Linux-libre 4.10.7-gnu
[librecmc/linux-libre.git] / drivers / gpu / drm / msm / adreno / a3xx_gpu.c
1 /*
2  * Copyright (C) 2013 Red Hat
3  * Author: Rob Clark <robdclark@gmail.com>
4  *
5  * Copyright (c) 2014 The Linux Foundation. All rights reserved.
6  *
7  * This program is free software; you can redistribute it and/or modify it
8  * under the terms of the GNU General Public License version 2 as published by
9  * the Free Software Foundation.
10  *
11  * This program is distributed in the hope that it will be useful, but WITHOUT
12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
14  * more details.
15  *
16  * You should have received a copy of the GNU General Public License along with
17  * this program.  If not, see <http://www.gnu.org/licenses/>.
18  */
19
20 #ifdef CONFIG_MSM_OCMEM
21 #  include <mach/ocmem.h>
22 #endif
23
24 #include "a3xx_gpu.h"
25
26 #define A3XX_INT0_MASK \
27         (A3XX_INT0_RBBM_AHB_ERROR |        \
28          A3XX_INT0_RBBM_ATB_BUS_OVERFLOW | \
29          A3XX_INT0_CP_T0_PACKET_IN_IB |    \
30          A3XX_INT0_CP_OPCODE_ERROR |       \
31          A3XX_INT0_CP_RESERVED_BIT_ERROR | \
32          A3XX_INT0_CP_HW_FAULT |           \
33          A3XX_INT0_CP_IB1_INT |            \
34          A3XX_INT0_CP_IB2_INT |            \
35          A3XX_INT0_CP_RB_INT |             \
36          A3XX_INT0_CP_REG_PROTECT_FAULT |  \
37          A3XX_INT0_CP_AHB_ERROR_HALT |     \
38          A3XX_INT0_UCHE_OOB_ACCESS)
39
40 extern bool hang_debug;
41
42 static void a3xx_dump(struct msm_gpu *gpu);
43
44 static bool a3xx_me_init(struct msm_gpu *gpu)
45 {
46         struct msm_ringbuffer *ring = gpu->rb;
47
48         OUT_PKT3(ring, CP_ME_INIT, 17);
49         OUT_RING(ring, 0x000003f7);
50         OUT_RING(ring, 0x00000000);
51         OUT_RING(ring, 0x00000000);
52         OUT_RING(ring, 0x00000000);
53         OUT_RING(ring, 0x00000080);
54         OUT_RING(ring, 0x00000100);
55         OUT_RING(ring, 0x00000180);
56         OUT_RING(ring, 0x00006600);
57         OUT_RING(ring, 0x00000150);
58         OUT_RING(ring, 0x0000014e);
59         OUT_RING(ring, 0x00000154);
60         OUT_RING(ring, 0x00000001);
61         OUT_RING(ring, 0x00000000);
62         OUT_RING(ring, 0x00000000);
63         OUT_RING(ring, 0x00000000);
64         OUT_RING(ring, 0x00000000);
65         OUT_RING(ring, 0x00000000);
66
67         gpu->funcs->flush(gpu);
68         return gpu->funcs->idle(gpu);
69 }
70
71 static int a3xx_hw_init(struct msm_gpu *gpu)
72 {
73         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
74         struct a3xx_gpu *a3xx_gpu = to_a3xx_gpu(adreno_gpu);
75         uint32_t *ptr, len;
76         int i, ret;
77
78         DBG("%s", gpu->name);
79
80         if (adreno_is_a305(adreno_gpu)) {
81                 /* Set up 16 deep read/write request queues: */
82                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF0, 0x10101010);
83                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF1, 0x10101010);
84                 gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x10101010);
85                 gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x10101010);
86                 gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
87                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF0, 0x10101010);
88                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF1, 0x10101010);
89                 /* Enable WR-REQ: */
90                 gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x0000ff);
91                 /* Set up round robin arbitration between both AXI ports: */
92                 gpu_write(gpu, REG_A3XX_VBIF_ARB_CTL, 0x00000030);
93                 /* Set up AOOO: */
94                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO_EN, 0x0000003c);
95                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO, 0x003c003c);
96         } else if (adreno_is_a306(adreno_gpu)) {
97                 gpu_write(gpu, REG_A3XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x0003);
98                 gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x0000000a);
99                 gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x0000000a);
100         } else if (adreno_is_a320(adreno_gpu)) {
101                 /* Set up 16 deep read/write request queues: */
102                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF0, 0x10101010);
103                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF1, 0x10101010);
104                 gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x10101010);
105                 gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x10101010);
106                 gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
107                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF0, 0x10101010);
108                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF1, 0x10101010);
109                 /* Enable WR-REQ: */
110                 gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x0000ff);
111                 /* Set up round robin arbitration between both AXI ports: */
112                 gpu_write(gpu, REG_A3XX_VBIF_ARB_CTL, 0x00000030);
113                 /* Set up AOOO: */
114                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO_EN, 0x0000003c);
115                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO, 0x003c003c);
116                 /* Enable 1K sort: */
117                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT, 0x000000ff);
118                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT_CONF, 0x000000a4);
119
120         } else if (adreno_is_a330v2(adreno_gpu)) {
121                 /*
122                  * Most of the VBIF registers on 8974v2 have the correct
123                  * values at power on, so we won't modify those if we don't
124                  * need to
125                  */
126                 /* Enable 1k sort: */
127                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT, 0x0001003f);
128                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT_CONF, 0x000000a4);
129                 /* Enable WR-REQ: */
130                 gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x00003f);
131                 gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
132                 /* Set up VBIF_ROUND_ROBIN_QOS_ARB: */
133                 gpu_write(gpu, REG_A3XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x0003);
134
135         } else if (adreno_is_a330(adreno_gpu)) {
136                 /* Set up 16 deep read/write request queues: */
137                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF0, 0x18181818);
138                 gpu_write(gpu, REG_A3XX_VBIF_IN_RD_LIM_CONF1, 0x18181818);
139                 gpu_write(gpu, REG_A3XX_VBIF_OUT_RD_LIM_CONF0, 0x18181818);
140                 gpu_write(gpu, REG_A3XX_VBIF_OUT_WR_LIM_CONF0, 0x18181818);
141                 gpu_write(gpu, REG_A3XX_VBIF_DDR_OUT_MAX_BURST, 0x0000303);
142                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF0, 0x18181818);
143                 gpu_write(gpu, REG_A3XX_VBIF_IN_WR_LIM_CONF1, 0x18181818);
144                 /* Enable WR-REQ: */
145                 gpu_write(gpu, REG_A3XX_VBIF_GATE_OFF_WRREQ_EN, 0x00003f);
146                 /* Set up round robin arbitration between both AXI ports: */
147                 gpu_write(gpu, REG_A3XX_VBIF_ARB_CTL, 0x00000030);
148                 /* Set up VBIF_ROUND_ROBIN_QOS_ARB: */
149                 gpu_write(gpu, REG_A3XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x0001);
150                 /* Set up AOOO: */
151                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO_EN, 0x0000003f);
152                 gpu_write(gpu, REG_A3XX_VBIF_OUT_AXI_AOOO, 0x003f003f);
153                 /* Enable 1K sort: */
154                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT, 0x0001003f);
155                 gpu_write(gpu, REG_A3XX_VBIF_ABIT_SORT_CONF, 0x000000a4);
156                 /* Disable VBIF clock gating. This is to enable AXI running
157                  * higher frequency than GPU:
158                  */
159                 gpu_write(gpu, REG_A3XX_VBIF_CLKON, 0x00000001);
160
161         } else {
162                 BUG();
163         }
164
165         /* Make all blocks contribute to the GPU BUSY perf counter: */
166         gpu_write(gpu, REG_A3XX_RBBM_GPU_BUSY_MASKED, 0xffffffff);
167
168         /* Tune the hystersis counters for SP and CP idle detection: */
169         gpu_write(gpu, REG_A3XX_RBBM_SP_HYST_CNT, 0x10);
170         gpu_write(gpu, REG_A3XX_RBBM_WAIT_IDLE_CLOCKS_CTL, 0x10);
171
172         /* Enable the RBBM error reporting bits.  This lets us get
173          * useful information on failure:
174          */
175         gpu_write(gpu, REG_A3XX_RBBM_AHB_CTL0, 0x00000001);
176
177         /* Enable AHB error reporting: */
178         gpu_write(gpu, REG_A3XX_RBBM_AHB_CTL1, 0xa6ffffff);
179
180         /* Turn on the power counters: */
181         gpu_write(gpu, REG_A3XX_RBBM_RBBM_CTL, 0x00030000);
182
183         /* Turn on hang detection - this spews a lot of useful information
184          * into the RBBM registers on a hang:
185          */
186         gpu_write(gpu, REG_A3XX_RBBM_INTERFACE_HANG_INT_CTL, 0x00010fff);
187
188         /* Enable 64-byte cacheline size. HW Default is 32-byte (0x000000E0): */
189         gpu_write(gpu, REG_A3XX_UCHE_CACHE_MODE_CONTROL_REG, 0x00000001);
190
191         /* Enable Clock gating: */
192         if (adreno_is_a306(adreno_gpu))
193                 gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xaaaaaaaa);
194         else if (adreno_is_a320(adreno_gpu))
195                 gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xbfffffff);
196         else if (adreno_is_a330v2(adreno_gpu))
197                 gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xaaaaaaaa);
198         else if (adreno_is_a330(adreno_gpu))
199                 gpu_write(gpu, REG_A3XX_RBBM_CLOCK_CTL, 0xbffcffff);
200
201         if (adreno_is_a330v2(adreno_gpu))
202                 gpu_write(gpu, REG_A3XX_RBBM_GPR0_CTL, 0x05515455);
203         else if (adreno_is_a330(adreno_gpu))
204                 gpu_write(gpu, REG_A3XX_RBBM_GPR0_CTL, 0x00000000);
205
206         /* Set the OCMEM base address for A330, etc */
207         if (a3xx_gpu->ocmem_hdl) {
208                 gpu_write(gpu, REG_A3XX_RB_GMEM_BASE_ADDR,
209                         (unsigned int)(a3xx_gpu->ocmem_base >> 14));
210         }
211
212         /* Turn on performance counters: */
213         gpu_write(gpu, REG_A3XX_RBBM_PERFCTR_CTL, 0x01);
214
215         /* Enable the perfcntrs that we use.. */
216         for (i = 0; i < gpu->num_perfcntrs; i++) {
217                 const struct msm_gpu_perfcntr *perfcntr = &gpu->perfcntrs[i];
218                 gpu_write(gpu, perfcntr->select_reg, perfcntr->select_val);
219         }
220
221         gpu_write(gpu, REG_A3XX_RBBM_INT_0_MASK, A3XX_INT0_MASK);
222
223         ret = adreno_hw_init(gpu);
224         if (ret)
225                 return ret;
226
227         /* setup access protection: */
228         gpu_write(gpu, REG_A3XX_CP_PROTECT_CTRL, 0x00000007);
229
230         /* RBBM registers */
231         gpu_write(gpu, REG_A3XX_CP_PROTECT(0), 0x63000040);
232         gpu_write(gpu, REG_A3XX_CP_PROTECT(1), 0x62000080);
233         gpu_write(gpu, REG_A3XX_CP_PROTECT(2), 0x600000cc);
234         gpu_write(gpu, REG_A3XX_CP_PROTECT(3), 0x60000108);
235         gpu_write(gpu, REG_A3XX_CP_PROTECT(4), 0x64000140);
236         gpu_write(gpu, REG_A3XX_CP_PROTECT(5), 0x66000400);
237
238         /* CP registers */
239         gpu_write(gpu, REG_A3XX_CP_PROTECT(6), 0x65000700);
240         gpu_write(gpu, REG_A3XX_CP_PROTECT(7), 0x610007d8);
241         gpu_write(gpu, REG_A3XX_CP_PROTECT(8), 0x620007e0);
242         gpu_write(gpu, REG_A3XX_CP_PROTECT(9), 0x61001178);
243         gpu_write(gpu, REG_A3XX_CP_PROTECT(10), 0x64001180);
244
245         /* RB registers */
246         gpu_write(gpu, REG_A3XX_CP_PROTECT(11), 0x60003300);
247
248         /* VBIF registers */
249         gpu_write(gpu, REG_A3XX_CP_PROTECT(12), 0x6b00c000);
250
251         /* NOTE: PM4/micro-engine firmware registers look to be the same
252          * for a2xx and a3xx.. we could possibly push that part down to
253          * adreno_gpu base class.  Or push both PM4 and PFP but
254          * parameterize the pfp ucode addr/data registers..
255          */
256
257         /* Load PM4: */
258         ptr = (uint32_t *)(adreno_gpu->pm4->data);
259         len = adreno_gpu->pm4->size / 4;
260         DBG("loading PM4 ucode version: %x", ptr[1]);
261
262         gpu_write(gpu, REG_AXXX_CP_DEBUG,
263                         AXXX_CP_DEBUG_DYNAMIC_CLK_DISABLE |
264                         AXXX_CP_DEBUG_MIU_128BIT_WRITE_ENABLE);
265         gpu_write(gpu, REG_AXXX_CP_ME_RAM_WADDR, 0);
266         for (i = 1; i < len; i++)
267                 gpu_write(gpu, REG_AXXX_CP_ME_RAM_DATA, ptr[i]);
268
269         /* Load PFP: */
270         ptr = (uint32_t *)(adreno_gpu->pfp->data);
271         len = adreno_gpu->pfp->size / 4;
272         DBG("loading PFP ucode version: %x", ptr[5]);
273
274         gpu_write(gpu, REG_A3XX_CP_PFP_UCODE_ADDR, 0);
275         for (i = 1; i < len; i++)
276                 gpu_write(gpu, REG_A3XX_CP_PFP_UCODE_DATA, ptr[i]);
277
278         /* CP ROQ queue sizes (bytes) - RB:16, ST:16, IB1:32, IB2:64 */
279         if (adreno_is_a305(adreno_gpu) || adreno_is_a306(adreno_gpu) ||
280                         adreno_is_a320(adreno_gpu)) {
281                 gpu_write(gpu, REG_AXXX_CP_QUEUE_THRESHOLDS,
282                                 AXXX_CP_QUEUE_THRESHOLDS_CSQ_IB1_START(2) |
283                                 AXXX_CP_QUEUE_THRESHOLDS_CSQ_IB2_START(6) |
284                                 AXXX_CP_QUEUE_THRESHOLDS_CSQ_ST_START(14));
285         } else if (adreno_is_a330(adreno_gpu)) {
286                 /* NOTE: this (value take from downstream android driver)
287                  * includes some bits outside of the known bitfields.  But
288                  * A330 has this "MERCIU queue" thing too, which might
289                  * explain a new bitfield or reshuffling:
290                  */
291                 gpu_write(gpu, REG_AXXX_CP_QUEUE_THRESHOLDS, 0x003e2008);
292         }
293
294         /* clear ME_HALT to start micro engine */
295         gpu_write(gpu, REG_AXXX_CP_ME_CNTL, 0);
296
297         return a3xx_me_init(gpu) ? 0 : -EINVAL;
298 }
299
300 static void a3xx_recover(struct msm_gpu *gpu)
301 {
302         int i;
303
304         adreno_dump_info(gpu);
305
306         for (i = 0; i < 8; i++) {
307                 printk("CP_SCRATCH_REG%d: %u\n", i,
308                         gpu_read(gpu, REG_AXXX_CP_SCRATCH_REG0 + i));
309         }
310
311         /* dump registers before resetting gpu, if enabled: */
312         if (hang_debug)
313                 a3xx_dump(gpu);
314
315         gpu_write(gpu, REG_A3XX_RBBM_SW_RESET_CMD, 1);
316         gpu_read(gpu, REG_A3XX_RBBM_SW_RESET_CMD);
317         gpu_write(gpu, REG_A3XX_RBBM_SW_RESET_CMD, 0);
318         adreno_recover(gpu);
319 }
320
321 static void a3xx_destroy(struct msm_gpu *gpu)
322 {
323         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
324         struct a3xx_gpu *a3xx_gpu = to_a3xx_gpu(adreno_gpu);
325
326         DBG("%s", gpu->name);
327
328         adreno_gpu_cleanup(adreno_gpu);
329
330 #ifdef CONFIG_MSM_OCMEM
331         if (a3xx_gpu->ocmem_base)
332                 ocmem_free(OCMEM_GRAPHICS, a3xx_gpu->ocmem_hdl);
333 #endif
334
335         kfree(a3xx_gpu);
336 }
337
338 static bool a3xx_idle(struct msm_gpu *gpu)
339 {
340         /* wait for ringbuffer to drain: */
341         if (!adreno_idle(gpu))
342                 return false;
343
344         /* then wait for GPU to finish: */
345         if (spin_until(!(gpu_read(gpu, REG_A3XX_RBBM_STATUS) &
346                         A3XX_RBBM_STATUS_GPU_BUSY))) {
347                 DRM_ERROR("%s: timeout waiting for GPU to idle!\n", gpu->name);
348
349                 /* TODO maybe we need to reset GPU here to recover from hang? */
350                 return false;
351         }
352
353         return true;
354 }
355
356 static irqreturn_t a3xx_irq(struct msm_gpu *gpu)
357 {
358         uint32_t status;
359
360         status = gpu_read(gpu, REG_A3XX_RBBM_INT_0_STATUS);
361         DBG("%s: %08x", gpu->name, status);
362
363         // TODO
364
365         gpu_write(gpu, REG_A3XX_RBBM_INT_CLEAR_CMD, status);
366
367         msm_gpu_retire(gpu);
368
369         return IRQ_HANDLED;
370 }
371
372 static const unsigned int a3xx_registers[] = {
373         0x0000, 0x0002, 0x0010, 0x0012, 0x0018, 0x0018, 0x0020, 0x0027,
374         0x0029, 0x002b, 0x002e, 0x0033, 0x0040, 0x0042, 0x0050, 0x005c,
375         0x0060, 0x006c, 0x0080, 0x0082, 0x0084, 0x0088, 0x0090, 0x00e5,
376         0x00ea, 0x00ed, 0x0100, 0x0100, 0x0110, 0x0123, 0x01c0, 0x01c1,
377         0x01c3, 0x01c5, 0x01c7, 0x01c7, 0x01d5, 0x01d9, 0x01dc, 0x01dd,
378         0x01ea, 0x01ea, 0x01ee, 0x01f1, 0x01f5, 0x01f5, 0x01fc, 0x01ff,
379         0x0440, 0x0440, 0x0443, 0x0443, 0x0445, 0x0445, 0x044d, 0x044f,
380         0x0452, 0x0452, 0x0454, 0x046f, 0x047c, 0x047c, 0x047f, 0x047f,
381         0x0578, 0x057f, 0x0600, 0x0602, 0x0605, 0x0607, 0x060a, 0x060e,
382         0x0612, 0x0614, 0x0c01, 0x0c02, 0x0c06, 0x0c1d, 0x0c3d, 0x0c3f,
383         0x0c48, 0x0c4b, 0x0c80, 0x0c80, 0x0c88, 0x0c8b, 0x0ca0, 0x0cb7,
384         0x0cc0, 0x0cc1, 0x0cc6, 0x0cc7, 0x0ce4, 0x0ce5, 0x0e00, 0x0e05,
385         0x0e0c, 0x0e0c, 0x0e22, 0x0e23, 0x0e41, 0x0e45, 0x0e64, 0x0e65,
386         0x0e80, 0x0e82, 0x0e84, 0x0e89, 0x0ea0, 0x0ea1, 0x0ea4, 0x0ea7,
387         0x0ec4, 0x0ecb, 0x0ee0, 0x0ee0, 0x0f00, 0x0f01, 0x0f03, 0x0f09,
388         0x2040, 0x2040, 0x2044, 0x2044, 0x2048, 0x204d, 0x2068, 0x2069,
389         0x206c, 0x206d, 0x2070, 0x2070, 0x2072, 0x2072, 0x2074, 0x2075,
390         0x2079, 0x207a, 0x20c0, 0x20d3, 0x20e4, 0x20ef, 0x2100, 0x2109,
391         0x210c, 0x210c, 0x210e, 0x210e, 0x2110, 0x2111, 0x2114, 0x2115,
392         0x21e4, 0x21e4, 0x21ea, 0x21ea, 0x21ec, 0x21ed, 0x21f0, 0x21f0,
393         0x2200, 0x2212, 0x2214, 0x2217, 0x221a, 0x221a, 0x2240, 0x227e,
394         0x2280, 0x228b, 0x22c0, 0x22c0, 0x22c4, 0x22ce, 0x22d0, 0x22d8,
395         0x22df, 0x22e6, 0x22e8, 0x22e9, 0x22ec, 0x22ec, 0x22f0, 0x22f7,
396         0x22ff, 0x22ff, 0x2340, 0x2343, 0x2348, 0x2349, 0x2350, 0x2356,
397         0x2360, 0x2360, 0x2440, 0x2440, 0x2444, 0x2444, 0x2448, 0x244d,
398         0x2468, 0x2469, 0x246c, 0x246d, 0x2470, 0x2470, 0x2472, 0x2472,
399         0x2474, 0x2475, 0x2479, 0x247a, 0x24c0, 0x24d3, 0x24e4, 0x24ef,
400         0x2500, 0x2509, 0x250c, 0x250c, 0x250e, 0x250e, 0x2510, 0x2511,
401         0x2514, 0x2515, 0x25e4, 0x25e4, 0x25ea, 0x25ea, 0x25ec, 0x25ed,
402         0x25f0, 0x25f0, 0x2600, 0x2612, 0x2614, 0x2617, 0x261a, 0x261a,
403         0x2640, 0x267e, 0x2680, 0x268b, 0x26c0, 0x26c0, 0x26c4, 0x26ce,
404         0x26d0, 0x26d8, 0x26df, 0x26e6, 0x26e8, 0x26e9, 0x26ec, 0x26ec,
405         0x26f0, 0x26f7, 0x26ff, 0x26ff, 0x2740, 0x2743, 0x2748, 0x2749,
406         0x2750, 0x2756, 0x2760, 0x2760, 0x300c, 0x300e, 0x301c, 0x301d,
407         0x302a, 0x302a, 0x302c, 0x302d, 0x3030, 0x3031, 0x3034, 0x3036,
408         0x303c, 0x303c, 0x305e, 0x305f,
409         ~0   /* sentinel */
410 };
411
412 #ifdef CONFIG_DEBUG_FS
413 static void a3xx_show(struct msm_gpu *gpu, struct seq_file *m)
414 {
415         gpu->funcs->pm_resume(gpu);
416         seq_printf(m, "status:   %08x\n",
417                         gpu_read(gpu, REG_A3XX_RBBM_STATUS));
418         gpu->funcs->pm_suspend(gpu);
419         adreno_show(gpu, m);
420 }
421 #endif
422
423 /* would be nice to not have to duplicate the _show() stuff with printk(): */
424 static void a3xx_dump(struct msm_gpu *gpu)
425 {
426         printk("status:   %08x\n",
427                         gpu_read(gpu, REG_A3XX_RBBM_STATUS));
428         adreno_dump(gpu);
429 }
430 /* Register offset defines for A3XX */
431 static const unsigned int a3xx_register_offsets[REG_ADRENO_REGISTER_MAX] = {
432         REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_BASE, REG_AXXX_CP_RB_BASE),
433         REG_ADRENO_SKIP(REG_ADRENO_CP_RB_BASE_HI),
434         REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR_ADDR, REG_AXXX_CP_RB_RPTR_ADDR),
435         REG_ADRENO_SKIP(REG_ADRENO_CP_RB_RPTR_ADDR_HI),
436         REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_RPTR, REG_AXXX_CP_RB_RPTR),
437         REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_WPTR, REG_AXXX_CP_RB_WPTR),
438         REG_ADRENO_DEFINE(REG_ADRENO_CP_RB_CNTL, REG_AXXX_CP_RB_CNTL),
439 };
440
441 static const struct adreno_gpu_funcs funcs = {
442         .base = {
443                 .get_param = adreno_get_param,
444                 .hw_init = a3xx_hw_init,
445                 .pm_suspend = msm_gpu_pm_suspend,
446                 .pm_resume = msm_gpu_pm_resume,
447                 .recover = a3xx_recover,
448                 .last_fence = adreno_last_fence,
449                 .submit = adreno_submit,
450                 .flush = adreno_flush,
451                 .idle = a3xx_idle,
452                 .irq = a3xx_irq,
453                 .destroy = a3xx_destroy,
454 #ifdef CONFIG_DEBUG_FS
455                 .show = a3xx_show,
456 #endif
457         },
458 };
459
460 static const struct msm_gpu_perfcntr perfcntrs[] = {
461         { REG_A3XX_SP_PERFCOUNTER6_SELECT, REG_A3XX_RBBM_PERFCTR_SP_6_LO,
462                         SP_ALU_ACTIVE_CYCLES, "ALUACTIVE" },
463         { REG_A3XX_SP_PERFCOUNTER7_SELECT, REG_A3XX_RBBM_PERFCTR_SP_7_LO,
464                         SP_FS_FULL_ALU_INSTRUCTIONS, "ALUFULL" },
465 };
466
467 struct msm_gpu *a3xx_gpu_init(struct drm_device *dev)
468 {
469         struct a3xx_gpu *a3xx_gpu = NULL;
470         struct adreno_gpu *adreno_gpu;
471         struct msm_gpu *gpu;
472         struct msm_drm_private *priv = dev->dev_private;
473         struct platform_device *pdev = priv->gpu_pdev;
474         int ret;
475
476         if (!pdev) {
477                 dev_err(dev->dev, "no a3xx device\n");
478                 ret = -ENXIO;
479                 goto fail;
480         }
481
482         a3xx_gpu = kzalloc(sizeof(*a3xx_gpu), GFP_KERNEL);
483         if (!a3xx_gpu) {
484                 ret = -ENOMEM;
485                 goto fail;
486         }
487
488         adreno_gpu = &a3xx_gpu->base;
489         gpu = &adreno_gpu->base;
490
491         a3xx_gpu->pdev = pdev;
492
493         gpu->perfcntrs = perfcntrs;
494         gpu->num_perfcntrs = ARRAY_SIZE(perfcntrs);
495
496         adreno_gpu->registers = a3xx_registers;
497         adreno_gpu->reg_offsets = a3xx_register_offsets;
498
499         ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs);
500         if (ret)
501                 goto fail;
502
503         /* if needed, allocate gmem: */
504         if (adreno_is_a330(adreno_gpu)) {
505 #ifdef CONFIG_MSM_OCMEM
506                 /* TODO this is different/missing upstream: */
507                 struct ocmem_buf *ocmem_hdl =
508                                 ocmem_allocate(OCMEM_GRAPHICS, adreno_gpu->gmem);
509
510                 a3xx_gpu->ocmem_hdl = ocmem_hdl;
511                 a3xx_gpu->ocmem_base = ocmem_hdl->addr;
512                 adreno_gpu->gmem = ocmem_hdl->len;
513                 DBG("using %dK of OCMEM at 0x%08x", adreno_gpu->gmem / 1024,
514                                 a3xx_gpu->ocmem_base);
515 #endif
516         }
517
518         if (!gpu->aspace) {
519                 /* TODO we think it is possible to configure the GPU to
520                  * restrict access to VRAM carveout.  But the required
521                  * registers are unknown.  For now just bail out and
522                  * limp along with just modesetting.  If it turns out
523                  * to not be possible to restrict access, then we must
524                  * implement a cmdstream validator.
525                  */
526                 dev_err(dev->dev, "No memory protection without IOMMU\n");
527                 ret = -ENXIO;
528                 goto fail;
529         }
530
531         return gpu;
532
533 fail:
534         if (a3xx_gpu)
535                 a3xx_destroy(&a3xx_gpu->base.base);
536
537         return ERR_PTR(ret);
538 }