1 /*
2 * Copyright 2018 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 */
23 #include "amdgpu.h"
24 #include "df_v3_6.h"
25
26 #include "df/df_3_6_default.h"
27 #include "df/df_3_6_offset.h"
28 #include "df/df_3_6_sh_mask.h"
29
30 #define DF_3_6_SMN_REG_INST_DIST 0x8
31 #define DF_3_6_INST_CNT 8
32
33 static u32 df_v3_6_channel_number[] = {1, 2, 0, 4, 0, 8, 0,
34 16, 32, 0, 0, 0, 2, 4, 8};
35
36 /* init df format attrs */
37 AMDGPU_PMU_ATTR(event, "config:0-7");
38 AMDGPU_PMU_ATTR(instance, "config:8-15");
39 AMDGPU_PMU_ATTR(umask, "config:16-23");
40
41 /* df format attributes */
42 static struct attribute *df_v3_6_format_attrs[] = {
43 &pmu_attr_event.attr,
44 &pmu_attr_instance.attr,
45 &pmu_attr_umask.attr,
46 NULL
47 };
48
49 /* df format attribute group */
50 static struct attribute_group df_v3_6_format_attr_group = {
51 .name = "format",
52 .attrs = df_v3_6_format_attrs,
53 };
54
55 /* df event attrs */
56 AMDGPU_PMU_ATTR(cake0_pcsout_txdata,
57 "event=0x7,instance=0x46,umask=0x2");
58 AMDGPU_PMU_ATTR(cake1_pcsout_txdata,
59 "event=0x7,instance=0x47,umask=0x2");
60 AMDGPU_PMU_ATTR(cake0_pcsout_txmeta,
61 "event=0x7,instance=0x46,umask=0x4");
62 AMDGPU_PMU_ATTR(cake1_pcsout_txmeta,
63 "event=0x7,instance=0x47,umask=0x4");
64 AMDGPU_PMU_ATTR(cake0_ftiinstat_reqalloc,
65 "event=0xb,instance=0x46,umask=0x4");
66 AMDGPU_PMU_ATTR(cake1_ftiinstat_reqalloc,
67 "event=0xb,instance=0x47,umask=0x4");
68 AMDGPU_PMU_ATTR(cake0_ftiinstat_rspalloc,
69 "event=0xb,instance=0x46,umask=0x8");
70 AMDGPU_PMU_ATTR(cake1_ftiinstat_rspalloc,
71 "event=0xb,instance=0x47,umask=0x8");
72
73 /* df event attributes */
74 static struct attribute *df_v3_6_event_attrs[] = {
75 &pmu_attr_cake0_pcsout_txdata.attr,
76 &pmu_attr_cake1_pcsout_txdata.attr,
77 &pmu_attr_cake0_pcsout_txmeta.attr,
78 &pmu_attr_cake1_pcsout_txmeta.attr,
79 &pmu_attr_cake0_ftiinstat_reqalloc.attr,
80 &pmu_attr_cake1_ftiinstat_reqalloc.attr,
81 &pmu_attr_cake0_ftiinstat_rspalloc.attr,
82 &pmu_attr_cake1_ftiinstat_rspalloc.attr,
83 NULL
84 };
85
86 /* df event attribute group */
87 static struct attribute_group df_v3_6_event_attr_group = {
88 .name = "events",
89 .attrs = df_v3_6_event_attrs
90 };
91
92 /* df event attr groups */
93 const struct attribute_group *df_v3_6_attr_groups[] = {
94 &df_v3_6_format_attr_group,
95 &df_v3_6_event_attr_group,
96 NULL
97 };
98
df_v3_6_get_fica(struct amdgpu_device * adev,uint32_t ficaa_val)99 static uint64_t df_v3_6_get_fica(struct amdgpu_device *adev,
100 uint32_t ficaa_val)
101 {
102 unsigned long flags, address, data;
103 uint32_t ficadl_val, ficadh_val;
104
105 address = adev->nbio.funcs->get_pcie_index_offset(adev);
106 data = adev->nbio.funcs->get_pcie_data_offset(adev);
107
108 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
109 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessAddress3);
110 WREG32(data, ficaa_val);
111
112 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessDataLo3);
113 ficadl_val = RREG32(data);
114
115 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessDataHi3);
116 ficadh_val = RREG32(data);
117
118 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
119
120 return (((ficadh_val & 0xFFFFFFFFFFFFFFFF) << 32) | ficadl_val);
121 }
122
df_v3_6_set_fica(struct amdgpu_device * adev,uint32_t ficaa_val,uint32_t ficadl_val,uint32_t ficadh_val)123 static void df_v3_6_set_fica(struct amdgpu_device *adev, uint32_t ficaa_val,
124 uint32_t ficadl_val, uint32_t ficadh_val)
125 {
126 unsigned long flags, address, data;
127
128 address = adev->nbio.funcs->get_pcie_index_offset(adev);
129 data = adev->nbio.funcs->get_pcie_data_offset(adev);
130
131 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
132 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessAddress3);
133 WREG32(data, ficaa_val);
134
135 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessDataLo3);
136 WREG32(data, ficadl_val);
137
138 WREG32(address, smnDF_PIE_AON_FabricIndirectConfigAccessDataHi3);
139 WREG32(data, ficadh_val);
140
141 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
142 }
143
144 /*
145 * df_v3_6_perfmon_rreg - read perfmon lo and hi
146 *
147 * required to be atomic. no mmio method provided so subsequent reads for lo
148 * and hi require to preserve df finite state machine
149 */
df_v3_6_perfmon_rreg(struct amdgpu_device * adev,uint32_t lo_addr,uint32_t * lo_val,uint32_t hi_addr,uint32_t * hi_val)150 static void df_v3_6_perfmon_rreg(struct amdgpu_device *adev,
151 uint32_t lo_addr, uint32_t *lo_val,
152 uint32_t hi_addr, uint32_t *hi_val)
153 {
154 unsigned long flags, address, data;
155
156 address = adev->nbio.funcs->get_pcie_index_offset(adev);
157 data = adev->nbio.funcs->get_pcie_data_offset(adev);
158
159 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
160 WREG32(address, lo_addr);
161 *lo_val = RREG32(data);
162 WREG32(address, hi_addr);
163 *hi_val = RREG32(data);
164 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
165 }
166
167 /*
168 * df_v3_6_perfmon_wreg - write to perfmon lo and hi
169 *
170 * required to be atomic. no mmio method provided so subsequent reads after
171 * data writes cannot occur to preserve data fabrics finite state machine.
172 */
df_v3_6_perfmon_wreg(struct amdgpu_device * adev,uint32_t lo_addr,uint32_t lo_val,uint32_t hi_addr,uint32_t hi_val)173 static void df_v3_6_perfmon_wreg(struct amdgpu_device *adev, uint32_t lo_addr,
174 uint32_t lo_val, uint32_t hi_addr, uint32_t hi_val)
175 {
176 unsigned long flags, address, data;
177
178 address = adev->nbio.funcs->get_pcie_index_offset(adev);
179 data = adev->nbio.funcs->get_pcie_data_offset(adev);
180
181 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
182 WREG32(address, lo_addr);
183 WREG32(data, lo_val);
184 WREG32(address, hi_addr);
185 WREG32(data, hi_val);
186 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
187 }
188
189 /* same as perfmon_wreg but return status on write value check */
df_v3_6_perfmon_arm_with_status(struct amdgpu_device * adev,uint32_t lo_addr,uint32_t lo_val,uint32_t hi_addr,uint32_t hi_val)190 static int df_v3_6_perfmon_arm_with_status(struct amdgpu_device *adev,
191 uint32_t lo_addr, uint32_t lo_val,
192 uint32_t hi_addr, uint32_t hi_val)
193 {
194 unsigned long flags, address, data;
195 uint32_t lo_val_rb, hi_val_rb;
196
197 address = adev->nbio.funcs->get_pcie_index_offset(adev);
198 data = adev->nbio.funcs->get_pcie_data_offset(adev);
199
200 spin_lock_irqsave(&adev->pcie_idx_lock, flags);
201 WREG32(address, lo_addr);
202 WREG32(data, lo_val);
203 WREG32(address, hi_addr);
204 WREG32(data, hi_val);
205
206 WREG32(address, lo_addr);
207 lo_val_rb = RREG32(data);
208 WREG32(address, hi_addr);
209 hi_val_rb = RREG32(data);
210 spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
211
212 if (!(lo_val == lo_val_rb && hi_val == hi_val_rb))
213 return -EBUSY;
214
215 return 0;
216 }
217
218
219 /*
220 * retry arming counters every 100 usecs within 1 millisecond interval.
221 * if retry fails after time out, return error.
222 */
223 #define ARM_RETRY_USEC_TIMEOUT 1000
224 #define ARM_RETRY_USEC_INTERVAL 100
df_v3_6_perfmon_arm_with_retry(struct amdgpu_device * adev,uint32_t lo_addr,uint32_t lo_val,uint32_t hi_addr,uint32_t hi_val)225 static int df_v3_6_perfmon_arm_with_retry(struct amdgpu_device *adev,
226 uint32_t lo_addr, uint32_t lo_val,
227 uint32_t hi_addr, uint32_t hi_val)
228 {
229 int countdown = ARM_RETRY_USEC_TIMEOUT;
230
231 while (countdown) {
232
233 if (!df_v3_6_perfmon_arm_with_status(adev, lo_addr, lo_val,
234 hi_addr, hi_val))
235 break;
236
237 countdown -= ARM_RETRY_USEC_INTERVAL;
238 udelay(ARM_RETRY_USEC_INTERVAL);
239 }
240
241 return countdown > 0 ? 0 : -ETIME;
242 }
243
244 /* get the number of df counters available */
df_v3_6_get_df_cntr_avail(struct device * dev,struct device_attribute * attr,char * buf)245 static ssize_t df_v3_6_get_df_cntr_avail(struct device *dev,
246 struct device_attribute *attr,
247 char *buf)
248 {
249 struct amdgpu_device *adev;
250 struct drm_device *ddev;
251 int i, count;
252
253 ddev = dev_get_drvdata(dev);
254 adev = drm_to_adev(ddev);
255 count = 0;
256
257 for (i = 0; i < DF_V3_6_MAX_COUNTERS; i++) {
258 if (adev->df_perfmon_config_assign_mask[i] == 0)
259 count++;
260 }
261
262 return snprintf(buf, PAGE_SIZE, "%i\n", count);
263 }
264
265 /* device attr for available perfmon counters */
266 static DEVICE_ATTR(df_cntr_avail, S_IRUGO, df_v3_6_get_df_cntr_avail, NULL);
267
df_v3_6_query_hashes(struct amdgpu_device * adev)268 static void df_v3_6_query_hashes(struct amdgpu_device *adev)
269 {
270 u32 tmp;
271
272 adev->df.hash_status.hash_64k = false;
273 adev->df.hash_status.hash_2m = false;
274 adev->df.hash_status.hash_1g = false;
275
276 if (adev->asic_type != CHIP_ARCTURUS)
277 return;
278
279 /* encoding for hash-enabled on Arcturus */
280 if (adev->df.funcs->get_fb_channel_number(adev) == 0xe) {
281 tmp = RREG32_SOC15(DF, 0, mmDF_CS_UMC_AON0_DfGlobalCtrl);
282 adev->df.hash_status.hash_64k = REG_GET_FIELD(tmp,
283 DF_CS_UMC_AON0_DfGlobalCtrl,
284 GlbHashIntlvCtl64K);
285 adev->df.hash_status.hash_2m = REG_GET_FIELD(tmp,
286 DF_CS_UMC_AON0_DfGlobalCtrl,
287 GlbHashIntlvCtl2M);
288 adev->df.hash_status.hash_1g = REG_GET_FIELD(tmp,
289 DF_CS_UMC_AON0_DfGlobalCtrl,
290 GlbHashIntlvCtl1G);
291 }
292 }
293
294 /* init perfmons */
df_v3_6_sw_init(struct amdgpu_device * adev)295 static void df_v3_6_sw_init(struct amdgpu_device *adev)
296 {
297 int i, ret;
298
299 ret = device_create_file(adev->dev, &dev_attr_df_cntr_avail);
300 if (ret)
301 DRM_ERROR("failed to create file for available df counters\n");
302
303 for (i = 0; i < AMDGPU_MAX_DF_PERFMONS; i++)
304 adev->df_perfmon_config_assign_mask[i] = 0;
305
306 df_v3_6_query_hashes(adev);
307 }
308
df_v3_6_sw_fini(struct amdgpu_device * adev)309 static void df_v3_6_sw_fini(struct amdgpu_device *adev)
310 {
311
312 device_remove_file(adev->dev, &dev_attr_df_cntr_avail);
313
314 }
315
df_v3_6_enable_broadcast_mode(struct amdgpu_device * adev,bool enable)316 static void df_v3_6_enable_broadcast_mode(struct amdgpu_device *adev,
317 bool enable)
318 {
319 u32 tmp;
320
321 if (enable) {
322 tmp = RREG32_SOC15(DF, 0, mmFabricConfigAccessControl);
323 tmp &= ~FabricConfigAccessControl__CfgRegInstAccEn_MASK;
324 WREG32_SOC15(DF, 0, mmFabricConfigAccessControl, tmp);
325 } else
326 WREG32_SOC15(DF, 0, mmFabricConfigAccessControl,
327 mmFabricConfigAccessControl_DEFAULT);
328 }
329
df_v3_6_get_fb_channel_number(struct amdgpu_device * adev)330 static u32 df_v3_6_get_fb_channel_number(struct amdgpu_device *adev)
331 {
332 u32 tmp;
333
334 tmp = RREG32_SOC15(DF, 0, mmDF_CS_UMC_AON0_DramBaseAddress0);
335 tmp &= DF_CS_UMC_AON0_DramBaseAddress0__IntLvNumChan_MASK;
336 tmp >>= DF_CS_UMC_AON0_DramBaseAddress0__IntLvNumChan__SHIFT;
337
338 return tmp;
339 }
340
df_v3_6_get_hbm_channel_number(struct amdgpu_device * adev)341 static u32 df_v3_6_get_hbm_channel_number(struct amdgpu_device *adev)
342 {
343 int fb_channel_number;
344
345 fb_channel_number = adev->df.funcs->get_fb_channel_number(adev);
346 if (fb_channel_number >= ARRAY_SIZE(df_v3_6_channel_number))
347 fb_channel_number = 0;
348
349 return df_v3_6_channel_number[fb_channel_number];
350 }
351
df_v3_6_update_medium_grain_clock_gating(struct amdgpu_device * adev,bool enable)352 static void df_v3_6_update_medium_grain_clock_gating(struct amdgpu_device *adev,
353 bool enable)
354 {
355 u32 tmp;
356
357 if (adev->cg_flags & AMD_CG_SUPPORT_DF_MGCG) {
358 /* Put DF on broadcast mode */
359 adev->df.funcs->enable_broadcast_mode(adev, true);
360
361 if (enable) {
362 tmp = RREG32_SOC15(DF, 0,
363 mmDF_PIE_AON0_DfGlobalClkGater);
364 tmp &= ~DF_PIE_AON0_DfGlobalClkGater__MGCGMode_MASK;
365 tmp |= DF_V3_6_MGCG_ENABLE_15_CYCLE_DELAY;
366 WREG32_SOC15(DF, 0,
367 mmDF_PIE_AON0_DfGlobalClkGater, tmp);
368 } else {
369 tmp = RREG32_SOC15(DF, 0,
370 mmDF_PIE_AON0_DfGlobalClkGater);
371 tmp &= ~DF_PIE_AON0_DfGlobalClkGater__MGCGMode_MASK;
372 tmp |= DF_V3_6_MGCG_DISABLE;
373 WREG32_SOC15(DF, 0,
374 mmDF_PIE_AON0_DfGlobalClkGater, tmp);
375 }
376
377 /* Exit broadcast mode */
378 adev->df.funcs->enable_broadcast_mode(adev, false);
379 }
380 }
381
df_v3_6_get_clockgating_state(struct amdgpu_device * adev,u32 * flags)382 static void df_v3_6_get_clockgating_state(struct amdgpu_device *adev,
383 u32 *flags)
384 {
385 u32 tmp;
386
387 /* AMD_CG_SUPPORT_DF_MGCG */
388 tmp = RREG32_SOC15(DF, 0, mmDF_PIE_AON0_DfGlobalClkGater);
389 if (tmp & DF_V3_6_MGCG_ENABLE_15_CYCLE_DELAY)
390 *flags |= AMD_CG_SUPPORT_DF_MGCG;
391 }
392
393 /* get assigned df perfmon ctr as int */
df_v3_6_pmc_config_2_cntr(struct amdgpu_device * adev,uint64_t config)394 static int df_v3_6_pmc_config_2_cntr(struct amdgpu_device *adev,
395 uint64_t config)
396 {
397 int i;
398
399 for (i = 0; i < DF_V3_6_MAX_COUNTERS; i++) {
400 if ((config & 0x0FFFFFFUL) ==
401 adev->df_perfmon_config_assign_mask[i])
402 return i;
403 }
404
405 return -EINVAL;
406 }
407
408 /* get address based on counter assignment */
df_v3_6_pmc_get_addr(struct amdgpu_device * adev,uint64_t config,int is_ctrl,uint32_t * lo_base_addr,uint32_t * hi_base_addr)409 static void df_v3_6_pmc_get_addr(struct amdgpu_device *adev,
410 uint64_t config,
411 int is_ctrl,
412 uint32_t *lo_base_addr,
413 uint32_t *hi_base_addr)
414 {
415 int target_cntr = df_v3_6_pmc_config_2_cntr(adev, config);
416
417 if (target_cntr < 0)
418 return;
419
420 switch (target_cntr) {
421
422 case 0:
423 *lo_base_addr = is_ctrl ? smnPerfMonCtlLo4 : smnPerfMonCtrLo4;
424 *hi_base_addr = is_ctrl ? smnPerfMonCtlHi4 : smnPerfMonCtrHi4;
425 break;
426 case 1:
427 *lo_base_addr = is_ctrl ? smnPerfMonCtlLo5 : smnPerfMonCtrLo5;
428 *hi_base_addr = is_ctrl ? smnPerfMonCtlHi5 : smnPerfMonCtrHi5;
429 break;
430 case 2:
431 *lo_base_addr = is_ctrl ? smnPerfMonCtlLo6 : smnPerfMonCtrLo6;
432 *hi_base_addr = is_ctrl ? smnPerfMonCtlHi6 : smnPerfMonCtrHi6;
433 break;
434 case 3:
435 *lo_base_addr = is_ctrl ? smnPerfMonCtlLo7 : smnPerfMonCtrLo7;
436 *hi_base_addr = is_ctrl ? smnPerfMonCtlHi7 : smnPerfMonCtrHi7;
437 break;
438
439 }
440
441 }
442
443 /* get read counter address */
df_v3_6_pmc_get_read_settings(struct amdgpu_device * adev,uint64_t config,uint32_t * lo_base_addr,uint32_t * hi_base_addr)444 static void df_v3_6_pmc_get_read_settings(struct amdgpu_device *adev,
445 uint64_t config,
446 uint32_t *lo_base_addr,
447 uint32_t *hi_base_addr)
448 {
449 df_v3_6_pmc_get_addr(adev, config, 0, lo_base_addr, hi_base_addr);
450 }
451
452 /* get control counter settings i.e. address and values to set */
df_v3_6_pmc_get_ctrl_settings(struct amdgpu_device * adev,uint64_t config,uint32_t * lo_base_addr,uint32_t * hi_base_addr,uint32_t * lo_val,uint32_t * hi_val,bool is_enable)453 static int df_v3_6_pmc_get_ctrl_settings(struct amdgpu_device *adev,
454 uint64_t config,
455 uint32_t *lo_base_addr,
456 uint32_t *hi_base_addr,
457 uint32_t *lo_val,
458 uint32_t *hi_val,
459 bool is_enable)
460 {
461
462 uint32_t eventsel, instance, unitmask;
463 uint32_t instance_10, instance_5432, instance_76;
464
465 df_v3_6_pmc_get_addr(adev, config, 1, lo_base_addr, hi_base_addr);
466
467 if ((*lo_base_addr == 0) || (*hi_base_addr == 0)) {
468 DRM_ERROR("[DF PMC] addressing not retrieved! Lo: %x, Hi: %x",
469 *lo_base_addr, *hi_base_addr);
470 return -ENXIO;
471 }
472
473 eventsel = DF_V3_6_GET_EVENT(config) & 0x3f;
474 unitmask = DF_V3_6_GET_UNITMASK(config) & 0xf;
475 instance = DF_V3_6_GET_INSTANCE(config);
476
477 instance_10 = instance & 0x3;
478 instance_5432 = (instance >> 2) & 0xf;
479 instance_76 = (instance >> 6) & 0x3;
480
481 *lo_val = (unitmask << 8) | (instance_10 << 6) | eventsel;
482 *lo_val = is_enable ? *lo_val | (1 << 22) : *lo_val & ~(1 << 22);
483 *hi_val = (instance_76 << 29) | instance_5432;
484
485 DRM_DEBUG_DRIVER("config=%llx addr=%08x:%08x val=%08x:%08x",
486 config, *lo_base_addr, *hi_base_addr, *lo_val, *hi_val);
487
488 return 0;
489 }
490
491 /* add df performance counters for read */
df_v3_6_pmc_add_cntr(struct amdgpu_device * adev,uint64_t config)492 static int df_v3_6_pmc_add_cntr(struct amdgpu_device *adev,
493 uint64_t config)
494 {
495 int i, target_cntr;
496
497 target_cntr = df_v3_6_pmc_config_2_cntr(adev, config);
498
499 if (target_cntr >= 0)
500 return 0;
501
502 for (i = 0; i < DF_V3_6_MAX_COUNTERS; i++) {
503 if (adev->df_perfmon_config_assign_mask[i] == 0U) {
504 adev->df_perfmon_config_assign_mask[i] =
505 config & 0x0FFFFFFUL;
506 return 0;
507 }
508 }
509
510 return -ENOSPC;
511 }
512
513 #define DEFERRED_ARM_MASK (1 << 31)
df_v3_6_pmc_set_deferred(struct amdgpu_device * adev,uint64_t config,bool is_deferred)514 static int df_v3_6_pmc_set_deferred(struct amdgpu_device *adev,
515 uint64_t config, bool is_deferred)
516 {
517 int target_cntr;
518
519 target_cntr = df_v3_6_pmc_config_2_cntr(adev, config);
520
521 if (target_cntr < 0)
522 return -EINVAL;
523
524 if (is_deferred)
525 adev->df_perfmon_config_assign_mask[target_cntr] |=
526 DEFERRED_ARM_MASK;
527 else
528 adev->df_perfmon_config_assign_mask[target_cntr] &=
529 ~DEFERRED_ARM_MASK;
530
531 return 0;
532 }
533
df_v3_6_pmc_is_deferred(struct amdgpu_device * adev,uint64_t config)534 static bool df_v3_6_pmc_is_deferred(struct amdgpu_device *adev,
535 uint64_t config)
536 {
537 int target_cntr;
538
539 target_cntr = df_v3_6_pmc_config_2_cntr(adev, config);
540
541 /*
542 * we never get target_cntr < 0 since this funciton is only called in
543 * pmc_count for now but we should check anyways.
544 */
545 return (target_cntr >= 0 &&
546 (adev->df_perfmon_config_assign_mask[target_cntr]
547 & DEFERRED_ARM_MASK));
548
549 }
550
551 /* release performance counter */
df_v3_6_pmc_release_cntr(struct amdgpu_device * adev,uint64_t config)552 static void df_v3_6_pmc_release_cntr(struct amdgpu_device *adev,
553 uint64_t config)
554 {
555 int target_cntr = df_v3_6_pmc_config_2_cntr(adev, config);
556
557 if (target_cntr >= 0)
558 adev->df_perfmon_config_assign_mask[target_cntr] = 0ULL;
559 }
560
561
df_v3_6_reset_perfmon_cntr(struct amdgpu_device * adev,uint64_t config)562 static void df_v3_6_reset_perfmon_cntr(struct amdgpu_device *adev,
563 uint64_t config)
564 {
565 uint32_t lo_base_addr = 0, hi_base_addr = 0;
566
567 df_v3_6_pmc_get_read_settings(adev, config, &lo_base_addr,
568 &hi_base_addr);
569
570 if ((lo_base_addr == 0) || (hi_base_addr == 0))
571 return;
572
573 df_v3_6_perfmon_wreg(adev, lo_base_addr, 0, hi_base_addr, 0);
574 }
575
df_v3_6_pmc_start(struct amdgpu_device * adev,uint64_t config,int is_add)576 static int df_v3_6_pmc_start(struct amdgpu_device *adev, uint64_t config,
577 int is_add)
578 {
579 uint32_t lo_base_addr, hi_base_addr, lo_val, hi_val;
580 int err = 0, ret = 0;
581
582 switch (adev->asic_type) {
583 case CHIP_VEGA20:
584 if (is_add)
585 return df_v3_6_pmc_add_cntr(adev, config);
586
587 df_v3_6_reset_perfmon_cntr(adev, config);
588
589 ret = df_v3_6_pmc_get_ctrl_settings(adev,
590 config,
591 &lo_base_addr,
592 &hi_base_addr,
593 &lo_val,
594 &hi_val,
595 true);
596
597 if (ret)
598 return ret;
599
600 err = df_v3_6_perfmon_arm_with_retry(adev,
601 lo_base_addr,
602 lo_val,
603 hi_base_addr,
604 hi_val);
605
606 if (err)
607 ret = df_v3_6_pmc_set_deferred(adev, config, true);
608
609 break;
610 default:
611 break;
612 }
613
614 return ret;
615 }
616
df_v3_6_pmc_stop(struct amdgpu_device * adev,uint64_t config,int is_remove)617 static int df_v3_6_pmc_stop(struct amdgpu_device *adev, uint64_t config,
618 int is_remove)
619 {
620 uint32_t lo_base_addr, hi_base_addr, lo_val, hi_val;
621 int ret = 0;
622
623 switch (adev->asic_type) {
624 case CHIP_VEGA20:
625 ret = df_v3_6_pmc_get_ctrl_settings(adev,
626 config,
627 &lo_base_addr,
628 &hi_base_addr,
629 &lo_val,
630 &hi_val,
631 false);
632
633 if (ret)
634 return ret;
635
636
637 if (is_remove) {
638 df_v3_6_reset_perfmon_cntr(adev, config);
639 df_v3_6_pmc_release_cntr(adev, config);
640 }
641
642 break;
643 default:
644 break;
645 }
646
647 return ret;
648 }
649
df_v3_6_pmc_get_count(struct amdgpu_device * adev,uint64_t config,uint64_t * count)650 static void df_v3_6_pmc_get_count(struct amdgpu_device *adev,
651 uint64_t config,
652 uint64_t *count)
653 {
654 uint32_t lo_base_addr = 0, hi_base_addr = 0, lo_val = 0, hi_val = 0;
655 *count = 0;
656
657 switch (adev->asic_type) {
658 case CHIP_VEGA20:
659 df_v3_6_pmc_get_read_settings(adev, config, &lo_base_addr,
660 &hi_base_addr);
661
662 if ((lo_base_addr == 0) || (hi_base_addr == 0))
663 return;
664
665 /* rearm the counter or throw away count value on failure */
666 if (df_v3_6_pmc_is_deferred(adev, config)) {
667 int rearm_err = df_v3_6_perfmon_arm_with_status(adev,
668 lo_base_addr, lo_val,
669 hi_base_addr, hi_val);
670
671 if (rearm_err)
672 return;
673
674 df_v3_6_pmc_set_deferred(adev, config, false);
675 }
676
677 df_v3_6_perfmon_rreg(adev, lo_base_addr, &lo_val,
678 hi_base_addr, &hi_val);
679
680 *count = ((hi_val | 0ULL) << 32) | (lo_val | 0ULL);
681
682 if (*count >= DF_V3_6_PERFMON_OVERFLOW)
683 *count = 0;
684
685 DRM_DEBUG_DRIVER("config=%llx addr=%08x:%08x val=%08x:%08x",
686 config, lo_base_addr, hi_base_addr, lo_val, hi_val);
687
688 break;
689 default:
690 break;
691 }
692 }
693
694 const struct amdgpu_df_funcs df_v3_6_funcs = {
695 .sw_init = df_v3_6_sw_init,
696 .sw_fini = df_v3_6_sw_fini,
697 .enable_broadcast_mode = df_v3_6_enable_broadcast_mode,
698 .get_fb_channel_number = df_v3_6_get_fb_channel_number,
699 .get_hbm_channel_number = df_v3_6_get_hbm_channel_number,
700 .update_medium_grain_clock_gating =
701 df_v3_6_update_medium_grain_clock_gating,
702 .get_clockgating_state = df_v3_6_get_clockgating_state,
703 .pmc_start = df_v3_6_pmc_start,
704 .pmc_stop = df_v3_6_pmc_stop,
705 .pmc_get_count = df_v3_6_pmc_get_count,
706 .get_fica = df_v3_6_get_fica,
707 .set_fica = df_v3_6_set_fica,
708 };
709