]> Git Repo - linux.git/blob - drivers/gpu/drm/i915/gt/intel_rc6.c
Merge tag 'amd-drm-next-6.5-2023-06-09' of https://gitlab.freedesktop.org/agd5f/linux...
[linux.git] / drivers / gpu / drm / i915 / gt / intel_rc6.c
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5
6 #include <linux/pm_runtime.h>
7 #include <linux/string_helpers.h>
8
9 #include "gem/i915_gem_region.h"
10 #include "i915_drv.h"
11 #include "i915_reg.h"
12 #include "i915_vgpu.h"
13 #include "intel_engine_regs.h"
14 #include "intel_gt.h"
15 #include "intel_gt_pm.h"
16 #include "intel_gt_regs.h"
17 #include "intel_pcode.h"
18 #include "intel_rc6.h"
19
20 /**
21  * DOC: RC6
22  *
23  * RC6 is a special power stage which allows the GPU to enter an very
24  * low-voltage mode when idle, using down to 0V while at this stage.  This
25  * stage is entered automatically when the GPU is idle when RC6 support is
26  * enabled, and as soon as new workload arises GPU wakes up automatically as
27  * well.
28  *
29  * There are different RC6 modes available in Intel GPU, which differentiate
30  * among each other with the latency required to enter and leave RC6 and
31  * voltage consumed by the GPU in different states.
32  *
33  * The combination of the following flags define which states GPU is allowed
34  * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
35  * RC6pp is deepest RC6. Their support by hardware varies according to the
36  * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
37  * which brings the most power savings; deeper states save more power, but
38  * require higher latency to switch to and wake up.
39  */
40
41 static struct intel_gt *rc6_to_gt(struct intel_rc6 *rc6)
42 {
43         return container_of(rc6, struct intel_gt, rc6);
44 }
45
46 static struct intel_uncore *rc6_to_uncore(struct intel_rc6 *rc)
47 {
48         return rc6_to_gt(rc)->uncore;
49 }
50
51 static struct drm_i915_private *rc6_to_i915(struct intel_rc6 *rc)
52 {
53         return rc6_to_gt(rc)->i915;
54 }
55
56 static void gen11_rc6_enable(struct intel_rc6 *rc6)
57 {
58         struct intel_gt *gt = rc6_to_gt(rc6);
59         struct intel_uncore *uncore = gt->uncore;
60         struct intel_engine_cs *engine;
61         enum intel_engine_id id;
62         u32 pg_enable;
63         int i;
64
65         /*
66          * With GuCRC, these parameters are set by GuC
67          */
68         if (!intel_uc_uses_guc_rc(&gt->uc)) {
69                 /* 2b: Program RC6 thresholds.*/
70                 intel_uncore_write_fw(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16 | 85);
71                 intel_uncore_write_fw(uncore, GEN10_MEDIA_WAKE_RATE_LIMIT, 150);
72
73                 intel_uncore_write_fw(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
74                 intel_uncore_write_fw(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
75                 for_each_engine(engine, rc6_to_gt(rc6), id)
76                         intel_uncore_write_fw(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
77
78                 intel_uncore_write_fw(uncore, GUC_MAX_IDLE_COUNT, 0xA);
79
80                 intel_uncore_write_fw(uncore, GEN6_RC_SLEEP, 0);
81
82                 intel_uncore_write_fw(uncore, GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
83         }
84
85         /*
86          * 2c: Program Coarse Power Gating Policies.
87          *
88          * Bspec's guidance is to use 25us (really 25 * 1280ns) here. What we
89          * use instead is a more conservative estimate for the maximum time
90          * it takes us to service a CS interrupt and submit a new ELSP - that
91          * is the time which the GPU is idle waiting for the CPU to select the
92          * next request to execute. If the idle hysteresis is less than that
93          * interrupt service latency, the hardware will automatically gate
94          * the power well and we will then incur the wake up cost on top of
95          * the service latency. A similar guide from plane_state is that we
96          * do not want the enable hysteresis to less than the wakeup latency.
97          *
98          * igt/gem_exec_nop/sequential provides a rough estimate for the
99          * service latency, and puts it under 10us for Icelake, similar to
100          * Broadwell+, To be conservative, we want to factor in a context
101          * switch on top (due to ksoftirqd).
102          */
103         intel_uncore_write_fw(uncore, GEN9_MEDIA_PG_IDLE_HYSTERESIS, 60);
104         intel_uncore_write_fw(uncore, GEN9_RENDER_PG_IDLE_HYSTERESIS, 60);
105
106         /* 3a: Enable RC6
107          *
108          * With GuCRC, we do not enable bit 31 of RC_CTL,
109          * thus allowing GuC to control RC6 entry/exit fully instead.
110          * We will not set the HW ENABLE and EI bits
111          */
112         if (!intel_guc_rc_enable(&gt->uc.guc))
113                 rc6->ctl_enable = GEN6_RC_CTL_RC6_ENABLE;
114         else
115                 rc6->ctl_enable =
116                         GEN6_RC_CTL_HW_ENABLE |
117                         GEN6_RC_CTL_RC6_ENABLE |
118                         GEN6_RC_CTL_EI_MODE(1);
119
120         /*
121          * Wa_16011777198 and BSpec 52698 - Render powergating must be off.
122          * FIXME BSpec is outdated, disabling powergating for MTL is just
123          * temporary wa and should be removed after fixing real cause
124          * of forcewake timeouts.
125          */
126         if (IS_METEORLAKE(gt->i915) ||
127             IS_DG2_GRAPHICS_STEP(gt->i915, G10, STEP_A0, STEP_C0) ||
128             IS_DG2_GRAPHICS_STEP(gt->i915, G11, STEP_A0, STEP_B0))
129                 pg_enable =
130                         GEN9_MEDIA_PG_ENABLE |
131                         GEN11_MEDIA_SAMPLER_PG_ENABLE;
132         else
133                 pg_enable =
134                         GEN9_RENDER_PG_ENABLE |
135                         GEN9_MEDIA_PG_ENABLE |
136                         GEN11_MEDIA_SAMPLER_PG_ENABLE;
137
138         if (GRAPHICS_VER(gt->i915) >= 12) {
139                 for (i = 0; i < I915_MAX_VCS; i++)
140                         if (HAS_ENGINE(gt, _VCS(i)))
141                                 pg_enable |= (VDN_HCP_POWERGATE_ENABLE(i) |
142                                               VDN_MFX_POWERGATE_ENABLE(i));
143         }
144
145         intel_uncore_write_fw(uncore, GEN9_PG_ENABLE, pg_enable);
146 }
147
148 static void gen9_rc6_enable(struct intel_rc6 *rc6)
149 {
150         struct intel_uncore *uncore = rc6_to_uncore(rc6);
151         struct intel_engine_cs *engine;
152         enum intel_engine_id id;
153
154         /* 2b: Program RC6 thresholds.*/
155         if (GRAPHICS_VER(rc6_to_i915(rc6)) >= 11) {
156                 intel_uncore_write_fw(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16 | 85);
157                 intel_uncore_write_fw(uncore, GEN10_MEDIA_WAKE_RATE_LIMIT, 150);
158         } else if (IS_SKYLAKE(rc6_to_i915(rc6))) {
159                 /*
160                  * WaRsDoubleRc6WrlWithCoarsePowerGating:skl Doubling WRL only
161                  * when CPG is enabled
162                  */
163                 intel_uncore_write_fw(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
164         } else {
165                 intel_uncore_write_fw(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
166         }
167
168         intel_uncore_write_fw(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
169         intel_uncore_write_fw(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
170         for_each_engine(engine, rc6_to_gt(rc6), id)
171                 intel_uncore_write_fw(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
172
173         intel_uncore_write_fw(uncore, GUC_MAX_IDLE_COUNT, 0xA);
174
175         intel_uncore_write_fw(uncore, GEN6_RC_SLEEP, 0);
176
177         /*
178          * 2c: Program Coarse Power Gating Policies.
179          *
180          * Bspec's guidance is to use 25us (really 25 * 1280ns) here. What we
181          * use instead is a more conservative estimate for the maximum time
182          * it takes us to service a CS interrupt and submit a new ELSP - that
183          * is the time which the GPU is idle waiting for the CPU to select the
184          * next request to execute. If the idle hysteresis is less than that
185          * interrupt service latency, the hardware will automatically gate
186          * the power well and we will then incur the wake up cost on top of
187          * the service latency. A similar guide from plane_state is that we
188          * do not want the enable hysteresis to less than the wakeup latency.
189          *
190          * igt/gem_exec_nop/sequential provides a rough estimate for the
191          * service latency, and puts it around 10us for Broadwell (and other
192          * big core) and around 40us for Broxton (and other low power cores).
193          * [Note that for legacy ringbuffer submission, this is less than 1us!]
194          * However, the wakeup latency on Broxton is closer to 100us. To be
195          * conservative, we have to factor in a context switch on top (due
196          * to ksoftirqd).
197          */
198         intel_uncore_write_fw(uncore, GEN9_MEDIA_PG_IDLE_HYSTERESIS, 250);
199         intel_uncore_write_fw(uncore, GEN9_RENDER_PG_IDLE_HYSTERESIS, 250);
200
201         /* 3a: Enable RC6 */
202         intel_uncore_write_fw(uncore, GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
203
204         rc6->ctl_enable =
205                 GEN6_RC_CTL_HW_ENABLE |
206                 GEN6_RC_CTL_RC6_ENABLE |
207                 GEN6_RC_CTL_EI_MODE(1);
208
209         /*
210          * WaRsDisableCoarsePowerGating:skl,cnl
211          *   - Render/Media PG need to be disabled with RC6.
212          */
213         if (!NEEDS_WaRsDisableCoarsePowerGating(rc6_to_i915(rc6)))
214                 intel_uncore_write_fw(uncore, GEN9_PG_ENABLE,
215                                       GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE);
216 }
217
218 static void gen8_rc6_enable(struct intel_rc6 *rc6)
219 {
220         struct intel_uncore *uncore = rc6_to_uncore(rc6);
221         struct intel_engine_cs *engine;
222         enum intel_engine_id id;
223
224         /* 2b: Program RC6 thresholds.*/
225         intel_uncore_write_fw(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
226         intel_uncore_write_fw(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
227         intel_uncore_write_fw(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
228         for_each_engine(engine, rc6_to_gt(rc6), id)
229                 intel_uncore_write_fw(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
230         intel_uncore_write_fw(uncore, GEN6_RC_SLEEP, 0);
231         intel_uncore_write_fw(uncore, GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
232
233         /* 3: Enable RC6 */
234         rc6->ctl_enable =
235             GEN6_RC_CTL_HW_ENABLE |
236             GEN7_RC_CTL_TO_MODE |
237             GEN6_RC_CTL_RC6_ENABLE;
238 }
239
240 static void gen6_rc6_enable(struct intel_rc6 *rc6)
241 {
242         struct intel_uncore *uncore = rc6_to_uncore(rc6);
243         struct drm_i915_private *i915 = rc6_to_i915(rc6);
244         struct intel_engine_cs *engine;
245         enum intel_engine_id id;
246         u32 rc6vids, rc6_mask;
247         int ret;
248
249         intel_uncore_write_fw(uncore, GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
250         intel_uncore_write_fw(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
251         intel_uncore_write_fw(uncore, GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
252         intel_uncore_write_fw(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000);
253         intel_uncore_write_fw(uncore, GEN6_RC_IDLE_HYSTERSIS, 25);
254
255         for_each_engine(engine, rc6_to_gt(rc6), id)
256                 intel_uncore_write_fw(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
257
258         intel_uncore_write_fw(uncore, GEN6_RC_SLEEP, 0);
259         intel_uncore_write_fw(uncore, GEN6_RC1e_THRESHOLD, 1000);
260         intel_uncore_write_fw(uncore, GEN6_RC6_THRESHOLD, 50000);
261         intel_uncore_write_fw(uncore, GEN6_RC6p_THRESHOLD, 150000);
262         intel_uncore_write_fw(uncore, GEN6_RC6pp_THRESHOLD, 64000); /* unused */
263
264         /* We don't use those on Haswell */
265         rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
266         if (HAS_RC6p(i915))
267                 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
268         if (HAS_RC6pp(i915))
269                 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
270         rc6->ctl_enable =
271             rc6_mask |
272             GEN6_RC_CTL_EI_MODE(1) |
273             GEN6_RC_CTL_HW_ENABLE;
274
275         rc6vids = 0;
276         ret = snb_pcode_read(rc6_to_gt(rc6)->uncore, GEN6_PCODE_READ_RC6VIDS, &rc6vids, NULL);
277         if (GRAPHICS_VER(i915) == 6 && ret) {
278                 drm_dbg(&i915->drm, "Couldn't check for BIOS workaround\n");
279         } else if (GRAPHICS_VER(i915) == 6 &&
280                    (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
281                 drm_dbg(&i915->drm,
282                         "You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
283                         GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
284                 rc6vids &= 0xffff00;
285                 rc6vids |= GEN6_ENCODE_RC6_VID(450);
286                 ret = snb_pcode_write(rc6_to_gt(rc6)->uncore, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
287                 if (ret)
288                         drm_err(&i915->drm,
289                                 "Couldn't fix incorrect rc6 voltage\n");
290         }
291 }
292
293 /* Check that the pcbr address is not empty. */
294 static int chv_rc6_init(struct intel_rc6 *rc6)
295 {
296         struct intel_uncore *uncore = rc6_to_uncore(rc6);
297         struct drm_i915_private *i915 = rc6_to_i915(rc6);
298         resource_size_t pctx_paddr, paddr;
299         resource_size_t pctx_size = 32 * SZ_1K;
300         u32 pcbr;
301
302         pcbr = intel_uncore_read(uncore, VLV_PCBR);
303         if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
304                 drm_dbg(&i915->drm, "BIOS didn't set up PCBR, fixing up\n");
305                 paddr = i915->dsm.stolen.end + 1 - pctx_size;
306                 GEM_BUG_ON(paddr > U32_MAX);
307
308                 pctx_paddr = (paddr & ~4095);
309                 intel_uncore_write(uncore, VLV_PCBR, pctx_paddr);
310         }
311
312         return 0;
313 }
314
315 static int vlv_rc6_init(struct intel_rc6 *rc6)
316 {
317         struct drm_i915_private *i915 = rc6_to_i915(rc6);
318         struct intel_uncore *uncore = rc6_to_uncore(rc6);
319         struct drm_i915_gem_object *pctx;
320         resource_size_t pctx_paddr;
321         resource_size_t pctx_size = 24 * SZ_1K;
322         u32 pcbr;
323
324         pcbr = intel_uncore_read(uncore, VLV_PCBR);
325         if (pcbr) {
326                 /* BIOS set it up already, grab the pre-alloc'd space */
327                 resource_size_t pcbr_offset;
328
329                 pcbr_offset = (pcbr & ~4095) - i915->dsm.stolen.start;
330                 pctx = i915_gem_object_create_region_at(i915->mm.stolen_region,
331                                                         pcbr_offset,
332                                                         pctx_size,
333                                                         0);
334                 if (IS_ERR(pctx))
335                         return PTR_ERR(pctx);
336
337                 goto out;
338         }
339
340         drm_dbg(&i915->drm, "BIOS didn't set up PCBR, fixing up\n");
341
342         /*
343          * From the Gunit register HAS:
344          * The Gfx driver is expected to program this register and ensure
345          * proper allocation within Gfx stolen memory.  For example, this
346          * register should be programmed such than the PCBR range does not
347          * overlap with other ranges, such as the frame buffer, protected
348          * memory, or any other relevant ranges.
349          */
350         pctx = i915_gem_object_create_stolen(i915, pctx_size);
351         if (IS_ERR(pctx)) {
352                 drm_dbg(&i915->drm,
353                         "not enough stolen space for PCTX, disabling\n");
354                 return PTR_ERR(pctx);
355         }
356
357         GEM_BUG_ON(range_overflows_end_t(u64,
358                                          i915->dsm.stolen.start,
359                                          pctx->stolen->start,
360                                          U32_MAX));
361         pctx_paddr = i915->dsm.stolen.start + pctx->stolen->start;
362         intel_uncore_write(uncore, VLV_PCBR, pctx_paddr);
363
364 out:
365         rc6->pctx = pctx;
366         return 0;
367 }
368
369 static void chv_rc6_enable(struct intel_rc6 *rc6)
370 {
371         struct intel_uncore *uncore = rc6_to_uncore(rc6);
372         struct intel_engine_cs *engine;
373         enum intel_engine_id id;
374
375         /* 2a: Program RC6 thresholds.*/
376         intel_uncore_write_fw(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
377         intel_uncore_write_fw(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
378         intel_uncore_write_fw(uncore, GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
379
380         for_each_engine(engine, rc6_to_gt(rc6), id)
381                 intel_uncore_write_fw(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
382         intel_uncore_write_fw(uncore, GEN6_RC_SLEEP, 0);
383
384         /* TO threshold set to 500 us (0x186 * 1.28 us) */
385         intel_uncore_write_fw(uncore, GEN6_RC6_THRESHOLD, 0x186);
386
387         /* Allows RC6 residency counter to work */
388         intel_uncore_write_fw(uncore, VLV_COUNTER_CONTROL,
389                               _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
390                                                  VLV_MEDIA_RC6_COUNT_EN |
391                                                  VLV_RENDER_RC6_COUNT_EN));
392
393         /* 3: Enable RC6 */
394         rc6->ctl_enable = GEN7_RC_CTL_TO_MODE;
395 }
396
397 static void vlv_rc6_enable(struct intel_rc6 *rc6)
398 {
399         struct intel_uncore *uncore = rc6_to_uncore(rc6);
400         struct intel_engine_cs *engine;
401         enum intel_engine_id id;
402
403         intel_uncore_write_fw(uncore, GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
404         intel_uncore_write_fw(uncore, GEN6_RC_EVALUATION_INTERVAL, 125000);
405         intel_uncore_write_fw(uncore, GEN6_RC_IDLE_HYSTERSIS, 25);
406
407         for_each_engine(engine, rc6_to_gt(rc6), id)
408                 intel_uncore_write_fw(uncore, RING_MAX_IDLE(engine->mmio_base), 10);
409
410         intel_uncore_write_fw(uncore, GEN6_RC6_THRESHOLD, 0x557);
411
412         /* Allows RC6 residency counter to work */
413         intel_uncore_write_fw(uncore, VLV_COUNTER_CONTROL,
414                               _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
415                                                  VLV_MEDIA_RC0_COUNT_EN |
416                                                  VLV_RENDER_RC0_COUNT_EN |
417                                                  VLV_MEDIA_RC6_COUNT_EN |
418                                                  VLV_RENDER_RC6_COUNT_EN));
419
420         rc6->ctl_enable =
421             GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
422 }
423
424 bool intel_check_bios_c6_setup(struct intel_rc6 *rc6)
425 {
426         if (!rc6->bios_state_captured) {
427                 struct intel_uncore *uncore = rc6_to_uncore(rc6);
428                 intel_wakeref_t wakeref;
429
430                 with_intel_runtime_pm(uncore->rpm, wakeref)
431                         rc6->bios_rc_state = intel_uncore_read(uncore, GEN6_RC_STATE);
432
433                 rc6->bios_state_captured = true;
434         }
435
436         return rc6->bios_rc_state & RC_SW_TARGET_STATE_MASK;
437 }
438
439 static bool bxt_check_bios_rc6_setup(struct intel_rc6 *rc6)
440 {
441         struct intel_uncore *uncore = rc6_to_uncore(rc6);
442         struct drm_i915_private *i915 = rc6_to_i915(rc6);
443         u32 rc6_ctx_base, rc_ctl, rc_sw_target;
444         bool enable_rc6 = true;
445
446         rc_ctl = intel_uncore_read(uncore, GEN6_RC_CONTROL);
447         rc_sw_target = intel_uncore_read(uncore, GEN6_RC_STATE);
448         rc_sw_target &= RC_SW_TARGET_STATE_MASK;
449         rc_sw_target >>= RC_SW_TARGET_STATE_SHIFT;
450         drm_dbg(&i915->drm, "BIOS enabled RC states: "
451                          "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
452                          str_on_off(rc_ctl & GEN6_RC_CTL_HW_ENABLE),
453                          str_on_off(rc_ctl & GEN6_RC_CTL_RC6_ENABLE),
454                          rc_sw_target);
455
456         if (!(intel_uncore_read(uncore, RC6_LOCATION) & RC6_CTX_IN_DRAM)) {
457                 drm_dbg(&i915->drm, "RC6 Base location not set properly.\n");
458                 enable_rc6 = false;
459         }
460
461         /*
462          * The exact context size is not known for BXT, so assume a page size
463          * for this check.
464          */
465         rc6_ctx_base =
466                 intel_uncore_read(uncore, RC6_CTX_BASE) & RC6_CTX_BASE_MASK;
467         if (!(rc6_ctx_base >= i915->dsm.reserved.start &&
468               rc6_ctx_base + PAGE_SIZE < i915->dsm.reserved.end)) {
469                 drm_dbg(&i915->drm, "RC6 Base address not as expected.\n");
470                 enable_rc6 = false;
471         }
472
473         if (!((intel_uncore_read(uncore, PWRCTX_MAXCNT(RENDER_RING_BASE)) & IDLE_TIME_MASK) > 1 &&
474               (intel_uncore_read(uncore, PWRCTX_MAXCNT(GEN6_BSD_RING_BASE)) & IDLE_TIME_MASK) > 1 &&
475               (intel_uncore_read(uncore, PWRCTX_MAXCNT(BLT_RING_BASE)) & IDLE_TIME_MASK) > 1 &&
476               (intel_uncore_read(uncore, PWRCTX_MAXCNT(VEBOX_RING_BASE)) & IDLE_TIME_MASK) > 1)) {
477                 drm_dbg(&i915->drm,
478                         "Engine Idle wait time not set properly.\n");
479                 enable_rc6 = false;
480         }
481
482         if (!intel_uncore_read(uncore, GEN8_PUSHBUS_CONTROL) ||
483             !intel_uncore_read(uncore, GEN8_PUSHBUS_ENABLE) ||
484             !intel_uncore_read(uncore, GEN8_PUSHBUS_SHIFT)) {
485                 drm_dbg(&i915->drm, "Pushbus not setup properly.\n");
486                 enable_rc6 = false;
487         }
488
489         if (!intel_uncore_read(uncore, GEN6_GFXPAUSE)) {
490                 drm_dbg(&i915->drm, "GFX pause not setup properly.\n");
491                 enable_rc6 = false;
492         }
493
494         if (!intel_uncore_read(uncore, GEN8_MISC_CTRL0)) {
495                 drm_dbg(&i915->drm, "GPM control not setup properly.\n");
496                 enable_rc6 = false;
497         }
498
499         return enable_rc6;
500 }
501
502 static bool rc6_supported(struct intel_rc6 *rc6)
503 {
504         struct drm_i915_private *i915 = rc6_to_i915(rc6);
505         struct intel_gt *gt = rc6_to_gt(rc6);
506
507         if (!HAS_RC6(i915))
508                 return false;
509
510         if (intel_vgpu_active(i915))
511                 return false;
512
513         if (is_mock_gt(rc6_to_gt(rc6)))
514                 return false;
515
516         if (IS_GEN9_LP(i915) && !bxt_check_bios_rc6_setup(rc6)) {
517                 drm_notice(&i915->drm,
518                            "RC6 and powersaving disabled by BIOS\n");
519                 return false;
520         }
521
522         if (IS_METEORLAKE(gt->i915) &&
523             !intel_check_bios_c6_setup(rc6)) {
524                 drm_notice(&i915->drm,
525                            "C6 disabled by BIOS\n");
526                 return false;
527         }
528
529         if (IS_MTL_MEDIA_STEP(gt->i915, STEP_A0, STEP_B0) &&
530             gt->type == GT_MEDIA) {
531                 drm_notice(&i915->drm,
532                            "Media RC6 disabled on A step\n");
533                 return false;
534         }
535
536         return true;
537 }
538
539 static void rpm_get(struct intel_rc6 *rc6)
540 {
541         GEM_BUG_ON(rc6->wakeref);
542         pm_runtime_get_sync(rc6_to_i915(rc6)->drm.dev);
543         rc6->wakeref = true;
544 }
545
546 static void rpm_put(struct intel_rc6 *rc6)
547 {
548         GEM_BUG_ON(!rc6->wakeref);
549         pm_runtime_put(rc6_to_i915(rc6)->drm.dev);
550         rc6->wakeref = false;
551 }
552
553 static bool pctx_corrupted(struct intel_rc6 *rc6)
554 {
555         struct drm_i915_private *i915 = rc6_to_i915(rc6);
556
557         if (!NEEDS_RC6_CTX_CORRUPTION_WA(i915))
558                 return false;
559
560         if (intel_uncore_read(rc6_to_uncore(rc6), GEN8_RC6_CTX_INFO))
561                 return false;
562
563         drm_notice(&i915->drm,
564                    "RC6 context corruption, disabling runtime power management\n");
565         return true;
566 }
567
568 static void __intel_rc6_disable(struct intel_rc6 *rc6)
569 {
570         struct drm_i915_private *i915 = rc6_to_i915(rc6);
571         struct intel_uncore *uncore = rc6_to_uncore(rc6);
572         struct intel_gt *gt = rc6_to_gt(rc6);
573
574         /* Take control of RC6 back from GuC */
575         intel_guc_rc_disable(&gt->uc.guc);
576
577         intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
578         if (GRAPHICS_VER(i915) >= 9)
579                 intel_uncore_write_fw(uncore, GEN9_PG_ENABLE, 0);
580         intel_uncore_write_fw(uncore, GEN6_RC_CONTROL, 0);
581         intel_uncore_write_fw(uncore, GEN6_RC_STATE, 0);
582         intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
583 }
584
585 static void rc6_res_reg_init(struct intel_rc6 *rc6)
586 {
587         memset(rc6->res_reg, INVALID_MMIO_REG.reg, sizeof(rc6->res_reg));
588
589         switch (rc6_to_gt(rc6)->type) {
590         case GT_MEDIA:
591                 rc6->res_reg[INTEL_RC6_RES_RC6] = MTL_MEDIA_MC6;
592                 break;
593         default:
594                 rc6->res_reg[INTEL_RC6_RES_RC6_LOCKED] = GEN6_GT_GFX_RC6_LOCKED;
595                 rc6->res_reg[INTEL_RC6_RES_RC6] = GEN6_GT_GFX_RC6;
596                 rc6->res_reg[INTEL_RC6_RES_RC6p] = GEN6_GT_GFX_RC6p;
597                 rc6->res_reg[INTEL_RC6_RES_RC6pp] = GEN6_GT_GFX_RC6pp;
598                 break;
599         }
600 }
601
602 void intel_rc6_init(struct intel_rc6 *rc6)
603 {
604         struct drm_i915_private *i915 = rc6_to_i915(rc6);
605         int err;
606
607         /* Disable runtime-pm until we can save the GPU state with rc6 pctx */
608         rpm_get(rc6);
609
610         if (!rc6_supported(rc6))
611                 return;
612
613         rc6_res_reg_init(rc6);
614
615         if (IS_CHERRYVIEW(i915))
616                 err = chv_rc6_init(rc6);
617         else if (IS_VALLEYVIEW(i915))
618                 err = vlv_rc6_init(rc6);
619         else
620                 err = 0;
621
622         /* Sanitize rc6, ensure it is disabled before we are ready. */
623         __intel_rc6_disable(rc6);
624
625         rc6->supported = err == 0;
626 }
627
628 void intel_rc6_sanitize(struct intel_rc6 *rc6)
629 {
630         memset(rc6->prev_hw_residency, 0, sizeof(rc6->prev_hw_residency));
631
632         if (rc6->enabled) { /* unbalanced suspend/resume */
633                 rpm_get(rc6);
634                 rc6->enabled = false;
635         }
636
637         if (rc6->supported)
638                 __intel_rc6_disable(rc6);
639 }
640
641 void intel_rc6_enable(struct intel_rc6 *rc6)
642 {
643         struct drm_i915_private *i915 = rc6_to_i915(rc6);
644         struct intel_uncore *uncore = rc6_to_uncore(rc6);
645
646         if (!rc6->supported)
647                 return;
648
649         GEM_BUG_ON(rc6->enabled);
650
651         intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
652
653         if (IS_CHERRYVIEW(i915))
654                 chv_rc6_enable(rc6);
655         else if (IS_VALLEYVIEW(i915))
656                 vlv_rc6_enable(rc6);
657         else if (GRAPHICS_VER(i915) >= 11)
658                 gen11_rc6_enable(rc6);
659         else if (GRAPHICS_VER(i915) >= 9)
660                 gen9_rc6_enable(rc6);
661         else if (IS_BROADWELL(i915))
662                 gen8_rc6_enable(rc6);
663         else if (GRAPHICS_VER(i915) >= 6)
664                 gen6_rc6_enable(rc6);
665
666         rc6->manual = rc6->ctl_enable & GEN6_RC_CTL_RC6_ENABLE;
667         if (NEEDS_RC6_CTX_CORRUPTION_WA(i915))
668                 rc6->ctl_enable = 0;
669
670         intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
671
672         if (unlikely(pctx_corrupted(rc6)))
673                 return;
674
675         /* rc6 is ready, runtime-pm is go! */
676         rpm_put(rc6);
677         rc6->enabled = true;
678 }
679
680 void intel_rc6_unpark(struct intel_rc6 *rc6)
681 {
682         struct intel_uncore *uncore = rc6_to_uncore(rc6);
683
684         if (!rc6->enabled)
685                 return;
686
687         /* Restore HW timers for automatic RC6 entry while busy */
688         intel_uncore_write_fw(uncore, GEN6_RC_CONTROL, rc6->ctl_enable);
689 }
690
691 void intel_rc6_park(struct intel_rc6 *rc6)
692 {
693         struct intel_uncore *uncore = rc6_to_uncore(rc6);
694         unsigned int target;
695
696         if (!rc6->enabled)
697                 return;
698
699         if (unlikely(pctx_corrupted(rc6))) {
700                 intel_rc6_disable(rc6);
701                 return;
702         }
703
704         if (!rc6->manual)
705                 return;
706
707         /* Turn off the HW timers and go directly to rc6 */
708         intel_uncore_write_fw(uncore, GEN6_RC_CONTROL, GEN6_RC_CTL_RC6_ENABLE);
709
710         if (HAS_RC6pp(rc6_to_i915(rc6)))
711                 target = 0x6; /* deepest rc6 */
712         else if (HAS_RC6p(rc6_to_i915(rc6)))
713                 target = 0x5; /* deep rc6 */
714         else
715                 target = 0x4; /* normal rc6 */
716         intel_uncore_write_fw(uncore, GEN6_RC_STATE, target << RC_SW_TARGET_STATE_SHIFT);
717 }
718
719 void intel_rc6_disable(struct intel_rc6 *rc6)
720 {
721         if (!rc6->enabled)
722                 return;
723
724         rpm_get(rc6);
725         rc6->enabled = false;
726
727         __intel_rc6_disable(rc6);
728 }
729
730 void intel_rc6_fini(struct intel_rc6 *rc6)
731 {
732         struct drm_i915_gem_object *pctx;
733         struct intel_uncore *uncore = rc6_to_uncore(rc6);
734
735         intel_rc6_disable(rc6);
736
737         /* We want the BIOS C6 state preserved across loads for MTL */
738         if (IS_METEORLAKE(rc6_to_i915(rc6)) && rc6->bios_state_captured)
739                 intel_uncore_write_fw(uncore, GEN6_RC_STATE, rc6->bios_rc_state);
740
741         pctx = fetch_and_zero(&rc6->pctx);
742         if (pctx)
743                 i915_gem_object_put(pctx);
744
745         if (rc6->wakeref)
746                 rpm_put(rc6);
747 }
748
749 static u64 vlv_residency_raw(struct intel_uncore *uncore, const i915_reg_t reg)
750 {
751         u32 lower, upper, tmp;
752         int loop = 2;
753
754         /*
755          * The register accessed do not need forcewake. We borrow
756          * uncore lock to prevent concurrent access to range reg.
757          */
758         lockdep_assert_held(&uncore->lock);
759
760         /*
761          * vlv and chv residency counters are 40 bits in width.
762          * With a control bit, we can choose between upper or lower
763          * 32bit window into this counter.
764          *
765          * Although we always use the counter in high-range mode elsewhere,
766          * userspace may attempt to read the value before rc6 is initialised,
767          * before we have set the default VLV_COUNTER_CONTROL value. So always
768          * set the high bit to be safe.
769          */
770         intel_uncore_write_fw(uncore, VLV_COUNTER_CONTROL,
771                               _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH));
772         upper = intel_uncore_read_fw(uncore, reg);
773         do {
774                 tmp = upper;
775
776                 intel_uncore_write_fw(uncore, VLV_COUNTER_CONTROL,
777                                       _MASKED_BIT_DISABLE(VLV_COUNT_RANGE_HIGH));
778                 lower = intel_uncore_read_fw(uncore, reg);
779
780                 intel_uncore_write_fw(uncore, VLV_COUNTER_CONTROL,
781                                       _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH));
782                 upper = intel_uncore_read_fw(uncore, reg);
783         } while (upper != tmp && --loop);
784
785         /*
786          * Everywhere else we always use VLV_COUNTER_CONTROL with the
787          * VLV_COUNT_RANGE_HIGH bit set - so it is safe to leave it set
788          * now.
789          */
790
791         return lower | (u64)upper << 8;
792 }
793
794 u64 intel_rc6_residency_ns(struct intel_rc6 *rc6, enum intel_rc6_res_type id)
795 {
796         struct drm_i915_private *i915 = rc6_to_i915(rc6);
797         struct intel_uncore *uncore = rc6_to_uncore(rc6);
798         u64 time_hw, prev_hw, overflow_hw;
799         i915_reg_t reg = rc6->res_reg[id];
800         unsigned int fw_domains;
801         unsigned long flags;
802         u32 mul, div;
803
804         if (!rc6->supported)
805                 return 0;
806
807         fw_domains = intel_uncore_forcewake_for_reg(uncore, reg, FW_REG_READ);
808
809         spin_lock_irqsave(&uncore->lock, flags);
810         intel_uncore_forcewake_get__locked(uncore, fw_domains);
811
812         /* On VLV and CHV, residency time is in CZ units rather than 1.28us */
813         if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915)) {
814                 mul = 1000000;
815                 div = i915->czclk_freq;
816                 overflow_hw = BIT_ULL(40);
817                 time_hw = vlv_residency_raw(uncore, reg);
818         } else {
819                 /* 833.33ns units on Gen9LP, 1.28us elsewhere. */
820                 if (IS_GEN9_LP(i915)) {
821                         mul = 10000;
822                         div = 12;
823                 } else {
824                         mul = 1280;
825                         div = 1;
826                 }
827
828                 overflow_hw = BIT_ULL(32);
829                 time_hw = intel_uncore_read_fw(uncore, reg);
830         }
831
832         /*
833          * Counter wrap handling.
834          *
835          * Store previous hw counter values for counter wrap-around handling. But
836          * relying on a sufficient frequency of queries otherwise counters can still wrap.
837          */
838         prev_hw = rc6->prev_hw_residency[id];
839         rc6->prev_hw_residency[id] = time_hw;
840
841         /* RC6 delta from last sample. */
842         if (time_hw >= prev_hw)
843                 time_hw -= prev_hw;
844         else
845                 time_hw += overflow_hw - prev_hw;
846
847         /* Add delta to RC6 extended raw driver copy. */
848         time_hw += rc6->cur_residency[id];
849         rc6->cur_residency[id] = time_hw;
850
851         intel_uncore_forcewake_put__locked(uncore, fw_domains);
852         spin_unlock_irqrestore(&uncore->lock, flags);
853
854         return mul_u64_u32_div(time_hw, mul, div);
855 }
856
857 u64 intel_rc6_residency_us(struct intel_rc6 *rc6, enum intel_rc6_res_type id)
858 {
859         return DIV_ROUND_UP_ULL(intel_rc6_residency_ns(rc6, id), 1000);
860 }
861
862 void intel_rc6_print_residency(struct seq_file *m, const char *title,
863                                enum intel_rc6_res_type id)
864 {
865         struct intel_gt *gt = m->private;
866         i915_reg_t reg = gt->rc6.res_reg[id];
867         intel_wakeref_t wakeref;
868
869         with_intel_runtime_pm(gt->uncore->rpm, wakeref)
870                 seq_printf(m, "%s %u (%llu us)\n", title,
871                            intel_uncore_read(gt->uncore, reg),
872                            intel_rc6_residency_us(&gt->rc6, id));
873 }
874
875 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
876 #include "selftest_rc6.c"
877 #endif
This page took 0.090489 seconds and 4 git commands to generate.