1 /*
2 * Copyright 2019 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 #include <linux/module.h>
23 #include <linux/uaccess.h>
24 #include <linux/firmware.h>
25 #include "amdgpu.h"
26 #include "amdgpu_amdkfd.h"
27 #include "amdgpu_amdkfd_arcturus.h"
28 #include "amdgpu_reset.h"
29 #include "sdma0/sdma0_4_2_2_offset.h"
30 #include "sdma0/sdma0_4_2_2_sh_mask.h"
31 #include "sdma1/sdma1_4_2_2_offset.h"
32 #include "sdma1/sdma1_4_2_2_sh_mask.h"
33 #include "sdma2/sdma2_4_2_2_offset.h"
34 #include "sdma2/sdma2_4_2_2_sh_mask.h"
35 #include "sdma3/sdma3_4_2_2_offset.h"
36 #include "sdma3/sdma3_4_2_2_sh_mask.h"
37 #include "sdma4/sdma4_4_2_2_offset.h"
38 #include "sdma4/sdma4_4_2_2_sh_mask.h"
39 #include "sdma5/sdma5_4_2_2_offset.h"
40 #include "sdma5/sdma5_4_2_2_sh_mask.h"
41 #include "sdma6/sdma6_4_2_2_offset.h"
42 #include "sdma6/sdma6_4_2_2_sh_mask.h"
43 #include "sdma7/sdma7_4_2_2_offset.h"
44 #include "sdma7/sdma7_4_2_2_sh_mask.h"
45 #include "v9_structs.h"
46 #include "soc15.h"
47 #include "soc15d.h"
48 #include "amdgpu_amdkfd_gfx_v9.h"
49 #include "gfxhub_v1_0.h"
50 #include "mmhub_v9_4.h"
51 #include "gc/gc_9_0_offset.h"
52 #include "gc/gc_9_0_sh_mask.h"
53
54 #define HQD_N_REGS 56
55 #define DUMP_REG(addr) do { \
56 if (WARN_ON_ONCE(i >= HQD_N_REGS)) \
57 break; \
58 (*dump)[i][0] = (addr) << 2; \
59 (*dump)[i++][1] = RREG32(addr); \
60 } while (0)
61
get_sdma_mqd(void * mqd)62 static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
63 {
64 return (struct v9_sdma_mqd *)mqd;
65 }
66
get_sdma_rlc_reg_offset(struct amdgpu_device * adev,unsigned int engine_id,unsigned int queue_id)67 static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
68 unsigned int engine_id,
69 unsigned int queue_id)
70 {
71 uint32_t sdma_engine_reg_base = 0;
72 uint32_t sdma_rlc_reg_offset;
73
74 switch (engine_id) {
75 default:
76 dev_warn(adev->dev,
77 "Invalid sdma engine id (%d), using engine id 0\n",
78 engine_id);
79 fallthrough;
80 case 0:
81 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA0, 0,
82 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL;
83 break;
84 case 1:
85 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA1, 0,
86 mmSDMA1_RLC0_RB_CNTL) - mmSDMA1_RLC0_RB_CNTL;
87 break;
88 case 2:
89 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA2, 0,
90 mmSDMA2_RLC0_RB_CNTL) - mmSDMA2_RLC0_RB_CNTL;
91 break;
92 case 3:
93 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA3, 0,
94 mmSDMA3_RLC0_RB_CNTL) - mmSDMA3_RLC0_RB_CNTL;
95 break;
96 case 4:
97 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA4, 0,
98 mmSDMA4_RLC0_RB_CNTL) - mmSDMA4_RLC0_RB_CNTL;
99 break;
100 case 5:
101 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA5, 0,
102 mmSDMA5_RLC0_RB_CNTL) - mmSDMA5_RLC0_RB_CNTL;
103 break;
104 case 6:
105 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA6, 0,
106 mmSDMA6_RLC0_RB_CNTL) - mmSDMA6_RLC0_RB_CNTL;
107 break;
108 case 7:
109 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA7, 0,
110 mmSDMA7_RLC0_RB_CNTL) - mmSDMA7_RLC0_RB_CNTL;
111 break;
112 }
113
114 sdma_rlc_reg_offset = sdma_engine_reg_base
115 + queue_id * (mmSDMA0_RLC1_RB_CNTL - mmSDMA0_RLC0_RB_CNTL);
116
117 pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
118 queue_id, sdma_rlc_reg_offset);
119
120 return sdma_rlc_reg_offset;
121 }
122
kgd_arcturus_hqd_sdma_load(struct amdgpu_device * adev,void * mqd,uint32_t __user * wptr,struct mm_struct * mm)123 int kgd_arcturus_hqd_sdma_load(struct amdgpu_device *adev, void *mqd,
124 uint32_t __user *wptr, struct mm_struct *mm)
125 {
126 struct v9_sdma_mqd *m;
127 uint32_t sdma_rlc_reg_offset;
128 unsigned long end_jiffies;
129 uint32_t data;
130 uint64_t data64;
131 uint64_t __user *wptr64 = (uint64_t __user *)wptr;
132
133 m = get_sdma_mqd(mqd);
134 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
135 m->sdma_queue_id);
136
137 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
138 m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
139
140 end_jiffies = msecs_to_jiffies(2000) + jiffies;
141 while (true) {
142 data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
143 if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
144 break;
145 if (time_after(jiffies, end_jiffies)) {
146 pr_err("SDMA RLC not idle in %s\n", __func__);
147 return -ETIME;
148 }
149 usleep_range(500, 1000);
150 }
151
152 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL_OFFSET,
153 m->sdmax_rlcx_doorbell_offset);
154
155 data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
156 ENABLE, 1);
157 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
158 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
159 m->sdmax_rlcx_rb_rptr);
160 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI,
161 m->sdmax_rlcx_rb_rptr_hi);
162
163 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
164 if (read_user_wptr(mm, wptr64, data64)) {
165 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
166 lower_32_bits(data64));
167 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
168 upper_32_bits(data64));
169 } else {
170 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
171 m->sdmax_rlcx_rb_rptr);
172 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
173 m->sdmax_rlcx_rb_rptr_hi);
174 }
175 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
176
177 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
178 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
179 m->sdmax_rlcx_rb_base_hi);
180 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
181 m->sdmax_rlcx_rb_rptr_addr_lo);
182 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
183 m->sdmax_rlcx_rb_rptr_addr_hi);
184
185 data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
186 RB_ENABLE, 1);
187 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
188
189 return 0;
190 }
191
kgd_arcturus_hqd_sdma_dump(struct amdgpu_device * adev,uint32_t engine_id,uint32_t queue_id,uint32_t (** dump)[2],uint32_t * n_regs)192 int kgd_arcturus_hqd_sdma_dump(struct amdgpu_device *adev,
193 uint32_t engine_id, uint32_t queue_id,
194 uint32_t (**dump)[2], uint32_t *n_regs)
195 {
196 uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
197 engine_id, queue_id);
198 uint32_t i = 0, reg;
199 #undef HQD_N_REGS
200 #define HQD_N_REGS (19+6+7+10)
201
202 *dump = kmalloc_array(HQD_N_REGS, sizeof(**dump), GFP_KERNEL);
203 if (*dump == NULL)
204 return -ENOMEM;
205
206 for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
207 DUMP_REG(sdma_rlc_reg_offset + reg);
208 for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
209 DUMP_REG(sdma_rlc_reg_offset + reg);
210 for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
211 reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
212 DUMP_REG(sdma_rlc_reg_offset + reg);
213 for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
214 reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
215 DUMP_REG(sdma_rlc_reg_offset + reg);
216
217 WARN_ON_ONCE(i != HQD_N_REGS);
218 *n_regs = i;
219
220 return 0;
221 }
222
kgd_arcturus_hqd_sdma_is_occupied(struct amdgpu_device * adev,void * mqd)223 bool kgd_arcturus_hqd_sdma_is_occupied(struct amdgpu_device *adev,
224 void *mqd)
225 {
226 struct v9_sdma_mqd *m;
227 uint32_t sdma_rlc_reg_offset;
228 uint32_t sdma_rlc_rb_cntl;
229
230 m = get_sdma_mqd(mqd);
231 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
232 m->sdma_queue_id);
233
234 sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
235
236 if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
237 return true;
238
239 return false;
240 }
241
kgd_arcturus_hqd_sdma_destroy(struct amdgpu_device * adev,void * mqd,unsigned int utimeout)242 int kgd_arcturus_hqd_sdma_destroy(struct amdgpu_device *adev, void *mqd,
243 unsigned int utimeout)
244 {
245 struct v9_sdma_mqd *m;
246 uint32_t sdma_rlc_reg_offset;
247 uint32_t temp;
248 unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
249
250 m = get_sdma_mqd(mqd);
251 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
252 m->sdma_queue_id);
253
254 temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
255 temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
256 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
257
258 while (true) {
259 temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
260 if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
261 break;
262 if (time_after(jiffies, end_jiffies)) {
263 pr_err("SDMA RLC not idle in %s\n", __func__);
264 return -ETIME;
265 }
266 usleep_range(500, 1000);
267 }
268
269 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
270 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
271 RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
272 SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
273
274 m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
275 m->sdmax_rlcx_rb_rptr_hi =
276 RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI);
277
278 return 0;
279 }
280
281 /*
282 * Helper used to suspend/resume gfx pipe for image post process work to set
283 * barrier behaviour.
284 */
suspend_resume_compute_scheduler(struct amdgpu_device * adev,bool suspend)285 static int suspend_resume_compute_scheduler(struct amdgpu_device *adev, bool suspend)
286 {
287 int i, r = 0;
288
289 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
290 struct amdgpu_ring *ring = &adev->gfx.compute_ring[i];
291
292 if (!amdgpu_ring_sched_ready(ring))
293 continue;
294
295 /* stop secheduler and drain ring. */
296 if (suspend) {
297 drm_sched_stop(&ring->sched, NULL);
298 r = amdgpu_fence_wait_empty(ring);
299 if (r)
300 goto out;
301 } else {
302 drm_sched_start(&ring->sched, 0);
303 }
304 }
305
306 out:
307 /* return on resume or failure to drain rings. */
308 if (!suspend || r)
309 return r;
310
311 return amdgpu_device_ip_wait_for_idle(adev, AMD_IP_BLOCK_TYPE_GFX);
312 }
313
set_barrier_auto_waitcnt(struct amdgpu_device * adev,bool enable_waitcnt)314 static void set_barrier_auto_waitcnt(struct amdgpu_device *adev, bool enable_waitcnt)
315 {
316 uint32_t data;
317
318 WRITE_ONCE(adev->barrier_has_auto_waitcnt, enable_waitcnt);
319
320 if (!down_read_trylock(&adev->reset_domain->sem))
321 return;
322
323 amdgpu_amdkfd_suspend(adev, false);
324
325 if (suspend_resume_compute_scheduler(adev, true))
326 goto out;
327
328 data = RREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CONFIG));
329 data = REG_SET_FIELD(data, SQ_CONFIG, DISABLE_BARRIER_WAITCNT,
330 !enable_waitcnt);
331 WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CONFIG), data);
332
333 out:
334 suspend_resume_compute_scheduler(adev, false);
335
336 amdgpu_amdkfd_resume(adev, false);
337
338 up_read(&adev->reset_domain->sem);
339 }
340
341 /*
342 * restore_dbg_registers is ignored here but is a general interface requirement
343 * for devices that support GFXOFF and where the RLC save/restore list
344 * does not support hw registers for debugging i.e. the driver has to manually
345 * initialize the debug mode registers after it has disabled GFX off during the
346 * debug session.
347 */
kgd_arcturus_enable_debug_trap(struct amdgpu_device * adev,bool restore_dbg_registers,uint32_t vmid)348 static uint32_t kgd_arcturus_enable_debug_trap(struct amdgpu_device *adev,
349 bool restore_dbg_registers,
350 uint32_t vmid)
351 {
352 mutex_lock(&adev->grbm_idx_mutex);
353
354 kgd_gfx_v9_set_wave_launch_stall(adev, vmid, true);
355
356 set_barrier_auto_waitcnt(adev, true);
357
358 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), 0);
359
360 kgd_gfx_v9_set_wave_launch_stall(adev, vmid, false);
361
362 mutex_unlock(&adev->grbm_idx_mutex);
363
364 return 0;
365 }
366
367 /*
368 * keep_trap_enabled is ignored here but is a general interface requirement
369 * for devices that support multi-process debugging where the performance
370 * overhead from trap temporary setup needs to be bypassed when the debug
371 * session has ended.
372 */
kgd_arcturus_disable_debug_trap(struct amdgpu_device * adev,bool keep_trap_enabled,uint32_t vmid)373 static uint32_t kgd_arcturus_disable_debug_trap(struct amdgpu_device *adev,
374 bool keep_trap_enabled,
375 uint32_t vmid)
376 {
377
378 mutex_lock(&adev->grbm_idx_mutex);
379
380 kgd_gfx_v9_set_wave_launch_stall(adev, vmid, true);
381
382 set_barrier_auto_waitcnt(adev, false);
383
384 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), 0);
385
386 kgd_gfx_v9_set_wave_launch_stall(adev, vmid, false);
387
388 mutex_unlock(&adev->grbm_idx_mutex);
389
390 return 0;
391 }
392 const struct kfd2kgd_calls arcturus_kfd2kgd = {
393 .program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings,
394 .set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping,
395 .init_interrupts = kgd_gfx_v9_init_interrupts,
396 .hqd_load = kgd_gfx_v9_hqd_load,
397 .hiq_mqd_load = kgd_gfx_v9_hiq_mqd_load,
398 .hqd_sdma_load = kgd_arcturus_hqd_sdma_load,
399 .hqd_dump = kgd_gfx_v9_hqd_dump,
400 .hqd_sdma_dump = kgd_arcturus_hqd_sdma_dump,
401 .hqd_is_occupied = kgd_gfx_v9_hqd_is_occupied,
402 .hqd_sdma_is_occupied = kgd_arcturus_hqd_sdma_is_occupied,
403 .hqd_destroy = kgd_gfx_v9_hqd_destroy,
404 .hqd_sdma_destroy = kgd_arcturus_hqd_sdma_destroy,
405 .wave_control_execute = kgd_gfx_v9_wave_control_execute,
406 .get_atc_vmid_pasid_mapping_info =
407 kgd_gfx_v9_get_atc_vmid_pasid_mapping_info,
408 .set_vm_context_page_table_base =
409 kgd_gfx_v9_set_vm_context_page_table_base,
410 .enable_debug_trap = kgd_arcturus_enable_debug_trap,
411 .disable_debug_trap = kgd_arcturus_disable_debug_trap,
412 .validate_trap_override_request = kgd_gfx_v9_validate_trap_override_request,
413 .set_wave_launch_trap_override = kgd_gfx_v9_set_wave_launch_trap_override,
414 .set_wave_launch_mode = kgd_gfx_v9_set_wave_launch_mode,
415 .set_address_watch = kgd_gfx_v9_set_address_watch,
416 .clear_address_watch = kgd_gfx_v9_clear_address_watch,
417 .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times,
418 .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info,
419 .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy,
420 .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings,
421 .hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
422 .hqd_reset = kgd_gfx_v9_hqd_reset
423 };
424