) {
187 			if (mode1->hdisplay > 2560)
188 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
189 			else
190 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
191 		} else if (mode2->hdisplay > mode1->hdisplay) {
192 			if (mode2->hdisplay > 2560)
193 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
194 			else
195 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
196 		} else
197 			tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
198 	} else if (mode1) {
199 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_ONLY;
200 	} else if (mode2) {
201 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
202 	}
203 	WREG32(R_006520_DC_LB_MEMORY_SPLIT, tmp);
204 }
205 
206 struct rs690_watermark {
207 	u32        lb_request_fifo_depth;
208 	fixed20_12 num_line_pair;
209 	fixed20_12 estimated_width;
210 	fixed20_12 worst_case_latency;
211 	fixed20_12 consumption_rate;
212 	fixed20_12 active_time;
213 	fixed20_12 dbpp;
214 	fixed20_12 priority_mark_max;
215 	fixed20_12 priority_mark;
216 	fixed20_12 sclk;
217 };
218 
219 void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
220 				  struct radeon_crtc *crtc,
221 				  struct rs690_watermark *wm)
222 {
223 	struct drm_display_mode *mode = &crtc->base.mode;
224 	fixed20_12 a, b, c;
225 	fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
226 	fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
227 	/* FIXME: detect IGP with sideport memory, i don't think there is any
228 	 * such product available
229 	 */
230 	bool sideport = false;
231 
232 	if (!crtc->base.enabled) {
233 		/* FIXME: wouldn't it better to set priority mark to maximum */
234 		wm->lb_request_fifo_depth = 4;
235 		return;
236 	}
237 
238 	if (crtc->vsc.full > rfixed_const(2))
239 		wm->num_line_pair.full = rfixed_const(2);
240 	else
241 		wm->num_line_pair.full = rfixed_const(1);
242 
243 	b.full = rfixed_const(mode->crtc_hdisplay);
244 	c.full = rfixed_const(256);
245 	a.full = rfixed_div(b, c);
246 	request_fifo_depth.full = rfixed_mul(a, wm->num_line_pair);
247 	request_fifo_depth.full = rfixed_ceil(request_fifo_depth);
248 	if (a.full < rfixed_const(4)) {
249 		wm->lb_request_fifo_depth = 4;
250 	} else {
251 		wm->lb_request_fifo_depth = rfixed_trunc(request_fifo_depth);
252 	}
253 
254 	/* Determine consumption rate
255 	 *  pclk = pixel clock period(ns) = 1000 / (mode.clock / 1000)
256 	 *  vtaps = number of vertical taps,
257 	 *  vsc = vertical scaling ratio, defined as source/destination
258 	 *  hsc = horizontal scaling ration, defined as source/destination
259 	 */
260 	a.full = rfixed_const(mode->clock);
261 	b.full = rfixed_const(1000);
262 	a.full = rfixed_div(a, b);
263 	pclk.full = rfixed_div(b, a);
264 	if (crtc->rmx_type != RMX_OFF) {
265 		b.full = rfixed_const(2);
266 		if (crtc->vsc.full > b.full)
267 			b.full = crtc->vsc.full;
268 		b.full = rfixed_mul(b, crtc->hsc);
269 		c.full = rfixed_const(2);
270 		b.full = rfixed_div(b, c);
271 		consumption_time.full = rfixed_div(pclk, b);
272 	} else {
273 		consumption_time.full = pclk.full;
274 	}
275 	a.full = rfixed_const(1);
276 	wm->consumption_rate.full = rfixed_div(a, consumption_time);
277 
278 
279 	/* Determine line time
280 	 *  LineTime = total time for one line of displayhtotal
281 	 *  LineTime = total number of horizontal pixels
282 	 *  pclk = pixel clock period(ns)
283 	 */
284 	a.full = rfixed_const(crtc->base.mode.crtc_htotal);
285 	line_time.full = rfixed_mul(a, pclk);
286 
287 	/* Determine active time
288 	 *  ActiveTime = time of active region of display within one line,
289 	 *  hactive = total number of horizontal active pixels
290 	 *  htotal = total number of horizontal pixels
291 	 */
292 	a.full = rfixed_const(crtc->base.mode.crtc_htotal);
293 	b.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
294 	wm->active_time.full = rfixed_mul(line_time, b);
295 	wm->active_time.full = rfixed_div(wm->active_time, a);
296 
297 	/* Maximun bandwidth is the minimun bandwidth of all component */
298 	rdev->pm.max_bandwidth = rdev->pm.core_bandwidth;
299 	if (sideport) {
300 		if (rdev->pm.max_bandwidth.full > rdev->pm.sideport_bandwidth.full &&
301 			rdev->pm.sideport_bandwidth.full)
302 			rdev->pm.max_bandwidth = rdev->pm.sideport_bandwidth;
303 		read_delay_latency.full = rfixed_const(370 * 800 * 1000);
304 		read_delay_latency.full = rfixed_div(read_delay_latency,
305 			rdev->pm.igp_sideport_mclk);
306 	} else {
307 		if (rdev->pm.max_bandwidth.full > rdev->pm.k8_bandwidth.full &&
308 			rdev->pm.k8_bandwidth.full)
309 			rdev->pm.max_bandwidth = rdev->pm.k8_bandwidth;
310 		if (rdev->pm.max_bandwidth.full > rdev->pm.ht_bandwidth.full &&
311 			rdev->pm.ht_bandwidth.full)
312 			rdev->pm.max_bandwidth = rdev->pm.ht_bandwidth;
313 		read_delay_latency.full = rfixed_const(5000);
314 	}
315 
316 	/* sclk = system clocks(ns) = 1000 / max_bandwidth / 16 */
317 	a.full = rfixed_const(16);
318 	rdev->pm.sclk.full = rfixed_mul(rdev->pm.max_bandwidth, a);
319 	a.full = rfixed_const(1000);
320 	rdev->pm.sclk.full = rfixed_div(a, rdev->pm.sclk);
321 	/* Determine chunk time
322 	 * ChunkTime = the time it takes the DCP to send one chunk of data
323 	 * to the LB which consists of pipeline delay and inter chunk gap
324 	 * sclk = system clock(ns)
325 	 */
326 	a.full = rfixed_const(256 * 13);
327 	chunk_time.full = rfixed_mul(rdev->pm.sclk, a);
328 	a.full = rfixed_const(10);
329 	chunk_time.full = rfixed_div(chunk_time, a);
330 
331 	/* Determine the worst case latency
332 	 * NumLinePair = Number of line pairs to request(1=2 lines, 2=4 lines)
333 	 * WorstCaseLatency = worst case time from urgent to when the MC starts
334 	 *                    to return data
335 	 * READ_DELAY_IDLE_MAX = constant of 1us
336 	 * ChunkTime = time it takes the DCP to send one chunk of data to the LB
337 	 *             which consists of pipeline delay and inter chunk gap
338 	 */
339 	if (rfixed_trunc(wm->num_line_pair) > 1) {
340 		a.full = rfixed_const(3);
341 		wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
342 		wm->worst_case_latency.full += read_delay_latency.full;
343 	} else {
344 		a.full = rfixed_const(2);
345 		wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
346 		wm->worst_case_latency.full += read_delay_latency.full;
347 	}
348 
349 	/* Determine the tolerable latency
350 	 * TolerableLatency = Any given request has only 1 line time
351 	 *                    for the data to be returned
352 	 * LBRequestFifoDepth = Number of chunk requests the LB can
353 	 *                      put into the request FIFO for a display
354 	 *  LineTime = total time for one line of display
355 	 *  ChunkTime = the time it takes the DCP to send one chunk
356 	 *              of data to the LB which consists of
357 	 *  pipeline delay and inter chunk gap
358 	 */
359 	if ((2+wm->lb_request_fifo_depth) >= rfixed_trunc(request_fifo_depth)) {
360 		tolerable_latency.full = line_time.full;
361 	} else {
362 		tolerable_latency.full = rfixed_const(wm->lb_request_fifo_depth - 2);
363 		tolerable_latency.full = request_fifo_depth.full - tolerable_latency.full;
364 		tolerable_latency.full = rfixed_mul(tolerable_latency, chunk_time);
365 		tolerable_latency.full = line_time.full - tolerable_latency.full;
366 	}
367 	/* We assume worst case 32bits (4 bytes) */
368 	wm->dbpp.full = rfixed_const(4 * 8);
369 
370 	/* Determine the maximum priority mark
371 	 *  width = viewport width in pixels
372 	 */
373 	a.full = rfixed_const(16);
374 	wm->priority_mark_max.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
375 	wm->priority_mark_max.full = rfixed_div(wm->priority_mark_max, a);
376 	wm->priority_mark_max.full = rfixed_ceil(wm->priority_mark_max);
377 
378 	/* Determine estimated width */
379 	estimated_width.full = tolerable_latency.full - wm->worst_case_latency.full;
380 	estimated_width.full = rfixed_div(estimated_width, consumption_time);
381 	if (rfixed_trunc(estimated_width) > crtc->base.mode.crtc_hdisplay) {
382 		wm->priority_mark.full = rfixed_const(10);
383 	} else {
384 		a.full = rfixed_const(16);
385 		wm->priority_mark.full = rfixed_div(estimated_width, a);
386 		wm->priority_mark.full = rfixed_ceil(wm->priority_mark);
387 		wm->priority_mark.full = wm->priority_mark_max.full - wm->priority_mark.full;
388 	}
389 }
390 
391 void rs690_bandwidth_update(struct radeon_device *rdev)
392 {
393 	struct drm_display_mode *mode0 = NULL;
394 	struct drm_display_mode *mode1 = NULL;
395 	struct rs690_watermark wm0;
396 	struct rs690_watermark wm1;
397 	u32 tmp;
398 	fixed20_12 priority_mark02, priority_mark12, fill_rate;
399 	fixed20_12 a, b;
400 
401 	if (rdev->mode_info.crtcs[0]->base.enabled)
402 		mode0 = &rdev->mode_info.crtcs[0]->base.mode;
403 	if (rdev->mode_info.crtcs[1]->base.enabled)
404 		mode1 = &rdev->mode_info.crtcs[1]->base.mode;
405 	/*
406 	 * Set display0/1 priority up in the memory controller for
407 	 * modes if the user specifies HIGH for displaypriority
408 	 * option.
409 	 */
410 	if (rdev->disp_priority == 2) {
411 		tmp = RREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER);
412 		tmp &= C_000104_MC_DISP0R_INIT_LAT;
413 		tmp &= C_000104_MC_DISP1R_INIT_LAT;
414 		if (mode0)
415 			tmp |= S_000104_MC_DISP0R_INIT_LAT(1);
416 		if (mode1)
417 			tmp |= S_000104_MC_DISP1R_INIT_LAT(1);
418 		WREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER, tmp);
419 	}
420 	rs690_line_buffer_adjust(rdev, mode0, mode1);
421 
422 	if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
423 		WREG32(R_006C9C_DCP_CONTROL, 0);
424 	if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
425 		WREG32(R_006C9C_DCP_CONTROL, 2);
426 
427 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0);
428 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1);
429 
430 	tmp = (wm0.lb_request_fifo_depth - 1);
431 	tmp |= (wm1.lb_request_fifo_depth - 1) << 16;
432 	WREG32(R_006D58_LB_MAX_REQ_OUTSTANDING, tmp);
433 
434 	if (mode0 && mode1) {
435 		if (rfixed_trunc(wm0.dbpp) > 64)
436 			a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
437 		else
438 			a.full = wm0.num_line_pair.full;
439 		if (rfixed_trunc(wm1.dbpp) > 64)
440 			b.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
441 		else
442 			b.full = wm1.num_line_pair.full;
443 		a.full += b.full;
444 		fill_rate.full = rfixed_div(wm0.sclk, a);
445 		if (wm0.consumption_rate.full > fill_rate.full) {
446 			b.full = wm0.consumption_rate.full - fill_rate.full;
447 			b.full = rfixed_mul(b, wm0.active_time);
448 			a.full = rfixed_mul(wm0.worst_case_latency,
449 						wm0.consumption_rate);
450 			a.full = a.full + b.full;
451 			b.full = rfixed_const(16 * 1000);
452 			priority_mark02.full = rfixed_div(a, b);
453 		} else {
454 			a.full = rfixed_mul(wm0.worst_case_latency,
455 						wm0.consumption_rate);
456 			b.full = rfixed_const(16 * 1000);
457 			priority_mark02.full = rfixed_div(a, b);
458 		}
459 		if (wm1.consumption_rate.full > fill_rate.full) {
460 			b.full = wm1.consumption_rate.full - fill_rate.full;
461 			b.full = rfixed_mul(b, wm1.active_time);
462 			a.full = rfixed_mul(wm1.worst_case_latency,
463 						wm1.consumption_rate);
464 			a.full = a.full + b.full;
465 			b.full = rfixed_const(16 * 1000);
466 			priority_mark12.full = rfixed_div(a, b);
467 		} else {
468 			a.full = rfixed_mul(wm1.worst_case_latency,
469 						wm1.consumption_rate);
470 			b.full = rfixed_const(16 * 1000);
471 			priority_mark12.full = rfixed_div(a, b);
472 		}
473 		if (wm0.priority_mark.full > priority_mark02.full)
474 			priority_mark02.full = wm0.priority_mark.full;
475 		if (rfixed_trunc(priority_mark02) < 0)
476 			priority_mark02.full = 0;
477 		if (wm0.priority_mark_max.full > priority_mark02.full)
478 			priority_mark02.full = wm0.priority_mark_max.full;
479 		if (wm1.priority_mark.full > priority_mark12.full)
480 			priority_mark12.full = wm1.priority_mark.full;
481 		if (rfixed_trunc(priority_mark12) < 0)
482 			priority_mark12.full = 0;
483 		if (wm1.priority_mark_max.full > priority_mark12.full)
484 			priority_mark12.full = wm1.priority_mark_max.full;
485 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
486 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
487 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
488 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
489 	} else if (mode0) {
490 		if (rfixed_trunc(wm0.dbpp) > 64)
491 			a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
492 		else
493 			a.full = wm0.num_line_pair.full;
494 		fill_rate.full = rfixed_div(wm0.sclk, a);
495 		if (wm0.consumption_rate.full > fill_rate.full) {
496 			b.full = wm0.consumption_rate.full - fill_rate.full;
497 			b.full = rfixed_mul(b, wm0.active_time);
498 			a.full = rfixed_mul(wm0.worst_case_latency,
499 						wm0.consumption_rate);
500 			a.full = a.full + b.full;
501 			b.full = rfixed_const(16 * 1000);
502 			priority_mark02.full = rfixed_div(a, b);
503 		} else {
504 			a.full = rfixed_mul(wm0.worst_case_latency,
505 						wm0.consumption_rate);
506 			b.full = rfixed_const(16 * 1000);
507 			priority_mark02.full = rfixed_div(a, b);
508 		}
509 		if (wm0.priority_mark.full > priority_mark02.full)
510 			priority_mark02.full = wm0.priority_mark.full;
511 		if (rfixed_trunc(priority_mark02) < 0)
512 			priority_mark02.full = 0;
513 		if (wm0.priority_mark_max.full > priority_mark02.full)
514 			priority_mark02.full = wm0.priority_mark_max.full;
515 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
516 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
517 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT,
518 			S_006D48_D2MODE_PRIORITY_A_OFF(1));
519 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT,
520 			S_006D4C_D2MODE_PRIORITY_B_OFF(1));
521 	} else {
522 		if (rfixed_trunc(wm1.dbpp) > 64)
523 			a.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
524 		else
525 			a.full = wm1.num_line_pair.full;
526 		fill_rate.full = rfixed_div(wm1.sclk, a);
527 		if (wm1.consumption_rate.full > fill_rate.full) {
528 			b.full = wm1.consumption_rate.full - fill_rate.full;
529 			b.full = rfixed_mul(b, wm1.active_time);
530 			a.full = rfixed_mul(wm1.worst_case_latency,
531 						wm1.consumption_rate);
532 			a.full = a.full + b.full;
533 			b.full = rfixed_const(16 * 1000);
534 			priority_mark12.full = rfixed_div(a, b);
535 		} else {
536 			a.full = rfixed_mul(wm1.worst_case_latency,
537 						wm1.consumption_rate);
538 			b.full = rfixed_const(16 * 1000);
539 			priority_mark12.full = rfixed_div(a, b);
540 		}
541 		if (wm1.priority_mark.full > priority_mark12.full)
542 			priority_mark12.full = wm1.priority_mark.full;
543 		if (rfixed_trunc(priority_mark12) < 0)
544 			priority_mark12.full = 0;
545 		if (wm1.priority_mark_max.full > priority_mark12.full)
546 			priority_mark12.full = wm1.priority_mark_max.full;
547 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT,
548 			S_006548_D1MODE_PRIORITY_A_OFF(1));
549 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT,
550 			S_00654C_D1MODE_PRIORITY_B_OFF(1));
551 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
552 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
553 	}
554 }
555 
556 uint32_t rs690_mc_rreg(struct radeon_device *rdev, uint32_t reg)
557 {
558 	uint32_t r;
559 
560 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg));
561 	r = RREG32(R_00007C_MC_DATA);
562 	WREG32(R_000078_MC_INDEX, ~C_000078_MC_IND_ADDR);
563 	return r;
564 }
565 
566 void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
567 {
568 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg) |
569 		S_000078_MC_IND_WR_EN(1));
570 	WREG32(R_00007C_MC_DATA, v);
571 	WREG32(R_000078_MC_INDEX, 0x7F);
572 }
573 
574 void rs690_mc_program(struct radeon_device *rdev)
575 {
576 	struct rv515_mc_save save;
577 
578 	/* Stops all mc clients */
579 	rv515_mc_stop(rdev, &save);
580 
581 	/* Wait for mc idle */
582 	if (rs690_mc_wait_for_idle(rdev))
583 		dev_warn(rdev->dev, "Wait MC idle timeout before updating MC.\n");
584 	/* Program MC, should be a 32bits limited address space */
585 	WREG32_MC(R_000100_MCCFG_FB_LOCATION,
586 			S_000100_MC_FB_START(rdev->mc.vram_start >> 16) |
587 			S_000100_MC_FB_TOP(rdev->mc.vram_end >> 16));
588 	WREG32(R_000134_HDP_FB_LOCATION,
589 		S_000134_HDP_FB_START(rdev->mc.vram_start >> 16));
590 
591 	rv515_mc_resume(rdev, &save);
592 }
593 
594 static int rs690_startup(struct radeon_device *rdev)
595 {
596 	int r;
597 
598 	rs690_mc_program(rdev);
599 	/* Resume clock */
600 	rv515_clock_startup(rdev);
601 	/* Initialize GPU configuration (# pipes, ...) */
602 	rs690_gpu_init(rdev);
603 	/* Initialize GART (initialize after TTM so we can allocate
604 	 * memory through TTM but finalize after TTM) */
605 	r = rs400_gart_enable(rdev);
606 	if (r)
607 		return r;
608 	/* Enable IRQ */
609 	rs600_irq_set(rdev);
610 	rdev->config.r300.hdp_cntl = RREG32(RADEON_HOST_PATH_CNTL);
611 	/* 1M ring buffer */
612 	r = r100_cp_init(rdev, 1024 * 1024);
613 	if (r) {
614 		dev_err(rdev->dev, "failled initializing CP (%d).\n", r);
615 		return r;
616 	}
617 	r = r100_wb_init(rdev);
618 	if (r)
619 		dev_err(rdev->dev, "failled initializing WB (%d).\n", r);
620 	r = r100_ib_init(rdev);
621 	if (r) {
622 		dev_err(rdev->dev, "failled initializing IB (%d).\n", r);
623 		return r;
624 	}
625 	return 0;
626 }
627 
628 int rs690_resume(struct radeon_device *rdev)
629 {
630 	/* Make sur GART are not working */
631 	rs400_gart_disable(rdev);
632 	/* Resume clock before doing reset */
633 	rv515_clock_startup(rdev);
634 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
635 	if (radeon_gpu_reset(rdev)) {
636 		dev_warn(rdev->dev, "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
637 			RREG32(R_000E40_RBBM_STATUS),
638 			RREG32(R_0007C0_CP_STAT));
639 	}
640 	/* post */
641 	atom_asic_init(rdev->mode_info.atom_context);
642 	/* Resume clock after posting */
643 	rv515_clock_startup(rdev);
644 	/* Initialize surface registers */
645 	radeon_surface_init(rdev);
646 	return rs690_startup(rdev);
647 }
648 
649 int rs690_suspend(struct radeon_device *rdev)
650 {
651 	r100_cp_disable(rdev);
652 	r100_wb_disable(rdev);
653 	rs600_irq_disable(rdev);
654 	rs400_gart_disable(rdev);
655 	return 0;
656 }
657 
658 void rs690_fini(struct radeon_device *rdev)
659 {
660 	r100_cp_fini(rdev);
661 	r100_wb_fini(rdev);
662 	r100_ib_fini(rdev);
663 	radeon_gem_fini(rdev);
664 	rs400_gart_fini(rdev);
665 	radeon_irq_kms_fini(rdev);
666 	radeon_fence_driver_fini(rdev);
667 	radeon_bo_fini(rdev);
668 	radeon_atombios_fini(rdev);
669 	kfree(rdev->bios);
670 	rdev->bios = NULL;
671 }
672 
673 int rs690_init(struct radeon_device *rdev)
674 {
675 	int r;
676 
677 	/* Disable VGA */
678 	rv515_vga_render_disable(rdev);
679 	/* Initialize scratch registers */
680 	radeon_scratch_init(rdev);
681 	/* Initialize surface registers */
682 	radeon_surface_init(rdev);
683 	/* TODO: disable VGA need to use VGA request */
684 	/* BIOS*/
685 	if (!radeon_get_bios(rdev)) {
686 		if (ASIC_IS_AVIVO(rdev))
687 			return -EINVAL;
688 	}
689 	if (rdev->is_atom_bios) {
690 		r = radeon_atombios_init(rdev);
691 		if (r)
692 			return r;
693 	} else {
694 		dev_err(rdev->dev, "Expecting atombios for RV515 GPU\n");
695 		return -EINVAL;
696 	}
697 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
698 	if (radeon_gpu_reset(rdev)) {
699 		dev_warn(rdev->dev,
700 			"GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
701 			RREG32(R_000E40_RBBM_STATUS),
702 			RREG32(R_0007C0_CP_STAT));
703 	}
704 	/* check if cards are posted or not */
705 	if (radeon_boot_test_post_card(rdev) == false)
706 		return -EINVAL;
707 
708 	/* Initialize clocks */
709 	radeon_get_clock_info(rdev->ddev);
710 	/* Initialize power management */
711 	radeon_pm_init(rdev);
712 	/* initialize memory controller */
713 	rs690_mc_init(rdev);
714 	rv515_debugfs(rdev);
715 	/* Fence driver */
716 	r = radeon_fence_driver_init(rdev);
717 	if (r)
718 		return r;
719 	r = radeon_irq_kms_init(rdev);
720 	if (r)
721 		return r;
722 	/* Memory manager */
723 	r = radeon_bo_init(rdev);
724 	if (r)
725 		return r;
726 	r = rs400_gart_init(rdev);
727 	if (r)
728 		return r;
729 	rs600_set_safe_registers(rdev);
730 	rdev->accel_working = true;
731 	r = rs690_startup(rdev);
732 	if (r) {
733 		/* Somethings want wront with the accel init stop accel */
734 		dev_err(rdev->dev, "Disabling GPU acceleration\n");
735 		r100_cp_fini(rdev);
736 		r100_wb_fini(rdev);
737 		r100_ib_fini(rdev);
738 		rs400_gart_fini(rdev);
739 		radeon_irq_kms_fini(rdev);
740 		rdev->accel_working = false;
741 	}
742 	return 0;
743 }
744