1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017-2019 The Linux Foundation. All rights reserved. */
3
4 #include <linux/clk.h>
5 #include <linux/interconnect.h>
6 #include <linux/of_platform.h>
7 #include <linux/platform_device.h>
8 #include <linux/pm_domain.h>
9 #include <linux/pm_opp.h>
10 #include <soc/qcom/cmd-db.h>
11 #include <drm/drm_gem.h>
12
13 #include "a6xx_gpu.h"
14 #include "a6xx_gmu.xml.h"
15 #include "msm_gem.h"
16 #include "msm_gpu_trace.h"
17 #include "msm_mmu.h"
18
a6xx_gmu_fault(struct a6xx_gmu * gmu)19 static void a6xx_gmu_fault(struct a6xx_gmu *gmu)
20 {
21 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
22 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
23 struct msm_gpu *gpu = &adreno_gpu->base;
24
25 /* FIXME: add a banner here */
26 gmu->hung = true;
27
28 /* Turn off the hangcheck timer while we are resetting */
29 del_timer(&gpu->hangcheck_timer);
30
31 /* Queue the GPU handler because we need to treat this as a recovery */
32 kthread_queue_work(gpu->worker, &gpu->recover_work);
33 }
34
a6xx_gmu_irq(int irq,void * data)35 static irqreturn_t a6xx_gmu_irq(int irq, void *data)
36 {
37 struct a6xx_gmu *gmu = data;
38 u32 status;
39
40 status = gmu_read(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_STATUS);
41 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, status);
42
43 if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE) {
44 dev_err_ratelimited(gmu->dev, "GMU watchdog expired\n");
45
46 a6xx_gmu_fault(gmu);
47 }
48
49 if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR)
50 dev_err_ratelimited(gmu->dev, "GMU AHB bus error\n");
51
52 if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
53 dev_err_ratelimited(gmu->dev, "GMU fence error: 0x%x\n",
54 gmu_read(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS));
55
56 return IRQ_HANDLED;
57 }
58
a6xx_hfi_irq(int irq,void * data)59 static irqreturn_t a6xx_hfi_irq(int irq, void *data)
60 {
61 struct a6xx_gmu *gmu = data;
62 u32 status;
63
64 status = gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO);
65 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, status);
66
67 if (status & A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) {
68 dev_err_ratelimited(gmu->dev, "GMU firmware fault\n");
69
70 a6xx_gmu_fault(gmu);
71 }
72
73 return IRQ_HANDLED;
74 }
75
a6xx_gmu_sptprac_is_on(struct a6xx_gmu * gmu)76 bool a6xx_gmu_sptprac_is_on(struct a6xx_gmu *gmu)
77 {
78 u32 val;
79
80 /* This can be called from gpu state code so make sure GMU is valid */
81 if (!gmu->initialized)
82 return false;
83
84 val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
85
86 return !(val &
87 (A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SPTPRAC_GDSC_POWER_OFF |
88 A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SP_CLOCK_OFF));
89 }
90
91 /* Check to see if the GX rail is still powered */
a6xx_gmu_gx_is_on(struct a6xx_gmu * gmu)92 bool a6xx_gmu_gx_is_on(struct a6xx_gmu *gmu)
93 {
94 u32 val;
95
96 /* This can be called from gpu state code so make sure GMU is valid */
97 if (!gmu->initialized)
98 return false;
99
100 val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
101
102 return !(val &
103 (A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_GDSC_POWER_OFF |
104 A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_CLK_OFF));
105 }
106
a6xx_gmu_set_freq(struct msm_gpu * gpu,struct dev_pm_opp * opp,bool suspended)107 void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp,
108 bool suspended)
109 {
110 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
111 struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
112 struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
113 u32 perf_index;
114 unsigned long gpu_freq;
115 int ret = 0;
116
117 gpu_freq = dev_pm_opp_get_freq(opp);
118
119 if (gpu_freq == gmu->freq)
120 return;
121
122 for (perf_index = 0; perf_index < gmu->nr_gpu_freqs - 1; perf_index++)
123 if (gpu_freq == gmu->gpu_freqs[perf_index])
124 break;
125
126 gmu->current_perf_index = perf_index;
127 gmu->freq = gmu->gpu_freqs[perf_index];
128
129 trace_msm_gmu_freq_change(gmu->freq, perf_index);
130
131 /*
132 * This can get called from devfreq while the hardware is idle. Don't
133 * bring up the power if it isn't already active. All we're doing here
134 * is updating the frequency so that when we come back online we're at
135 * the right rate.
136 */
137 if (suspended)
138 return;
139
140 if (!gmu->legacy) {
141 a6xx_hfi_set_freq(gmu, perf_index);
142 dev_pm_opp_set_opp(&gpu->pdev->dev, opp);
143 return;
144 }
145
146 gmu_write(gmu, REG_A6XX_GMU_DCVS_ACK_OPTION, 0);
147
148 gmu_write(gmu, REG_A6XX_GMU_DCVS_PERF_SETTING,
149 ((3 & 0xf) << 28) | perf_index);
150
151 /*
152 * Send an invalid index as a vote for the bus bandwidth and let the
153 * firmware decide on the right vote
154 */
155 gmu_write(gmu, REG_A6XX_GMU_DCVS_BW_SETTING, 0xff);
156
157 /* Set and clear the OOB for DCVS to trigger the GMU */
158 a6xx_gmu_set_oob(gmu, GMU_OOB_DCVS_SET);
159 a6xx_gmu_clear_oob(gmu, GMU_OOB_DCVS_SET);
160
161 ret = gmu_read(gmu, REG_A6XX_GMU_DCVS_RETURN);
162 if (ret)
163 dev_err(gmu->dev, "GMU set GPU frequency error: %d\n", ret);
164
165 dev_pm_opp_set_opp(&gpu->pdev->dev, opp);
166 }
167
a6xx_gmu_get_freq(struct msm_gpu * gpu)168 unsigned long a6xx_gmu_get_freq(struct msm_gpu *gpu)
169 {
170 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
171 struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
172 struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
173
174 return gmu->freq;
175 }
176
a6xx_gmu_check_idle_level(struct a6xx_gmu * gmu)177 static bool a6xx_gmu_check_idle_level(struct a6xx_gmu *gmu)
178 {
179 u32 val;
180 int local = gmu->idle_level;
181
182 /* SPTP and IFPC both report as IFPC */
183 if (gmu->idle_level == GMU_IDLE_STATE_SPTP)
184 local = GMU_IDLE_STATE_IFPC;
185
186 val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
187
188 if (val == local) {
189 if (gmu->idle_level != GMU_IDLE_STATE_IFPC ||
190 !a6xx_gmu_gx_is_on(gmu))
191 return true;
192 }
193
194 return false;
195 }
196
197 /* Wait for the GMU to get to its most idle state */
a6xx_gmu_wait_for_idle(struct a6xx_gmu * gmu)198 int a6xx_gmu_wait_for_idle(struct a6xx_gmu *gmu)
199 {
200 return spin_until(a6xx_gmu_check_idle_level(gmu));
201 }
202
a6xx_gmu_start(struct a6xx_gmu * gmu)203 static int a6xx_gmu_start(struct a6xx_gmu *gmu)
204 {
205 int ret;
206 u32 val;
207 u32 mask, reset_val;
208
209 val = gmu_read(gmu, REG_A6XX_GMU_CM3_DTCM_START + 0xff8);
210 if (val <= 0x20010004) {
211 mask = 0xffffffff;
212 reset_val = 0xbabeface;
213 } else {
214 mask = 0x1ff;
215 reset_val = 0x100;
216 }
217
218 gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1);
219
220 /* Set the log wptr index
221 * note: downstream saves the value in poweroff and restores it here
222 */
223 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_RESP, 0);
224
225 gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 0);
226
227 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, val,
228 (val & mask) == reset_val, 100, 10000);
229
230 if (ret)
231 DRM_DEV_ERROR(gmu->dev, "GMU firmware initialization timed out\n");
232
233 return ret;
234 }
235
a6xx_gmu_hfi_start(struct a6xx_gmu * gmu)236 static int a6xx_gmu_hfi_start(struct a6xx_gmu *gmu)
237 {
238 u32 val;
239 int ret;
240
241 gmu_write(gmu, REG_A6XX_GMU_HFI_CTRL_INIT, 1);
242
243 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_HFI_CTRL_STATUS, val,
244 val & 1, 100, 10000);
245 if (ret)
246 DRM_DEV_ERROR(gmu->dev, "Unable to start the HFI queues\n");
247
248 return ret;
249 }
250
251 struct a6xx_gmu_oob_bits {
252 int set, ack, set_new, ack_new, clear, clear_new;
253 const char *name;
254 };
255
256 /* These are the interrupt / ack bits for each OOB request that are set
257 * in a6xx_gmu_set_oob and a6xx_clear_oob
258 */
259 static const struct a6xx_gmu_oob_bits a6xx_gmu_oob_bits[] = {
260 [GMU_OOB_GPU_SET] = {
261 .name = "GPU_SET",
262 .set = 16,
263 .ack = 24,
264 .set_new = 30,
265 .ack_new = 31,
266 .clear = 24,
267 .clear_new = 31,
268 },
269
270 [GMU_OOB_PERFCOUNTER_SET] = {
271 .name = "PERFCOUNTER",
272 .set = 17,
273 .ack = 25,
274 .set_new = 28,
275 .ack_new = 30,
276 .clear = 25,
277 .clear_new = 29,
278 },
279
280 [GMU_OOB_BOOT_SLUMBER] = {
281 .name = "BOOT_SLUMBER",
282 .set = 22,
283 .ack = 30,
284 .clear = 30,
285 },
286
287 [GMU_OOB_DCVS_SET] = {
288 .name = "GPU_DCVS",
289 .set = 23,
290 .ack = 31,
291 .clear = 31,
292 },
293 };
294
295 /* Trigger a OOB (out of band) request to the GMU */
a6xx_gmu_set_oob(struct a6xx_gmu * gmu,enum a6xx_gmu_oob_state state)296 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
297 {
298 int ret;
299 u32 val;
300 int request, ack;
301
302 WARN_ON_ONCE(!mutex_is_locked(&gmu->lock));
303
304 if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits))
305 return -EINVAL;
306
307 if (gmu->legacy) {
308 request = a6xx_gmu_oob_bits[state].set;
309 ack = a6xx_gmu_oob_bits[state].ack;
310 } else {
311 request = a6xx_gmu_oob_bits[state].set_new;
312 ack = a6xx_gmu_oob_bits[state].ack_new;
313 if (!request || !ack) {
314 DRM_DEV_ERROR(gmu->dev,
315 "Invalid non-legacy GMU request %s\n",
316 a6xx_gmu_oob_bits[state].name);
317 return -EINVAL;
318 }
319 }
320
321 /* Trigger the equested OOB operation */
322 gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request);
323
324 /* Wait for the acknowledge interrupt */
325 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
326 val & (1 << ack), 100, 10000);
327
328 if (ret)
329 DRM_DEV_ERROR(gmu->dev,
330 "Timeout waiting for GMU OOB set %s: 0x%x\n",
331 a6xx_gmu_oob_bits[state].name,
332 gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
333
334 /* Clear the acknowledge interrupt */
335 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack);
336
337 return ret;
338 }
339
340 /* Clear a pending OOB state in the GMU */
a6xx_gmu_clear_oob(struct a6xx_gmu * gmu,enum a6xx_gmu_oob_state state)341 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
342 {
343 int bit;
344
345 WARN_ON_ONCE(!mutex_is_locked(&gmu->lock));
346
347 if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits))
348 return;
349
350 if (gmu->legacy)
351 bit = a6xx_gmu_oob_bits[state].clear;
352 else
353 bit = a6xx_gmu_oob_bits[state].clear_new;
354
355 gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << bit);
356 }
357
358 /* Enable CPU control of SPTP power power collapse */
a6xx_sptprac_enable(struct a6xx_gmu * gmu)359 int a6xx_sptprac_enable(struct a6xx_gmu *gmu)
360 {
361 int ret;
362 u32 val;
363
364 if (!gmu->legacy)
365 return 0;
366
367 gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000);
368
369 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
370 (val & 0x38) == 0x28, 1, 100);
371
372 if (ret) {
373 DRM_DEV_ERROR(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n",
374 gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
375 }
376
377 return 0;
378 }
379
380 /* Disable CPU control of SPTP power power collapse */
a6xx_sptprac_disable(struct a6xx_gmu * gmu)381 void a6xx_sptprac_disable(struct a6xx_gmu *gmu)
382 {
383 u32 val;
384 int ret;
385
386 if (!gmu->legacy)
387 return;
388
389 /* Make sure retention is on */
390 gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11));
391
392 gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001);
393
394 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
395 (val & 0x04), 100, 10000);
396
397 if (ret)
398 DRM_DEV_ERROR(gmu->dev, "failed to power off SPTPRAC: 0x%x\n",
399 gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
400 }
401
402 /* Let the GMU know we are starting a boot sequence */
a6xx_gmu_gfx_rail_on(struct a6xx_gmu * gmu)403 static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu)
404 {
405 u32 vote;
406
407 /* Let the GMU know we are getting ready for boot */
408 gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0);
409
410 /* Choose the "default" power level as the highest available */
411 vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1];
412
413 gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff);
414 gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff);
415
416 /* Let the GMU know the boot sequence has started */
417 return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
418 }
419
420 /* Let the GMU know that we are about to go into slumber */
a6xx_gmu_notify_slumber(struct a6xx_gmu * gmu)421 static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu)
422 {
423 int ret;
424
425 /* Disable the power counter so the GMU isn't busy */
426 gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0);
427
428 /* Disable SPTP_PC if the CPU is responsible for it */
429 if (gmu->idle_level < GMU_IDLE_STATE_SPTP)
430 a6xx_sptprac_disable(gmu);
431
432 if (!gmu->legacy) {
433 ret = a6xx_hfi_send_prep_slumber(gmu);
434 goto out;
435 }
436
437 /* Tell the GMU to get ready to slumber */
438 gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1);
439
440 ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
441 a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
442
443 if (!ret) {
444 /* Check to see if the GMU really did slumber */
445 if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE)
446 != 0x0f) {
447 DRM_DEV_ERROR(gmu->dev, "The GMU did not go into slumber\n");
448 ret = -ETIMEDOUT;
449 }
450 }
451
452 out:
453 /* Put fence into allow mode */
454 gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
455 return ret;
456 }
457
a6xx_rpmh_start(struct a6xx_gmu * gmu)458 static int a6xx_rpmh_start(struct a6xx_gmu *gmu)
459 {
460 int ret;
461 u32 val;
462
463 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1);
464 /* Wait for the register to finish posting */
465 wmb();
466
467 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val,
468 val & (1 << 1), 100, 10000);
469 if (ret) {
470 DRM_DEV_ERROR(gmu->dev, "Unable to power on the GPU RSC\n");
471 return ret;
472 }
473
474 ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val,
475 !val, 100, 10000);
476
477 if (ret) {
478 DRM_DEV_ERROR(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n");
479 return ret;
480 }
481
482 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
483
484 return 0;
485 }
486
a6xx_rpmh_stop(struct a6xx_gmu * gmu)487 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu)
488 {
489 int ret;
490 u32 val;
491
492 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1);
493
494 ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0,
495 val, val & (1 << 16), 100, 10000);
496 if (ret)
497 DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n");
498
499 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
500 }
501
pdc_write(void __iomem * ptr,u32 offset,u32 value)502 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value)
503 {
504 msm_writel(value, ptr + (offset << 2));
505 }
506
507 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
508 const char *name);
509
a6xx_gmu_rpmh_init(struct a6xx_gmu * gmu)510 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
511 {
512 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
513 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
514 struct platform_device *pdev = to_platform_device(gmu->dev);
515 void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc");
516 void __iomem *seqptr = NULL;
517 uint32_t pdc_address_offset;
518 bool pdc_in_aop = false;
519
520 if (IS_ERR(pdcptr))
521 goto err;
522
523 if (adreno_is_a650(adreno_gpu) || adreno_is_a660_family(adreno_gpu))
524 pdc_in_aop = true;
525 else if (adreno_is_a618(adreno_gpu) || adreno_is_a640_family(adreno_gpu))
526 pdc_address_offset = 0x30090;
527 else if (adreno_is_a619(adreno_gpu))
528 pdc_address_offset = 0x300a0;
529 else
530 pdc_address_offset = 0x30080;
531
532 if (!pdc_in_aop) {
533 seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq");
534 if (IS_ERR(seqptr))
535 goto err;
536 }
537
538 /* Disable SDE clock gating */
539 gmu_write_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24));
540
541 /* Setup RSC PDC handshake for sleep and wakeup */
542 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1);
543 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0);
544 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0);
545 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0);
546 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0);
547 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000);
548 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0);
549 gmu_write_rscc(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0);
550 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520);
551 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510);
552 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514);
553
554 /* Load RSC sequencer uCode for sleep and wakeup */
555 if (adreno_is_a650_family(adreno_gpu)) {
556 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xeaaae5a0);
557 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xe1a1ebab);
558 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e0a581);
559 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xecac82e2);
560 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020edad);
561 } else {
562 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0);
563 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7);
564 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1);
565 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2);
566 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8);
567 }
568
569 if (pdc_in_aop)
570 goto setup_pdc;
571
572 /* Load PDC sequencer uCode for power up and power down sequence */
573 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1);
574 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2);
575 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0);
576 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284);
577 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc);
578
579 /* Set TCS commands used by PDC sequence for low power modes */
580 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7);
581 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0);
582 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0);
583 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108);
584 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010);
585 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1);
586 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108);
587 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000);
588 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0);
589
590 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108);
591 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, pdc_address_offset);
592 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0);
593
594 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7);
595 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0);
596 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0);
597 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108);
598 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010);
599 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2);
600
601 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108);
602 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000);
603 if (adreno_is_a618(adreno_gpu) || adreno_is_a619(adreno_gpu) ||
604 adreno_is_a650_family(adreno_gpu))
605 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2);
606 else
607 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3);
608 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108);
609 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, pdc_address_offset);
610 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3);
611
612 /* Setup GPU PDC */
613 setup_pdc:
614 pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0);
615 pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001);
616
617 /* ensure no writes happen before the uCode is fully written */
618 wmb();
619
620 a6xx_rpmh_stop(gmu);
621
622 err:
623 if (!IS_ERR_OR_NULL(pdcptr))
624 iounmap(pdcptr);
625 if (!IS_ERR_OR_NULL(seqptr))
626 iounmap(seqptr);
627 }
628
629 /*
630 * The lowest 16 bits of this value are the number of XO clock cycles for main
631 * hysteresis which is set at 0x1680 cycles (300 us). The higher 16 bits are
632 * for the shorter hysteresis that happens after main - this is 0xa (.5 us)
633 */
634
635 #define GMU_PWR_COL_HYST 0x000a1680
636
637 /* Set up the idle state for the GMU */
a6xx_gmu_power_config(struct a6xx_gmu * gmu)638 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu)
639 {
640 /* Disable GMU WB/RB buffer */
641 gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1);
642 gmu_write(gmu, REG_A6XX_GMU_ICACHE_CONFIG, 0x1);
643 gmu_write(gmu, REG_A6XX_GMU_DCACHE_CONFIG, 0x1);
644
645 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400);
646
647 switch (gmu->idle_level) {
648 case GMU_IDLE_STATE_IFPC:
649 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST,
650 GMU_PWR_COL_HYST);
651 gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
652 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
653 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE);
654 fallthrough;
655 case GMU_IDLE_STATE_SPTP:
656 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST,
657 GMU_PWR_COL_HYST);
658 gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
659 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
660 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE);
661 }
662
663 /* Enable RPMh GPU client */
664 gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0,
665 A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE |
666 A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE |
667 A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE |
668 A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE |
669 A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE |
670 A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE);
671 }
672
673 struct block_header {
674 u32 addr;
675 u32 size;
676 u32 type;
677 u32 value;
678 u32 data[];
679 };
680
fw_block_mem(struct a6xx_gmu_bo * bo,const struct block_header * blk)681 static bool fw_block_mem(struct a6xx_gmu_bo *bo, const struct block_header *blk)
682 {
683 if (!in_range(blk->addr, bo->iova, bo->size))
684 return false;
685
686 memcpy(bo->virt + blk->addr - bo->iova, blk->data, blk->size);
687 return true;
688 }
689
a6xx_gmu_fw_load(struct a6xx_gmu * gmu)690 static int a6xx_gmu_fw_load(struct a6xx_gmu *gmu)
691 {
692 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
693 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
694 const struct firmware *fw_image = adreno_gpu->fw[ADRENO_FW_GMU];
695 const struct block_header *blk;
696 u32 reg_offset;
697
698 u32 itcm_base = 0x00000000;
699 u32 dtcm_base = 0x00040000;
700
701 if (adreno_is_a650_family(adreno_gpu))
702 dtcm_base = 0x10004000;
703
704 if (gmu->legacy) {
705 /* Sanity check the size of the firmware that was loaded */
706 if (fw_image->size > 0x8000) {
707 DRM_DEV_ERROR(gmu->dev,
708 "GMU firmware is bigger than the available region\n");
709 return -EINVAL;
710 }
711
712 gmu_write_bulk(gmu, REG_A6XX_GMU_CM3_ITCM_START,
713 (u32*) fw_image->data, fw_image->size);
714 return 0;
715 }
716
717
718 for (blk = (const struct block_header *) fw_image->data;
719 (const u8*) blk < fw_image->data + fw_image->size;
720 blk = (const struct block_header *) &blk->data[blk->size >> 2]) {
721 if (blk->size == 0)
722 continue;
723
724 if (in_range(blk->addr, itcm_base, SZ_16K)) {
725 reg_offset = (blk->addr - itcm_base) >> 2;
726 gmu_write_bulk(gmu,
727 REG_A6XX_GMU_CM3_ITCM_START + reg_offset,
728 blk->data, blk->size);
729 } else if (in_range(blk->addr, dtcm_base, SZ_16K)) {
730 reg_offset = (blk->addr - dtcm_base) >> 2;
731 gmu_write_bulk(gmu,
732 REG_A6XX_GMU_CM3_DTCM_START + reg_offset,
733 blk->data, blk->size);
734 } else if (!fw_block_mem(&gmu->icache, blk) &&
735 !fw_block_mem(&gmu->dcache, blk) &&
736 !fw_block_mem(&gmu->dummy, blk)) {
737 DRM_DEV_ERROR(gmu->dev,
738 "failed to match fw block (addr=%.8x size=%d data[0]=%.8x)\n",
739 blk->addr, blk->size, blk->data[0]);
740 }
741 }
742
743 return 0;
744 }
745
a6xx_gmu_fw_start(struct a6xx_gmu * gmu,unsigned int state)746 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state)
747 {
748 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
749 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
750 int ret;
751 u32 chipid;
752
753 if (adreno_is_a650_family(adreno_gpu)) {
754 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FALNEXT_INTF, 1);
755 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FAL_INTF, 1);
756 }
757
758 if (state == GMU_WARM_BOOT) {
759 ret = a6xx_rpmh_start(gmu);
760 if (ret)
761 return ret;
762 } else {
763 if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU],
764 "GMU firmware is not loaded\n"))
765 return -ENOENT;
766
767 /* Turn on register retention */
768 gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1);
769
770 ret = a6xx_rpmh_start(gmu);
771 if (ret)
772 return ret;
773
774 ret = a6xx_gmu_fw_load(gmu);
775 if (ret)
776 return ret;
777 }
778
779 gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0);
780 gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02);
781
782 /* Write the iova of the HFI table */
783 gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi.iova);
784 gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1);
785
786 gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0,
787 (1 << 31) | (0xa << 18) | (0xa0));
788
789 /*
790 * Snapshots toggle the NMI bit which will result in a jump to the NMI
791 * handler instead of __main. Set the M3 config value to avoid that.
792 */
793 gmu_write(gmu, REG_A6XX_GMU_CM3_CFG, 0x4052);
794
795 /*
796 * Note that the GMU has a slightly different layout for
797 * chip_id, for whatever reason, so a bit of massaging
798 * is needed. The upper 16b are the same, but minor and
799 * patchid are packed in four bits each with the lower
800 * 8b unused:
801 */
802 chipid = adreno_gpu->chip_id & 0xffff0000;
803 chipid |= (adreno_gpu->chip_id << 4) & 0xf000; /* minor */
804 chipid |= (adreno_gpu->chip_id << 8) & 0x0f00; /* patchid */
805
806 gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid);
807
808 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_MSG,
809 gmu->log.iova | (gmu->log.size / SZ_4K - 1));
810
811 /* Set up the lowest idle level on the GMU */
812 a6xx_gmu_power_config(gmu);
813
814 ret = a6xx_gmu_start(gmu);
815 if (ret)
816 return ret;
817
818 if (gmu->legacy) {
819 ret = a6xx_gmu_gfx_rail_on(gmu);
820 if (ret)
821 return ret;
822 }
823
824 /* Enable SPTP_PC if the CPU is responsible for it */
825 if (gmu->idle_level < GMU_IDLE_STATE_SPTP) {
826 ret = a6xx_sptprac_enable(gmu);
827 if (ret)
828 return ret;
829 }
830
831 ret = a6xx_gmu_hfi_start(gmu);
832 if (ret)
833 return ret;
834
835 /* FIXME: Do we need this wmb() here? */
836 wmb();
837
838 return 0;
839 }
840
841 #define A6XX_HFI_IRQ_MASK \
842 (A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT)
843
844 #define A6XX_GMU_IRQ_MASK \
845 (A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \
846 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \
847 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
848
a6xx_gmu_irq_disable(struct a6xx_gmu * gmu)849 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu)
850 {
851 disable_irq(gmu->gmu_irq);
852 disable_irq(gmu->hfi_irq);
853
854 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0);
855 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0);
856 }
857
a6xx_gmu_rpmh_off(struct a6xx_gmu * gmu)858 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu)
859 {
860 u32 val;
861
862 /* Make sure there are no outstanding RPMh votes */
863 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val,
864 (val & 1), 100, 10000);
865 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val,
866 (val & 1), 100, 10000);
867 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val,
868 (val & 1), 100, 10000);
869 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val,
870 (val & 1), 100, 1000);
871 }
872
873 /* Force the GMU off in case it isn't responsive */
a6xx_gmu_force_off(struct a6xx_gmu * gmu)874 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu)
875 {
876 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
877 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
878 struct msm_gpu *gpu = &adreno_gpu->base;
879
880 /*
881 * Turn off keep alive that might have been enabled by the hang
882 * interrupt
883 */
884 gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0);
885
886 /* Flush all the queues */
887 a6xx_hfi_stop(gmu);
888
889 /* Stop the interrupts */
890 a6xx_gmu_irq_disable(gmu);
891
892 /* Force off SPTP in case the GMU is managing it */
893 a6xx_sptprac_disable(gmu);
894
895 /* Make sure there are no outstanding RPMh votes */
896 a6xx_gmu_rpmh_off(gmu);
897
898 /* Clear the WRITEDROPPED fields and put fence into allow mode */
899 gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS_CLR, 0x7);
900 gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
901
902 /* Make sure the above writes go through */
903 wmb();
904
905 /* Halt the gmu cm3 core */
906 gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1);
907
908 a6xx_bus_clear_pending_transactions(adreno_gpu, true);
909
910 /* Reset GPU core blocks */
911 a6xx_gpu_sw_reset(gpu, true);
912 }
913
a6xx_gmu_set_initial_freq(struct msm_gpu * gpu,struct a6xx_gmu * gmu)914 static void a6xx_gmu_set_initial_freq(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
915 {
916 struct dev_pm_opp *gpu_opp;
917 unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
918
919 gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
920 if (IS_ERR(gpu_opp))
921 return;
922
923 gmu->freq = 0; /* so a6xx_gmu_set_freq() doesn't exit early */
924 a6xx_gmu_set_freq(gpu, gpu_opp, false);
925 dev_pm_opp_put(gpu_opp);
926 }
927
a6xx_gmu_set_initial_bw(struct msm_gpu * gpu,struct a6xx_gmu * gmu)928 static void a6xx_gmu_set_initial_bw(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
929 {
930 struct dev_pm_opp *gpu_opp;
931 unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
932
933 gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
934 if (IS_ERR(gpu_opp))
935 return;
936
937 dev_pm_opp_set_opp(&gpu->pdev->dev, gpu_opp);
938 dev_pm_opp_put(gpu_opp);
939 }
940
a6xx_gmu_resume(struct a6xx_gpu * a6xx_gpu)941 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
942 {
943 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
944 struct msm_gpu *gpu = &adreno_gpu->base;
945 struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
946 int status, ret;
947
948 if (WARN(!gmu->initialized, "The GMU is not set up yet\n"))
949 return -EINVAL;
950
951 gmu->hung = false;
952
953 /* Turn on the resources */
954 pm_runtime_get_sync(gmu->dev);
955
956 /*
957 * "enable" the GX power domain which won't actually do anything but it
958 * will make sure that the refcounting is correct in case we need to
959 * bring down the GX after a GMU failure
960 */
961 if (!IS_ERR_OR_NULL(gmu->gxpd))
962 pm_runtime_get_sync(gmu->gxpd);
963
964 /* Use a known rate to bring up the GMU */
965 clk_set_rate(gmu->core_clk, 200000000);
966 clk_set_rate(gmu->hub_clk, 150000000);
967 ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
968 if (ret) {
969 pm_runtime_put(gmu->gxpd);
970 pm_runtime_put(gmu->dev);
971 return ret;
972 }
973
974 /* Set the bus quota to a reasonable value for boot */
975 a6xx_gmu_set_initial_bw(gpu, gmu);
976
977 /* Enable the GMU interrupt */
978 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0);
979 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK);
980 enable_irq(gmu->gmu_irq);
981
982 /* Check to see if we are doing a cold or warm boot */
983 status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ?
984 GMU_WARM_BOOT : GMU_COLD_BOOT;
985
986 /*
987 * Warm boot path does not work on newer GPUs
988 * Presumably this is because icache/dcache regions must be restored
989 */
990 if (!gmu->legacy)
991 status = GMU_COLD_BOOT;
992
993 ret = a6xx_gmu_fw_start(gmu, status);
994 if (ret)
995 goto out;
996
997 ret = a6xx_hfi_start(gmu, status);
998 if (ret)
999 goto out;
1000
1001 /*
1002 * Turn on the GMU firmware fault interrupt after we know the boot
1003 * sequence is successful
1004 */
1005 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0);
1006 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK);
1007 enable_irq(gmu->hfi_irq);
1008
1009 /* Set the GPU to the current freq */
1010 a6xx_gmu_set_initial_freq(gpu, gmu);
1011
1012 out:
1013 /* On failure, shut down the GMU to leave it in a good state */
1014 if (ret) {
1015 disable_irq(gmu->gmu_irq);
1016 a6xx_rpmh_stop(gmu);
1017 pm_runtime_put(gmu->gxpd);
1018 pm_runtime_put(gmu->dev);
1019 }
1020
1021 return ret;
1022 }
1023
a6xx_gmu_isidle(struct a6xx_gmu * gmu)1024 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
1025 {
1026 u32 reg;
1027
1028 if (!gmu->initialized)
1029 return true;
1030
1031 reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS);
1032
1033 if (reg & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB)
1034 return false;
1035
1036 return true;
1037 }
1038
1039 /* Gracefully try to shut down the GMU and by extension the GPU */
a6xx_gmu_shutdown(struct a6xx_gmu * gmu)1040 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
1041 {
1042 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1043 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1044 u32 val;
1045 int ret;
1046
1047 /*
1048 * GMU firmware's internal power state gets messed up if we send "prepare_slumber" hfi when
1049 * oob_gpu handshake wasn't done after the last wake up. So do a dummy handshake here when
1050 * required
1051 */
1052 if (adreno_gpu->base.needs_hw_init) {
1053 if (a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET))
1054 goto force_off;
1055
1056 a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET);
1057 }
1058
1059 ret = a6xx_gmu_wait_for_idle(gmu);
1060
1061 /* If the GMU isn't responding assume it is hung */
1062 if (ret)
1063 goto force_off;
1064
1065 a6xx_bus_clear_pending_transactions(adreno_gpu, a6xx_gpu->hung);
1066
1067 /* tell the GMU we want to slumber */
1068 ret = a6xx_gmu_notify_slumber(gmu);
1069 if (ret)
1070 goto force_off;
1071
1072 ret = gmu_poll_timeout(gmu,
1073 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
1074 !(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
1075 100, 10000);
1076
1077 /*
1078 * Let the user know we failed to slumber but don't worry too
1079 * much because we are powering down anyway
1080 */
1081
1082 if (ret)
1083 DRM_DEV_ERROR(gmu->dev,
1084 "Unable to slumber GMU: status = 0%x/0%x\n",
1085 gmu_read(gmu,
1086 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS),
1087 gmu_read(gmu,
1088 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
1089
1090 /* Turn off HFI */
1091 a6xx_hfi_stop(gmu);
1092
1093 /* Stop the interrupts and mask the hardware */
1094 a6xx_gmu_irq_disable(gmu);
1095
1096 /* Tell RPMh to power off the GPU */
1097 a6xx_rpmh_stop(gmu);
1098
1099 return;
1100
1101 force_off:
1102 a6xx_gmu_force_off(gmu);
1103 }
1104
1105
a6xx_gmu_stop(struct a6xx_gpu * a6xx_gpu)1106 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu)
1107 {
1108 struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1109 struct msm_gpu *gpu = &a6xx_gpu->base.base;
1110
1111 if (!pm_runtime_active(gmu->dev))
1112 return 0;
1113
1114 /*
1115 * Force the GMU off if we detected a hang, otherwise try to shut it
1116 * down gracefully
1117 */
1118 if (gmu->hung)
1119 a6xx_gmu_force_off(gmu);
1120 else
1121 a6xx_gmu_shutdown(gmu);
1122
1123 /* Remove the bus vote */
1124 dev_pm_opp_set_opp(&gpu->pdev->dev, NULL);
1125
1126 /*
1127 * Make sure the GX domain is off before turning off the GMU (CX)
1128 * domain. Usually the GMU does this but only if the shutdown sequence
1129 * was successful
1130 */
1131 if (!IS_ERR_OR_NULL(gmu->gxpd))
1132 pm_runtime_put_sync(gmu->gxpd);
1133
1134 clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
1135
1136 pm_runtime_put_sync(gmu->dev);
1137
1138 return 0;
1139 }
1140
a6xx_gmu_memory_free(struct a6xx_gmu * gmu)1141 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu)
1142 {
1143 msm_gem_kernel_put(gmu->hfi.obj, gmu->aspace);
1144 msm_gem_kernel_put(gmu->debug.obj, gmu->aspace);
1145 msm_gem_kernel_put(gmu->icache.obj, gmu->aspace);
1146 msm_gem_kernel_put(gmu->dcache.obj, gmu->aspace);
1147 msm_gem_kernel_put(gmu->dummy.obj, gmu->aspace);
1148 msm_gem_kernel_put(gmu->log.obj, gmu->aspace);
1149
1150 gmu->aspace->mmu->funcs->detach(gmu->aspace->mmu);
1151 msm_gem_address_space_put(gmu->aspace);
1152 }
1153
a6xx_gmu_memory_alloc(struct a6xx_gmu * gmu,struct a6xx_gmu_bo * bo,size_t size,u64 iova,const char * name)1154 static int a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo,
1155 size_t size, u64 iova, const char *name)
1156 {
1157 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1158 struct drm_device *dev = a6xx_gpu->base.base.dev;
1159 uint32_t flags = MSM_BO_WC;
1160 u64 range_start, range_end;
1161 int ret;
1162
1163 size = PAGE_ALIGN(size);
1164 if (!iova) {
1165 /* no fixed address - use GMU's uncached range */
1166 range_start = 0x60000000 + PAGE_SIZE; /* skip dummy page */
1167 range_end = 0x80000000;
1168 } else {
1169 /* range for fixed address */
1170 range_start = iova;
1171 range_end = iova + size;
1172 /* use IOMMU_PRIV for icache/dcache */
1173 flags |= MSM_BO_MAP_PRIV;
1174 }
1175
1176 bo->obj = msm_gem_new(dev, size, flags);
1177 if (IS_ERR(bo->obj))
1178 return PTR_ERR(bo->obj);
1179
1180 ret = msm_gem_get_and_pin_iova_range(bo->obj, gmu->aspace, &bo->iova,
1181 range_start, range_end);
1182 if (ret) {
1183 drm_gem_object_put(bo->obj);
1184 return ret;
1185 }
1186
1187 bo->virt = msm_gem_get_vaddr(bo->obj);
1188 bo->size = size;
1189
1190 msm_gem_object_set_name(bo->obj, name);
1191
1192 return 0;
1193 }
1194
a6xx_gmu_memory_probe(struct a6xx_gmu * gmu)1195 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu)
1196 {
1197 struct msm_mmu *mmu;
1198
1199 mmu = msm_iommu_new(gmu->dev, 0);
1200 if (!mmu)
1201 return -ENODEV;
1202 if (IS_ERR(mmu))
1203 return PTR_ERR(mmu);
1204
1205 gmu->aspace = msm_gem_address_space_create(mmu, "gmu", 0x0, 0x80000000);
1206 if (IS_ERR(gmu->aspace))
1207 return PTR_ERR(gmu->aspace);
1208
1209 return 0;
1210 }
1211
1212 /* Return the 'arc-level' for the given frequency */
a6xx_gmu_get_arc_level(struct device * dev,unsigned long freq)1213 static unsigned int a6xx_gmu_get_arc_level(struct device *dev,
1214 unsigned long freq)
1215 {
1216 struct dev_pm_opp *opp;
1217 unsigned int val;
1218
1219 if (!freq)
1220 return 0;
1221
1222 opp = dev_pm_opp_find_freq_exact(dev, freq, true);
1223 if (IS_ERR(opp))
1224 return 0;
1225
1226 val = dev_pm_opp_get_level(opp);
1227
1228 dev_pm_opp_put(opp);
1229
1230 return val;
1231 }
1232
a6xx_gmu_rpmh_arc_votes_init(struct device * dev,u32 * votes,unsigned long * freqs,int freqs_count,const char * id)1233 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
1234 unsigned long *freqs, int freqs_count, const char *id)
1235 {
1236 int i, j;
1237 const u16 *pri, *sec;
1238 size_t pri_count, sec_count;
1239
1240 pri = cmd_db_read_aux_data(id, &pri_count);
1241 if (IS_ERR(pri))
1242 return PTR_ERR(pri);
1243 /*
1244 * The data comes back as an array of unsigned shorts so adjust the
1245 * count accordingly
1246 */
1247 pri_count >>= 1;
1248 if (!pri_count)
1249 return -EINVAL;
1250
1251 sec = cmd_db_read_aux_data("mx.lvl", &sec_count);
1252 if (IS_ERR(sec))
1253 return PTR_ERR(sec);
1254
1255 sec_count >>= 1;
1256 if (!sec_count)
1257 return -EINVAL;
1258
1259 /* Construct a vote for each frequency */
1260 for (i = 0; i < freqs_count; i++) {
1261 u8 pindex = 0, sindex = 0;
1262 unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]);
1263
1264 /* Get the primary index that matches the arc level */
1265 for (j = 0; j < pri_count; j++) {
1266 if (pri[j] >= level) {
1267 pindex = j;
1268 break;
1269 }
1270 }
1271
1272 if (j == pri_count) {
1273 DRM_DEV_ERROR(dev,
1274 "Level %u not found in the RPMh list\n",
1275 level);
1276 DRM_DEV_ERROR(dev, "Available levels:\n");
1277 for (j = 0; j < pri_count; j++)
1278 DRM_DEV_ERROR(dev, " %u\n", pri[j]);
1279
1280 return -EINVAL;
1281 }
1282
1283 /*
1284 * Look for a level in in the secondary list that matches. If
1285 * nothing fits, use the maximum non zero vote
1286 */
1287
1288 for (j = 0; j < sec_count; j++) {
1289 if (sec[j] >= level) {
1290 sindex = j;
1291 break;
1292 } else if (sec[j]) {
1293 sindex = j;
1294 }
1295 }
1296
1297 /* Construct the vote */
1298 votes[i] = ((pri[pindex] & 0xffff) << 16) |
1299 (sindex << 8) | pindex;
1300 }
1301
1302 return 0;
1303 }
1304
1305 /*
1306 * The GMU votes with the RPMh for itself and on behalf of the GPU but we need
1307 * to construct the list of votes on the CPU and send it over. Query the RPMh
1308 * voltage levels and build the votes
1309 */
1310
a6xx_gmu_rpmh_votes_init(struct a6xx_gmu * gmu)1311 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
1312 {
1313 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1314 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1315 struct msm_gpu *gpu = &adreno_gpu->base;
1316 int ret;
1317
1318 /* Build the GX votes */
1319 ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes,
1320 gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl");
1321
1322 /* Build the CX votes */
1323 ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
1324 gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl");
1325
1326 return ret;
1327 }
1328
a6xx_gmu_build_freq_table(struct device * dev,unsigned long * freqs,u32 size)1329 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs,
1330 u32 size)
1331 {
1332 int count = dev_pm_opp_get_opp_count(dev);
1333 struct dev_pm_opp *opp;
1334 int i, index = 0;
1335 unsigned long freq = 1;
1336
1337 /*
1338 * The OPP table doesn't contain the "off" frequency level so we need to
1339 * add 1 to the table size to account for it
1340 */
1341
1342 if (WARN(count + 1 > size,
1343 "The GMU frequency table is being truncated\n"))
1344 count = size - 1;
1345
1346 /* Set the "off" frequency */
1347 freqs[index++] = 0;
1348
1349 for (i = 0; i < count; i++) {
1350 opp = dev_pm_opp_find_freq_ceil(dev, &freq);
1351 if (IS_ERR(opp))
1352 break;
1353
1354 dev_pm_opp_put(opp);
1355 freqs[index++] = freq++;
1356 }
1357
1358 return index;
1359 }
1360
a6xx_gmu_pwrlevels_probe(struct a6xx_gmu * gmu)1361 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
1362 {
1363 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1364 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1365 struct msm_gpu *gpu = &adreno_gpu->base;
1366
1367 int ret = 0;
1368
1369 /*
1370 * The GMU handles its own frequency switching so build a list of
1371 * available frequencies to send during initialization
1372 */
1373 ret = devm_pm_opp_of_add_table(gmu->dev);
1374 if (ret) {
1375 DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n");
1376 return ret;
1377 }
1378
1379 gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev,
1380 gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs));
1381
1382 /*
1383 * The GMU also handles GPU frequency switching so build a list
1384 * from the GPU OPP table
1385 */
1386 gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev,
1387 gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs));
1388
1389 gmu->current_perf_index = gmu->nr_gpu_freqs - 1;
1390
1391 /* Build the list of RPMh votes that we'll send to the GMU */
1392 return a6xx_gmu_rpmh_votes_init(gmu);
1393 }
1394
a6xx_gmu_clocks_probe(struct a6xx_gmu * gmu)1395 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu)
1396 {
1397 int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks);
1398
1399 if (ret < 1)
1400 return ret;
1401
1402 gmu->nr_clocks = ret;
1403
1404 gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks,
1405 gmu->nr_clocks, "gmu");
1406
1407 gmu->hub_clk = msm_clk_bulk_get_clock(gmu->clocks,
1408 gmu->nr_clocks, "hub");
1409
1410 return 0;
1411 }
1412
a6xx_gmu_get_mmio(struct platform_device * pdev,const char * name)1413 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
1414 const char *name)
1415 {
1416 void __iomem *ret;
1417 struct resource *res = platform_get_resource_byname(pdev,
1418 IORESOURCE_MEM, name);
1419
1420 if (!res) {
1421 DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name);
1422 return ERR_PTR(-EINVAL);
1423 }
1424
1425 ret = ioremap(res->start, resource_size(res));
1426 if (!ret) {
1427 DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name);
1428 return ERR_PTR(-EINVAL);
1429 }
1430
1431 return ret;
1432 }
1433
a6xx_gmu_get_irq(struct a6xx_gmu * gmu,struct platform_device * pdev,const char * name,irq_handler_t handler)1434 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
1435 const char *name, irq_handler_t handler)
1436 {
1437 int irq, ret;
1438
1439 irq = platform_get_irq_byname(pdev, name);
1440
1441 ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH | IRQF_NO_AUTOEN, name, gmu);
1442 if (ret) {
1443 DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n",
1444 name, ret);
1445 return ret;
1446 }
1447
1448 return irq;
1449 }
1450
a6xx_gmu_remove(struct a6xx_gpu * a6xx_gpu)1451 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
1452 {
1453 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1454 struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1455 struct platform_device *pdev = to_platform_device(gmu->dev);
1456
1457 mutex_lock(&gmu->lock);
1458 if (!gmu->initialized) {
1459 mutex_unlock(&gmu->lock);
1460 return;
1461 }
1462
1463 gmu->initialized = false;
1464
1465 mutex_unlock(&gmu->lock);
1466
1467 pm_runtime_force_suspend(gmu->dev);
1468
1469 /*
1470 * Since cxpd is a virt device, the devlink with gmu-dev will be removed
1471 * automatically when we do detach
1472 */
1473 dev_pm_domain_detach(gmu->cxpd, false);
1474
1475 if (!IS_ERR_OR_NULL(gmu->gxpd)) {
1476 pm_runtime_disable(gmu->gxpd);
1477 dev_pm_domain_detach(gmu->gxpd, false);
1478 }
1479
1480 iounmap(gmu->mmio);
1481 if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1482 iounmap(gmu->rscc);
1483 gmu->mmio = NULL;
1484 gmu->rscc = NULL;
1485
1486 if (!adreno_has_gmu_wrapper(adreno_gpu)) {
1487 a6xx_gmu_memory_free(gmu);
1488
1489 free_irq(gmu->gmu_irq, gmu);
1490 free_irq(gmu->hfi_irq, gmu);
1491 }
1492
1493 /* Drop reference taken in of_find_device_by_node */
1494 put_device(gmu->dev);
1495 }
1496
cxpd_notifier_cb(struct notifier_block * nb,unsigned long action,void * data)1497 static int cxpd_notifier_cb(struct notifier_block *nb,
1498 unsigned long action, void *data)
1499 {
1500 struct a6xx_gmu *gmu = container_of(nb, struct a6xx_gmu, pd_nb);
1501
1502 if (action == GENPD_NOTIFY_OFF)
1503 complete_all(&gmu->pd_gate);
1504
1505 return 0;
1506 }
1507
a6xx_gmu_wrapper_init(struct a6xx_gpu * a6xx_gpu,struct device_node * node)1508 int a6xx_gmu_wrapper_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1509 {
1510 struct platform_device *pdev = of_find_device_by_node(node);
1511 struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1512 int ret;
1513
1514 if (!pdev)
1515 return -ENODEV;
1516
1517 gmu->dev = &pdev->dev;
1518
1519 ret = of_dma_configure(gmu->dev, node, true);
1520 if (ret)
1521 return ret;
1522
1523 pm_runtime_enable(gmu->dev);
1524
1525 /* Mark legacy for manual SPTPRAC control */
1526 gmu->legacy = true;
1527
1528 /* Map the GMU registers */
1529 gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1530 if (IS_ERR(gmu->mmio)) {
1531 ret = PTR_ERR(gmu->mmio);
1532 goto err_mmio;
1533 }
1534
1535 gmu->cxpd = dev_pm_domain_attach_by_name(gmu->dev, "cx");
1536 if (IS_ERR(gmu->cxpd)) {
1537 ret = PTR_ERR(gmu->cxpd);
1538 goto err_mmio;
1539 }
1540
1541 if (!device_link_add(gmu->dev, gmu->cxpd, DL_FLAG_PM_RUNTIME)) {
1542 ret = -ENODEV;
1543 goto detach_cxpd;
1544 }
1545
1546 init_completion(&gmu->pd_gate);
1547 complete_all(&gmu->pd_gate);
1548 gmu->pd_nb.notifier_call = cxpd_notifier_cb;
1549
1550 /* Get a link to the GX power domain to reset the GPU */
1551 gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1552 if (IS_ERR(gmu->gxpd)) {
1553 ret = PTR_ERR(gmu->gxpd);
1554 goto err_mmio;
1555 }
1556
1557 gmu->initialized = true;
1558
1559 return 0;
1560
1561 detach_cxpd:
1562 dev_pm_domain_detach(gmu->cxpd, false);
1563
1564 err_mmio:
1565 iounmap(gmu->mmio);
1566
1567 /* Drop reference taken in of_find_device_by_node */
1568 put_device(gmu->dev);
1569
1570 return ret;
1571 }
1572
a6xx_gmu_init(struct a6xx_gpu * a6xx_gpu,struct device_node * node)1573 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1574 {
1575 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1576 struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1577 struct platform_device *pdev = of_find_device_by_node(node);
1578 int ret;
1579
1580 if (!pdev)
1581 return -ENODEV;
1582
1583 gmu->dev = &pdev->dev;
1584
1585 ret = of_dma_configure(gmu->dev, node, true);
1586 if (ret)
1587 return ret;
1588
1589 /* Fow now, don't do anything fancy until we get our feet under us */
1590 gmu->idle_level = GMU_IDLE_STATE_ACTIVE;
1591
1592 pm_runtime_enable(gmu->dev);
1593
1594 /* Get the list of clocks */
1595 ret = a6xx_gmu_clocks_probe(gmu);
1596 if (ret)
1597 goto err_put_device;
1598
1599 ret = a6xx_gmu_memory_probe(gmu);
1600 if (ret)
1601 goto err_put_device;
1602
1603
1604 /* A660 now requires handling "prealloc requests" in GMU firmware
1605 * For now just hardcode allocations based on the known firmware.
1606 * note: there is no indication that these correspond to "dummy" or
1607 * "debug" regions, but this "guess" allows reusing these BOs which
1608 * are otherwise unused by a660.
1609 */
1610 gmu->dummy.size = SZ_4K;
1611 if (adreno_is_a660_family(adreno_gpu)) {
1612 ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_4K * 7,
1613 0x60400000, "debug");
1614 if (ret)
1615 goto err_memory;
1616
1617 gmu->dummy.size = SZ_8K;
1618 }
1619
1620 /* Allocate memory for the GMU dummy page */
1621 ret = a6xx_gmu_memory_alloc(gmu, &gmu->dummy, gmu->dummy.size,
1622 0x60000000, "dummy");
1623 if (ret)
1624 goto err_memory;
1625
1626 /* Note that a650 family also includes a660 family: */
1627 if (adreno_is_a650_family(adreno_gpu)) {
1628 ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1629 SZ_16M - SZ_16K, 0x04000, "icache");
1630 if (ret)
1631 goto err_memory;
1632 /*
1633 * NOTE: when porting legacy ("pre-650-family") GPUs you may be tempted to add a condition
1634 * to allocate icache/dcache here, as per downstream code flow, but it may not actually be
1635 * necessary. If you omit this step and you don't get random pagefaults, you are likely
1636 * good to go without this!
1637 */
1638 } else if (adreno_is_a640_family(adreno_gpu)) {
1639 ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1640 SZ_256K - SZ_16K, 0x04000, "icache");
1641 if (ret)
1642 goto err_memory;
1643
1644 ret = a6xx_gmu_memory_alloc(gmu, &gmu->dcache,
1645 SZ_256K - SZ_16K, 0x44000, "dcache");
1646 if (ret)
1647 goto err_memory;
1648 } else if (adreno_is_a630_family(adreno_gpu)) {
1649 /* HFI v1, has sptprac */
1650 gmu->legacy = true;
1651
1652 /* Allocate memory for the GMU debug region */
1653 ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_16K, 0, "debug");
1654 if (ret)
1655 goto err_memory;
1656 }
1657
1658 /* Allocate memory for the GMU log region */
1659 ret = a6xx_gmu_memory_alloc(gmu, &gmu->log, SZ_16K, 0, "log");
1660 if (ret)
1661 goto err_memory;
1662
1663 /* Allocate memory for for the HFI queues */
1664 ret = a6xx_gmu_memory_alloc(gmu, &gmu->hfi, SZ_16K, 0, "hfi");
1665 if (ret)
1666 goto err_memory;
1667
1668 /* Map the GMU registers */
1669 gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1670 if (IS_ERR(gmu->mmio)) {
1671 ret = PTR_ERR(gmu->mmio);
1672 goto err_memory;
1673 }
1674
1675 if (adreno_is_a650_family(adreno_gpu)) {
1676 gmu->rscc = a6xx_gmu_get_mmio(pdev, "rscc");
1677 if (IS_ERR(gmu->rscc)) {
1678 ret = -ENODEV;
1679 goto err_mmio;
1680 }
1681 } else {
1682 gmu->rscc = gmu->mmio + 0x23000;
1683 }
1684
1685 /* Get the HFI and GMU interrupts */
1686 gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq);
1687 gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq);
1688
1689 if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0) {
1690 ret = -ENODEV;
1691 goto err_mmio;
1692 }
1693
1694 gmu->cxpd = dev_pm_domain_attach_by_name(gmu->dev, "cx");
1695 if (IS_ERR(gmu->cxpd)) {
1696 ret = PTR_ERR(gmu->cxpd);
1697 goto err_mmio;
1698 }
1699
1700 if (!device_link_add(gmu->dev, gmu->cxpd,
1701 DL_FLAG_PM_RUNTIME)) {
1702 ret = -ENODEV;
1703 goto detach_cxpd;
1704 }
1705
1706 init_completion(&gmu->pd_gate);
1707 complete_all(&gmu->pd_gate);
1708 gmu->pd_nb.notifier_call = cxpd_notifier_cb;
1709
1710 /*
1711 * Get a link to the GX power domain to reset the GPU in case of GMU
1712 * crash
1713 */
1714 gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1715
1716 /* Get the power levels for the GMU and GPU */
1717 a6xx_gmu_pwrlevels_probe(gmu);
1718
1719 /* Set up the HFI queues */
1720 a6xx_hfi_init(gmu);
1721
1722 /* Initialize RPMh */
1723 a6xx_gmu_rpmh_init(gmu);
1724
1725 gmu->initialized = true;
1726
1727 return 0;
1728
1729 detach_cxpd:
1730 dev_pm_domain_detach(gmu->cxpd, false);
1731
1732 err_mmio:
1733 iounmap(gmu->mmio);
1734 if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1735 iounmap(gmu->rscc);
1736 free_irq(gmu->gmu_irq, gmu);
1737 free_irq(gmu->hfi_irq, gmu);
1738
1739 err_memory:
1740 a6xx_gmu_memory_free(gmu);
1741 err_put_device:
1742 /* Drop reference taken in of_find_device_by_node */
1743 put_device(gmu->dev);
1744
1745 return ret;
1746 }
1747