xref: /openbmc/linux/drivers/gpu/drm/radeon/rs690.c (revision f7777dcc)
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 <drm/drmP.h>
29 #include "radeon.h"
30 #include "radeon_asic.h"
31 #include "atom.h"
32 #include "rs690d.h"
33 
34 int rs690_mc_wait_for_idle(struct radeon_device *rdev)
35 {
36 	unsigned i;
37 	uint32_t tmp;
38 
39 	for (i = 0; i < rdev->usec_timeout; i++) {
40 		/* read MC_STATUS */
41 		tmp = RREG32_MC(R_000090_MC_SYSTEM_STATUS);
42 		if (G_000090_MC_SYSTEM_IDLE(tmp))
43 			return 0;
44 		udelay(1);
45 	}
46 	return -1;
47 }
48 
49 static void rs690_gpu_init(struct radeon_device *rdev)
50 {
51 	/* FIXME: is this correct ? */
52 	r420_pipes_init(rdev);
53 	if (rs690_mc_wait_for_idle(rdev)) {
54 		printk(KERN_WARNING "Failed to wait MC idle while "
55 		       "programming pipes. Bad things might happen.\n");
56 	}
57 }
58 
59 union igp_info {
60 	struct _ATOM_INTEGRATED_SYSTEM_INFO info;
61 	struct _ATOM_INTEGRATED_SYSTEM_INFO_V2 info_v2;
62 };
63 
64 void rs690_pm_info(struct radeon_device *rdev)
65 {
66 	int index = GetIndexIntoMasterTable(DATA, IntegratedSystemInfo);
67 	union igp_info *info;
68 	uint16_t data_offset;
69 	uint8_t frev, crev;
70 	fixed20_12 tmp;
71 
72 	if (atom_parse_data_header(rdev->mode_info.atom_context, index, NULL,
73 				   &frev, &crev, &data_offset)) {
74 		info = (union igp_info *)(rdev->mode_info.atom_context->bios + data_offset);
75 
76 		/* Get various system informations from bios */
77 		switch (crev) {
78 		case 1:
79 			tmp.full = dfixed_const(100);
80 			rdev->pm.igp_sideport_mclk.full = dfixed_const(le32_to_cpu(info->info.ulBootUpMemoryClock));
81 			rdev->pm.igp_sideport_mclk.full = dfixed_div(rdev->pm.igp_sideport_mclk, tmp);
82 			if (le16_to_cpu(info->info.usK8MemoryClock))
83 				rdev->pm.igp_system_mclk.full = dfixed_const(le16_to_cpu(info->info.usK8MemoryClock));
84 			else if (rdev->clock.default_mclk) {
85 				rdev->pm.igp_system_mclk.full = dfixed_const(rdev->clock.default_mclk);
86 				rdev->pm.igp_system_mclk.full = dfixed_div(rdev->pm.igp_system_mclk, tmp);
87 			} else
88 				rdev->pm.igp_system_mclk.full = dfixed_const(400);
89 			rdev->pm.igp_ht_link_clk.full = dfixed_const(le16_to_cpu(info->info.usFSBClock));
90 			rdev->pm.igp_ht_link_width.full = dfixed_const(info->info.ucHTLinkWidth);
91 			break;
92 		case 2:
93 			tmp.full = dfixed_const(100);
94 			rdev->pm.igp_sideport_mclk.full = dfixed_const(le32_to_cpu(info->info_v2.ulBootUpSidePortClock));
95 			rdev->pm.igp_sideport_mclk.full = dfixed_div(rdev->pm.igp_sideport_mclk, tmp);
96 			if (le32_to_cpu(info->info_v2.ulBootUpUMAClock))
97 				rdev->pm.igp_system_mclk.full = dfixed_const(le32_to_cpu(info->info_v2.ulBootUpUMAClock));
98 			else if (rdev->clock.default_mclk)
99 				rdev->pm.igp_system_mclk.full = dfixed_const(rdev->clock.default_mclk);
100 			else
101 				rdev->pm.igp_system_mclk.full = dfixed_const(66700);
102 			rdev->pm.igp_system_mclk.full = dfixed_div(rdev->pm.igp_system_mclk, tmp);
103 			rdev->pm.igp_ht_link_clk.full = dfixed_const(le32_to_cpu(info->info_v2.ulHTLinkFreq));
104 			rdev->pm.igp_ht_link_clk.full = dfixed_div(rdev->pm.igp_ht_link_clk, tmp);
105 			rdev->pm.igp_ht_link_width.full = dfixed_const(le16_to_cpu(info->info_v2.usMinHTLinkWidth));
106 			break;
107 		default:
108 			/* We assume the slower possible clock ie worst case */
109 			rdev->pm.igp_sideport_mclk.full = dfixed_const(200);
110 			rdev->pm.igp_system_mclk.full = dfixed_const(200);
111 			rdev->pm.igp_ht_link_clk.full = dfixed_const(1000);
112 			rdev->pm.igp_ht_link_width.full = dfixed_const(8);
113 			DRM_ERROR("No integrated system info for your GPU, using safe default\n");
114 			break;
115 		}
116 	} else {
117 		/* We assume the slower possible clock ie worst case */
118 		rdev->pm.igp_sideport_mclk.full = dfixed_const(200);
119 		rdev->pm.igp_system_mclk.full = dfixed_const(200);
120 		rdev->pm.igp_ht_link_clk.full = dfixed_const(1000);
121 		rdev->pm.igp_ht_link_width.full = dfixed_const(8);
122 		DRM_ERROR("No integrated system info for your GPU, using safe default\n");
123 	}
124 	/* Compute various bandwidth */
125 	/* k8_bandwidth = (memory_clk / 2) * 2 * 8 * 0.5 = memory_clk * 4  */
126 	tmp.full = dfixed_const(4);
127 	rdev->pm.k8_bandwidth.full = dfixed_mul(rdev->pm.igp_system_mclk, tmp);
128 	/* ht_bandwidth = ht_clk * 2 * ht_width / 8 * 0.8
129 	 *              = ht_clk * ht_width / 5
130 	 */
131 	tmp.full = dfixed_const(5);
132 	rdev->pm.ht_bandwidth.full = dfixed_mul(rdev->pm.igp_ht_link_clk,
133 						rdev->pm.igp_ht_link_width);
134 	rdev->pm.ht_bandwidth.full = dfixed_div(rdev->pm.ht_bandwidth, tmp);
135 	if (tmp.full < rdev->pm.max_bandwidth.full) {
136 		/* HT link is a limiting factor */
137 		rdev->pm.max_bandwidth.full = tmp.full;
138 	}
139 	/* sideport_bandwidth = (sideport_clk / 2) * 2 * 2 * 0.7
140 	 *                    = (sideport_clk * 14) / 10
141 	 */
142 	tmp.full = dfixed_const(14);
143 	rdev->pm.sideport_bandwidth.full = dfixed_mul(rdev->pm.igp_sideport_mclk, tmp);
144 	tmp.full = dfixed_const(10);
145 	rdev->pm.sideport_bandwidth.full = dfixed_div(rdev->pm.sideport_bandwidth, tmp);
146 }
147 
148 static void rs690_mc_init(struct radeon_device *rdev)
149 {
150 	u64 base;
151 	uint32_t h_addr, l_addr;
152 	unsigned long long k8_addr;
153 
154 	rs400_gart_adjust_size(rdev);
155 	rdev->mc.vram_is_ddr = true;
156 	rdev->mc.vram_width = 128;
157 	rdev->mc.real_vram_size = RREG32(RADEON_CONFIG_MEMSIZE);
158 	rdev->mc.mc_vram_size = rdev->mc.real_vram_size;
159 	rdev->mc.aper_base = pci_resource_start(rdev->pdev, 0);
160 	rdev->mc.aper_size = pci_resource_len(rdev->pdev, 0);
161 	rdev->mc.visible_vram_size = rdev->mc.aper_size;
162 	base = RREG32_MC(R_000100_MCCFG_FB_LOCATION);
163 	base = G_000100_MC_FB_START(base) << 16;
164 	rdev->mc.igp_sideport_enabled = radeon_atombios_sideport_present(rdev);
165 
166 	/* Use K8 direct mapping for fast fb access. */
167 	rdev->fastfb_working = false;
168 	h_addr = G_00005F_K8_ADDR_EXT(RREG32_MC(R_00005F_MC_MISC_UMA_CNTL));
169 	l_addr = RREG32_MC(R_00001E_K8_FB_LOCATION);
170 	k8_addr = ((unsigned long long)h_addr) << 32 | l_addr;
171 #if defined(CONFIG_X86_32) && !defined(CONFIG_X86_PAE)
172 	if (k8_addr + rdev->mc.visible_vram_size < 0x100000000ULL)
173 #endif
174 	{
175 		/* FastFB shall be used with UMA memory. Here it is simply disabled when sideport
176 		 * memory is present.
177 		 */
178 		if (rdev->mc.igp_sideport_enabled == false && radeon_fastfb == 1) {
179 			DRM_INFO("Direct mapping: aper base at 0x%llx, replaced by direct mapping base 0x%llx.\n",
180 					(unsigned long long)rdev->mc.aper_base, k8_addr);
181 			rdev->mc.aper_base = (resource_size_t)k8_addr;
182 			rdev->fastfb_working = true;
183 		}
184 	}
185 
186 	rs690_pm_info(rdev);
187 	radeon_vram_location(rdev, &rdev->mc, base);
188 	rdev->mc.gtt_base_align = rdev->mc.gtt_size - 1;
189 	radeon_gtt_location(rdev, &rdev->mc);
190 	radeon_update_bandwidth_info(rdev);
191 }
192 
193 void rs690_line_buffer_adjust(struct radeon_device *rdev,
194 			      struct drm_display_mode *mode1,
195 			      struct drm_display_mode *mode2)
196 {
197 	u32 tmp;
198 
199 	/*
200 	 * Line Buffer Setup
201 	 * There is a single line buffer shared by both display controllers.
202 	 * R_006520_DC_LB_MEMORY_SPLIT controls how that line buffer is shared between
203 	 * the display controllers.  The paritioning can either be done
204 	 * manually or via one of four preset allocations specified in bits 1:0:
205 	 *  0 - line buffer is divided in half and shared between crtc
206 	 *  1 - D1 gets 3/4 of the line buffer, D2 gets 1/4
207 	 *  2 - D1 gets the whole buffer
208 	 *  3 - D1 gets 1/4 of the line buffer, D2 gets 3/4
209 	 * Setting bit 2 of R_006520_DC_LB_MEMORY_SPLIT controls switches to manual
210 	 * allocation mode. In manual allocation mode, D1 always starts at 0,
211 	 * D1 end/2 is specified in bits 14:4; D2 allocation follows D1.
212 	 */
213 	tmp = RREG32(R_006520_DC_LB_MEMORY_SPLIT) & C_006520_DC_LB_MEMORY_SPLIT;
214 	tmp &= ~C_006520_DC_LB_MEMORY_SPLIT_MODE;
215 	/* auto */
216 	if (mode1 && mode2) {
217 		if (mode1->hdisplay > mode2->hdisplay) {
218 			if (mode1->hdisplay > 2560)
219 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_3Q_D2_1Q;
220 			else
221 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
222 		} else if (mode2->hdisplay > mode1->hdisplay) {
223 			if (mode2->hdisplay > 2560)
224 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
225 			else
226 				tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
227 		} else
228 			tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1HALF_D2HALF;
229 	} else if (mode1) {
230 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_ONLY;
231 	} else if (mode2) {
232 		tmp |= V_006520_DC_LB_MEMORY_SPLIT_D1_1Q_D2_3Q;
233 	}
234 	WREG32(R_006520_DC_LB_MEMORY_SPLIT, tmp);
235 }
236 
237 struct rs690_watermark {
238 	u32        lb_request_fifo_depth;
239 	fixed20_12 num_line_pair;
240 	fixed20_12 estimated_width;
241 	fixed20_12 worst_case_latency;
242 	fixed20_12 consumption_rate;
243 	fixed20_12 active_time;
244 	fixed20_12 dbpp;
245 	fixed20_12 priority_mark_max;
246 	fixed20_12 priority_mark;
247 	fixed20_12 sclk;
248 };
249 
250 static void rs690_crtc_bandwidth_compute(struct radeon_device *rdev,
251 					 struct radeon_crtc *crtc,
252 					 struct rs690_watermark *wm,
253 					 bool low)
254 {
255 	struct drm_display_mode *mode = &crtc->base.mode;
256 	fixed20_12 a, b, c;
257 	fixed20_12 pclk, request_fifo_depth, tolerable_latency, estimated_width;
258 	fixed20_12 consumption_time, line_time, chunk_time, read_delay_latency;
259 	fixed20_12 sclk, core_bandwidth, max_bandwidth;
260 	u32 selected_sclk;
261 
262 	if (!crtc->base.enabled) {
263 		/* FIXME: wouldn't it better to set priority mark to maximum */
264 		wm->lb_request_fifo_depth = 4;
265 		return;
266 	}
267 
268 	if (((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880)) &&
269 	    (rdev->pm.pm_method == PM_METHOD_DPM) && rdev->pm.dpm_enabled)
270 		selected_sclk = radeon_dpm_get_sclk(rdev, low);
271 	else
272 		selected_sclk = rdev->pm.current_sclk;
273 
274 	/* sclk in Mhz */
275 	a.full = dfixed_const(100);
276 	sclk.full = dfixed_const(selected_sclk);
277 	sclk.full = dfixed_div(sclk, a);
278 
279 	/* core_bandwidth = sclk(Mhz) * 16 */
280 	a.full = dfixed_const(16);
281 	core_bandwidth.full = dfixed_div(rdev->pm.sclk, a);
282 
283 	if (crtc->vsc.full > dfixed_const(2))
284 		wm->num_line_pair.full = dfixed_const(2);
285 	else
286 		wm->num_line_pair.full = dfixed_const(1);
287 
288 	b.full = dfixed_const(mode->crtc_hdisplay);
289 	c.full = dfixed_const(256);
290 	a.full = dfixed_div(b, c);
291 	request_fifo_depth.full = dfixed_mul(a, wm->num_line_pair);
292 	request_fifo_depth.full = dfixed_ceil(request_fifo_depth);
293 	if (a.full < dfixed_const(4)) {
294 		wm->lb_request_fifo_depth = 4;
295 	} else {
296 		wm->lb_request_fifo_depth = dfixed_trunc(request_fifo_depth);
297 	}
298 
299 	/* Determine consumption rate
300 	 *  pclk = pixel clock period(ns) = 1000 / (mode.clock / 1000)
301 	 *  vtaps = number of vertical taps,
302 	 *  vsc = vertical scaling ratio, defined as source/destination
303 	 *  hsc = horizontal scaling ration, defined as source/destination
304 	 */
305 	a.full = dfixed_const(mode->clock);
306 	b.full = dfixed_const(1000);
307 	a.full = dfixed_div(a, b);
308 	pclk.full = dfixed_div(b, a);
309 	if (crtc->rmx_type != RMX_OFF) {
310 		b.full = dfixed_const(2);
311 		if (crtc->vsc.full > b.full)
312 			b.full = crtc->vsc.full;
313 		b.full = dfixed_mul(b, crtc->hsc);
314 		c.full = dfixed_const(2);
315 		b.full = dfixed_div(b, c);
316 		consumption_time.full = dfixed_div(pclk, b);
317 	} else {
318 		consumption_time.full = pclk.full;
319 	}
320 	a.full = dfixed_const(1);
321 	wm->consumption_rate.full = dfixed_div(a, consumption_time);
322 
323 
324 	/* Determine line time
325 	 *  LineTime = total time for one line of displayhtotal
326 	 *  LineTime = total number of horizontal pixels
327 	 *  pclk = pixel clock period(ns)
328 	 */
329 	a.full = dfixed_const(crtc->base.mode.crtc_htotal);
330 	line_time.full = dfixed_mul(a, pclk);
331 
332 	/* Determine active time
333 	 *  ActiveTime = time of active region of display within one line,
334 	 *  hactive = total number of horizontal active pixels
335 	 *  htotal = total number of horizontal pixels
336 	 */
337 	a.full = dfixed_const(crtc->base.mode.crtc_htotal);
338 	b.full = dfixed_const(crtc->base.mode.crtc_hdisplay);
339 	wm->active_time.full = dfixed_mul(line_time, b);
340 	wm->active_time.full = dfixed_div(wm->active_time, a);
341 
342 	/* Maximun bandwidth is the minimun bandwidth of all component */
343 	max_bandwidth = core_bandwidth;
344 	if (rdev->mc.igp_sideport_enabled) {
345 		if (max_bandwidth.full > rdev->pm.sideport_bandwidth.full &&
346 			rdev->pm.sideport_bandwidth.full)
347 			max_bandwidth = rdev->pm.sideport_bandwidth;
348 		read_delay_latency.full = dfixed_const(370 * 800 * 1000);
349 		read_delay_latency.full = dfixed_div(read_delay_latency,
350 			rdev->pm.igp_sideport_mclk);
351 	} else {
352 		if (max_bandwidth.full > rdev->pm.k8_bandwidth.full &&
353 			rdev->pm.k8_bandwidth.full)
354 			max_bandwidth = rdev->pm.k8_bandwidth;
355 		if (max_bandwidth.full > rdev->pm.ht_bandwidth.full &&
356 			rdev->pm.ht_bandwidth.full)
357 			max_bandwidth = rdev->pm.ht_bandwidth;
358 		read_delay_latency.full = dfixed_const(5000);
359 	}
360 
361 	/* sclk = system clocks(ns) = 1000 / max_bandwidth / 16 */
362 	a.full = dfixed_const(16);
363 	sclk.full = dfixed_mul(max_bandwidth, a);
364 	a.full = dfixed_const(1000);
365 	sclk.full = dfixed_div(a, sclk);
366 	/* Determine chunk time
367 	 * ChunkTime = the time it takes the DCP to send one chunk of data
368 	 * to the LB which consists of pipeline delay and inter chunk gap
369 	 * sclk = system clock(ns)
370 	 */
371 	a.full = dfixed_const(256 * 13);
372 	chunk_time.full = dfixed_mul(sclk, a);
373 	a.full = dfixed_const(10);
374 	chunk_time.full = dfixed_div(chunk_time, a);
375 
376 	/* Determine the worst case latency
377 	 * NumLinePair = Number of line pairs to request(1=2 lines, 2=4 lines)
378 	 * WorstCaseLatency = worst case time from urgent to when the MC starts
379 	 *                    to return data
380 	 * READ_DELAY_IDLE_MAX = constant of 1us
381 	 * ChunkTime = time it takes the DCP to send one chunk of data to the LB
382 	 *             which consists of pipeline delay and inter chunk gap
383 	 */
384 	if (dfixed_trunc(wm->num_line_pair) > 1) {
385 		a.full = dfixed_const(3);
386 		wm->worst_case_latency.full = dfixed_mul(a, chunk_time);
387 		wm->worst_case_latency.full += read_delay_latency.full;
388 	} else {
389 		a.full = dfixed_const(2);
390 		wm->worst_case_latency.full = dfixed_mul(a, chunk_time);
391 		wm->worst_case_latency.full += read_delay_latency.full;
392 	}
393 
394 	/* Determine the tolerable latency
395 	 * TolerableLatency = Any given request has only 1 line time
396 	 *                    for the data to be returned
397 	 * LBRequestFifoDepth = Number of chunk requests the LB can
398 	 *                      put into the request FIFO for a display
399 	 *  LineTime = total time for one line of display
400 	 *  ChunkTime = the time it takes the DCP to send one chunk
401 	 *              of data to the LB which consists of
402 	 *  pipeline delay and inter chunk gap
403 	 */
404 	if ((2+wm->lb_request_fifo_depth) >= dfixed_trunc(request_fifo_depth)) {
405 		tolerable_latency.full = line_time.full;
406 	} else {
407 		tolerable_latency.full = dfixed_const(wm->lb_request_fifo_depth - 2);
408 		tolerable_latency.full = request_fifo_depth.full - tolerable_latency.full;
409 		tolerable_latency.full = dfixed_mul(tolerable_latency, chunk_time);
410 		tolerable_latency.full = line_time.full - tolerable_latency.full;
411 	}
412 	/* We assume worst case 32bits (4 bytes) */
413 	wm->dbpp.full = dfixed_const(4 * 8);
414 
415 	/* Determine the maximum priority mark
416 	 *  width = viewport width in pixels
417 	 */
418 	a.full = dfixed_const(16);
419 	wm->priority_mark_max.full = dfixed_const(crtc->base.mode.crtc_hdisplay);
420 	wm->priority_mark_max.full = dfixed_div(wm->priority_mark_max, a);
421 	wm->priority_mark_max.full = dfixed_ceil(wm->priority_mark_max);
422 
423 	/* Determine estimated width */
424 	estimated_width.full = tolerable_latency.full - wm->worst_case_latency.full;
425 	estimated_width.full = dfixed_div(estimated_width, consumption_time);
426 	if (dfixed_trunc(estimated_width) > crtc->base.mode.crtc_hdisplay) {
427 		wm->priority_mark.full = dfixed_const(10);
428 	} else {
429 		a.full = dfixed_const(16);
430 		wm->priority_mark.full = dfixed_div(estimated_width, a);
431 		wm->priority_mark.full = dfixed_ceil(wm->priority_mark);
432 		wm->priority_mark.full = wm->priority_mark_max.full - wm->priority_mark.full;
433 	}
434 }
435 
436 static void rs690_compute_mode_priority(struct radeon_device *rdev,
437 					struct rs690_watermark *wm0,
438 					struct rs690_watermark *wm1,
439 					struct drm_display_mode *mode0,
440 					struct drm_display_mode *mode1,
441 					u32 *d1mode_priority_a_cnt,
442 					u32 *d2mode_priority_a_cnt)
443 {
444 	fixed20_12 priority_mark02, priority_mark12, fill_rate;
445 	fixed20_12 a, b;
446 
447 	*d1mode_priority_a_cnt = S_006548_D1MODE_PRIORITY_A_OFF(1);
448 	*d2mode_priority_a_cnt = S_006548_D1MODE_PRIORITY_A_OFF(1);
449 
450 	if (mode0 && mode1) {
451 		if (dfixed_trunc(wm0->dbpp) > 64)
452 			a.full = dfixed_mul(wm0->dbpp, wm0->num_line_pair);
453 		else
454 			a.full = wm0->num_line_pair.full;
455 		if (dfixed_trunc(wm1->dbpp) > 64)
456 			b.full = dfixed_mul(wm1->dbpp, wm1->num_line_pair);
457 		else
458 			b.full = wm1->num_line_pair.full;
459 		a.full += b.full;
460 		fill_rate.full = dfixed_div(wm0->sclk, a);
461 		if (wm0->consumption_rate.full > fill_rate.full) {
462 			b.full = wm0->consumption_rate.full - fill_rate.full;
463 			b.full = dfixed_mul(b, wm0->active_time);
464 			a.full = dfixed_mul(wm0->worst_case_latency,
465 						wm0->consumption_rate);
466 			a.full = a.full + b.full;
467 			b.full = dfixed_const(16 * 1000);
468 			priority_mark02.full = dfixed_div(a, b);
469 		} else {
470 			a.full = dfixed_mul(wm0->worst_case_latency,
471 						wm0->consumption_rate);
472 			b.full = dfixed_const(16 * 1000);
473 			priority_mark02.full = dfixed_div(a, b);
474 		}
475 		if (wm1->consumption_rate.full > fill_rate.full) {
476 			b.full = wm1->consumption_rate.full - fill_rate.full;
477 			b.full = dfixed_mul(b, wm1->active_time);
478 			a.full = dfixed_mul(wm1->worst_case_latency,
479 						wm1->consumption_rate);
480 			a.full = a.full + b.full;
481 			b.full = dfixed_const(16 * 1000);
482 			priority_mark12.full = dfixed_div(a, b);
483 		} else {
484 			a.full = dfixed_mul(wm1->worst_case_latency,
485 						wm1->consumption_rate);
486 			b.full = dfixed_const(16 * 1000);
487 			priority_mark12.full = dfixed_div(a, b);
488 		}
489 		if (wm0->priority_mark.full > priority_mark02.full)
490 			priority_mark02.full = wm0->priority_mark.full;
491 		if (dfixed_trunc(priority_mark02) < 0)
492 			priority_mark02.full = 0;
493 		if (wm0->priority_mark_max.full > priority_mark02.full)
494 			priority_mark02.full = wm0->priority_mark_max.full;
495 		if (wm1->priority_mark.full > priority_mark12.full)
496 			priority_mark12.full = wm1->priority_mark.full;
497 		if (dfixed_trunc(priority_mark12) < 0)
498 			priority_mark12.full = 0;
499 		if (wm1->priority_mark_max.full > priority_mark12.full)
500 			priority_mark12.full = wm1->priority_mark_max.full;
501 		*d1mode_priority_a_cnt = dfixed_trunc(priority_mark02);
502 		*d2mode_priority_a_cnt = dfixed_trunc(priority_mark12);
503 		if (rdev->disp_priority == 2) {
504 			*d1mode_priority_a_cnt |= S_006548_D1MODE_PRIORITY_A_ALWAYS_ON(1);
505 			*d2mode_priority_a_cnt |= S_006D48_D2MODE_PRIORITY_A_ALWAYS_ON(1);
506 		}
507 	} else if (mode0) {
508 		if (dfixed_trunc(wm0->dbpp) > 64)
509 			a.full = dfixed_mul(wm0->dbpp, wm0->num_line_pair);
510 		else
511 			a.full = wm0->num_line_pair.full;
512 		fill_rate.full = dfixed_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 = dfixed_mul(b, wm0->active_time);
516 			a.full = dfixed_mul(wm0->worst_case_latency,
517 						wm0->consumption_rate);
518 			a.full = a.full + b.full;
519 			b.full = dfixed_const(16 * 1000);
520 			priority_mark02.full = dfixed_div(a, b);
521 		} else {
522 			a.full = dfixed_mul(wm0->worst_case_latency,
523 						wm0->consumption_rate);
524 			b.full = dfixed_const(16 * 1000);
525 			priority_mark02.full = dfixed_div(a, b);
526 		}
527 		if (wm0->priority_mark.full > priority_mark02.full)
528 			priority_mark02.full = wm0->priority_mark.full;
529 		if (dfixed_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 		*d1mode_priority_a_cnt = dfixed_trunc(priority_mark02);
534 		if (rdev->disp_priority == 2)
535 			*d1mode_priority_a_cnt |= S_006548_D1MODE_PRIORITY_A_ALWAYS_ON(1);
536 	} else if (mode1) {
537 		if (dfixed_trunc(wm1->dbpp) > 64)
538 			a.full = dfixed_mul(wm1->dbpp, wm1->num_line_pair);
539 		else
540 			a.full = wm1->num_line_pair.full;
541 		fill_rate.full = dfixed_div(wm1->sclk, a);
542 		if (wm1->consumption_rate.full > fill_rate.full) {
543 			b.full = wm1->consumption_rate.full - fill_rate.full;
544 			b.full = dfixed_mul(b, wm1->active_time);
545 			a.full = dfixed_mul(wm1->worst_case_latency,
546 						wm1->consumption_rate);
547 			a.full = a.full + b.full;
548 			b.full = dfixed_const(16 * 1000);
549 			priority_mark12.full = dfixed_div(a, b);
550 		} else {
551 			a.full = dfixed_mul(wm1->worst_case_latency,
552 						wm1->consumption_rate);
553 			b.full = dfixed_const(16 * 1000);
554 			priority_mark12.full = dfixed_div(a, b);
555 		}
556 		if (wm1->priority_mark.full > priority_mark12.full)
557 			priority_mark12.full = wm1->priority_mark.full;
558 		if (dfixed_trunc(priority_mark12) < 0)
559 			priority_mark12.full = 0;
560 		if (wm1->priority_mark_max.full > priority_mark12.full)
561 			priority_mark12.full = wm1->priority_mark_max.full;
562 		*d2mode_priority_a_cnt = dfixed_trunc(priority_mark12);
563 		if (rdev->disp_priority == 2)
564 			*d2mode_priority_a_cnt |= S_006D48_D2MODE_PRIORITY_A_ALWAYS_ON(1);
565 	}
566 }
567 
568 void rs690_bandwidth_update(struct radeon_device *rdev)
569 {
570 	struct drm_display_mode *mode0 = NULL;
571 	struct drm_display_mode *mode1 = NULL;
572 	struct rs690_watermark wm0_high, wm0_low;
573 	struct rs690_watermark wm1_high, wm1_low;
574 	u32 tmp;
575 	u32 d1mode_priority_a_cnt, d1mode_priority_b_cnt;
576 	u32 d2mode_priority_a_cnt, d2mode_priority_b_cnt;
577 
578 	radeon_update_display_priority(rdev);
579 
580 	if (rdev->mode_info.crtcs[0]->base.enabled)
581 		mode0 = &rdev->mode_info.crtcs[0]->base.mode;
582 	if (rdev->mode_info.crtcs[1]->base.enabled)
583 		mode1 = &rdev->mode_info.crtcs[1]->base.mode;
584 	/*
585 	 * Set display0/1 priority up in the memory controller for
586 	 * modes if the user specifies HIGH for displaypriority
587 	 * option.
588 	 */
589 	if ((rdev->disp_priority == 2) &&
590 	    ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))) {
591 		tmp = RREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER);
592 		tmp &= C_000104_MC_DISP0R_INIT_LAT;
593 		tmp &= C_000104_MC_DISP1R_INIT_LAT;
594 		if (mode0)
595 			tmp |= S_000104_MC_DISP0R_INIT_LAT(1);
596 		if (mode1)
597 			tmp |= S_000104_MC_DISP1R_INIT_LAT(1);
598 		WREG32_MC(R_000104_MC_INIT_MISC_LAT_TIMER, tmp);
599 	}
600 	rs690_line_buffer_adjust(rdev, mode0, mode1);
601 
602 	if ((rdev->family == CHIP_RS690) || (rdev->family == CHIP_RS740))
603 		WREG32(R_006C9C_DCP_CONTROL, 0);
604 	if ((rdev->family == CHIP_RS780) || (rdev->family == CHIP_RS880))
605 		WREG32(R_006C9C_DCP_CONTROL, 2);
606 
607 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0_high, false);
608 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1_high, false);
609 
610 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[0], &wm0_low, true);
611 	rs690_crtc_bandwidth_compute(rdev, rdev->mode_info.crtcs[1], &wm1_low, true);
612 
613 	tmp = (wm0_high.lb_request_fifo_depth - 1);
614 	tmp |= (wm1_high.lb_request_fifo_depth - 1) << 16;
615 	WREG32(R_006D58_LB_MAX_REQ_OUTSTANDING, tmp);
616 
617 	rs690_compute_mode_priority(rdev,
618 				    &wm0_high, &wm1_high,
619 				    mode0, mode1,
620 				    &d1mode_priority_a_cnt, &d2mode_priority_a_cnt);
621 	rs690_compute_mode_priority(rdev,
622 				    &wm0_low, &wm1_low,
623 				    mode0, mode1,
624 				    &d1mode_priority_b_cnt, &d2mode_priority_b_cnt);
625 
626 	WREG32(R_006548_D1MODE_PRIORITY_A_CNT, d1mode_priority_a_cnt);
627 	WREG32(R_00654C_D1MODE_PRIORITY_B_CNT, d1mode_priority_b_cnt);
628 	WREG32(R_006D48_D2MODE_PRIORITY_A_CNT, d2mode_priority_a_cnt);
629 	WREG32(R_006D4C_D2MODE_PRIORITY_B_CNT, d2mode_priority_b_cnt);
630 }
631 
632 uint32_t rs690_mc_rreg(struct radeon_device *rdev, uint32_t reg)
633 {
634 	unsigned long flags;
635 	uint32_t r;
636 
637 	spin_lock_irqsave(&rdev->mc_idx_lock, flags);
638 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg));
639 	r = RREG32(R_00007C_MC_DATA);
640 	WREG32(R_000078_MC_INDEX, ~C_000078_MC_IND_ADDR);
641 	spin_unlock_irqrestore(&rdev->mc_idx_lock, flags);
642 	return r;
643 }
644 
645 void rs690_mc_wreg(struct radeon_device *rdev, uint32_t reg, uint32_t v)
646 {
647 	unsigned long flags;
648 
649 	spin_lock_irqsave(&rdev->mc_idx_lock, flags);
650 	WREG32(R_000078_MC_INDEX, S_000078_MC_IND_ADDR(reg) |
651 		S_000078_MC_IND_WR_EN(1));
652 	WREG32(R_00007C_MC_DATA, v);
653 	WREG32(R_000078_MC_INDEX, 0x7F);
654 	spin_unlock_irqrestore(&rdev->mc_idx_lock, flags);
655 }
656 
657 static void rs690_mc_program(struct radeon_device *rdev)
658 {
659 	struct rv515_mc_save save;
660 
661 	/* Stops all mc clients */
662 	rv515_mc_stop(rdev, &save);
663 
664 	/* Wait for mc idle */
665 	if (rs690_mc_wait_for_idle(rdev))
666 		dev_warn(rdev->dev, "Wait MC idle timeout before updating MC.\n");
667 	/* Program MC, should be a 32bits limited address space */
668 	WREG32_MC(R_000100_MCCFG_FB_LOCATION,
669 			S_000100_MC_FB_START(rdev->mc.vram_start >> 16) |
670 			S_000100_MC_FB_TOP(rdev->mc.vram_end >> 16));
671 	WREG32(R_000134_HDP_FB_LOCATION,
672 		S_000134_HDP_FB_START(rdev->mc.vram_start >> 16));
673 
674 	rv515_mc_resume(rdev, &save);
675 }
676 
677 static int rs690_startup(struct radeon_device *rdev)
678 {
679 	int r;
680 
681 	rs690_mc_program(rdev);
682 	/* Resume clock */
683 	rv515_clock_startup(rdev);
684 	/* Initialize GPU configuration (# pipes, ...) */
685 	rs690_gpu_init(rdev);
686 	/* Initialize GART (initialize after TTM so we can allocate
687 	 * memory through TTM but finalize after TTM) */
688 	r = rs400_gart_enable(rdev);
689 	if (r)
690 		return r;
691 
692 	/* allocate wb buffer */
693 	r = radeon_wb_init(rdev);
694 	if (r)
695 		return r;
696 
697 	r = radeon_fence_driver_start_ring(rdev, RADEON_RING_TYPE_GFX_INDEX);
698 	if (r) {
699 		dev_err(rdev->dev, "failed initializing CP fences (%d).\n", r);
700 		return r;
701 	}
702 
703 	/* Enable IRQ */
704 	if (!rdev->irq.installed) {
705 		r = radeon_irq_kms_init(rdev);
706 		if (r)
707 			return r;
708 	}
709 
710 	rs600_irq_set(rdev);
711 	rdev->config.r300.hdp_cntl = RREG32(RADEON_HOST_PATH_CNTL);
712 	/* 1M ring buffer */
713 	r = r100_cp_init(rdev, 1024 * 1024);
714 	if (r) {
715 		dev_err(rdev->dev, "failed initializing CP (%d).\n", r);
716 		return r;
717 	}
718 
719 	r = radeon_ib_pool_init(rdev);
720 	if (r) {
721 		dev_err(rdev->dev, "IB initialization failed (%d).\n", r);
722 		return r;
723 	}
724 
725 	r = r600_audio_init(rdev);
726 	if (r) {
727 		dev_err(rdev->dev, "failed initializing audio\n");
728 		return r;
729 	}
730 
731 	return 0;
732 }
733 
734 int rs690_resume(struct radeon_device *rdev)
735 {
736 	int r;
737 
738 	/* Make sur GART are not working */
739 	rs400_gart_disable(rdev);
740 	/* Resume clock before doing reset */
741 	rv515_clock_startup(rdev);
742 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
743 	if (radeon_asic_reset(rdev)) {
744 		dev_warn(rdev->dev, "GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
745 			RREG32(R_000E40_RBBM_STATUS),
746 			RREG32(R_0007C0_CP_STAT));
747 	}
748 	/* post */
749 	atom_asic_init(rdev->mode_info.atom_context);
750 	/* Resume clock after posting */
751 	rv515_clock_startup(rdev);
752 	/* Initialize surface registers */
753 	radeon_surface_init(rdev);
754 
755 	rdev->accel_working = true;
756 	r = rs690_startup(rdev);
757 	if (r) {
758 		rdev->accel_working = false;
759 	}
760 	return r;
761 }
762 
763 int rs690_suspend(struct radeon_device *rdev)
764 {
765 	r600_audio_fini(rdev);
766 	r100_cp_disable(rdev);
767 	radeon_wb_disable(rdev);
768 	rs600_irq_disable(rdev);
769 	rs400_gart_disable(rdev);
770 	return 0;
771 }
772 
773 void rs690_fini(struct radeon_device *rdev)
774 {
775 	r600_audio_fini(rdev);
776 	r100_cp_fini(rdev);
777 	radeon_wb_fini(rdev);
778 	radeon_ib_pool_fini(rdev);
779 	radeon_gem_fini(rdev);
780 	rs400_gart_fini(rdev);
781 	radeon_irq_kms_fini(rdev);
782 	radeon_fence_driver_fini(rdev);
783 	radeon_bo_fini(rdev);
784 	radeon_atombios_fini(rdev);
785 	kfree(rdev->bios);
786 	rdev->bios = NULL;
787 }
788 
789 int rs690_init(struct radeon_device *rdev)
790 {
791 	int r;
792 
793 	/* Disable VGA */
794 	rv515_vga_render_disable(rdev);
795 	/* Initialize scratch registers */
796 	radeon_scratch_init(rdev);
797 	/* Initialize surface registers */
798 	radeon_surface_init(rdev);
799 	/* restore some register to sane defaults */
800 	r100_restore_sanity(rdev);
801 	/* TODO: disable VGA need to use VGA request */
802 	/* BIOS*/
803 	if (!radeon_get_bios(rdev)) {
804 		if (ASIC_IS_AVIVO(rdev))
805 			return -EINVAL;
806 	}
807 	if (rdev->is_atom_bios) {
808 		r = radeon_atombios_init(rdev);
809 		if (r)
810 			return r;
811 	} else {
812 		dev_err(rdev->dev, "Expecting atombios for RV515 GPU\n");
813 		return -EINVAL;
814 	}
815 	/* Reset gpu before posting otherwise ATOM will enter infinite loop */
816 	if (radeon_asic_reset(rdev)) {
817 		dev_warn(rdev->dev,
818 			"GPU reset failed ! (0xE40=0x%08X, 0x7C0=0x%08X)\n",
819 			RREG32(R_000E40_RBBM_STATUS),
820 			RREG32(R_0007C0_CP_STAT));
821 	}
822 	/* check if cards are posted or not */
823 	if (radeon_boot_test_post_card(rdev) == false)
824 		return -EINVAL;
825 
826 	/* Initialize clocks */
827 	radeon_get_clock_info(rdev->ddev);
828 	/* initialize memory controller */
829 	rs690_mc_init(rdev);
830 	rv515_debugfs(rdev);
831 	/* Fence driver */
832 	r = radeon_fence_driver_init(rdev);
833 	if (r)
834 		return r;
835 	/* Memory manager */
836 	r = radeon_bo_init(rdev);
837 	if (r)
838 		return r;
839 	r = rs400_gart_init(rdev);
840 	if (r)
841 		return r;
842 	rs600_set_safe_registers(rdev);
843 
844 	rdev->accel_working = true;
845 	r = rs690_startup(rdev);
846 	if (r) {
847 		/* Somethings want wront with the accel init stop accel */
848 		dev_err(rdev->dev, "Disabling GPU acceleration\n");
849 		r100_cp_fini(rdev);
850 		radeon_wb_fini(rdev);
851 		radeon_ib_pool_fini(rdev);
852 		rs400_gart_fini(rdev);
853 		radeon_irq_kms_fini(rdev);
854 		rdev->accel_working = false;
855 	}
856 	return 0;
857 }
858