1 /*
2 * Copyright 2021 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/mmu_context.h>
23 #include "amdgpu.h"
24 #include "amdgpu_amdkfd.h"
25 #include "gc/gc_11_0_0_offset.h"
26 #include "gc/gc_11_0_0_sh_mask.h"
27 #include "oss/osssys_6_0_0_offset.h"
28 #include "oss/osssys_6_0_0_sh_mask.h"
29 #include "soc15_common.h"
30 #include "soc15d.h"
31 #include "v11_structs.h"
32 #include "soc21.h"
33 #include <uapi/linux/kfd_ioctl.h>
34
35 enum hqd_dequeue_request_type {
36 NO_ACTION = 0,
37 DRAIN_PIPE,
38 RESET_WAVES,
39 SAVE_WAVES
40 };
41
lock_srbm(struct amdgpu_device * adev,uint32_t mec,uint32_t pipe,uint32_t queue,uint32_t vmid)42 static void lock_srbm(struct amdgpu_device *adev, uint32_t mec, uint32_t pipe,
43 uint32_t queue, uint32_t vmid)
44 {
45 mutex_lock(&adev->srbm_mutex);
46 soc21_grbm_select(adev, mec, pipe, queue, vmid);
47 }
48
unlock_srbm(struct amdgpu_device * adev)49 static void unlock_srbm(struct amdgpu_device *adev)
50 {
51 soc21_grbm_select(adev, 0, 0, 0, 0);
52 mutex_unlock(&adev->srbm_mutex);
53 }
54
acquire_queue(struct amdgpu_device * adev,uint32_t pipe_id,uint32_t queue_id)55 static void acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id,
56 uint32_t queue_id)
57 {
58 uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
59 uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
60
61 lock_srbm(adev, mec, pipe, queue_id, 0);
62 }
63
get_queue_mask(struct amdgpu_device * adev,uint32_t pipe_id,uint32_t queue_id)64 static uint64_t get_queue_mask(struct amdgpu_device *adev,
65 uint32_t pipe_id, uint32_t queue_id)
66 {
67 unsigned int bit = pipe_id * adev->gfx.mec.num_queue_per_pipe +
68 queue_id;
69
70 return 1ull << bit;
71 }
72
release_queue(struct amdgpu_device * adev)73 static void release_queue(struct amdgpu_device *adev)
74 {
75 unlock_srbm(adev);
76 }
77
program_sh_mem_settings_v11(struct amdgpu_device * adev,uint32_t vmid,uint32_t sh_mem_config,uint32_t sh_mem_ape1_base,uint32_t sh_mem_ape1_limit,uint32_t sh_mem_bases,uint32_t inst)78 static void program_sh_mem_settings_v11(struct amdgpu_device *adev, uint32_t vmid,
79 uint32_t sh_mem_config,
80 uint32_t sh_mem_ape1_base,
81 uint32_t sh_mem_ape1_limit,
82 uint32_t sh_mem_bases, uint32_t inst)
83 {
84 lock_srbm(adev, 0, 0, 0, vmid);
85
86 WREG32(SOC15_REG_OFFSET(GC, 0, regSH_MEM_CONFIG), sh_mem_config);
87 WREG32(SOC15_REG_OFFSET(GC, 0, regSH_MEM_BASES), sh_mem_bases);
88
89 unlock_srbm(adev);
90 }
91
set_pasid_vmid_mapping_v11(struct amdgpu_device * adev,unsigned int pasid,unsigned int vmid,uint32_t inst)92 static int set_pasid_vmid_mapping_v11(struct amdgpu_device *adev, unsigned int pasid,
93 unsigned int vmid, uint32_t inst)
94 {
95 uint32_t value = pasid << IH_VMID_0_LUT__PASID__SHIFT;
96
97 /* Mapping vmid to pasid also for IH block */
98 pr_debug("mapping vmid %d -> pasid %d in IH block for GFX client\n",
99 vmid, pasid);
100 WREG32(SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid, value);
101
102 return 0;
103 }
104
init_interrupts_v11(struct amdgpu_device * adev,uint32_t pipe_id,uint32_t inst)105 static int init_interrupts_v11(struct amdgpu_device *adev, uint32_t pipe_id,
106 uint32_t inst)
107 {
108 uint32_t mec;
109 uint32_t pipe;
110
111 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
112 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
113
114 lock_srbm(adev, mec, pipe, 0, 0);
115
116 WREG32_SOC15(GC, 0, regCPC_INT_CNTL,
117 CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
118 CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
119
120 unlock_srbm(adev);
121
122 return 0;
123 }
124
get_sdma_rlc_reg_offset(struct amdgpu_device * adev,unsigned int engine_id,unsigned int queue_id)125 static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
126 unsigned int engine_id,
127 unsigned int queue_id)
128 {
129 uint32_t sdma_engine_reg_base = 0;
130 uint32_t sdma_rlc_reg_offset;
131
132 switch (engine_id) {
133 case 0:
134 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA0, 0,
135 regSDMA0_QUEUE0_RB_CNTL) - regSDMA0_QUEUE0_RB_CNTL;
136 break;
137 case 1:
138 sdma_engine_reg_base = SOC15_REG_OFFSET(SDMA1, 0,
139 regSDMA1_QUEUE0_RB_CNTL) - regSDMA0_QUEUE0_RB_CNTL;
140 break;
141 default:
142 BUG();
143 }
144
145 sdma_rlc_reg_offset = sdma_engine_reg_base
146 + queue_id * (regSDMA0_QUEUE1_RB_CNTL - regSDMA0_QUEUE0_RB_CNTL);
147
148 pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
149 queue_id, sdma_rlc_reg_offset);
150
151 return sdma_rlc_reg_offset;
152 }
153
get_mqd(void * mqd)154 static inline struct v11_compute_mqd *get_mqd(void *mqd)
155 {
156 return (struct v11_compute_mqd *)mqd;
157 }
158
get_sdma_mqd(void * mqd)159 static inline struct v11_sdma_mqd *get_sdma_mqd(void *mqd)
160 {
161 return (struct v11_sdma_mqd *)mqd;
162 }
163
hqd_load_v11(struct amdgpu_device * adev,void * mqd,uint32_t pipe_id,uint32_t queue_id,uint32_t __user * wptr,uint32_t wptr_shift,uint32_t wptr_mask,struct mm_struct * mm,uint32_t inst)164 static int hqd_load_v11(struct amdgpu_device *adev, void *mqd, uint32_t pipe_id,
165 uint32_t queue_id, uint32_t __user *wptr,
166 uint32_t wptr_shift, uint32_t wptr_mask,
167 struct mm_struct *mm, uint32_t inst)
168 {
169 struct v11_compute_mqd *m;
170 uint32_t *mqd_hqd;
171 uint32_t reg, hqd_base, data;
172
173 m = get_mqd(mqd);
174
175 pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id);
176 acquire_queue(adev, pipe_id, queue_id);
177
178 /* HIQ is set during driver init period with vmid set to 0*/
179 if (m->cp_hqd_vmid == 0) {
180 uint32_t value, mec, pipe;
181
182 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
183 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
184
185 pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
186 mec, pipe, queue_id);
187 value = RREG32(SOC15_REG_OFFSET(GC, 0, regRLC_CP_SCHEDULERS));
188 value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
189 ((mec << 5) | (pipe << 3) | queue_id | 0x80));
190 WREG32(SOC15_REG_OFFSET(GC, 0, regRLC_CP_SCHEDULERS), value);
191 }
192
193 /* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
194 mqd_hqd = &m->cp_mqd_base_addr_lo;
195 hqd_base = SOC15_REG_OFFSET(GC, 0, regCP_MQD_BASE_ADDR);
196
197 for (reg = hqd_base;
198 reg <= SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_HI); reg++)
199 WREG32(reg, mqd_hqd[reg - hqd_base]);
200
201
202 /* Activate doorbell logic before triggering WPTR poll. */
203 data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
204 CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
205 WREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL), data);
206
207 if (wptr) {
208 /* Don't read wptr with get_user because the user
209 * context may not be accessible (if this function
210 * runs in a work queue). Instead trigger a one-shot
211 * polling read from memory in the CP. This assumes
212 * that wptr is GPU-accessible in the queue's VMID via
213 * ATC or SVM. WPTR==RPTR before starting the poll so
214 * the CP starts fetching new commands from the right
215 * place.
216 *
217 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
218 * tricky. Assume that the queue didn't overflow. The
219 * number of valid bits in the 32-bit RPTR depends on
220 * the queue size. The remaining bits are taken from
221 * the saved 64-bit WPTR. If the WPTR wrapped, add the
222 * queue size.
223 */
224 uint32_t queue_size =
225 2 << REG_GET_FIELD(m->cp_hqd_pq_control,
226 CP_HQD_PQ_CONTROL, QUEUE_SIZE);
227 uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
228
229 if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
230 guessed_wptr += queue_size;
231 guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
232 guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
233
234 WREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_LO),
235 lower_32_bits(guessed_wptr));
236 WREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_HI),
237 upper_32_bits(guessed_wptr));
238 WREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR),
239 lower_32_bits((uint64_t)wptr));
240 WREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI),
241 upper_32_bits((uint64_t)wptr));
242 pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__,
243 (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
244 WREG32(SOC15_REG_OFFSET(GC, 0, regCP_PQ_WPTR_POLL_CNTL1),
245 (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
246 }
247
248 /* Start the EOP fetcher */
249 WREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_EOP_RPTR),
250 REG_SET_FIELD(m->cp_hqd_eop_rptr,
251 CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
252
253 data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
254 WREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_ACTIVE), data);
255
256 release_queue(adev);
257
258 return 0;
259 }
260
hiq_mqd_load_v11(struct amdgpu_device * adev,void * mqd,uint32_t pipe_id,uint32_t queue_id,uint32_t doorbell_off,uint32_t inst)261 static int hiq_mqd_load_v11(struct amdgpu_device *adev, void *mqd,
262 uint32_t pipe_id, uint32_t queue_id,
263 uint32_t doorbell_off, uint32_t inst)
264 {
265 struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring;
266 struct v11_compute_mqd *m;
267 uint32_t mec, pipe;
268 int r;
269
270 m = get_mqd(mqd);
271
272 acquire_queue(adev, pipe_id, queue_id);
273
274 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
275 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
276
277 pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
278 mec, pipe, queue_id);
279
280 spin_lock(&adev->gfx.kiq[0].ring_lock);
281 r = amdgpu_ring_alloc(kiq_ring, 7);
282 if (r) {
283 pr_err("Failed to alloc KIQ (%d).\n", r);
284 goto out_unlock;
285 }
286
287 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
288 amdgpu_ring_write(kiq_ring,
289 PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */
290 PACKET3_MAP_QUEUES_VMID(m->cp_hqd_vmid) | /* VMID */
291 PACKET3_MAP_QUEUES_QUEUE(queue_id) |
292 PACKET3_MAP_QUEUES_PIPE(pipe) |
293 PACKET3_MAP_QUEUES_ME((mec - 1)) |
294 PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */
295 PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */
296 PACKET3_MAP_QUEUES_ENGINE_SEL(1) | /* engine_sel: hiq */
297 PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */
298 amdgpu_ring_write(kiq_ring,
299 PACKET3_MAP_QUEUES_DOORBELL_OFFSET(doorbell_off));
300 amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_lo);
301 amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_hi);
302 amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_lo);
303 amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_hi);
304 amdgpu_ring_commit(kiq_ring);
305
306 out_unlock:
307 spin_unlock(&adev->gfx.kiq[0].ring_lock);
308 release_queue(adev);
309
310 return r;
311 }
312
hqd_dump_v11(struct amdgpu_device * adev,uint32_t pipe_id,uint32_t queue_id,uint32_t (** dump)[2],uint32_t * n_regs,uint32_t inst)313 static int hqd_dump_v11(struct amdgpu_device *adev,
314 uint32_t pipe_id, uint32_t queue_id,
315 uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst)
316 {
317 uint32_t i = 0, reg;
318 #define HQD_N_REGS 56
319 #define DUMP_REG(addr) do { \
320 if (WARN_ON_ONCE(i >= HQD_N_REGS)) \
321 break; \
322 (*dump)[i][0] = (addr) << 2; \
323 (*dump)[i++][1] = RREG32(addr); \
324 } while (0)
325
326 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
327 if (*dump == NULL)
328 return -ENOMEM;
329
330 acquire_queue(adev, pipe_id, queue_id);
331
332 for (reg = SOC15_REG_OFFSET(GC, 0, regCP_MQD_BASE_ADDR);
333 reg <= SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_WPTR_HI); reg++)
334 DUMP_REG(reg);
335
336 release_queue(adev);
337
338 WARN_ON_ONCE(i != HQD_N_REGS);
339 *n_regs = i;
340
341 return 0;
342 }
343
hqd_sdma_load_v11(struct amdgpu_device * adev,void * mqd,uint32_t __user * wptr,struct mm_struct * mm)344 static int hqd_sdma_load_v11(struct amdgpu_device *adev, void *mqd,
345 uint32_t __user *wptr, struct mm_struct *mm)
346 {
347 struct v11_sdma_mqd *m;
348 uint32_t sdma_rlc_reg_offset;
349 unsigned long end_jiffies;
350 uint32_t data;
351 uint64_t data64;
352 uint64_t __user *wptr64 = (uint64_t __user *)wptr;
353
354 m = get_sdma_mqd(mqd);
355 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
356 m->sdma_queue_id);
357
358 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_CNTL,
359 m->sdmax_rlcx_rb_cntl & (~SDMA0_QUEUE0_RB_CNTL__RB_ENABLE_MASK));
360
361 end_jiffies = msecs_to_jiffies(2000) + jiffies;
362 while (true) {
363 data = RREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_CONTEXT_STATUS);
364 if (data & SDMA0_QUEUE0_CONTEXT_STATUS__IDLE_MASK)
365 break;
366 if (time_after(jiffies, end_jiffies)) {
367 pr_err("SDMA RLC not idle in %s\n", __func__);
368 return -ETIME;
369 }
370 usleep_range(500, 1000);
371 }
372
373 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_DOORBELL_OFFSET,
374 m->sdmax_rlcx_doorbell_offset);
375
376 data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_QUEUE0_DOORBELL,
377 ENABLE, 1);
378 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_DOORBELL, data);
379 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_RPTR,
380 m->sdmax_rlcx_rb_rptr);
381 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_RPTR_HI,
382 m->sdmax_rlcx_rb_rptr_hi);
383
384 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_MINOR_PTR_UPDATE, 1);
385 if (read_user_wptr(mm, wptr64, data64)) {
386 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_WPTR,
387 lower_32_bits(data64));
388 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_WPTR_HI,
389 upper_32_bits(data64));
390 } else {
391 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_WPTR,
392 m->sdmax_rlcx_rb_rptr);
393 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_WPTR_HI,
394 m->sdmax_rlcx_rb_rptr_hi);
395 }
396 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_MINOR_PTR_UPDATE, 0);
397
398 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_BASE, m->sdmax_rlcx_rb_base);
399 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_BASE_HI,
400 m->sdmax_rlcx_rb_base_hi);
401 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_RPTR_ADDR_LO,
402 m->sdmax_rlcx_rb_rptr_addr_lo);
403 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_RPTR_ADDR_HI,
404 m->sdmax_rlcx_rb_rptr_addr_hi);
405
406 data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_QUEUE0_RB_CNTL,
407 RB_ENABLE, 1);
408 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_CNTL, data);
409
410 return 0;
411 }
412
hqd_sdma_dump_v11(struct amdgpu_device * adev,uint32_t engine_id,uint32_t queue_id,uint32_t (** dump)[2],uint32_t * n_regs)413 static int hqd_sdma_dump_v11(struct amdgpu_device *adev,
414 uint32_t engine_id, uint32_t queue_id,
415 uint32_t (**dump)[2], uint32_t *n_regs)
416 {
417 uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
418 engine_id, queue_id);
419 uint32_t i = 0, reg;
420 #undef HQD_N_REGS
421 #define HQD_N_REGS (7+11+1+12+12)
422
423 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
424 if (*dump == NULL)
425 return -ENOMEM;
426
427 for (reg = regSDMA0_QUEUE0_RB_CNTL;
428 reg <= regSDMA0_QUEUE0_RB_WPTR_HI; reg++)
429 DUMP_REG(sdma_rlc_reg_offset + reg);
430 for (reg = regSDMA0_QUEUE0_RB_RPTR_ADDR_HI;
431 reg <= regSDMA0_QUEUE0_DOORBELL; reg++)
432 DUMP_REG(sdma_rlc_reg_offset + reg);
433 for (reg = regSDMA0_QUEUE0_DOORBELL_LOG;
434 reg <= regSDMA0_QUEUE0_DOORBELL_LOG; reg++)
435 DUMP_REG(sdma_rlc_reg_offset + reg);
436 for (reg = regSDMA0_QUEUE0_DOORBELL_OFFSET;
437 reg <= regSDMA0_QUEUE0_RB_PREEMPT; reg++)
438 DUMP_REG(sdma_rlc_reg_offset + reg);
439 for (reg = regSDMA0_QUEUE0_MIDCMD_DATA0;
440 reg <= regSDMA0_QUEUE0_MIDCMD_CNTL; reg++)
441 DUMP_REG(sdma_rlc_reg_offset + reg);
442
443 WARN_ON_ONCE(i != HQD_N_REGS);
444 *n_regs = i;
445
446 return 0;
447 }
448
hqd_is_occupied_v11(struct amdgpu_device * adev,uint64_t queue_address,uint32_t pipe_id,uint32_t queue_id,uint32_t inst)449 static bool hqd_is_occupied_v11(struct amdgpu_device *adev, uint64_t queue_address,
450 uint32_t pipe_id, uint32_t queue_id, uint32_t inst)
451 {
452 uint32_t act;
453 bool retval = false;
454 uint32_t low, high;
455
456 acquire_queue(adev, pipe_id, queue_id);
457 act = RREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_ACTIVE));
458 if (act) {
459 low = lower_32_bits(queue_address >> 8);
460 high = upper_32_bits(queue_address >> 8);
461
462 if (low == RREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_BASE)) &&
463 high == RREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_PQ_BASE_HI)))
464 retval = true;
465 }
466 release_queue(adev);
467 return retval;
468 }
469
hqd_sdma_is_occupied_v11(struct amdgpu_device * adev,void * mqd)470 static bool hqd_sdma_is_occupied_v11(struct amdgpu_device *adev, void *mqd)
471 {
472 struct v11_sdma_mqd *m;
473 uint32_t sdma_rlc_reg_offset;
474 uint32_t sdma_rlc_rb_cntl;
475
476 m = get_sdma_mqd(mqd);
477 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
478 m->sdma_queue_id);
479
480 sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_CNTL);
481
482 if (sdma_rlc_rb_cntl & SDMA0_QUEUE0_RB_CNTL__RB_ENABLE_MASK)
483 return true;
484
485 return false;
486 }
487
hqd_destroy_v11(struct amdgpu_device * adev,void * mqd,enum kfd_preempt_type reset_type,unsigned int utimeout,uint32_t pipe_id,uint32_t queue_id,uint32_t inst)488 static int hqd_destroy_v11(struct amdgpu_device *adev, void *mqd,
489 enum kfd_preempt_type reset_type,
490 unsigned int utimeout, uint32_t pipe_id,
491 uint32_t queue_id, uint32_t inst)
492 {
493 enum hqd_dequeue_request_type type;
494 unsigned long end_jiffies;
495 uint32_t temp;
496 struct v11_compute_mqd *m = get_mqd(mqd);
497
498 acquire_queue(adev, pipe_id, queue_id);
499
500 if (m->cp_hqd_vmid == 0)
501 WREG32_FIELD15_PREREG(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
502
503 switch (reset_type) {
504 case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
505 type = DRAIN_PIPE;
506 break;
507 case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
508 type = RESET_WAVES;
509 break;
510 default:
511 type = DRAIN_PIPE;
512 break;
513 }
514
515 WREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_DEQUEUE_REQUEST), type);
516
517 end_jiffies = (utimeout * HZ / 1000) + jiffies;
518 while (true) {
519 temp = RREG32(SOC15_REG_OFFSET(GC, 0, regCP_HQD_ACTIVE));
520 if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
521 break;
522 if (time_after(jiffies, end_jiffies)) {
523 pr_err("cp queue pipe %d queue %d preemption failed\n",
524 pipe_id, queue_id);
525 release_queue(adev);
526 return -ETIME;
527 }
528 usleep_range(500, 1000);
529 }
530
531 release_queue(adev);
532 return 0;
533 }
534
hqd_sdma_destroy_v11(struct amdgpu_device * adev,void * mqd,unsigned int utimeout)535 static int hqd_sdma_destroy_v11(struct amdgpu_device *adev, void *mqd,
536 unsigned int utimeout)
537 {
538 struct v11_sdma_mqd *m;
539 uint32_t sdma_rlc_reg_offset;
540 uint32_t temp;
541 unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
542
543 m = get_sdma_mqd(mqd);
544 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
545 m->sdma_queue_id);
546
547 temp = RREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_CNTL);
548 temp = temp & ~SDMA0_QUEUE0_RB_CNTL__RB_ENABLE_MASK;
549 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_CNTL, temp);
550
551 while (true) {
552 temp = RREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_CONTEXT_STATUS);
553 if (temp & SDMA0_QUEUE0_CONTEXT_STATUS__IDLE_MASK)
554 break;
555 if (time_after(jiffies, end_jiffies)) {
556 pr_err("SDMA RLC not idle in %s\n", __func__);
557 return -ETIME;
558 }
559 usleep_range(500, 1000);
560 }
561
562 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_DOORBELL, 0);
563 WREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_CNTL,
564 RREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_CNTL) |
565 SDMA0_QUEUE0_RB_CNTL__RB_ENABLE_MASK);
566
567 m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_RPTR);
568 m->sdmax_rlcx_rb_rptr_hi =
569 RREG32(sdma_rlc_reg_offset + regSDMA0_QUEUE0_RB_RPTR_HI);
570
571 return 0;
572 }
573
wave_control_execute_v11(struct amdgpu_device * adev,uint32_t gfx_index_val,uint32_t sq_cmd,uint32_t inst)574 static int wave_control_execute_v11(struct amdgpu_device *adev,
575 uint32_t gfx_index_val,
576 uint32_t sq_cmd, uint32_t inst)
577 {
578 uint32_t data = 0;
579
580 mutex_lock(&adev->grbm_idx_mutex);
581
582 WREG32(SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_INDEX), gfx_index_val);
583 WREG32(SOC15_REG_OFFSET(GC, 0, regSQ_CMD), sq_cmd);
584
585 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
586 INSTANCE_BROADCAST_WRITES, 1);
587 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
588 SA_BROADCAST_WRITES, 1);
589 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
590 SE_BROADCAST_WRITES, 1);
591
592 WREG32(SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_INDEX), data);
593 mutex_unlock(&adev->grbm_idx_mutex);
594
595 return 0;
596 }
597
set_vm_context_page_table_base_v11(struct amdgpu_device * adev,uint32_t vmid,uint64_t page_table_base)598 static void set_vm_context_page_table_base_v11(struct amdgpu_device *adev,
599 uint32_t vmid, uint64_t page_table_base)
600 {
601 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
602 pr_err("trying to set page table base for wrong VMID %u\n",
603 vmid);
604 return;
605 }
606
607 /* SDMA is on gfxhub as well for gfx11 adapters */
608 adev->gfxhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
609 }
610
611 /*
612 * Returns TRAP_EN, EXCP_EN and EXCP_REPLACE.
613 *
614 * restore_dbg_registers is ignored here but is a general interface requirement
615 * for devices that support GFXOFF and where the RLC save/restore list
616 * does not support hw registers for debugging i.e. the driver has to manually
617 * initialize the debug mode registers after it has disabled GFX off during the
618 * debug session.
619 */
kgd_gfx_v11_enable_debug_trap(struct amdgpu_device * adev,bool restore_dbg_registers,uint32_t vmid)620 static uint32_t kgd_gfx_v11_enable_debug_trap(struct amdgpu_device *adev,
621 bool restore_dbg_registers,
622 uint32_t vmid)
623 {
624 uint32_t data = 0;
625
626 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1);
627 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_EN, 0);
628 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_REPLACE, 0);
629
630 return data;
631 }
632
633 /* Returns TRAP_EN, EXCP_EN and EXCP_REPLACE. */
kgd_gfx_v11_disable_debug_trap(struct amdgpu_device * adev,bool keep_trap_enabled,uint32_t vmid)634 static uint32_t kgd_gfx_v11_disable_debug_trap(struct amdgpu_device *adev,
635 bool keep_trap_enabled,
636 uint32_t vmid)
637 {
638 uint32_t data = 0;
639
640 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1);
641 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_EN, 0);
642 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_REPLACE, 0);
643
644 return data;
645 }
646
kgd_gfx_v11_validate_trap_override_request(struct amdgpu_device * adev,uint32_t trap_override,uint32_t * trap_mask_supported)647 static int kgd_gfx_v11_validate_trap_override_request(struct amdgpu_device *adev,
648 uint32_t trap_override,
649 uint32_t *trap_mask_supported)
650 {
651 *trap_mask_supported &= KFD_DBG_TRAP_MASK_FP_INVALID |
652 KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL |
653 KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO |
654 KFD_DBG_TRAP_MASK_FP_OVERFLOW |
655 KFD_DBG_TRAP_MASK_FP_UNDERFLOW |
656 KFD_DBG_TRAP_MASK_FP_INEXACT |
657 KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO |
658 KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH |
659 KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION;
660
661 if (adev->ip_versions[GC_HWIP][0] >= IP_VERSION(11, 0, 4))
662 *trap_mask_supported |= KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START |
663 KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END;
664
665 if (trap_override != KFD_DBG_TRAP_OVERRIDE_OR &&
666 trap_override != KFD_DBG_TRAP_OVERRIDE_REPLACE)
667 return -EPERM;
668
669 return 0;
670 }
671
trap_mask_map_sw_to_hw(uint32_t mask)672 static uint32_t trap_mask_map_sw_to_hw(uint32_t mask)
673 {
674 uint32_t trap_on_start = (mask & KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START) ? 1 : 0;
675 uint32_t trap_on_end = (mask & KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END) ? 1 : 0;
676 uint32_t excp_en = mask & (KFD_DBG_TRAP_MASK_FP_INVALID |
677 KFD_DBG_TRAP_MASK_FP_INPUT_DENORMAL |
678 KFD_DBG_TRAP_MASK_FP_DIVIDE_BY_ZERO |
679 KFD_DBG_TRAP_MASK_FP_OVERFLOW |
680 KFD_DBG_TRAP_MASK_FP_UNDERFLOW |
681 KFD_DBG_TRAP_MASK_FP_INEXACT |
682 KFD_DBG_TRAP_MASK_INT_DIVIDE_BY_ZERO |
683 KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH |
684 KFD_DBG_TRAP_MASK_DBG_MEMORY_VIOLATION);
685 uint32_t ret;
686
687 ret = REG_SET_FIELD(0, SPI_GDBG_PER_VMID_CNTL, EXCP_EN, excp_en);
688 ret = REG_SET_FIELD(ret, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_START, trap_on_start);
689 ret = REG_SET_FIELD(ret, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_END, trap_on_end);
690
691 return ret;
692 }
693
trap_mask_map_hw_to_sw(uint32_t mask)694 static uint32_t trap_mask_map_hw_to_sw(uint32_t mask)
695 {
696 uint32_t ret = REG_GET_FIELD(mask, SPI_GDBG_PER_VMID_CNTL, EXCP_EN);
697
698 if (REG_GET_FIELD(mask, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_START))
699 ret |= KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_START;
700
701 if (REG_GET_FIELD(mask, SPI_GDBG_PER_VMID_CNTL, TRAP_ON_END))
702 ret |= KFD_DBG_TRAP_MASK_TRAP_ON_WAVE_END;
703
704 return ret;
705 }
706
707 /* Returns TRAP_EN, EXCP_EN and EXCP_REPLACE. */
kgd_gfx_v11_set_wave_launch_trap_override(struct amdgpu_device * adev,uint32_t vmid,uint32_t trap_override,uint32_t trap_mask_bits,uint32_t trap_mask_request,uint32_t * trap_mask_prev,uint32_t kfd_dbg_trap_cntl_prev)708 static uint32_t kgd_gfx_v11_set_wave_launch_trap_override(struct amdgpu_device *adev,
709 uint32_t vmid,
710 uint32_t trap_override,
711 uint32_t trap_mask_bits,
712 uint32_t trap_mask_request,
713 uint32_t *trap_mask_prev,
714 uint32_t kfd_dbg_trap_cntl_prev)
715 {
716 uint32_t data = 0;
717
718 *trap_mask_prev = trap_mask_map_hw_to_sw(kfd_dbg_trap_cntl_prev);
719
720 data = (trap_mask_bits & trap_mask_request) | (*trap_mask_prev & ~trap_mask_request);
721 data = trap_mask_map_sw_to_hw(data);
722
723 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1);
724 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, EXCP_REPLACE, trap_override);
725
726 return data;
727 }
728
kgd_gfx_v11_set_wave_launch_mode(struct amdgpu_device * adev,uint8_t wave_launch_mode,uint32_t vmid)729 static uint32_t kgd_gfx_v11_set_wave_launch_mode(struct amdgpu_device *adev,
730 uint8_t wave_launch_mode,
731 uint32_t vmid)
732 {
733 uint32_t data = 0;
734
735 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, LAUNCH_MODE, wave_launch_mode);
736
737 return data;
738 }
739
740 #define TCP_WATCH_STRIDE (regTCP_WATCH1_ADDR_H - regTCP_WATCH0_ADDR_H)
kgd_gfx_v11_set_address_watch(struct amdgpu_device * adev,uint64_t watch_address,uint32_t watch_address_mask,uint32_t watch_id,uint32_t watch_mode,uint32_t debug_vmid,uint32_t inst)741 static uint32_t kgd_gfx_v11_set_address_watch(struct amdgpu_device *adev,
742 uint64_t watch_address,
743 uint32_t watch_address_mask,
744 uint32_t watch_id,
745 uint32_t watch_mode,
746 uint32_t debug_vmid,
747 uint32_t inst)
748 {
749 uint32_t watch_address_high;
750 uint32_t watch_address_low;
751 uint32_t watch_address_cntl;
752
753 watch_address_cntl = 0;
754 watch_address_low = lower_32_bits(watch_address);
755 watch_address_high = upper_32_bits(watch_address) & 0xffff;
756
757 watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
758 TCP_WATCH0_CNTL,
759 MODE,
760 watch_mode);
761
762 watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
763 TCP_WATCH0_CNTL,
764 MASK,
765 watch_address_mask >> 7);
766
767 watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
768 TCP_WATCH0_CNTL,
769 VALID,
770 1);
771
772 WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_H) +
773 (watch_id * TCP_WATCH_STRIDE)),
774 watch_address_high);
775
776 WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_L) +
777 (watch_id * TCP_WATCH_STRIDE)),
778 watch_address_low);
779
780 return watch_address_cntl;
781 }
782
kgd_gfx_v11_clear_address_watch(struct amdgpu_device * adev,uint32_t watch_id)783 static uint32_t kgd_gfx_v11_clear_address_watch(struct amdgpu_device *adev,
784 uint32_t watch_id)
785 {
786 return 0;
787 }
788
789 const struct kfd2kgd_calls gfx_v11_kfd2kgd = {
790 .program_sh_mem_settings = program_sh_mem_settings_v11,
791 .set_pasid_vmid_mapping = set_pasid_vmid_mapping_v11,
792 .init_interrupts = init_interrupts_v11,
793 .hqd_load = hqd_load_v11,
794 .hiq_mqd_load = hiq_mqd_load_v11,
795 .hqd_sdma_load = hqd_sdma_load_v11,
796 .hqd_dump = hqd_dump_v11,
797 .hqd_sdma_dump = hqd_sdma_dump_v11,
798 .hqd_is_occupied = hqd_is_occupied_v11,
799 .hqd_sdma_is_occupied = hqd_sdma_is_occupied_v11,
800 .hqd_destroy = hqd_destroy_v11,
801 .hqd_sdma_destroy = hqd_sdma_destroy_v11,
802 .wave_control_execute = wave_control_execute_v11,
803 .get_atc_vmid_pasid_mapping_info = NULL,
804 .set_vm_context_page_table_base = set_vm_context_page_table_base_v11,
805 .enable_debug_trap = kgd_gfx_v11_enable_debug_trap,
806 .disable_debug_trap = kgd_gfx_v11_disable_debug_trap,
807 .validate_trap_override_request = kgd_gfx_v11_validate_trap_override_request,
808 .set_wave_launch_trap_override = kgd_gfx_v11_set_wave_launch_trap_override,
809 .set_wave_launch_mode = kgd_gfx_v11_set_wave_launch_mode,
810 .set_address_watch = kgd_gfx_v11_set_address_watch,
811 .clear_address_watch = kgd_gfx_v11_clear_address_watch
812 };
813