xref: /openbmc/linux/drivers/gpu/drm/radeon/rs690.c (revision 1614f8b17b8cc3ad143541d41569623d30dbc9ec)
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 	uint32_t tmp;
135 	fixed20_12 a;
136 
137 	rs400_gart_adjust_size(rdev);
138 	/* DDR for all card after R300 & IGP */
139 	rdev->mc.vram_is_ddr = true;
140 	/* FIXME: is this correct for RS690/RS740 ? */
141 	tmp = RREG32(RADEON_MEM_CNTL);
142 	if (tmp & R300_MEM_NUM_CHANNELS_MASK) {
143 		rdev->mc.vram_width = 128;
144 	} else {
145 		rdev->mc.vram_width = 64;
146 	}
147 	rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
148 	rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
149 
150 	rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0);
151 	rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0);
152 	rs690_pm_info(rdev);
153 	/* FIXME: we should enforce default clock in case GPU is not in
154 	 * default setup
155 	 */
156 	a.full = rfixed_const(100);
157 	rdev->pm.sclk.full = rfixed_const(rdev->clock.default_sclk);
158 	rdev->pm.sclk.full = rfixed_div(rdev->pm.sclk, a);
159 	a.full = rfixed_const(16);
160 	/* core_bandwidth = sclk(Mhz) * 16 */
161 	rdev->pm.core_bandwidth.full = rfixed_div(rdev->pm.sclk, a);
162 }
163 
164 void rs690_line_buffer_adjust(struct radeon_device *rdev,
165 			      struct drm_display_mode *mode1,
166 			      struct drm_display_mode *mode2)
167 {
168 	u32 tmp;
169 
170 	/*
171 	 * Line Buffer Setup
172 	 * There is a single line buffer shared by both display controllers.
173 	 * R_006520_DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
174 	 * the display controllers.  The paritioning can either be done
175 	 * manually or via one of four preset allocations specified in bits 1:0:
176 	 *  0 - line buffer is divided in half and shared between crtc
177 	 *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
178 	 *  2 - D1 gets the whole buffer
179 	 *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
180 	 * Setting bit 2 of R_006520_DC_LB_MEMORY_SPLIT controls switches to manual
181 	 * allocation mode. In manual allocation mode, D1 always starts at 0,
182 	 * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
183 	 */
184 	tmp = RREG32(R_006520_DC_LB_MEMORY_SPLIT) & C_006520_DC_LB_MEMORY_SPLIT;
185 	tmp &= ~C_006520_DC_LB_MEMORY_SPLIT_MODE;
186 	/* auto */
187 	if (mode1 && mode2) {
188 		if (mode1->hdisplay > mode2->hdisplay) {
189 			if (mode1->hdisplay > 2560)
190 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
191 			else
192 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
193 		} else if (mode2->hdisplay > mode1->hdisplay) {
194 			if (mode2->hdisplay > 2560)
195 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
196 			else
197 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
198 		} else
199 			tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
200 	} else if (mode1) {
201 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_ONLY;
202 	} else if (mode2) {
203 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
204 	}
205 	WREG32(R_006520_DC_LB_MEMORY_SPLIT, tmp);
206 }
207 
208 struct rs690_watermark {
209 	u32        lb_request_fifo_depth;
210 	fixed20_12 num_line_pair;
211 	fixed20_12 estimated_width;
212 	fixed20_12 worst_case_latency;
213 	fixed20_12 consumption_rate;
214 	fixed20_12 active_time;
215 	fixed20_12 dbpp;
216 	fixed20_12 priority_mark_max;
217 	fixed20_12 priority_mark;
218 	fixed20_12 sclk;
219 };
220 
221 void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
222 				  struct radeon_crtc *crtc,
223 				  struct rs690_watermark *wm)
224 {
225 	struct drm_display_mode *mode = &crtc->base.mode;
226 	fixed20_12 a, b, c;
227 	fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
228 	fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
229 	/* FIXME: detect IGP with sideport memory, i don't think there is any
230 	 * such product available
231 	 */
232 	bool sideport = false;
233 
234 	if (!crtc->base.enabled) {
235 		/* FIXME: wouldn't it better to set priority mark to maximum */
236 		wm->lb_request_fifo_depth = 4;
237 		return;
238 	}
239 
240 	if (crtc->vsc.full > rfixed_const(2))
241 		wm->num_line_pair.full = rfixed_const(2);
242 	else
243 		wm->num_line_pair.full = rfixed_const(1);
244 
245 	b.full = rfixed_const(mode->crtc_hdisplay);
246 	c.full = rfixed_const(256);
247 	a.full = rfixed_mul(wm->num_line_pair, b);
248 	request_fifo_depth.full = rfixed_div(a, c);
249 	if (a.full < rfixed_const(4)) {
250 		wm->lb_request_fifo_depth = 4;
251 	} else {
252 		wm->lb_request_fifo_depth = rfixed_trunc(request_fifo_depth);
253 	}
254 
255 	/* Determine consumption rate
256 	 *  pclk = pixel clock period(ns) = 1000 / (mode.clock / 1000)
257 	 *  vtaps = number of vertical taps,
258 	 *  vsc = vertical scaling ratio, defined as source/destination
259 	 *  hsc = horizontal scaling ration, defined as source/destination
260 	 */
261 	a.full = rfixed_const(mode->clock);
262 	b.full = rfixed_const(1000);
263 	a.full = rfixed_div(a, b);
264 	pclk.full = rfixed_div(b, a);
265 	if (crtc->rmx_type != RMX_OFF) {
266 		b.full = rfixed_const(2);
267 		if (crtc->vsc.full > b.full)
268 			b.full = crtc->vsc.full;
269 		b.full = rfixed_mul(b, crtc->hsc);
270 		c.full = rfixed_const(2);
271 		b.full = rfixed_div(b, c);
272 		consumption_time.full = rfixed_div(pclk, b);
273 	} else {
274 		consumption_time.full = pclk.full;
275 	}
276 	a.full = rfixed_const(1);
277 	wm->consumption_rate.full = rfixed_div(a, consumption_time);
278 
279 
280 	/* Determine line time
281 	 *  LineTime = total time for one line of displayhtotal
282 	 *  LineTime = total number of horizontal pixels
283 	 *  pclk = pixel clock period(ns)
284 	 */
285 	a.full = rfixed_const(crtc->base.mode.crtc_htotal);
286 	line_time.full = rfixed_mul(a, pclk);
287 
288 	/* Determine active time
289 	 *  ActiveTime = time of active region of display within one line,
290 	 *  hactive = total number of horizontal active pixels
291 	 *  htotal = total number of horizontal pixels
292 	 */
293 	a.full = rfixed_const(crtc->base.mode.crtc_htotal);
294 	b.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
295 	wm->active_time.full = rfixed_mul(line_time, b);
296 	wm->active_time.full = rfixed_div(wm->active_time, a);
297 
298 	/* Maximun bandwidth is the minimun bandwidth of all component */
299 	rdev->pm.max_bandwidth = rdev->pm.core_bandwidth;
300 	if (sideport) {
301 		if (rdev->pm.max_bandwidth.full > rdev->pm.sideport_bandwidth.full &&
302 			rdev->pm.sideport_bandwidth.full)
303 			rdev->pm.max_bandwidth = rdev->pm.sideport_bandwidth;
304 		read_delay_latency.full = rfixed_const(370 * 800 * 1000);
305 		read_delay_latency.full = rfixed_div(read_delay_latency,
306 			rdev->pm.igp_sideport_mclk);
307 	} else {
308 		if (rdev->pm.max_bandwidth.full > rdev->pm.k8_bandwidth.full &&
309 			rdev->pm.k8_bandwidth.full)
310 			rdev->pm.max_bandwidth = rdev->pm.k8_bandwidth;
311 		if (rdev->pm.max_bandwidth.full > rdev->pm.ht_bandwidth.full &&
312 			rdev->pm.ht_bandwidth.full)
313 			rdev->pm.max_bandwidth = rdev->pm.ht_bandwidth;
314 		read_delay_latency.full = rfixed_const(5000);
315 	}
316 
317 	/* sclk = system clocks(ns) = 1000 / max_bandwidth / 16 */
318 	a.full = rfixed_const(16);
319 	rdev->pm.sclk.full = rfixed_mul(rdev->pm.max_bandwidth, a);
320 	a.full = rfixed_const(1000);
321 	rdev->pm.sclk.full = rfixed_div(a, rdev->pm.sclk);
322 	/* Determine chunk time
323 	 * ChunkTime = the time it takes the DCP to send one chunk of data
324 	 * to the LB which consists of pipeline delay and inter chunk gap
325 	 * sclk = system clock(ns)
326 	 */
327 	a.full = rfixed_const(256 * 13);
328 	chunk_time.full = rfixed_mul(rdev->pm.sclk, a);
329 	a.full = rfixed_const(10);
330 	chunk_time.full = rfixed_div(chunk_time, a);
331 
332 	/* Determine the worst case latency
333 	 * NumLinePair = Number of line pairs to request(1=2 lines, 2=4 lines)
334 	 * WorstCaseLatency = worst case time from urgent to when the MC starts
335 	 *                    to return data
336 	 * READ_DELAY_IDLE_MAX = constant of 1us
337 	 * ChunkTime = time it takes the DCP to send one chunk of data to the LB
338 	 *             which consists of pipeline delay and inter chunk gap
339 	 */
340 	if (rfixed_trunc(wm->num_line_pair) > 1) {
341 		a.full = rfixed_const(3);
342 		wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
343 		wm->worst_case_latency.full += read_delay_latency.full;
344 	} else {
345 		a.full = rfixed_const(2);
346 		wm->worst_case_latency.full = rfixed_mul(a, chunk_time);
347 		wm->worst_case_latency.full += read_delay_latency.full;
348 	}
349 
350 	/* Determine the tolerable latency
351 	 * TolerableLatency = Any given request has only 1 line time
352 	 *                    for the data to be returned
353 	 * LBRequestFifoDepth = Number of chunk requests the LB can
354 	 *                      put into the request FIFO for a display
355 	 *  LineTime = total time for one line of display
356 	 *  ChunkTime = the time it takes the DCP to send one chunk
357 	 *              of data to the LB which consists of
358 	 *  pipeline delay and inter chunk gap
359 	 */
360 	if ((2+wm->lb_request_fifo_depth) >= rfixed_trunc(request_fifo_depth)) {
361 		tolerable_latency.full = line_time.full;
362 	} else {
363 		tolerable_latency.full = rfixed_const(wm->lb_request_fifo_depth - 2);
364 		tolerable_latency.full = request_fifo_depth.full - tolerable_latency.full;
365 		tolerable_latency.full = rfixed_mul(tolerable_latency, chunk_time);
366 		tolerable_latency.full = line_time.full - tolerable_latency.full;
367 	}
368 	/* We assume worst case 32bits (4 bytes) */
369 	wm->dbpp.full = rfixed_const(4 * 8);
370 
371 	/* Determine the maximum priority mark
372 	 *  width = viewport width in pixels
373 	 */
374 	a.full = rfixed_const(16);
375 	wm->priority_mark_max.full = rfixed_const(crtc->base.mode.crtc_hdisplay);
376 	wm->priority_mark_max.full = rfixed_div(wm->priority_mark_max, a);
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 = wm->priority_mark_max.full - wm->priority_mark.full;
387 	}
388 }
389 
390 void rs690_bandwidth_update(struct radeon_device *rdev)
391 {
392 	struct drm_display_mode *mode0 = NULL;
393 	struct drm_display_mode *mode1 = NULL;
394 	struct rs690_watermark wm0;
395 	struct rs690_watermark wm1;
396 	u32 tmp;
397 	fixed20_12 priority_mark02, priority_mark12, fill_rate;
398 	fixed20_12 a, b;
399 
400 	if (rdev->mode_info.crtcs[0]->base.enabled)
401 		mode0 = &rdev->mode_info.crtcs[0]->base.mode;
402 	if (rdev->mode_info.crtcs[1]->base.enabled)
403 		mode1 = &rdev->mode_info.crtcs[1]->base.mode;
404 	/*
405 	 * Set display0/1 priority up in the memory controller for
406 	 * modes if the user specifies HIGH for displaypriority
407 	 * option.
408 	 */
409 	if (rdev->disp_priority == 2) {
410 		tmp = RREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER);
411 		tmp &= C_000104_MC_DISP0R_INIT_LAT;
412 		tmp &= C_000104_MC_DISP1R_INIT_LAT;
413 		if (mode0)
414 			tmp |= S_000104_MC_DISP0R_INIT_LAT(1);
415 		if (mode1)
416 			tmp |= S_000104_MC_DISP1R_INIT_LAT(1);
417 		WREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER, tmp);
418 	}
419 	rs690_line_buffer_adjust(rdev, mode0, mode1);
420 
421 	if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
422 		WREG32(R_006C9C_DCP_CONTROL, 0);
423 	if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
424 		WREG32(R_006C9C_DCP_CONTROL, 2);
425 
426 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0);
427 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1);
428 
429 	tmp = (wm0.lb_request_fifo_depth - 1);
430 	tmp |= (wm1.lb_request_fifo_depth - 1) << 16;
431 	WREG32(R_006D58_LB_MAX_REQ_OUTSTANDING, tmp);
432 
433 	if (mode0 && mode1) {
434 		if (rfixed_trunc(wm0.dbpp) > 64)
435 			a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
436 		else
437 			a.full = wm0.num_line_pair.full;
438 		if (rfixed_trunc(wm1.dbpp) > 64)
439 			b.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
440 		else
441 			b.full = wm1.num_line_pair.full;
442 		a.full += b.full;
443 		fill_rate.full = rfixed_div(wm0.sclk, a);
444 		if (wm0.consumption_rate.full > fill_rate.full) {
445 			b.full = wm0.consumption_rate.full - fill_rate.full;
446 			b.full = rfixed_mul(b, wm0.active_time);
447 			a.full = rfixed_mul(wm0.worst_case_latency,
448 						wm0.consumption_rate);
449 			a.full = a.full + b.full;
450 			b.full = rfixed_const(16 * 1000);
451 			priority_mark02.full = rfixed_div(a, b);
452 		} else {
453 			a.full = rfixed_mul(wm0.worst_case_latency,
454 						wm0.consumption_rate);
455 			b.full = rfixed_const(16 * 1000);
456 			priority_mark02.full = rfixed_div(a, b);
457 		}
458 		if (wm1.consumption_rate.full > fill_rate.full) {
459 			b.full = wm1.consumption_rate.full - fill_rate.full;
460 			b.full = rfixed_mul(b, wm1.active_time);
461 			a.full = rfixed_mul(wm1.worst_case_latency,
462 						wm1.consumption_rate);
463 			a.full = a.full + b.full;
464 			b.full = rfixed_const(16 * 1000);
465 			priority_mark12.full = rfixed_div(a, b);
466 		} else {
467 			a.full = rfixed_mul(wm1.worst_case_latency,
468 						wm1.consumption_rate);
469 			b.full = rfixed_const(16 * 1000);
470 			priority_mark12.full = rfixed_div(a, b);
471 		}
472 		if (wm0.priority_mark.full > priority_mark02.full)
473 			priority_mark02.full = wm0.priority_mark.full;
474 		if (rfixed_trunc(priority_mark02) < 0)
475 			priority_mark02.full = 0;
476 		if (wm0.priority_mark_max.full > priority_mark02.full)
477 			priority_mark02.full = wm0.priority_mark_max.full;
478 		if (wm1.priority_mark.full > priority_mark12.full)
479 			priority_mark12.full = wm1.priority_mark.full;
480 		if (rfixed_trunc(priority_mark12) < 0)
481 			priority_mark12.full = 0;
482 		if (wm1.priority_mark_max.full > priority_mark12.full)
483 			priority_mark12.full = wm1.priority_mark_max.full;
484 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
485 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
486 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
487 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
488 	} else if (mode0) {
489 		if (rfixed_trunc(wm0.dbpp) > 64)
490 			a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
491 		else
492 			a.full = wm0.num_line_pair.full;
493 		fill_rate.full = rfixed_div(wm0.sclk, a);
494 		if (wm0.consumption_rate.full > fill_rate.full) {
495 			b.full = wm0.consumption_rate.full - fill_rate.full;
496 			b.full = rfixed_mul(b, wm0.active_time);
497 			a.full = rfixed_mul(wm0.worst_case_latency,
498 						wm0.consumption_rate);
499 			a.full = a.full + b.full;
500 			b.full = rfixed_const(16 * 1000);
501 			priority_mark02.full = rfixed_div(a, b);
502 		} else {
503 			a.full = rfixed_mul(wm0.worst_case_latency,
504 						wm0.consumption_rate);
505 			b.full = rfixed_const(16 * 1000);
506 			priority_mark02.full = rfixed_div(a, b);
507 		}
508 		if (wm0.priority_mark.full > priority_mark02.full)
509 			priority_mark02.full = wm0.priority_mark.full;
510 		if (rfixed_trunc(priority_mark02) < 0)
511 			priority_mark02.full = 0;
512 		if (wm0.priority_mark_max.full > priority_mark02.full)
513 			priority_mark02.full = wm0.priority_mark_max.full;
514 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
515 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
516 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT,
517 			S_006D48_D2MODE_PRIORITY_A_OFF(1));
518 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT,
519 			S_006D4C_D2MODE_PRIORITY_B_OFF(1));
520 	} else {
521 		if (rfixed_trunc(wm1.dbpp) > 64)
522 			a.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
523 		else
524 			a.full = wm1.num_line_pair.full;
525 		fill_rate.full = rfixed_div(wm1.sclk, a);
526 		if (wm1.consumption_rate.full > fill_rate.full) {
527 			b.full = wm1.consumption_rate.full - fill_rate.full;
528 			b.full = rfixed_mul(b, wm1.active_time);
529 			a.full = rfixed_mul(wm1.worst_case_latency,
530 						wm1.consumption_rate);
531 			a.full = a.full + b.full;
532 			b.full = rfixed_const(16 * 1000);
533 			priority_mark12.full = rfixed_div(a, b);
534 		} else {
535 			a.full = rfixed_mul(wm1.worst_case_latency,
536 						wm1.consumption_rate);
537 			b.full = rfixed_const(16 * 1000);
538 			priority_mark12.full = rfixed_div(a, b);
539 		}
540 		if (wm1.priority_mark.full > priority_mark12.full)
541 			priority_mark12.full = wm1.priority_mark.full;
542 		if (rfixed_trunc(priority_mark12) < 0)
543 			priority_mark12.full = 0;
544 		if (wm1.priority_mark_max.full > priority_mark12.full)
545 			priority_mark12.full = wm1.priority_mark_max.full;
546 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT,
547 			S_006548_D1MODE_PRIORITY_A_OFF(1));
548 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT,
549 			S_00654C_D1MODE_PRIORITY_B_OFF(1));
550 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
551 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
552 	}
553 }
554 
555 uint32_t rs690_mc_rreg(struct radeon_device *rdev, uint32_t reg)
556 {
557 	uint32_t r;
558 
559 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg));
560 	r = RREG32(R_00007C_MC_DATA);
561 	WREG32(R_000078_MC_INDEX, ~C_000078_MC_IND_ADDR);
562 	return r;
563 }
564 
565 void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
566 {
567 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg) |
568 		S_000078_MC_IND_WR_EN(1));
569 	WREG32(R_00007C_MC_DATA, v);
570 	WREG32(R_000078_MC_INDEX, 0x7F);
571 }
572 
573 void rs690_mc_program(struct radeon_device *rdev)
574 {
575 	struct rv515_mc_save save;
576 
577 	/* Stops all mc clients */
578 	rv515_mc_stop(rdev, &save);
579 
580 	/* Wait for mc idle */
581 	if (rs690_mc_wait_for_idle(rdev))
582 		dev_warn(rdev->dev, "Wait MC idle timeout before updating MC.\n");
583 	/* Program MC, should be a 32bits limited address space */
584 	WREG32_MC(R_000100_MCCFG_FB_LOCATION,
585 			S_000100_MC_FB_START(rdev->mc.vram_start >> 16) |
586 			S_000100_MC_FB_TOP(rdev->mc.vram_end >> 16));
587 	WREG32(R_000134_HDP_FB_LOCATION,
588 		S_000134_HDP_FB_START(rdev->mc.vram_start >> 16));
589 
590 	rv515_mc_resume(rdev, &save);
591 }
592 
593 static int rs690_startup(struct radeon_device *rdev)
594 {
595 	int r;
596 
597 	rs690_mc_program(rdev);
598 	/* Resume clock */
599 	rv515_clock_startup(rdev);
600 	/* Initialize GPU configuration (# pipes, ...) */
601 	rs690_gpu_init(rdev);
602 	/* Initialize GART (initialize after TTM so we can allocate
603 	 * memory through TTM but finalize after TTM) */
604 	r = rs400_gart_enable(rdev);
605 	if (r)
606 		return r;
607 	/* Enable IRQ */
608 	rs600_irq_set(rdev);
609 	/* 1M ring buffer */
610 	r = r100_cp_init(rdev, 1024 * 1024);
611 	if (r) {
612 		dev_err(rdev->dev, "failled initializing CP (%d).\n", r);
613 		return r;
614 	}
615 	r = r100_wb_init(rdev);
616 	if (r)
617 		dev_err(rdev->dev, "failled initializing WB (%d).\n", r);
618 	r = r100_ib_init(rdev);
619 	if (r) {
620 		dev_err(rdev->dev, "failled initializing IB (%d).\n", r);
621 		return r;
622 	}
623 	return 0;
624 }
625 
626 int rs690_resume(struct radeon_device *rdev)
627 {
628 	/* Make sur GART are not working */
629 	rs400_gart_disable(rdev);
630 	/* Resume clock before doing reset */
631 	rv515_clock_startup(rdev);
632 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
633 	if (radeon_gpu_reset(rdev)) {
634 		dev_warn(rdev->dev, "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
635 			RREG32(R_000E40_RBBM_STATUS),
636 			RREG32(R_0007C0_CP_STAT));
637 	}
638 	/* post */
639 	atom_asic_init(rdev->mode_info.atom_context);
640 	/* Resume clock after posting */
641 	rv515_clock_startup(rdev);
642 	return rs690_startup(rdev);
643 }
644 
645 int rs690_suspend(struct radeon_device *rdev)
646 {
647 	r100_cp_disable(rdev);
648 	r100_wb_disable(rdev);
649 	rs600_irq_disable(rdev);
650 	rs400_gart_disable(rdev);
651 	return 0;
652 }
653 
654 void rs690_fini(struct radeon_device *rdev)
655 {
656 	rs690_suspend(rdev);
657 	r100_cp_fini(rdev);
658 	r100_wb_fini(rdev);
659 	r100_ib_fini(rdev);
660 	radeon_gem_fini(rdev);
661 	rs400_gart_fini(rdev);
662 	radeon_irq_kms_fini(rdev);
663 	radeon_fence_driver_fini(rdev);
664 	radeon_object_fini(rdev);
665 	radeon_atombios_fini(rdev);
666 	kfree(rdev->bios);
667 	rdev->bios = NULL;
668 }
669 
670 int rs690_init(struct radeon_device *rdev)
671 {
672 	int r;
673 
674 	/* Disable VGA */
675 	rv515_vga_render_disable(rdev);
676 	/* Initialize scratch registers */
677 	radeon_scratch_init(rdev);
678 	/* Initialize surface registers */
679 	radeon_surface_init(rdev);
680 	/* TODO: disable VGA need to use VGA request */
681 	/* BIOS*/
682 	if (!radeon_get_bios(rdev)) {
683 		if (ASIC_IS_AVIVO(rdev))
684 			return -EINVAL;
685 	}
686 	if (rdev->is_atom_bios) {
687 		r = radeon_atombios_init(rdev);
688 		if (r)
689 			return r;
690 	} else {
691 		dev_err(rdev->dev, "Expecting atombios for RV515 GPU\n");
692 		return -EINVAL;
693 	}
694 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
695 	if (radeon_gpu_reset(rdev)) {
696 		dev_warn(rdev->dev,
697 			"GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
698 			RREG32(R_000E40_RBBM_STATUS),
699 			RREG32(R_0007C0_CP_STAT));
700 	}
701 	/* check if cards are posted or not */
702 	if (radeon_boot_test_post_card(rdev) == false)
703 		return -EINVAL;
704 
705 	/* Initialize clocks */
706 	radeon_get_clock_info(rdev->ddev);
707 	/* Initialize power management */
708 	radeon_pm_init(rdev);
709 	/* Get vram informations */
710 	rs690_vram_info(rdev);
711 	/* Initialize memory controller (also test AGP) */
712 	r = r420_mc_init(rdev);
713 	if (r)
714 		return r;
715 	rv515_debugfs(rdev);
716 	/* Fence driver */
717 	r = radeon_fence_driver_init(rdev);
718 	if (r)
719 		return r;
720 	r = radeon_irq_kms_init(rdev);
721 	if (r)
722 		return r;
723 	/* Memory manager */
724 	r = radeon_object_init(rdev);
725 	if (r)
726 		return r;
727 	r = rs400_gart_init(rdev);
728 	if (r)
729 		return r;
730 	rs600_set_safe_registers(rdev);
731 	rdev->accel_working = true;
732 	r = rs690_startup(rdev);
733 	if (r) {
734 		/* Somethings want wront with the accel init stop accel */
735 		dev_err(rdev->dev, "Disabling GPU acceleration\n");
736 		rs690_suspend(rdev);
737 		r100_cp_fini(rdev);
738 		r100_wb_fini(rdev);
739 		r100_ib_fini(rdev);
740 		rs400_gart_fini(rdev);
741 		radeon_irq_kms_fini(rdev);
742 		rdev->accel_working = false;
743 	}
744 	return 0;
745 }
746