xref: /linux/drivers/gpu/drm/radeon/rs690.c (revision 273b281fa22c293963ee3e6eec418f5dda2dbc83)
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_vram_info(struct radeon_device *rdev)
133 {
134 	fixed20_12 a;
135 
136 	rs400_gart_adjust_size(rdev);
137 
138 	rdev->mc.vram_is_ddr = true;
139 	rdev->mc.vram_width = 128;
140 
141 	rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
142 	rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
143 
144 	rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0);
145 	rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0);
146 
147 	if (rdev->mc.mc_vram_size > rdev->mc.aper_size)
148 		rdev->mc.mc_vram_size = rdev->mc.aper_size;
149 
150 	if (rdev->mc.real_vram_size > rdev->mc.aper_size)
151 		rdev->mc.real_vram_size = rdev->mc.aper_size;
152 
153 	rs690_pm_info(rdev);
154 	/* FIXME: we should enforce default clock in case GPU is not in
155 	 * default setup
156 	 */
157 	a.full = rfixed_const(100);
158 	rdev->pm.sclk.full = rfixed_const(rdev->clock.default_sclk);
159 	rdev->pm.sclk.full = rfixed_div(rdev->pm.sclk, a);
160 	a.full = rfixed_const(16);
161 	/* core_bandwidth = sclk(Mhz) * 16 */
162 	rdev->pm.core_bandwidth.full = rfixed_div(rdev->pm.sclk, a);
163 }
164 
165 static int rs690_mc_init(struct radeon_device *rdev)
166 {
167 	int r;
168 	u32 tmp;
169 
170 	/* Setup GPU memory space */
171 	tmp = RREG32_MC(R_000100_MCCFG_FB_LOCATION);
172 	rdev->mc.vram_location = G_000100_MC_FB_START(tmp) << 16;
173 	rdev->mc.gtt_location = 0xFFFFFFFFUL;
174 	r = radeon_mc_setup(rdev);
175 	if (r)
176 		return r;
177 	return 0;
178 }
179 
180 void rs690_line_buffer_adjust(struct radeon_device *rdev,
181 			      struct drm_display_mode *mode1,
182 			      struct drm_display_mode *mode2)
183 {
184 	u32 tmp;
185 
186 	/*
187 	 * Line Buffer Setup
188 	 * There is a single line buffer shared by both display controllers.
189 	 * R_006520_DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
190 	 * the display controllers.  The paritioning can either be done
191 	 * manually or via one of four preset allocations specified in bits 1:0:
192 	 *  0 - line buffer is divided in half and shared between crtc
193 	 *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
194 	 *  2 - D1 gets the whole buffer
195 	 *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
196 	 * Setting bit 2 of R_006520_DC_LB_MEMORY_SPLIT controls switches to manual
197 	 * allocation mode. In manual allocation mode, D1 always starts at 0,
198 	 * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
199 	 */
200 	tmp = RREG32(R_006520_DC_LB_MEMORY_SPLIT) & C_006520_DC_LB_MEMORY_SPLIT;
201 	tmp &= ~C_006520_DC_LB_MEMORY_SPLIT_MODE;
202 	/* auto */
203 	if (mode1 && mode2) {
204 		if (mode1->hdisplay > mode2->hdisplay) {
205 			if (mode1->hdisplay > 2560)
206 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
207 			else
208 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
209 		} else if (mode2->hdisplay > mode1->hdisplay) {
210 			if (mode2->hdisplay > 2560)
211 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
212 			else
213 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
214 		} else
215 			tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
216 	} else if (mode1) {
217 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_ONLY;
218 	} else if (mode2) {
219 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
220 	}
221 	WREG32(R_006520_DC_LB_MEMORY_SPLIT, tmp);
222 }
223 
224 struct rs690_watermark {
225 	u32        lb_request_fifo_depth;
226 	fixed20_12 num_line_pair;
227 	fixed20_12 estimated_width;
228 	fixed20_12 worst_case_latency;
229 	fixed20_12 consumption_rate;
230 	fixed20_12 active_time;
231 	fixed20_12 dbpp;
232 	fixed20_12 priority_mark_max;
233 	fixed20_12 priority_mark;
234 	fixed20_12 sclk;
235 };
236 
237 void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
238 				  struct radeon_crtc *crtc,
239 				  struct rs690_watermark *wm)
240 {
241 	struct drm_display_mode *mode = &crtc->base.mode;
242 	fixed20_12 a, b, c;
243 	fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
244 	fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
245 	/* FIXME: detect IGP with sideport memory, i don't think there is any
246 	 * such product available
247 	 */
248 	bool sideport = false;
249 
250 	if (!crtc->base.enabled) {
251 		/* FIXME: wouldn't it better to set priority mark to maximum */
252 		wm->lb_request_fifo_depth = 4;
253 		return;
254 	}
255 
256 	if (crtc->vsc.full > rfixed_const(2))
257 		wm->num_line_pair.full = rfixed_const(2);
258 	else
259 		wm->num_line_pair.full = rfixed_const(1);
260 
261 	b.full = rfixed_const(mode->crtc_hdisplay);
262 	c.full = rfixed_const(256);
263 	a.full = rfixed_div(b, c);
264 	request_fifo_depth.full = rfixed_mul(a, wm->num_line_pair);
265 	request_fifo_depth.full = rfixed_ceil(request_fifo_depth);
266 	if (a.full < rfixed_const(4)) {
267 		wm->lb_request_fifo_depth = 4;
268 	} else {
269 		wm->lb_request_fifo_depth = rfixed_trunc(request_fifo_depth);
270 	}
271 
272 	/* Determine consumption rate
273 	 *  pclk = pixel clock period(ns) = 1000 / (mode.clock / 1000)
274 	 *  vtaps = number of vertical taps,
275 	 *  vsc = vertical scaling ratio, defined as source/destination
276 	 *  hsc = horizontal scaling ration, defined as source/destination
277 	 */
278 	a.full = rfixed_const(mode->clock);
279 	b.full = rfixed_const(1000);
280 	a.full = rfixed_div(a, b);
281 	pclk.full = rfixed_div(b, a);
282 	if (crtc->rmx_type != RMX_OFF) {
283 		b.full = rfixed_const(2);
284 		if (crtc->vsc.full > b.full)
285 			b.full = crtc->vsc.full;
286 		b.full = rfixed_mul(b, crtc->hsc);
287 		c.full = rfixed_const(2);
288 		b.full = rfixed_div(b, c);
289 		consumption_time.full = rfixed_div(pclk, b);
290 	} else {
291 		consumption_time.full = pclk.full;
292 	}
293 	a.full = rfixed_const(1);
294 	wm->consumption_rate.full = rfixed_div(a, consumption_time);
295 
296 
297 	/* Determine line time
298 	 *  LineTime = total time for one line of displayhtotal
299 	 *  LineTime = total number of horizontal pixels
300 	 *  pclk = pixel clock period(ns)
301 	 */
302 	a.full = rfixed_const(crtc->base.mode.crtc_htotal);
303 	line_time.full = rfixed_mul(a, pclk);
304 
305 	/* Determine active time
306 	 *  ActiveTime = time of active region of display within one line,
307 	 *  hactive = total number of horizontal active pixels
308 	 *  htotal = total number of horizontal pixels
309 	 */
310 	a.full = rfixed_const(crtc->base.mode.crtc_htotal);
311 	b.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
312 	wm->active_time.full = rfixed_mul(line_time, b);
313 	wm->active_time.full = rfixed_div(wm->active_time, a);
314 
315 	/* Maximun bandwidth is the minimun bandwidth of all component */
316 	rdev->pm.max_bandwidth = rdev->pm.core_bandwidth;
317 	if (sideport) {
318 		if (rdev->pm.max_bandwidth.full > rdev->pm.sideport_bandwidth.full &&
319 			rdev->pm.sideport_bandwidth.full)
320 			rdev->pm.max_bandwidth = rdev->pm.sideport_bandwidth;
321 		read_delay_latency.full = rfixed_const(370 * 800 * 1000);
322 		read_delay_latency.full = rfixed_div(read_delay_latency,
323 			rdev->pm.igp_sideport_mclk);
324 	} else {
325 		if (rdev->pm.max_bandwidth.full > rdev->pm.k8_bandwidth.full &&
326 			rdev->pm.k8_bandwidth.full)
327 			rdev->pm.max_bandwidth = rdev->pm.k8_bandwidth;
328 		if (rdev->pm.max_bandwidth.full > rdev->pm.ht_bandwidth.full &&
329 			rdev->pm.ht_bandwidth.full)
330 			rdev->pm.max_bandwidth = rdev->pm.ht_bandwidth;
331 		read_delay_latency.full = rfixed_const(5000);
332 	}
333 
334 	/* sclk = system clocks(ns) = 1000 / max_bandwidth / 16 */
335 	a.full = rfixed_const(16);
336 	rdev->pm.sclk.full = rfixed_mul(rdev->pm.max_bandwidth, a);
337 	a.full = rfixed_const(1000);
338 	rdev->pm.sclk.full = rfixed_div(a, rdev->pm.sclk);
339 	/* Determine chunk time
340 	 * ChunkTime = the time it takes the DCP to send one chunk of data
341 	 * to the LB which consists of pipeline delay and inter chunk gap
342 	 * sclk = system clock(ns)
343 	 */
344 	a.full = rfixed_const(256 * 13);
345 	chunk_time.full = rfixed_mul(rdev->pm.sclk, a);
346 	a.full = rfixed_const(10);
347 	chunk_time.full = rfixed_div(chunk_time, a);
348 
349 	/* Determine the worst case latency
350 	 * NumLinePair = Number of line pairs to request(1=2 lines, 2=4 lines)
351 	 * WorstCaseLatency = worst case time from urgent to when the MC starts
352 	 *                    to return data
353 	 * READ_DELAY_IDLE_MAX = constant of 1us
354 	 * ChunkTime = time it takes the DCP to send one chunk of data to the LB
355 	 *             which consists of pipeline delay and inter chunk gap
356 	 */
357 	if (rfixed_trunc(wm->num_line_pair) > 1) {
358 		a.full = rfixed_const(3);
359 		wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
360 		wm->worst_case_latency.full += read_delay_latency.full;
361 	} else {
362 		a.full = rfixed_const(2);
363 		wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
364 		wm->worst_case_latency.full += read_delay_latency.full;
365 	}
366 
367 	/* Determine the tolerable latency
368 	 * TolerableLatency = Any given request has only 1 line time
369 	 *                    for the data to be returned
370 	 * LBRequestFifoDepth = Number of chunk requests the LB can
371 	 *                      put into the request FIFO for a display
372 	 *  LineTime = total time for one line of display
373 	 *  ChunkTime = the time it takes the DCP to send one chunk
374 	 *              of data to the LB which consists of
375 	 *  pipeline delay and inter chunk gap
376 	 */
377 	if ((2+wm->lb_request_fifo_depth) >= rfixed_trunc(request_fifo_depth)) {
378 		tolerable_latency.full = line_time.full;
379 	} else {
380 		tolerable_latency.full = rfixed_const(wm->lb_request_fifo_depth - 2);
381 		tolerable_latency.full = request_fifo_depth.full - tolerable_latency.full;
382 		tolerable_latency.full = rfixed_mul(tolerable_latency, chunk_time);
383 		tolerable_latency.full = line_time.full - tolerable_latency.full;
384 	}
385 	/* We assume worst case 32bits (4 bytes) */
386 	wm->dbpp.full = rfixed_const(4 * 8);
387 
388 	/* Determine the maximum priority mark
389 	 *  width = viewport width in pixels
390 	 */
391 	a.full = rfixed_const(16);
392 	wm->priority_mark_max.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
393 	wm->priority_mark_max.full = rfixed_div(wm->priority_mark_max, a);
394 	wm->priority_mark_max.full = rfixed_ceil(wm->priority_mark_max);
395 
396 	/* Determine estimated width */
397 	estimated_width.full = tolerable_latency.full - wm->worst_case_latency.full;
398 	estimated_width.full = rfixed_div(estimated_width, consumption_time);
399 	if (rfixed_trunc(estimated_width) > crtc->base.mode.crtc_hdisplay) {
400 		wm->priority_mark.full = rfixed_const(10);
401 	} else {
402 		a.full = rfixed_const(16);
403 		wm->priority_mark.full = rfixed_div(estimated_width, a);
404 		wm->priority_mark.full = rfixed_ceil(wm->priority_mark);
405 		wm->priority_mark.full = wm->priority_mark_max.full - wm->priority_mark.full;
406 	}
407 }
408 
409 void rs690_bandwidth_update(struct radeon_device *rdev)
410 {
411 	struct drm_display_mode *mode0 = NULL;
412 	struct drm_display_mode *mode1 = NULL;
413 	struct rs690_watermark wm0;
414 	struct rs690_watermark wm1;
415 	u32 tmp;
416 	fixed20_12 priority_mark02, priority_mark12, fill_rate;
417 	fixed20_12 a, b;
418 
419 	if (rdev->mode_info.crtcs[0]->base.enabled)
420 		mode0 = &rdev->mode_info.crtcs[0]->base.mode;
421 	if (rdev->mode_info.crtcs[1]->base.enabled)
422 		mode1 = &rdev->mode_info.crtcs[1]->base.mode;
423 	/*
424 	 * Set display0/1 priority up in the memory controller for
425 	 * modes if the user specifies HIGH for displaypriority
426 	 * option.
427 	 */
428 	if (rdev->disp_priority == 2) {
429 		tmp = RREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER);
430 		tmp &= C_000104_MC_DISP0R_INIT_LAT;
431 		tmp &= C_000104_MC_DISP1R_INIT_LAT;
432 		if (mode0)
433 			tmp |= S_000104_MC_DISP0R_INIT_LAT(1);
434 		if (mode1)
435 			tmp |= S_000104_MC_DISP1R_INIT_LAT(1);
436 		WREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER, tmp);
437 	}
438 	rs690_line_buffer_adjust(rdev, mode0, mode1);
439 
440 	if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
441 		WREG32(R_006C9C_DCP_CONTROL, 0);
442 	if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
443 		WREG32(R_006C9C_DCP_CONTROL, 2);
444 
445 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0);
446 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1);
447 
448 	tmp = (wm0.lb_request_fifo_depth - 1);
449 	tmp |= (wm1.lb_request_fifo_depth - 1) << 16;
450 	WREG32(R_006D58_LB_MAX_REQ_OUTSTANDING, tmp);
451 
452 	if (mode0 && mode1) {
453 		if (rfixed_trunc(wm0.dbpp) > 64)
454 			a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
455 		else
456 			a.full = wm0.num_line_pair.full;
457 		if (rfixed_trunc(wm1.dbpp) > 64)
458 			b.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
459 		else
460 			b.full = wm1.num_line_pair.full;
461 		a.full += b.full;
462 		fill_rate.full = rfixed_div(wm0.sclk, a);
463 		if (wm0.consumption_rate.full > fill_rate.full) {
464 			b.full = wm0.consumption_rate.full - fill_rate.full;
465 			b.full = rfixed_mul(b, wm0.active_time);
466 			a.full = rfixed_mul(wm0.worst_case_latency,
467 						wm0.consumption_rate);
468 			a.full = a.full + b.full;
469 			b.full = rfixed_const(16 * 1000);
470 			priority_mark02.full = rfixed_div(a, b);
471 		} else {
472 			a.full = rfixed_mul(wm0.worst_case_latency,
473 						wm0.consumption_rate);
474 			b.full = rfixed_const(16 * 1000);
475 			priority_mark02.full = rfixed_div(a, b);
476 		}
477 		if (wm1.consumption_rate.full > fill_rate.full) {
478 			b.full = wm1.consumption_rate.full - fill_rate.full;
479 			b.full = rfixed_mul(b, wm1.active_time);
480 			a.full = rfixed_mul(wm1.worst_case_latency,
481 						wm1.consumption_rate);
482 			a.full = a.full + b.full;
483 			b.full = rfixed_const(16 * 1000);
484 			priority_mark12.full = rfixed_div(a, b);
485 		} else {
486 			a.full = rfixed_mul(wm1.worst_case_latency,
487 						wm1.consumption_rate);
488 			b.full = rfixed_const(16 * 1000);
489 			priority_mark12.full = rfixed_div(a, b);
490 		}
491 		if (wm0.priority_mark.full > priority_mark02.full)
492 			priority_mark02.full = wm0.priority_mark.full;
493 		if (rfixed_trunc(priority_mark02) < 0)
494 			priority_mark02.full = 0;
495 		if (wm0.priority_mark_max.full > priority_mark02.full)
496 			priority_mark02.full = wm0.priority_mark_max.full;
497 		if (wm1.priority_mark.full > priority_mark12.full)
498 			priority_mark12.full = wm1.priority_mark.full;
499 		if (rfixed_trunc(priority_mark12) < 0)
500 			priority_mark12.full = 0;
501 		if (wm1.priority_mark_max.full > priority_mark12.full)
502 			priority_mark12.full = wm1.priority_mark_max.full;
503 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
504 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
505 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
506 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
507 	} else if (mode0) {
508 		if (rfixed_trunc(wm0.dbpp) > 64)
509 			a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
510 		else
511 			a.full = wm0.num_line_pair.full;
512 		fill_rate.full = rfixed_div(wm0.sclk, a);
513 		if (wm0.consumption_rate.full > fill_rate.full) {
514 			b.full = wm0.consumption_rate.full - fill_rate.full;
515 			b.full = rfixed_mul(b, wm0.active_time);
516 			a.full = rfixed_mul(wm0.worst_case_latency,
517 						wm0.consumption_rate);
518 			a.full = a.full + b.full;
519 			b.full = rfixed_const(16 * 1000);
520 			priority_mark02.full = rfixed_div(a, b);
521 		} else {
522 			a.full = rfixed_mul(wm0.worst_case_latency,
523 						wm0.consumption_rate);
524 			b.full = rfixed_const(16 * 1000);
525 			priority_mark02.full = rfixed_div(a, b);
526 		}
527 		if (wm0.priority_mark.full > priority_mark02.full)
528 			priority_mark02.full = wm0.priority_mark.full;
529 		if (rfixed_trunc(priority_mark02) < 0)
530 			priority_mark02.full = 0;
531 		if (wm0.priority_mark_max.full > priority_mark02.full)
532 			priority_mark02.full = wm0.priority_mark_max.full;
533 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
534 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
535 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT,
536 			S_006D48_D2MODE_PRIORITY_A_OFF(1));
537 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT,
538 			S_006D4C_D2MODE_PRIORITY_B_OFF(1));
539 	} else {
540 		if (rfixed_trunc(wm1.dbpp) > 64)
541 			a.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
542 		else
543 			a.full = wm1.num_line_pair.full;
544 		fill_rate.full = rfixed_div(wm1.sclk, a);
545 		if (wm1.consumption_rate.full > fill_rate.full) {
546 			b.full = wm1.consumption_rate.full - fill_rate.full;
547 			b.full = rfixed_mul(b, wm1.active_time);
548 			a.full = rfixed_mul(wm1.worst_case_latency,
549 						wm1.consumption_rate);
550 			a.full = a.full + b.full;
551 			b.full = rfixed_const(16 * 1000);
552 			priority_mark12.full = rfixed_div(a, b);
553 		} else {
554 			a.full = rfixed_mul(wm1.worst_case_latency,
555 						wm1.consumption_rate);
556 			b.full = rfixed_const(16 * 1000);
557 			priority_mark12.full = rfixed_div(a, b);
558 		}
559 		if (wm1.priority_mark.full > priority_mark12.full)
560 			priority_mark12.full = wm1.priority_mark.full;
561 		if (rfixed_trunc(priority_mark12) < 0)
562 			priority_mark12.full = 0;
563 		if (wm1.priority_mark_max.full > priority_mark12.full)
564 			priority_mark12.full = wm1.priority_mark_max.full;
565 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT,
566 			S_006548_D1MODE_PRIORITY_A_OFF(1));
567 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT,
568 			S_00654C_D1MODE_PRIORITY_B_OFF(1));
569 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
570 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
571 	}
572 }
573 
574 uint32_t rs690_mc_rreg(struct radeon_device *rdev, uint32_t reg)
575 {
576 	uint32_t r;
577 
578 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg));
579 	r = RREG32(R_00007C_MC_DATA);
580 	WREG32(R_000078_MC_INDEX, ~C_000078_MC_IND_ADDR);
581 	return r;
582 }
583 
584 void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
585 {
586 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg) |
587 		S_000078_MC_IND_WR_EN(1));
588 	WREG32(R_00007C_MC_DATA, v);
589 	WREG32(R_000078_MC_INDEX, 0x7F);
590 }
591 
592 void rs690_mc_program(struct radeon_device *rdev)
593 {
594 	struct rv515_mc_save save;
595 
596 	/* Stops all mc clients */
597 	rv515_mc_stop(rdev, &save);
598 
599 	/* Wait for mc idle */
600 	if (rs690_mc_wait_for_idle(rdev))
601 		dev_warn(rdev->dev, "Wait MC idle timeout before updating MC.\n");
602 	/* Program MC, should be a 32bits limited address space */
603 	WREG32_MC(R_000100_MCCFG_FB_LOCATION,
604 			S_000100_MC_FB_START(rdev->mc.vram_start >> 16) |
605 			S_000100_MC_FB_TOP(rdev->mc.vram_end >> 16));
606 	WREG32(R_000134_HDP_FB_LOCATION,
607 		S_000134_HDP_FB_START(rdev->mc.vram_start >> 16));
608 
609 	rv515_mc_resume(rdev, &save);
610 }
611 
612 static int rs690_startup(struct radeon_device *rdev)
613 {
614 	int r;
615 
616 	rs690_mc_program(rdev);
617 	/* Resume clock */
618 	rv515_clock_startup(rdev);
619 	/* Initialize GPU configuration (# pipes, ...) */
620 	rs690_gpu_init(rdev);
621 	/* Initialize GART (initialize after TTM so we can allocate
622 	 * memory through TTM but finalize after TTM) */
623 	r = rs400_gart_enable(rdev);
624 	if (r)
625 		return r;
626 	/* Enable IRQ */
627 	rs600_irq_set(rdev);
628 	/* 1M ring buffer */
629 	r = r100_cp_init(rdev, 1024 * 1024);
630 	if (r) {
631 		dev_err(rdev->dev, "failled initializing CP (%d).\n", r);
632 		return r;
633 	}
634 	r = r100_wb_init(rdev);
635 	if (r)
636 		dev_err(rdev->dev, "failled initializing WB (%d).\n", r);
637 	r = r100_ib_init(rdev);
638 	if (r) {
639 		dev_err(rdev->dev, "failled initializing IB (%d).\n", r);
640 		return r;
641 	}
642 	return 0;
643 }
644 
645 int rs690_resume(struct radeon_device *rdev)
646 {
647 	/* Make sur GART are not working */
648 	rs400_gart_disable(rdev);
649 	/* Resume clock before doing reset */
650 	rv515_clock_startup(rdev);
651 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
652 	if (radeon_gpu_reset(rdev)) {
653 		dev_warn(rdev->dev, "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
654 			RREG32(R_000E40_RBBM_STATUS),
655 			RREG32(R_0007C0_CP_STAT));
656 	}
657 	/* post */
658 	atom_asic_init(rdev->mode_info.atom_context);
659 	/* Resume clock after posting */
660 	rv515_clock_startup(rdev);
661 	/* Initialize surface registers */
662 	radeon_surface_init(rdev);
663 	return rs690_startup(rdev);
664 }
665 
666 int rs690_suspend(struct radeon_device *rdev)
667 {
668 	r100_cp_disable(rdev);
669 	r100_wb_disable(rdev);
670 	rs600_irq_disable(rdev);
671 	rs400_gart_disable(rdev);
672 	return 0;
673 }
674 
675 void rs690_fini(struct radeon_device *rdev)
676 {
677 	rs690_suspend(rdev);
678 	r100_cp_fini(rdev);
679 	r100_wb_fini(rdev);
680 	r100_ib_fini(rdev);
681 	radeon_gem_fini(rdev);
682 	rs400_gart_fini(rdev);
683 	radeon_irq_kms_fini(rdev);
684 	radeon_fence_driver_fini(rdev);
685 	radeon_bo_fini(rdev);
686 	radeon_atombios_fini(rdev);
687 	kfree(rdev->bios);
688 	rdev->bios = NULL;
689 }
690 
691 int rs690_init(struct radeon_device *rdev)
692 {
693 	int r;
694 
695 	/* Disable VGA */
696 	rv515_vga_render_disable(rdev);
697 	/* Initialize scratch registers */
698 	radeon_scratch_init(rdev);
699 	/* Initialize surface registers */
700 	radeon_surface_init(rdev);
701 	/* TODO: disable VGA need to use VGA request */
702 	/* BIOS*/
703 	if (!radeon_get_bios(rdev)) {
704 		if (ASIC_IS_AVIVO(rdev))
705 			return -EINVAL;
706 	}
707 	if (rdev->is_atom_bios) {
708 		r = radeon_atombios_init(rdev);
709 		if (r)
710 			return r;
711 	} else {
712 		dev_err(rdev->dev, "Expecting atombios for RV515 GPU\n");
713 		return -EINVAL;
714 	}
715 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
716 	if (radeon_gpu_reset(rdev)) {
717 		dev_warn(rdev->dev,
718 			"GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
719 			RREG32(R_000E40_RBBM_STATUS),
720 			RREG32(R_0007C0_CP_STAT));
721 	}
722 	/* check if cards are posted or not */
723 	if (radeon_boot_test_post_card(rdev) == false)
724 		return -EINVAL;
725 
726 	/* Initialize clocks */
727 	radeon_get_clock_info(rdev->ddev);
728 	/* Initialize power management */
729 	radeon_pm_init(rdev);
730 	/* Get vram informations */
731 	rs690_vram_info(rdev);
732 	/* Initialize memory controller (also test AGP) */
733 	r = rs690_mc_init(rdev);
734 	if (r)
735 		return r;
736 	rv515_debugfs(rdev);
737 	/* Fence driver */
738 	r = radeon_fence_driver_init(rdev);
739 	if (r)
740 		return r;
741 	r = radeon_irq_kms_init(rdev);
742 	if (r)
743 		return r;
744 	/* Memory manager */
745 	r = radeon_bo_init(rdev);
746 	if (r)
747 		return r;
748 	r = rs400_gart_init(rdev);
749 	if (r)
750 		return r;
751 	rs600_set_safe_registers(rdev);
752 	rdev->accel_working = true;
753 	r = rs690_startup(rdev);
754 	if (r) {
755 		/* Somethings want wront with the accel init stop accel */
756 		dev_err(rdev->dev, "Disabling GPU acceleration\n");
757 		rs690_suspend(rdev);
758 		r100_cp_fini(rdev);
759 		r100_wb_fini(rdev);
760 		r100_ib_fini(rdev);
761 		rs400_gart_fini(rdev);
762 		radeon_irq_kms_fini(rdev);
763 		rdev->accel_working = false;
764 	}
765 	return 0;
766 }
767