xref: /linux/drivers/gpu/drm/radeon/rs690.c (revision e6990375ef4ec449994991034238f1ffab8a3a1a)
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 static 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: HDP same place on rs690 ? */
52 	r100_hdp_reset(rdev);
53 	/* FIXME: is this correct ? */
54 	r420_pipes_init(rdev);
55 	if (rs690_mc_wait_for_idle(rdev)) {
56 		printk(KERN_WARNING "Failed to wait MC idle while "
57 		       "programming pipes. Bad things might happen.\n");
58 	}
59 }
60 
61 void rs690_pm_info(struct radeon_device *rdev)
62 {
63 	int index = GetIndexIntoMasterTable(DATA, IntegratedSystemInfo);
64 	struct _ATOM_INTEGRATED_SYSTEM_INFO *info;
65 	struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 *info_v2;
66 	void *ptr;
67 	uint16_t data_offset;
68 	uint8_t frev, crev;
69 	fixed20_12 tmp;
70 
71 	atom_parse_data_header(rdev->mode_info.atom_context, index, NULL,
72 			       &frev, &crev, &data_offset);
73 	ptr = rdev->mode_info.atom_context->bios + data_offset;
74 	info = (struct _ATOM_INTEGRATED_SYSTEM_INFO *)ptr;
75 	info_v2 = (struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 *)ptr;
76 	/* Get various system informations from bios */
77 	switch (crev) {
78 	case 1:
79 		tmp.full = rfixed_const(100);
80 		rdev->pm.igp_sideport_mclk.full = rfixed_const(info->ulBootUpMemoryClock);
81 		rdev->pm.igp_sideport_mclk.full = rfixed_div(rdev->pm.igp_sideport_mclk, tmp);
82 		rdev->pm.igp_system_mclk.full = rfixed_const(le16_to_cpu(info->usK8MemoryClock));
83 		rdev->pm.igp_ht_link_clk.full = rfixed_const(le16_to_cpu(info->usFSBClock));
84 		rdev->pm.igp_ht_link_width.full = rfixed_const(info->ucHTLinkWidth);
85 		break;
86 	case 2:
87 		tmp.full = rfixed_const(100);
88 		rdev->pm.igp_sideport_mclk.full = rfixed_const(info_v2->ulBootUpSidePortClock);
89 		rdev->pm.igp_sideport_mclk.full = rfixed_div(rdev->pm.igp_sideport_mclk, tmp);
90 		rdev->pm.igp_system_mclk.full = rfixed_const(info_v2->ulBootUpUMAClock);
91 		rdev->pm.igp_system_mclk.full = rfixed_div(rdev->pm.igp_system_mclk, tmp);
92 		rdev->pm.igp_ht_link_clk.full = rfixed_const(info_v2->ulHTLinkFreq);
93 		rdev->pm.igp_ht_link_clk.full = rfixed_div(rdev->pm.igp_ht_link_clk, tmp);
94 		rdev->pm.igp_ht_link_width.full = rfixed_const(le16_to_cpu(info_v2->usMinHTLinkWidth));
95 		break;
96 	default:
97 		tmp.full = rfixed_const(100);
98 		/* We assume the slower possible clock ie worst case */
99 		/* DDR 333Mhz */
100 		rdev->pm.igp_sideport_mclk.full = rfixed_const(333);
101 		/* FIXME: system clock ? */
102 		rdev->pm.igp_system_mclk.full = rfixed_const(100);
103 		rdev->pm.igp_system_mclk.full = rfixed_div(rdev->pm.igp_system_mclk, tmp);
104 		rdev->pm.igp_ht_link_clk.full = rfixed_const(200);
105 		rdev->pm.igp_ht_link_width.full = rfixed_const(8);
106 		DRM_ERROR("No integrated system info for your GPU, using safe default\n");
107 		break;
108 	}
109 	/* Compute various bandwidth */
110 	/* k8_bandwidth = (memory_clk / 2) * 2 * 8 * 0.5 = memory_clk * 4  */
111 	tmp.full = rfixed_const(4);
112 	rdev->pm.k8_bandwidth.full = rfixed_mul(rdev->pm.igp_system_mclk, tmp);
113 	/* ht_bandwidth = ht_clk * 2 * ht_width / 8 * 0.8
114 	 *              = ht_clk * ht_width / 5
115 	 */
116 	tmp.full = rfixed_const(5);
117 	rdev->pm.ht_bandwidth.full = rfixed_mul(rdev->pm.igp_ht_link_clk,
118 						rdev->pm.igp_ht_link_width);
119 	rdev->pm.ht_bandwidth.full = rfixed_div(rdev->pm.ht_bandwidth, tmp);
120 	if (tmp.full < rdev->pm.max_bandwidth.full) {
121 		/* HT link is a limiting factor */
122 		rdev->pm.max_bandwidth.full = tmp.full;
123 	}
124 	/* sideport_bandwidth = (sideport_clk / 2) * 2 * 2 * 0.7
125 	 *                    = (sideport_clk * 14) / 10
126 	 */
127 	tmp.full = rfixed_const(14);
128 	rdev->pm.sideport_bandwidth.full = rfixed_mul(rdev->pm.igp_sideport_mclk, tmp);
129 	tmp.full = rfixed_const(10);
130 	rdev->pm.sideport_bandwidth.full = rfixed_div(rdev->pm.sideport_bandwidth, tmp);
131 }
132 
133 void rs690_mc_init(struct radeon_device *rdev)
134 {
135 	fixed20_12 a;
136 	u64 base;
137 
138 	rs400_gart_adjust_size(rdev);
139 	rdev->mc.vram_is_ddr = true;
140 	rdev->mc.vram_width = 128;
141 	rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
142 	rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
143 	rdev->mc.aper_base = drm_get_resource_start(rdev->ddev, 0);
144 	rdev->mc.aper_size = drm_get_resource_len(rdev->ddev, 0);
145 	rdev->mc.visible_vram_size = rdev->mc.aper_size;
146 	base = RREG32_MC(R_000100_MCCFG_FB_LOCATION);
147 	base = G_000100_MC_FB_START(base) << 16;
148 	rs690_pm_info(rdev);
149 	/* FIXME: we should enforce default clock in case GPU is not in
150 	 * default setup
151 	 */
152 	a.full = rfixed_const(100);
153 	rdev->pm.sclk.full = rfixed_const(rdev->clock.default_sclk);
154 	rdev->pm.sclk.full = rfixed_div(rdev->pm.sclk, a);
155 	a.full = rfixed_const(16);
156 	/* core_bandwidth = sclk(Mhz) * 16 */
157 	rdev->pm.core_bandwidth.full = rfixed_div(rdev->pm.sclk, a);
158 	rdev->mc.igp_sideport_enabled = radeon_atombios_sideport_present(rdev);
159 	radeon_vram_location(rdev, &rdev->mc, base);
160 	radeon_gtt_location(rdev, &rdev->mc);
161 }
162 
163 void rs690_line_buffer_adjust(struct radeon_device *rdev,
164 			      struct drm_display_mode *mode1,
165 			      struct drm_display_mode *mode2)
166 {
167 	u32 tmp;
168 
169 	/*
170 	 * Line Buffer Setup
171 	 * There is a single line buffer shared by both display controllers.
172 	 * R_006520_DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
173 	 * the display controllers.  The paritioning can either be done
174 	 * manually or via one of four preset allocations specified in bits 1:0:
175 	 *  0 - line buffer is divided in half and shared between crtc
176 	 *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
177 	 *  2 - D1 gets the whole buffer
178 	 *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
179 	 * Setting bit 2 of R_006520_DC_LB_MEMORY_SPLIT controls switches to manual
180 	 * allocation mode. In manual allocation mode, D1 always starts at 0,
181 	 * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
182 	 */
183 	tmp = RREG32(R_006520_DC_LB_MEMORY_SPLIT) & C_006520_DC_LB_MEMORY_SPLIT;
184 	tmp &= ~C_006520_DC_LB_MEMORY_SPLIT_MODE;
185 	/* auto */
186 	if (mode1 && mode2) {
187 		if (mode1->hdisplay > mode2->hdisplay) {
188 			if (mode1->hdisplay > 2560)
189 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
190 			else
191 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
192 		} else if (mode2->hdisplay > mode1->hdisplay) {
193 			if (mode2->hdisplay > 2560)
194 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
195 			else
196 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
197 		} else
198 			tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
199 	} else if (mode1) {
200 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_ONLY;
201 	} else if (mode2) {
202 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
203 	}
204 	WREG32(R_006520_DC_LB_MEMORY_SPLIT, tmp);
205 }
206 
207 struct rs690_watermark {
208 	u32        lb_request_fifo_depth;
209 	fixed20_12 num_line_pair;
210 	fixed20_12 estimated_width;
211 	fixed20_12 worst_case_latency;
212 	fixed20_12 consumption_rate;
213 	fixed20_12 active_time;
214 	fixed20_12 dbpp;
215 	fixed20_12 priority_mark_max;
216 	fixed20_12 priority_mark;
217 	fixed20_12 sclk;
218 };
219 
220 void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
221 				  struct radeon_crtc *crtc,
222 				  struct rs690_watermark *wm)
223 {
224 	struct drm_display_mode *mode = &crtc->base.mode;
225 	fixed20_12 a, b, c;
226 	fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
227 	fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
228 	/* FIXME: detect IGP with sideport memory, i don't think there is any
229 	 * such product available
230 	 */
231 	bool sideport = false;
232 
233 	if (!crtc->base.enabled) {
234 		/* FIXME: wouldn't it better to set priority mark to maximum */
235 		wm->lb_request_fifo_depth = 4;
236 		return;
237 	}
238 
239 	if (crtc->vsc.full > rfixed_const(2))
240 		wm->num_line_pair.full = rfixed_const(2);
241 	else
242 		wm->num_line_pair.full = rfixed_const(1);
243 
244 	b.full = rfixed_const(mode->crtc_hdisplay);
245 	c.full = rfixed_const(256);
246 	a.full = rfixed_div(b, c);
247 	request_fifo_depth.full = rfixed_mul(a, wm->num_line_pair);
248 	request_fifo_depth.full = rfixed_ceil(request_fifo_depth);
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 	wm->priority_mark_max.full = rfixed_ceil(wm->priority_mark_max);
378 
379 	/* Determine estimated width */
380 	estimated_width.full = tolerable_latency.full - wm->worst_case_latency.full;
381 	estimated_width.full = rfixed_div(estimated_width, consumption_time);
382 	if (rfixed_trunc(estimated_width) > crtc->base.mode.crtc_hdisplay) {
383 		wm->priority_mark.full = rfixed_const(10);
384 	} else {
385 		a.full = rfixed_const(16);
386 		wm->priority_mark.full = rfixed_div(estimated_width, a);
387 		wm->priority_mark.full = rfixed_ceil(wm->priority_mark);
388 		wm->priority_mark.full = wm->priority_mark_max.full - wm->priority_mark.full;
389 	}
390 }
391 
392 void rs690_bandwidth_update(struct radeon_device *rdev)
393 {
394 	struct drm_display_mode *mode0 = NULL;
395 	struct drm_display_mode *mode1 = NULL;
396 	struct rs690_watermark wm0;
397 	struct rs690_watermark wm1;
398 	u32 tmp;
399 	fixed20_12 priority_mark02, priority_mark12, fill_rate;
400 	fixed20_12 a, b;
401 
402 	if (rdev->mode_info.crtcs[0]->base.enabled)
403 		mode0 = &rdev->mode_info.crtcs[0]->base.mode;
404 	if (rdev->mode_info.crtcs[1]->base.enabled)
405 		mode1 = &rdev->mode_info.crtcs[1]->base.mode;
406 	/*
407 	 * Set display0/1 priority up in the memory controller for
408 	 * modes if the user specifies HIGH for displaypriority
409 	 * option.
410 	 */
411 	if (rdev->disp_priority == 2) {
412 		tmp = RREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER);
413 		tmp &= C_000104_MC_DISP0R_INIT_LAT;
414 		tmp &= C_000104_MC_DISP1R_INIT_LAT;
415 		if (mode0)
416 			tmp |= S_000104_MC_DISP0R_INIT_LAT(1);
417 		if (mode1)
418 			tmp |= S_000104_MC_DISP1R_INIT_LAT(1);
419 		WREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER, tmp);
420 	}
421 	rs690_line_buffer_adjust(rdev, mode0, mode1);
422 
423 	if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
424 		WREG32(R_006C9C_DCP_CONTROL, 0);
425 	if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
426 		WREG32(R_006C9C_DCP_CONTROL, 2);
427 
428 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0);
429 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1);
430 
431 	tmp = (wm0.lb_request_fifo_depth - 1);
432 	tmp |= (wm1.lb_request_fifo_depth - 1) << 16;
433 	WREG32(R_006D58_LB_MAX_REQ_OUTSTANDING, tmp);
434 
435 	if (mode0 && mode1) {
436 		if (rfixed_trunc(wm0.dbpp) > 64)
437 			a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
438 		else
439 			a.full = wm0.num_line_pair.full;
440 		if (rfixed_trunc(wm1.dbpp) > 64)
441 			b.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
442 		else
443 			b.full = wm1.num_line_pair.full;
444 		a.full += b.full;
445 		fill_rate.full = rfixed_div(wm0.sclk, a);
446 		if (wm0.consumption_rate.full > fill_rate.full) {
447 			b.full = wm0.consumption_rate.full - fill_rate.full;
448 			b.full = rfixed_mul(b, wm0.active_time);
449 			a.full = rfixed_mul(wm0.worst_case_latency,
450 						wm0.consumption_rate);
451 			a.full = a.full + b.full;
452 			b.full = rfixed_const(16 * 1000);
453 			priority_mark02.full = rfixed_div(a, b);
454 		} else {
455 			a.full = rfixed_mul(wm0.worst_case_latency,
456 						wm0.consumption_rate);
457 			b.full = rfixed_const(16 * 1000);
458 			priority_mark02.full = rfixed_div(a, b);
459 		}
460 		if (wm1.consumption_rate.full > fill_rate.full) {
461 			b.full = wm1.consumption_rate.full - fill_rate.full;
462 			b.full = rfixed_mul(b, wm1.active_time);
463 			a.full = rfixed_mul(wm1.worst_case_latency,
464 						wm1.consumption_rate);
465 			a.full = a.full + b.full;
466 			b.full = rfixed_const(16 * 1000);
467 			priority_mark12.full = rfixed_div(a, b);
468 		} else {
469 			a.full = rfixed_mul(wm1.worst_case_latency,
470 						wm1.consumption_rate);
471 			b.full = rfixed_const(16 * 1000);
472 			priority_mark12.full = rfixed_div(a, b);
473 		}
474 		if (wm0.priority_mark.full > priority_mark02.full)
475 			priority_mark02.full = wm0.priority_mark.full;
476 		if (rfixed_trunc(priority_mark02) < 0)
477 			priority_mark02.full = 0;
478 		if (wm0.priority_mark_max.full > priority_mark02.full)
479 			priority_mark02.full = wm0.priority_mark_max.full;
480 		if (wm1.priority_mark.full > priority_mark12.full)
481 			priority_mark12.full = wm1.priority_mark.full;
482 		if (rfixed_trunc(priority_mark12) < 0)
483 			priority_mark12.full = 0;
484 		if (wm1.priority_mark_max.full > priority_mark12.full)
485 			priority_mark12.full = wm1.priority_mark_max.full;
486 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
487 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
488 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
489 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
490 	} else if (mode0) {
491 		if (rfixed_trunc(wm0.dbpp) > 64)
492 			a.full = rfixed_mul(wm0.dbpp, wm0.num_line_pair);
493 		else
494 			a.full = wm0.num_line_pair.full;
495 		fill_rate.full = rfixed_div(wm0.sclk, a);
496 		if (wm0.consumption_rate.full > fill_rate.full) {
497 			b.full = wm0.consumption_rate.full - fill_rate.full;
498 			b.full = rfixed_mul(b, wm0.active_time);
499 			a.full = rfixed_mul(wm0.worst_case_latency,
500 						wm0.consumption_rate);
501 			a.full = a.full + b.full;
502 			b.full = rfixed_const(16 * 1000);
503 			priority_mark02.full = rfixed_div(a, b);
504 		} else {
505 			a.full = rfixed_mul(wm0.worst_case_latency,
506 						wm0.consumption_rate);
507 			b.full = rfixed_const(16 * 1000);
508 			priority_mark02.full = rfixed_div(a, b);
509 		}
510 		if (wm0.priority_mark.full > priority_mark02.full)
511 			priority_mark02.full = wm0.priority_mark.full;
512 		if (rfixed_trunc(priority_mark02) < 0)
513 			priority_mark02.full = 0;
514 		if (wm0.priority_mark_max.full > priority_mark02.full)
515 			priority_mark02.full = wm0.priority_mark_max.full;
516 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark02));
517 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark02));
518 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT,
519 			S_006D48_D2MODE_PRIORITY_A_OFF(1));
520 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT,
521 			S_006D4C_D2MODE_PRIORITY_B_OFF(1));
522 	} else {
523 		if (rfixed_trunc(wm1.dbpp) > 64)
524 			a.full = rfixed_mul(wm1.dbpp, wm1.num_line_pair);
525 		else
526 			a.full = wm1.num_line_pair.full;
527 		fill_rate.full = rfixed_div(wm1.sclk, a);
528 		if (wm1.consumption_rate.full > fill_rate.full) {
529 			b.full = wm1.consumption_rate.full - fill_rate.full;
530 			b.full = rfixed_mul(b, wm1.active_time);
531 			a.full = rfixed_mul(wm1.worst_case_latency,
532 						wm1.consumption_rate);
533 			a.full = a.full + b.full;
534 			b.full = rfixed_const(16 * 1000);
535 			priority_mark12.full = rfixed_div(a, b);
536 		} else {
537 			a.full = rfixed_mul(wm1.worst_case_latency,
538 						wm1.consumption_rate);
539 			b.full = rfixed_const(16 * 1000);
540 			priority_mark12.full = rfixed_div(a, b);
541 		}
542 		if (wm1.priority_mark.full > priority_mark12.full)
543 			priority_mark12.full = wm1.priority_mark.full;
544 		if (rfixed_trunc(priority_mark12) < 0)
545 			priority_mark12.full = 0;
546 		if (wm1.priority_mark_max.full > priority_mark12.full)
547 			priority_mark12.full = wm1.priority_mark_max.full;
548 		WREG32(R_006548_D1MODE_PRIORITY_A_CNT,
549 			S_006548_D1MODE_PRIORITY_A_OFF(1));
550 		WREG32(R_00654C_D1MODE_PRIORITY_B_CNT,
551 			S_00654C_D1MODE_PRIORITY_B_OFF(1));
552 		WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, rfixed_trunc(priority_mark12));
553 		WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, rfixed_trunc(priority_mark12));
554 	}
555 }
556 
557 uint32_t rs690_mc_rreg(struct radeon_device *rdev, uint32_t reg)
558 {
559 	uint32_t r;
560 
561 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg));
562 	r = RREG32(R_00007C_MC_DATA);
563 	WREG32(R_000078_MC_INDEX, ~C_000078_MC_IND_ADDR);
564 	return r;
565 }
566 
567 void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
568 {
569 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg) |
570 		S_000078_MC_IND_WR_EN(1));
571 	WREG32(R_00007C_MC_DATA, v);
572 	WREG32(R_000078_MC_INDEX, 0x7F);
573 }
574 
575 void rs690_mc_program(struct radeon_device *rdev)
576 {
577 	struct rv515_mc_save save;
578 
579 	/* Stops all mc clients */
580 	rv515_mc_stop(rdev, &save);
581 
582 	/* Wait for mc idle */
583 	if (rs690_mc_wait_for_idle(rdev))
584 		dev_warn(rdev->dev, "Wait MC idle timeout before updating MC.\n");
585 	/* Program MC, should be a 32bits limited address space */
586 	WREG32_MC(R_000100_MCCFG_FB_LOCATION,
587 			S_000100_MC_FB_START(rdev->mc.vram_start >> 16) |
588 			S_000100_MC_FB_TOP(rdev->mc.vram_end >> 16));
589 	WREG32(R_000134_HDP_FB_LOCATION,
590 		S_000134_HDP_FB_START(rdev->mc.vram_start >> 16));
591 
592 	rv515_mc_resume(rdev, &save);
593 }
594 
595 static int rs690_startup(struct radeon_device *rdev)
596 {
597 	int r;
598 
599 	rs690_mc_program(rdev);
600 	/* Resume clock */
601 	rv515_clock_startup(rdev);
602 	/* Initialize GPU configuration (# pipes, ...) */
603 	rs690_gpu_init(rdev);
604 	/* Initialize GART (initialize after TTM so we can allocate
605 	 * memory through TTM but finalize after TTM) */
606 	r = rs400_gart_enable(rdev);
607 	if (r)
608 		return r;
609 	/* Enable IRQ */
610 	rs600_irq_set(rdev);
611 	rdev->config.r300.hdp_cntl = RREG32(RADEON_HOST_PATH_CNTL);
612 	/* 1M ring buffer */
613 	r = r100_cp_init(rdev, 1024 * 1024);
614 	if (r) {
615 		dev_err(rdev->dev, "failled initializing CP (%d).\n", r);
616 		return r;
617 	}
618 	r = r100_wb_init(rdev);
619 	if (r)
620 		dev_err(rdev->dev, "failled initializing WB (%d).\n", r);
621 	r = r100_ib_init(rdev);
622 	if (r) {
623 		dev_err(rdev->dev, "failled initializing IB (%d).\n", r);
624 		return r;
625 	}
626 	return 0;
627 }
628 
629 int rs690_resume(struct radeon_device *rdev)
630 {
631 	/* Make sur GART are not working */
632 	rs400_gart_disable(rdev);
633 	/* Resume clock before doing reset */
634 	rv515_clock_startup(rdev);
635 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
636 	if (radeon_gpu_reset(rdev)) {
637 		dev_warn(rdev->dev, "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
638 			RREG32(R_000E40_RBBM_STATUS),
639 			RREG32(R_0007C0_CP_STAT));
640 	}
641 	/* post */
642 	atom_asic_init(rdev->mode_info.atom_context);
643 	/* Resume clock after posting */
644 	rv515_clock_startup(rdev);
645 	/* Initialize surface registers */
646 	radeon_surface_init(rdev);
647 	return rs690_startup(rdev);
648 }
649 
650 int rs690_suspend(struct radeon_device *rdev)
651 {
652 	r100_cp_disable(rdev);
653 	r100_wb_disable(rdev);
654 	rs600_irq_disable(rdev);
655 	rs400_gart_disable(rdev);
656 	return 0;
657 }
658 
659 void rs690_fini(struct radeon_device *rdev)
660 {
661 	r100_cp_fini(rdev);
662 	r100_wb_fini(rdev);
663 	r100_ib_fini(rdev);
664 	radeon_gem_fini(rdev);
665 	rs400_gart_fini(rdev);
666 	radeon_irq_kms_fini(rdev);
667 	radeon_fence_driver_fini(rdev);
668 	radeon_bo_fini(rdev);
669 	radeon_atombios_fini(rdev);
670 	kfree(rdev->bios);
671 	rdev->bios = NULL;
672 }
673 
674 int rs690_init(struct radeon_device *rdev)
675 {
676 	int r;
677 
678 	/* Disable VGA */
679 	rv515_vga_render_disable(rdev);
680 	/* Initialize scratch registers */
681 	radeon_scratch_init(rdev);
682 	/* Initialize surface registers */
683 	radeon_surface_init(rdev);
684 	/* TODO: disable VGA need to use VGA request */
685 	/* BIOS*/
686 	if (!radeon_get_bios(rdev)) {
687 		if (ASIC_IS_AVIVO(rdev))
688 			return -EINVAL;
689 	}
690 	if (rdev->is_atom_bios) {
691 		r = radeon_atombios_init(rdev);
692 		if (r)
693 			return r;
694 	} else {
695 		dev_err(rdev->dev, "Expecting atombios for RV515 GPU\n");
696 		return -EINVAL;
697 	}
698 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
699 	if (radeon_gpu_reset(rdev)) {
700 		dev_warn(rdev->dev,
701 			"GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
702 			RREG32(R_000E40_RBBM_STATUS),
703 			RREG32(R_0007C0_CP_STAT));
704 	}
705 	/* check if cards are posted or not */
706 	if (radeon_boot_test_post_card(rdev) == false)
707 		return -EINVAL;
708 
709 	/* Initialize clocks */
710 	radeon_get_clock_info(rdev->ddev);
711 	/* Initialize power management */
712 	radeon_pm_init(rdev);
713 	/* initialize memory controller */
714 	rs690_mc_init(rdev);
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_bo_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 		r100_cp_fini(rdev);
737 		r100_wb_fini(rdev);
738 		r100_ib_fini(rdev);
739 		rs400_gart_fini(rdev);
740 		radeon_irq_kms_fini(rdev);
741 		rdev->accel_working = false;
742 	}
743 	return 0;
744 }
745