Loading...
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 "amdgpu.h"
23#include "amdgpu_amdkfd.h"
24#include "amdgpu_amdkfd_gfx_v10.h"
25#include "gc/gc_10_1_0_offset.h"
26#include "gc/gc_10_1_0_sh_mask.h"
27#include "athub/athub_2_0_0_offset.h"
28#include "athub/athub_2_0_0_sh_mask.h"
29#include "oss/osssys_5_0_0_offset.h"
30#include "oss/osssys_5_0_0_sh_mask.h"
31#include "soc15_common.h"
32#include "v10_structs.h"
33#include "nv.h"
34#include "nvd.h"
35#include <uapi/linux/kfd_ioctl.h>
36
37enum hqd_dequeue_request_type {
38 NO_ACTION = 0,
39 DRAIN_PIPE,
40 RESET_WAVES,
41 SAVE_WAVES
42};
43
44static void lock_srbm(struct amdgpu_device *adev, uint32_t mec, uint32_t pipe,
45 uint32_t queue, uint32_t vmid)
46{
47 mutex_lock(&adev->srbm_mutex);
48 nv_grbm_select(adev, mec, pipe, queue, vmid);
49}
50
51static void unlock_srbm(struct amdgpu_device *adev)
52{
53 nv_grbm_select(adev, 0, 0, 0, 0);
54 mutex_unlock(&adev->srbm_mutex);
55}
56
57static void acquire_queue(struct amdgpu_device *adev, uint32_t pipe_id,
58 uint32_t queue_id)
59{
60 uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
61 uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
62
63 lock_srbm(adev, mec, pipe, queue_id, 0);
64}
65
66static uint64_t get_queue_mask(struct amdgpu_device *adev,
67 uint32_t pipe_id, uint32_t queue_id)
68{
69 unsigned int bit = pipe_id * adev->gfx.mec.num_queue_per_pipe +
70 queue_id;
71
72 return 1ull << bit;
73}
74
75static void release_queue(struct amdgpu_device *adev)
76{
77 unlock_srbm(adev);
78}
79
80static void kgd_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmid,
81 uint32_t sh_mem_config,
82 uint32_t sh_mem_ape1_base,
83 uint32_t sh_mem_ape1_limit,
84 uint32_t sh_mem_bases, uint32_t inst)
85{
86 lock_srbm(adev, 0, 0, 0, vmid);
87
88 WREG32_SOC15(GC, 0, mmSH_MEM_CONFIG, sh_mem_config);
89 WREG32_SOC15(GC, 0, mmSH_MEM_BASES, sh_mem_bases);
90 /* APE1 no longer exists on GFX9 */
91
92 unlock_srbm(adev);
93}
94
95static int kgd_set_pasid_vmid_mapping(struct amdgpu_device *adev, u32 pasid,
96 unsigned int vmid, uint32_t inst)
97{
98 /*
99 * We have to assume that there is no outstanding mapping.
100 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
101 * a mapping is in progress or because a mapping finished
102 * and the SW cleared it.
103 * So the protocol is to always wait & clear.
104 */
105 uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
106 ATC_VMID0_PASID_MAPPING__VALID_MASK;
107
108 pr_debug("pasid 0x%x vmid %d, reg value %x\n", pasid, vmid, pasid_mapping);
109
110 pr_debug("ATHUB, reg %x\n", SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid);
111 WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
112 pasid_mapping);
113
114#if 0
115 /* TODO: uncomment this code when the hardware support is ready. */
116 while (!(RREG32(SOC15_REG_OFFSET(
117 ATHUB, 0,
118 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
119 (1U << vmid)))
120 cpu_relax();
121
122 pr_debug("ATHUB mapping update finished\n");
123 WREG32(SOC15_REG_OFFSET(ATHUB, 0,
124 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
125 1U << vmid);
126#endif
127
128 /* Mapping vmid to pasid also for IH block */
129 pr_debug("update mapping for IH block and mmhub");
130 WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
131 pasid_mapping);
132
133 return 0;
134}
135
136/* TODO - RING0 form of field is obsolete, seems to date back to SI
137 * but still works
138 */
139
140static int kgd_init_interrupts(struct amdgpu_device *adev, uint32_t pipe_id,
141 uint32_t inst)
142{
143 uint32_t mec;
144 uint32_t pipe;
145
146 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
147 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
148
149 lock_srbm(adev, mec, pipe, 0, 0);
150
151 WREG32_SOC15(GC, 0, mmCPC_INT_CNTL,
152 CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
153 CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
154
155 unlock_srbm(adev);
156
157 return 0;
158}
159
160static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
161 unsigned int engine_id,
162 unsigned int queue_id)
163{
164 uint32_t sdma_engine_reg_base[2] = {
165 SOC15_REG_OFFSET(SDMA0, 0,
166 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
167 /* On gfx10, mmSDMA1_xxx registers are defined NOT based
168 * on SDMA1 base address (dw 0x1860) but based on SDMA0
169 * base address (dw 0x1260). Therefore use mmSDMA0_RLC0_RB_CNTL
170 * instead of mmSDMA1_RLC0_RB_CNTL for the base address calc
171 * below
172 */
173 SOC15_REG_OFFSET(SDMA1, 0,
174 mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL
175 };
176
177 uint32_t retval = sdma_engine_reg_base[engine_id]
178 + queue_id * (mmSDMA0_RLC1_RB_CNTL - mmSDMA0_RLC0_RB_CNTL);
179
180 pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
181 queue_id, retval);
182
183 return retval;
184}
185
186#if 0
187static uint32_t get_watch_base_addr(struct amdgpu_device *adev)
188{
189 uint32_t retval = SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) -
190 mmTCP_WATCH0_ADDR_H;
191
192 pr_debug("kfd: reg watch base address: 0x%x\n", retval);
193
194 return retval;
195}
196#endif
197
198static inline struct v10_compute_mqd *get_mqd(void *mqd)
199{
200 return (struct v10_compute_mqd *)mqd;
201}
202
203static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
204{
205 return (struct v10_sdma_mqd *)mqd;
206}
207
208static int kgd_hqd_load(struct amdgpu_device *adev, void *mqd,
209 uint32_t pipe_id, uint32_t queue_id,
210 uint32_t __user *wptr, uint32_t wptr_shift,
211 uint32_t wptr_mask, struct mm_struct *mm, uint32_t inst)
212{
213 struct v10_compute_mqd *m;
214 uint32_t *mqd_hqd;
215 uint32_t reg, hqd_base, data;
216
217 m = get_mqd(mqd);
218
219 pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id);
220 acquire_queue(adev, pipe_id, queue_id);
221
222 /* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
223 mqd_hqd = &m->cp_mqd_base_addr_lo;
224 hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
225
226 for (reg = hqd_base;
227 reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
228 WREG32_SOC15_IP(GC, reg, mqd_hqd[reg - hqd_base]);
229
230
231 /* Activate doorbell logic before triggering WPTR poll. */
232 data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
233 CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
234 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL, data);
235
236 if (wptr) {
237 /* Don't read wptr with get_user because the user
238 * context may not be accessible (if this function
239 * runs in a work queue). Instead trigger a one-shot
240 * polling read from memory in the CP. This assumes
241 * that wptr is GPU-accessible in the queue's VMID via
242 * ATC or SVM. WPTR==RPTR before starting the poll so
243 * the CP starts fetching new commands from the right
244 * place.
245 *
246 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
247 * tricky. Assume that the queue didn't overflow. The
248 * number of valid bits in the 32-bit RPTR depends on
249 * the queue size. The remaining bits are taken from
250 * the saved 64-bit WPTR. If the WPTR wrapped, add the
251 * queue size.
252 */
253 uint32_t queue_size =
254 2 << REG_GET_FIELD(m->cp_hqd_pq_control,
255 CP_HQD_PQ_CONTROL, QUEUE_SIZE);
256 uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
257
258 if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
259 guessed_wptr += queue_size;
260 guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
261 guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
262
263 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_LO,
264 lower_32_bits(guessed_wptr));
265 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_HI,
266 upper_32_bits(guessed_wptr));
267 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR,
268 lower_32_bits((uint64_t)wptr));
269 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI,
270 upper_32_bits((uint64_t)wptr));
271 pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__,
272 (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
273 WREG32_SOC15(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1,
274 (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
275 }
276
277 /* Start the EOP fetcher */
278 WREG32_SOC15(GC, 0, mmCP_HQD_EOP_RPTR,
279 REG_SET_FIELD(m->cp_hqd_eop_rptr,
280 CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
281
282 data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
283 WREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE, data);
284
285 release_queue(adev);
286
287 return 0;
288}
289
290static int kgd_hiq_mqd_load(struct amdgpu_device *adev, void *mqd,
291 uint32_t pipe_id, uint32_t queue_id,
292 uint32_t doorbell_off, uint32_t inst)
293{
294 struct amdgpu_ring *kiq_ring = &adev->gfx.kiq[0].ring;
295 struct v10_compute_mqd *m;
296 uint32_t mec, pipe;
297 int r;
298
299 m = get_mqd(mqd);
300
301 acquire_queue(adev, pipe_id, queue_id);
302
303 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
304 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
305
306 pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
307 mec, pipe, queue_id);
308
309 spin_lock(&adev->gfx.kiq[0].ring_lock);
310 r = amdgpu_ring_alloc(kiq_ring, 7);
311 if (r) {
312 pr_err("Failed to alloc KIQ (%d).\n", r);
313 goto out_unlock;
314 }
315
316 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
317 amdgpu_ring_write(kiq_ring,
318 PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */
319 PACKET3_MAP_QUEUES_VMID(m->cp_hqd_vmid) | /* VMID */
320 PACKET3_MAP_QUEUES_QUEUE(queue_id) |
321 PACKET3_MAP_QUEUES_PIPE(pipe) |
322 PACKET3_MAP_QUEUES_ME((mec - 1)) |
323 PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */
324 PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */
325 PACKET3_MAP_QUEUES_ENGINE_SEL(1) | /* engine_sel: hiq */
326 PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */
327 amdgpu_ring_write(kiq_ring,
328 PACKET3_MAP_QUEUES_DOORBELL_OFFSET(doorbell_off));
329 amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_lo);
330 amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_hi);
331 amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_lo);
332 amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_hi);
333 amdgpu_ring_commit(kiq_ring);
334
335out_unlock:
336 spin_unlock(&adev->gfx.kiq[0].ring_lock);
337 release_queue(adev);
338
339 return r;
340}
341
342static int kgd_hqd_dump(struct amdgpu_device *adev,
343 uint32_t pipe_id, uint32_t queue_id,
344 uint32_t (**dump)[2], uint32_t *n_regs, uint32_t inst)
345{
346 uint32_t i = 0, reg;
347#define HQD_N_REGS 56
348#define DUMP_REG(addr) do { \
349 if (WARN_ON_ONCE(i >= HQD_N_REGS)) \
350 break; \
351 (*dump)[i][0] = (addr) << 2; \
352 (*dump)[i++][1] = RREG32_SOC15_IP(GC, addr); \
353 } while (0)
354
355 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
356 if (*dump == NULL)
357 return -ENOMEM;
358
359 acquire_queue(adev, pipe_id, queue_id);
360
361 for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
362 reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
363 DUMP_REG(reg);
364
365 release_queue(adev);
366
367 WARN_ON_ONCE(i != HQD_N_REGS);
368 *n_regs = i;
369
370 return 0;
371}
372
373static int kgd_hqd_sdma_load(struct amdgpu_device *adev, void *mqd,
374 uint32_t __user *wptr, struct mm_struct *mm)
375{
376 struct v10_sdma_mqd *m;
377 uint32_t sdma_rlc_reg_offset;
378 unsigned long end_jiffies;
379 uint32_t data;
380 uint64_t data64;
381 uint64_t __user *wptr64 = (uint64_t __user *)wptr;
382
383 m = get_sdma_mqd(mqd);
384 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
385 m->sdma_queue_id);
386
387 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
388 m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
389
390 end_jiffies = msecs_to_jiffies(2000) + jiffies;
391 while (true) {
392 data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
393 if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
394 break;
395 if (time_after(jiffies, end_jiffies)) {
396 pr_err("SDMA RLC not idle in %s\n", __func__);
397 return -ETIME;
398 }
399 usleep_range(500, 1000);
400 }
401
402 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL_OFFSET,
403 m->sdmax_rlcx_doorbell_offset);
404
405 data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
406 ENABLE, 1);
407 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
408 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
409 m->sdmax_rlcx_rb_rptr);
410 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI,
411 m->sdmax_rlcx_rb_rptr_hi);
412
413 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
414 if (read_user_wptr(mm, wptr64, data64)) {
415 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
416 lower_32_bits(data64));
417 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
418 upper_32_bits(data64));
419 } else {
420 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
421 m->sdmax_rlcx_rb_rptr);
422 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
423 m->sdmax_rlcx_rb_rptr_hi);
424 }
425 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
426
427 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
428 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
429 m->sdmax_rlcx_rb_base_hi);
430 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
431 m->sdmax_rlcx_rb_rptr_addr_lo);
432 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
433 m->sdmax_rlcx_rb_rptr_addr_hi);
434
435 data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
436 RB_ENABLE, 1);
437 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
438
439 return 0;
440}
441
442static int kgd_hqd_sdma_dump(struct amdgpu_device *adev,
443 uint32_t engine_id, uint32_t queue_id,
444 uint32_t (**dump)[2], uint32_t *n_regs)
445{
446 uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
447 engine_id, queue_id);
448 uint32_t i = 0, reg;
449#undef HQD_N_REGS
450#define HQD_N_REGS (19+6+7+10)
451
452 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
453 if (*dump == NULL)
454 return -ENOMEM;
455
456 for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
457 DUMP_REG(sdma_rlc_reg_offset + reg);
458 for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
459 DUMP_REG(sdma_rlc_reg_offset + reg);
460 for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
461 reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
462 DUMP_REG(sdma_rlc_reg_offset + reg);
463 for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
464 reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
465 DUMP_REG(sdma_rlc_reg_offset + reg);
466
467 WARN_ON_ONCE(i != HQD_N_REGS);
468 *n_regs = i;
469
470 return 0;
471}
472
473static bool kgd_hqd_is_occupied(struct amdgpu_device *adev,
474 uint64_t queue_address, uint32_t pipe_id,
475 uint32_t queue_id, uint32_t inst)
476{
477 uint32_t act;
478 bool retval = false;
479 uint32_t low, high;
480
481 acquire_queue(adev, pipe_id, queue_id);
482 act = RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE);
483 if (act) {
484 low = lower_32_bits(queue_address >> 8);
485 high = upper_32_bits(queue_address >> 8);
486
487 if (low == RREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE) &&
488 high == RREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE_HI))
489 retval = true;
490 }
491 release_queue(adev);
492 return retval;
493}
494
495static bool kgd_hqd_sdma_is_occupied(struct amdgpu_device *adev, void *mqd)
496{
497 struct v10_sdma_mqd *m;
498 uint32_t sdma_rlc_reg_offset;
499 uint32_t sdma_rlc_rb_cntl;
500
501 m = get_sdma_mqd(mqd);
502 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
503 m->sdma_queue_id);
504
505 sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
506
507 if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
508 return true;
509
510 return false;
511}
512
513static int kgd_hqd_destroy(struct amdgpu_device *adev, void *mqd,
514 enum kfd_preempt_type reset_type,
515 unsigned int utimeout, uint32_t pipe_id,
516 uint32_t queue_id, uint32_t inst)
517{
518 enum hqd_dequeue_request_type type;
519 unsigned long end_jiffies;
520 uint32_t temp;
521 struct v10_compute_mqd *m = get_mqd(mqd);
522
523 if (amdgpu_in_reset(adev))
524 return -EIO;
525
526#if 0
527 unsigned long flags;
528 int retry;
529#endif
530
531 acquire_queue(adev, pipe_id, queue_id);
532
533 if (m->cp_hqd_vmid == 0)
534 WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
535
536 switch (reset_type) {
537 case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
538 type = DRAIN_PIPE;
539 break;
540 case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
541 type = RESET_WAVES;
542 break;
543 case KFD_PREEMPT_TYPE_WAVEFRONT_SAVE:
544 type = SAVE_WAVES;
545 break;
546 default:
547 type = DRAIN_PIPE;
548 break;
549 }
550
551#if 0 /* Is this still needed? */
552 /* Workaround: If IQ timer is active and the wait time is close to or
553 * equal to 0, dequeueing is not safe. Wait until either the wait time
554 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
555 * cleared before continuing. Also, ensure wait times are set to at
556 * least 0x3.
557 */
558 local_irq_save(flags);
559 preempt_disable();
560 retry = 5000; /* wait for 500 usecs at maximum */
561 while (true) {
562 temp = RREG32(mmCP_HQD_IQ_TIMER);
563 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
564 pr_debug("HW is processing IQ\n");
565 goto loop;
566 }
567 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
568 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
569 == 3) /* SEM-rearm is safe */
570 break;
571 /* Wait time 3 is safe for CP, but our MMIO read/write
572 * time is close to 1 microsecond, so check for 10 to
573 * leave more buffer room
574 */
575 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
576 >= 10)
577 break;
578 pr_debug("IQ timer is active\n");
579 } else
580 break;
581loop:
582 if (!retry) {
583 pr_err("CP HQD IQ timer status time out\n");
584 break;
585 }
586 ndelay(100);
587 --retry;
588 }
589 retry = 1000;
590 while (true) {
591 temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
592 if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
593 break;
594 pr_debug("Dequeue request is pending\n");
595
596 if (!retry) {
597 pr_err("CP HQD dequeue request time out\n");
598 break;
599 }
600 ndelay(100);
601 --retry;
602 }
603 local_irq_restore(flags);
604 preempt_enable();
605#endif
606
607 WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST, type);
608
609 end_jiffies = (utimeout * HZ / 1000) + jiffies;
610 while (true) {
611 temp = RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE);
612 if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
613 break;
614 if (time_after(jiffies, end_jiffies)) {
615 pr_err("cp queue preemption time out.\n");
616 release_queue(adev);
617 return -ETIME;
618 }
619 usleep_range(500, 1000);
620 }
621
622 release_queue(adev);
623 return 0;
624}
625
626static int kgd_hqd_sdma_destroy(struct amdgpu_device *adev, void *mqd,
627 unsigned int utimeout)
628{
629 struct v10_sdma_mqd *m;
630 uint32_t sdma_rlc_reg_offset;
631 uint32_t temp;
632 unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
633
634 m = get_sdma_mqd(mqd);
635 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
636 m->sdma_queue_id);
637
638 temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
639 temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
640 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
641
642 while (true) {
643 temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
644 if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
645 break;
646 if (time_after(jiffies, end_jiffies)) {
647 pr_err("SDMA RLC not idle in %s\n", __func__);
648 return -ETIME;
649 }
650 usleep_range(500, 1000);
651 }
652
653 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
654 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
655 RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
656 SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
657
658 m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
659 m->sdmax_rlcx_rb_rptr_hi =
660 RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI);
661
662 return 0;
663}
664
665static bool get_atc_vmid_pasid_mapping_info(struct amdgpu_device *adev,
666 uint8_t vmid, uint16_t *p_pasid)
667{
668 uint32_t value;
669
670 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
671 + vmid);
672 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
673
674 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
675}
676
677static int kgd_wave_control_execute(struct amdgpu_device *adev,
678 uint32_t gfx_index_val,
679 uint32_t sq_cmd, uint32_t inst)
680{
681 uint32_t data = 0;
682
683 mutex_lock(&adev->grbm_idx_mutex);
684
685 WREG32_SOC15(GC, 0, mmGRBM_GFX_INDEX, gfx_index_val);
686 WREG32_SOC15(GC, 0, mmSQ_CMD, sq_cmd);
687
688 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
689 INSTANCE_BROADCAST_WRITES, 1);
690 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
691 SA_BROADCAST_WRITES, 1);
692 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
693 SE_BROADCAST_WRITES, 1);
694
695 WREG32_SOC15(GC, 0, mmGRBM_GFX_INDEX, data);
696 mutex_unlock(&adev->grbm_idx_mutex);
697
698 return 0;
699}
700
701static void set_vm_context_page_table_base(struct amdgpu_device *adev,
702 uint32_t vmid, uint64_t page_table_base)
703{
704 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
705 pr_err("trying to set page table base for wrong VMID %u\n",
706 vmid);
707 return;
708 }
709
710 /* SDMA is on gfxhub as well for Navi1* series */
711 adev->gfxhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
712}
713
714/*
715 * GFX10 helper for wave launch stall requirements on debug trap setting.
716 *
717 * vmid:
718 * Target VMID to stall/unstall.
719 *
720 * stall:
721 * 0-unstall wave launch (enable), 1-stall wave launch (disable).
722 * After wavefront launch has been stalled, allocated waves must drain from
723 * SPI in order for debug trap settings to take effect on those waves.
724 * This is roughly a ~3500 clock cycle wait on SPI where a read on
725 * SPI_GDBG_WAVE_CNTL translates to ~32 clock cycles.
726 * KGD_GFX_V10_WAVE_LAUNCH_SPI_DRAIN_LATENCY indicates the number of reads required.
727 *
728 * NOTE: We can afford to clear the entire STALL_VMID field on unstall
729 * because current GFX10 chips cannot support multi-process debugging due to
730 * trap configuration and masking being limited to global scope. Always
731 * assume single process conditions.
732 *
733 */
734
735#define KGD_GFX_V10_WAVE_LAUNCH_SPI_DRAIN_LATENCY 110
736static void kgd_gfx_v10_set_wave_launch_stall(struct amdgpu_device *adev, uint32_t vmid, bool stall)
737{
738 uint32_t data = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL));
739 int i;
740
741 data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL, STALL_VMID,
742 stall ? 1 << vmid : 0);
743
744 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), data);
745
746 if (!stall)
747 return;
748
749 for (i = 0; i < KGD_GFX_V10_WAVE_LAUNCH_SPI_DRAIN_LATENCY; i++)
750 RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL));
751}
752
753uint32_t kgd_gfx_v10_enable_debug_trap(struct amdgpu_device *adev,
754 bool restore_dbg_registers,
755 uint32_t vmid)
756{
757
758 mutex_lock(&adev->grbm_idx_mutex);
759
760 kgd_gfx_v10_set_wave_launch_stall(adev, vmid, true);
761
762 /* assume gfx off is disabled for the debug session if rlc restore not supported. */
763 if (restore_dbg_registers) {
764 uint32_t data = 0;
765
766 data = REG_SET_FIELD(data, SPI_GDBG_TRAP_CONFIG,
767 VMID_SEL, 1 << vmid);
768 data = REG_SET_FIELD(data, SPI_GDBG_TRAP_CONFIG,
769 TRAP_EN, 1);
770 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_CONFIG), data);
771 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_DATA0), 0);
772 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_DATA1), 0);
773
774 kgd_gfx_v10_set_wave_launch_stall(adev, vmid, false);
775
776 mutex_unlock(&adev->grbm_idx_mutex);
777
778 return 0;
779 }
780
781 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), 0);
782
783 kgd_gfx_v10_set_wave_launch_stall(adev, vmid, false);
784
785 mutex_unlock(&adev->grbm_idx_mutex);
786
787 return 0;
788}
789
790uint32_t kgd_gfx_v10_disable_debug_trap(struct amdgpu_device *adev,
791 bool keep_trap_enabled,
792 uint32_t vmid)
793{
794 mutex_lock(&adev->grbm_idx_mutex);
795
796 kgd_gfx_v10_set_wave_launch_stall(adev, vmid, true);
797
798 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), 0);
799
800 kgd_gfx_v10_set_wave_launch_stall(adev, vmid, false);
801
802 mutex_unlock(&adev->grbm_idx_mutex);
803
804 return 0;
805}
806
807int kgd_gfx_v10_validate_trap_override_request(struct amdgpu_device *adev,
808 uint32_t trap_override,
809 uint32_t *trap_mask_supported)
810{
811 *trap_mask_supported &= KFD_DBG_TRAP_MASK_DBG_ADDRESS_WATCH;
812
813 /* The SPI_GDBG_TRAP_MASK register is global and affects all
814 * processes. Only allow OR-ing the address-watch bit, since
815 * this only affects processes under the debugger. Other bits
816 * should stay 0 to avoid the debugger interfering with other
817 * processes.
818 */
819 if (trap_override != KFD_DBG_TRAP_OVERRIDE_OR)
820 return -EINVAL;
821
822 return 0;
823}
824
825uint32_t kgd_gfx_v10_set_wave_launch_trap_override(struct amdgpu_device *adev,
826 uint32_t vmid,
827 uint32_t trap_override,
828 uint32_t trap_mask_bits,
829 uint32_t trap_mask_request,
830 uint32_t *trap_mask_prev,
831 uint32_t kfd_dbg_trap_cntl_prev)
832{
833 uint32_t data, wave_cntl_prev;
834
835 mutex_lock(&adev->grbm_idx_mutex);
836
837 wave_cntl_prev = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL));
838
839 kgd_gfx_v10_set_wave_launch_stall(adev, vmid, true);
840
841 data = RREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK));
842 *trap_mask_prev = REG_GET_FIELD(data, SPI_GDBG_TRAP_MASK, EXCP_EN);
843
844 trap_mask_bits = (trap_mask_bits & trap_mask_request) |
845 (*trap_mask_prev & ~trap_mask_request);
846
847 data = REG_SET_FIELD(data, SPI_GDBG_TRAP_MASK, EXCP_EN, trap_mask_bits);
848 data = REG_SET_FIELD(data, SPI_GDBG_TRAP_MASK, REPLACE, trap_override);
849 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_TRAP_MASK), data);
850
851 /* We need to preserve wave launch mode stall settings. */
852 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL), wave_cntl_prev);
853
854 mutex_unlock(&adev->grbm_idx_mutex);
855
856 return 0;
857}
858
859uint32_t kgd_gfx_v10_set_wave_launch_mode(struct amdgpu_device *adev,
860 uint8_t wave_launch_mode,
861 uint32_t vmid)
862{
863 uint32_t data = 0;
864 bool is_mode_set = !!wave_launch_mode;
865
866 mutex_lock(&adev->grbm_idx_mutex);
867
868 kgd_gfx_v10_set_wave_launch_stall(adev, vmid, true);
869
870 data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL2,
871 VMID_MASK, is_mode_set ? 1 << vmid : 0);
872 data = REG_SET_FIELD(data, SPI_GDBG_WAVE_CNTL2,
873 MODE, is_mode_set ? wave_launch_mode : 0);
874 WREG32(SOC15_REG_OFFSET(GC, 0, mmSPI_GDBG_WAVE_CNTL2), data);
875
876 kgd_gfx_v10_set_wave_launch_stall(adev, vmid, false);
877
878 mutex_unlock(&adev->grbm_idx_mutex);
879
880 return 0;
881}
882
883#define TCP_WATCH_STRIDE (mmTCP_WATCH1_ADDR_H - mmTCP_WATCH0_ADDR_H)
884uint32_t kgd_gfx_v10_set_address_watch(struct amdgpu_device *adev,
885 uint64_t watch_address,
886 uint32_t watch_address_mask,
887 uint32_t watch_id,
888 uint32_t watch_mode,
889 uint32_t debug_vmid,
890 uint32_t inst)
891{
892 uint32_t watch_address_high;
893 uint32_t watch_address_low;
894 uint32_t watch_address_cntl;
895
896 watch_address_cntl = 0;
897
898 watch_address_low = lower_32_bits(watch_address);
899 watch_address_high = upper_32_bits(watch_address) & 0xffff;
900
901 watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
902 TCP_WATCH0_CNTL,
903 VMID,
904 debug_vmid);
905 watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
906 TCP_WATCH0_CNTL,
907 MODE,
908 watch_mode);
909 watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
910 TCP_WATCH0_CNTL,
911 MASK,
912 watch_address_mask >> 7);
913
914 /* Turning off this watch point until we set all the registers */
915 watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
916 TCP_WATCH0_CNTL,
917 VALID,
918 0);
919
920 WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
921 (watch_id * TCP_WATCH_STRIDE)),
922 watch_address_cntl);
923
924 WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) +
925 (watch_id * TCP_WATCH_STRIDE)),
926 watch_address_high);
927
928 WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_L) +
929 (watch_id * TCP_WATCH_STRIDE)),
930 watch_address_low);
931
932 /* Enable the watch point */
933 watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
934 TCP_WATCH0_CNTL,
935 VALID,
936 1);
937
938 WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
939 (watch_id * TCP_WATCH_STRIDE)),
940 watch_address_cntl);
941
942 return 0;
943}
944
945uint32_t kgd_gfx_v10_clear_address_watch(struct amdgpu_device *adev,
946 uint32_t watch_id)
947{
948 uint32_t watch_address_cntl;
949
950 watch_address_cntl = 0;
951
952 WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
953 (watch_id * TCP_WATCH_STRIDE)),
954 watch_address_cntl);
955
956 return 0;
957}
958
959
960/* kgd_gfx_v10_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values
961 * The values read are:
962 * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads.
963 * atomic_offload_wait_time -- Wait Count for L2 and GDS Atomics Offloads.
964 * wrm_offload_wait_time -- Wait Count for WAIT_REG_MEM Offloads.
965 * gws_wait_time -- Wait Count for Global Wave Syncs.
966 * que_sleep_wait_time -- Wait Count for Dequeue Retry.
967 * sch_wave_wait_time -- Wait Count for Scheduling Wave Message.
968 * sem_rearm_wait_time -- Wait Count for Semaphore re-arm.
969 * deq_retry_wait_time -- Wait Count for Global Wave Syncs.
970 */
971void kgd_gfx_v10_get_iq_wait_times(struct amdgpu_device *adev,
972 uint32_t *wait_times,
973 uint32_t inst)
974
975{
976 *wait_times = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2));
977}
978
979void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev,
980 uint32_t wait_times,
981 uint32_t grace_period,
982 uint32_t *reg_offset,
983 uint32_t *reg_data)
984{
985 *reg_data = wait_times;
986
987 /*
988 * The CP cannont handle a 0 grace period input and will result in
989 * an infinite grace period being set so set to 1 to prevent this.
990 */
991 if (grace_period == 0)
992 grace_period = 1;
993
994 *reg_data = REG_SET_FIELD(*reg_data,
995 CP_IQ_WAIT_TIME2,
996 SCH_WAVE,
997 grace_period);
998
999 *reg_offset = SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2);
1000}
1001
1002static void program_trap_handler_settings(struct amdgpu_device *adev,
1003 uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr,
1004 uint32_t inst)
1005{
1006 lock_srbm(adev, 0, 0, 0, vmid);
1007
1008 /*
1009 * Program TBA registers
1010 */
1011 WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TBA_LO),
1012 lower_32_bits(tba_addr >> 8));
1013 WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TBA_HI),
1014 upper_32_bits(tba_addr >> 8) |
1015 (1 << SQ_SHADER_TBA_HI__TRAP_EN__SHIFT));
1016
1017 /*
1018 * Program TMA registers
1019 */
1020 WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TMA_LO),
1021 lower_32_bits(tma_addr >> 8));
1022 WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_SHADER_TMA_HI),
1023 upper_32_bits(tma_addr >> 8));
1024
1025 unlock_srbm(adev);
1026}
1027
1028const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
1029 .program_sh_mem_settings = kgd_program_sh_mem_settings,
1030 .set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
1031 .init_interrupts = kgd_init_interrupts,
1032 .hqd_load = kgd_hqd_load,
1033 .hiq_mqd_load = kgd_hiq_mqd_load,
1034 .hqd_sdma_load = kgd_hqd_sdma_load,
1035 .hqd_dump = kgd_hqd_dump,
1036 .hqd_sdma_dump = kgd_hqd_sdma_dump,
1037 .hqd_is_occupied = kgd_hqd_is_occupied,
1038 .hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
1039 .hqd_destroy = kgd_hqd_destroy,
1040 .hqd_sdma_destroy = kgd_hqd_sdma_destroy,
1041 .wave_control_execute = kgd_wave_control_execute,
1042 .get_atc_vmid_pasid_mapping_info =
1043 get_atc_vmid_pasid_mapping_info,
1044 .set_vm_context_page_table_base = set_vm_context_page_table_base,
1045 .enable_debug_trap = kgd_gfx_v10_enable_debug_trap,
1046 .disable_debug_trap = kgd_gfx_v10_disable_debug_trap,
1047 .validate_trap_override_request = kgd_gfx_v10_validate_trap_override_request,
1048 .set_wave_launch_trap_override = kgd_gfx_v10_set_wave_launch_trap_override,
1049 .set_wave_launch_mode = kgd_gfx_v10_set_wave_launch_mode,
1050 .set_address_watch = kgd_gfx_v10_set_address_watch,
1051 .clear_address_watch = kgd_gfx_v10_clear_address_watch,
1052 .get_iq_wait_times = kgd_gfx_v10_get_iq_wait_times,
1053 .build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info,
1054 .program_trap_handler_settings = program_trap_handler_settings,
1055};
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#undef pr_fmt
23#define pr_fmt(fmt) "kfd2kgd: " fmt
24
25#include <linux/module.h>
26#include <linux/fdtable.h>
27#include <linux/uaccess.h>
28#include <linux/firmware.h>
29#include <linux/mmu_context.h>
30#include "amdgpu.h"
31#include "amdgpu_amdkfd.h"
32#include "amdgpu_ucode.h"
33#include "soc15_hw_ip.h"
34#include "gc/gc_10_1_0_offset.h"
35#include "gc/gc_10_1_0_sh_mask.h"
36#include "navi10_enum.h"
37#include "athub/athub_2_0_0_offset.h"
38#include "athub/athub_2_0_0_sh_mask.h"
39#include "oss/osssys_5_0_0_offset.h"
40#include "oss/osssys_5_0_0_sh_mask.h"
41#include "soc15_common.h"
42#include "v10_structs.h"
43#include "nv.h"
44#include "nvd.h"
45
46enum hqd_dequeue_request_type {
47 NO_ACTION = 0,
48 DRAIN_PIPE,
49 RESET_WAVES,
50 SAVE_WAVES
51};
52
53/*
54 * Register access functions
55 */
56
57static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
58 uint32_t sh_mem_config,
59 uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit,
60 uint32_t sh_mem_bases);
61static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
62 unsigned int vmid);
63static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
64static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
65 uint32_t queue_id, uint32_t __user *wptr,
66 uint32_t wptr_shift, uint32_t wptr_mask,
67 struct mm_struct *mm);
68static int kgd_hqd_dump(struct kgd_dev *kgd,
69 uint32_t pipe_id, uint32_t queue_id,
70 uint32_t (**dump)[2], uint32_t *n_regs);
71static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
72 uint32_t __user *wptr, struct mm_struct *mm);
73static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
74 uint32_t engine_id, uint32_t queue_id,
75 uint32_t (**dump)[2], uint32_t *n_regs);
76static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
77 uint32_t pipe_id, uint32_t queue_id);
78static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
79static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
80 enum kfd_preempt_type reset_type,
81 unsigned int utimeout, uint32_t pipe_id,
82 uint32_t queue_id);
83static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
84 unsigned int utimeout);
85#if 0
86static uint32_t get_watch_base_addr(struct amdgpu_device *adev);
87#endif
88static int kgd_address_watch_disable(struct kgd_dev *kgd);
89static int kgd_address_watch_execute(struct kgd_dev *kgd,
90 unsigned int watch_point_id,
91 uint32_t cntl_val,
92 uint32_t addr_hi,
93 uint32_t addr_lo);
94static int kgd_wave_control_execute(struct kgd_dev *kgd,
95 uint32_t gfx_index_val,
96 uint32_t sq_cmd);
97static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
98 unsigned int watch_point_id,
99 unsigned int reg_offset);
100
101static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
102 uint8_t vmid);
103static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
104 uint8_t vmid);
105static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
106 uint64_t page_table_base);
107static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
108static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
109
110/* Because of REG_GET_FIELD() being used, we put this function in the
111 * asic specific file.
112 */
113static int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
114 struct tile_config *config)
115{
116 struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
117
118 config->gb_addr_config = adev->gfx.config.gb_addr_config;
119#if 0
120/* TODO - confirm REG_GET_FIELD x2, should be OK as is... but
121 * MC_ARB_RAMCFG register doesn't exist on Vega10 - initial amdgpu
122 * changes commented out related code, doing the same here for now but
123 * need to sync with Ken et al
124 */
125 config->num_banks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
126 MC_ARB_RAMCFG, NOOFBANK);
127 config->num_ranks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
128 MC_ARB_RAMCFG, NOOFRANKS);
129#endif
130
131 config->tile_config_ptr = adev->gfx.config.tile_mode_array;
132 config->num_tile_configs =
133 ARRAY_SIZE(adev->gfx.config.tile_mode_array);
134 config->macro_tile_config_ptr =
135 adev->gfx.config.macrotile_mode_array;
136 config->num_macro_tile_configs =
137 ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
138
139 return 0;
140}
141
142static const struct kfd2kgd_calls kfd2kgd = {
143 .program_sh_mem_settings = kgd_program_sh_mem_settings,
144 .set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
145 .init_interrupts = kgd_init_interrupts,
146 .hqd_load = kgd_hqd_load,
147 .hqd_sdma_load = kgd_hqd_sdma_load,
148 .hqd_dump = kgd_hqd_dump,
149 .hqd_sdma_dump = kgd_hqd_sdma_dump,
150 .hqd_is_occupied = kgd_hqd_is_occupied,
151 .hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
152 .hqd_destroy = kgd_hqd_destroy,
153 .hqd_sdma_destroy = kgd_hqd_sdma_destroy,
154 .address_watch_disable = kgd_address_watch_disable,
155 .address_watch_execute = kgd_address_watch_execute,
156 .wave_control_execute = kgd_wave_control_execute,
157 .address_watch_get_offset = kgd_address_watch_get_offset,
158 .get_atc_vmid_pasid_mapping_pasid =
159 get_atc_vmid_pasid_mapping_pasid,
160 .get_atc_vmid_pasid_mapping_valid =
161 get_atc_vmid_pasid_mapping_valid,
162 .invalidate_tlbs = invalidate_tlbs,
163 .invalidate_tlbs_vmid = invalidate_tlbs_vmid,
164 .set_vm_context_page_table_base = set_vm_context_page_table_base,
165 .get_tile_config = amdgpu_amdkfd_get_tile_config,
166};
167
168struct kfd2kgd_calls *amdgpu_amdkfd_gfx_10_0_get_functions()
169{
170 return (struct kfd2kgd_calls *)&kfd2kgd;
171}
172
173static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
174{
175 return (struct amdgpu_device *)kgd;
176}
177
178static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
179 uint32_t queue, uint32_t vmid)
180{
181 struct amdgpu_device *adev = get_amdgpu_device(kgd);
182
183 mutex_lock(&adev->srbm_mutex);
184 nv_grbm_select(adev, mec, pipe, queue, vmid);
185}
186
187static void unlock_srbm(struct kgd_dev *kgd)
188{
189 struct amdgpu_device *adev = get_amdgpu_device(kgd);
190
191 nv_grbm_select(adev, 0, 0, 0, 0);
192 mutex_unlock(&adev->srbm_mutex);
193}
194
195static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
196 uint32_t queue_id)
197{
198 struct amdgpu_device *adev = get_amdgpu_device(kgd);
199
200 uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
201 uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
202
203 lock_srbm(kgd, mec, pipe, queue_id, 0);
204}
205
206static uint32_t get_queue_mask(struct amdgpu_device *adev,
207 uint32_t pipe_id, uint32_t queue_id)
208{
209 unsigned int bit = (pipe_id * adev->gfx.mec.num_queue_per_pipe +
210 queue_id) & 31;
211
212 return ((uint32_t)1) << bit;
213}
214
215static void release_queue(struct kgd_dev *kgd)
216{
217 unlock_srbm(kgd);
218}
219
220static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
221 uint32_t sh_mem_config,
222 uint32_t sh_mem_ape1_base,
223 uint32_t sh_mem_ape1_limit,
224 uint32_t sh_mem_bases)
225{
226 struct amdgpu_device *adev = get_amdgpu_device(kgd);
227
228 lock_srbm(kgd, 0, 0, 0, vmid);
229
230 WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
231 WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
232 /* APE1 no longer exists on GFX9 */
233
234 unlock_srbm(kgd);
235}
236
237static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
238 unsigned int vmid)
239{
240 struct amdgpu_device *adev = get_amdgpu_device(kgd);
241
242 /*
243 * We have to assume that there is no outstanding mapping.
244 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
245 * a mapping is in progress or because a mapping finished
246 * and the SW cleared it.
247 * So the protocol is to always wait & clear.
248 */
249 uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
250 ATC_VMID0_PASID_MAPPING__VALID_MASK;
251
252 pr_debug("pasid 0x%x vmid %d, reg value %x\n", pasid, vmid, pasid_mapping);
253 /*
254 * need to do this twice, once for gfx and once for mmhub
255 * for ATC add 16 to VMID for mmhub, for IH different registers.
256 * ATC_VMID0..15 registers are separate from ATC_VMID16..31.
257 */
258
259 pr_debug("ATHUB, reg %x\n", SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid);
260 WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
261 pasid_mapping);
262
263#if 0
264 /* TODO: uncomment this code when the hardware support is ready. */
265 while (!(RREG32(SOC15_REG_OFFSET(
266 ATHUB, 0,
267 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
268 (1U << vmid)))
269 cpu_relax();
270
271 pr_debug("ATHUB mapping update finished\n");
272 WREG32(SOC15_REG_OFFSET(ATHUB, 0,
273 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
274 1U << vmid);
275#endif
276
277 /* Mapping vmid to pasid also for IH block */
278 pr_debug("update mapping for IH block and mmhub");
279 WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
280 pasid_mapping);
281
282 return 0;
283}
284
285/* TODO - RING0 form of field is obsolete, seems to date back to SI
286 * but still works
287 */
288
289static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
290{
291 struct amdgpu_device *adev = get_amdgpu_device(kgd);
292 uint32_t mec;
293 uint32_t pipe;
294
295 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
296 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
297
298 lock_srbm(kgd, mec, pipe, 0, 0);
299
300 WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
301 CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
302 CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
303
304 unlock_srbm(kgd);
305
306 return 0;
307}
308
309static uint32_t get_sdma_base_addr(struct amdgpu_device *adev,
310 unsigned int engine_id,
311 unsigned int queue_id)
312{
313 uint32_t base[2] = {
314 SOC15_REG_OFFSET(SDMA0, 0,
315 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
316 /* On gfx10, mmSDMA1_xxx registers are defined NOT based
317 * on SDMA1 base address (dw 0x1860) but based on SDMA0
318 * base address (dw 0x1260). Therefore use mmSDMA0_RLC0_RB_CNTL
319 * instead of mmSDMA1_RLC0_RB_CNTL for the base address calc
320 * below
321 */
322 SOC15_REG_OFFSET(SDMA1, 0,
323 mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL
324 };
325 uint32_t retval;
326
327 retval = base[engine_id] + queue_id * (mmSDMA0_RLC1_RB_CNTL -
328 mmSDMA0_RLC0_RB_CNTL);
329
330 pr_debug("sdma base address: 0x%x\n", retval);
331
332 return retval;
333}
334
335#if 0
336static uint32_t get_watch_base_addr(struct amdgpu_device *adev)
337{
338 uint32_t retval = SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) -
339 mmTCP_WATCH0_ADDR_H;
340
341 pr_debug("kfd: reg watch base address: 0x%x\n", retval);
342
343 return retval;
344}
345#endif
346
347static inline struct v10_compute_mqd *get_mqd(void *mqd)
348{
349 return (struct v10_compute_mqd *)mqd;
350}
351
352static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
353{
354 return (struct v10_sdma_mqd *)mqd;
355}
356
357static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
358 uint32_t queue_id, uint32_t __user *wptr,
359 uint32_t wptr_shift, uint32_t wptr_mask,
360 struct mm_struct *mm)
361{
362 struct amdgpu_device *adev = get_amdgpu_device(kgd);
363 struct v10_compute_mqd *m;
364 uint32_t *mqd_hqd;
365 uint32_t reg, hqd_base, data;
366
367 m = get_mqd(mqd);
368
369 pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id);
370 acquire_queue(kgd, pipe_id, queue_id);
371
372 /* HIQ is set during driver init period with vmid set to 0*/
373 if (m->cp_hqd_vmid == 0) {
374 uint32_t value, mec, pipe;
375
376 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
377 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
378
379 pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
380 mec, pipe, queue_id);
381 value = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS));
382 value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
383 ((mec << 5) | (pipe << 3) | queue_id | 0x80));
384 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS), value);
385 }
386
387 /* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
388 mqd_hqd = &m->cp_mqd_base_addr_lo;
389 hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
390
391 for (reg = hqd_base;
392 reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
393 WREG32(reg, mqd_hqd[reg - hqd_base]);
394
395
396 /* Activate doorbell logic before triggering WPTR poll. */
397 data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
398 CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
399 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data);
400
401 if (wptr) {
402 /* Don't read wptr with get_user because the user
403 * context may not be accessible (if this function
404 * runs in a work queue). Instead trigger a one-shot
405 * polling read from memory in the CP. This assumes
406 * that wptr is GPU-accessible in the queue's VMID via
407 * ATC or SVM. WPTR==RPTR before starting the poll so
408 * the CP starts fetching new commands from the right
409 * place.
410 *
411 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
412 * tricky. Assume that the queue didn't overflow. The
413 * number of valid bits in the 32-bit RPTR depends on
414 * the queue size. The remaining bits are taken from
415 * the saved 64-bit WPTR. If the WPTR wrapped, add the
416 * queue size.
417 */
418 uint32_t queue_size =
419 2 << REG_GET_FIELD(m->cp_hqd_pq_control,
420 CP_HQD_PQ_CONTROL, QUEUE_SIZE);
421 uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
422
423 if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
424 guessed_wptr += queue_size;
425 guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
426 guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
427
428 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
429 lower_32_bits(guessed_wptr));
430 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
431 upper_32_bits(guessed_wptr));
432 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
433 lower_32_bits((uint64_t)wptr));
434 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
435 upper_32_bits((uint64_t)wptr));
436 pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__, get_queue_mask(adev, pipe_id, queue_id));
437 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
438 get_queue_mask(adev, pipe_id, queue_id));
439 }
440
441 /* Start the EOP fetcher */
442 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
443 REG_SET_FIELD(m->cp_hqd_eop_rptr,
444 CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
445
446 data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
447 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
448
449 release_queue(kgd);
450
451 return 0;
452}
453
454static int kgd_hqd_dump(struct kgd_dev *kgd,
455 uint32_t pipe_id, uint32_t queue_id,
456 uint32_t (**dump)[2], uint32_t *n_regs)
457{
458 struct amdgpu_device *adev = get_amdgpu_device(kgd);
459 uint32_t i = 0, reg;
460#define HQD_N_REGS 56
461#define DUMP_REG(addr) do { \
462 if (WARN_ON_ONCE(i >= HQD_N_REGS)) \
463 break; \
464 (*dump)[i][0] = (addr) << 2; \
465 (*dump)[i++][1] = RREG32(addr); \
466 } while (0)
467
468 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
469 if (*dump == NULL)
470 return -ENOMEM;
471
472 acquire_queue(kgd, pipe_id, queue_id);
473
474 for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
475 reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
476 DUMP_REG(reg);
477
478 release_queue(kgd);
479
480 WARN_ON_ONCE(i != HQD_N_REGS);
481 *n_regs = i;
482
483 return 0;
484}
485
486static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
487 uint32_t __user *wptr, struct mm_struct *mm)
488{
489 struct amdgpu_device *adev = get_amdgpu_device(kgd);
490 struct v10_sdma_mqd *m;
491 uint32_t sdma_base_addr, sdmax_gfx_context_cntl;
492 unsigned long end_jiffies;
493 uint32_t data;
494 uint64_t data64;
495 uint64_t __user *wptr64 = (uint64_t __user *)wptr;
496
497 m = get_sdma_mqd(mqd);
498 sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
499 m->sdma_queue_id);
500 pr_debug("sdma load base addr %x for engine %d, queue %d\n", sdma_base_addr, m->sdma_engine_id, m->sdma_queue_id);
501 sdmax_gfx_context_cntl = m->sdma_engine_id ?
502 SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GFX_CONTEXT_CNTL) :
503 SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GFX_CONTEXT_CNTL);
504
505 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
506 m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
507
508 end_jiffies = msecs_to_jiffies(2000) + jiffies;
509 while (true) {
510 data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
511 if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
512 break;
513 if (time_after(jiffies, end_jiffies))
514 return -ETIME;
515 usleep_range(500, 1000);
516 }
517 data = RREG32(sdmax_gfx_context_cntl);
518 data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
519 RESUME_CTX, 0);
520 WREG32(sdmax_gfx_context_cntl, data);
521
522 WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL_OFFSET,
523 m->sdmax_rlcx_doorbell_offset);
524
525 data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
526 ENABLE, 1);
527 WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data);
528 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdmax_rlcx_rb_rptr);
529 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI,
530 m->sdmax_rlcx_rb_rptr_hi);
531
532 WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
533 if (read_user_wptr(mm, wptr64, data64)) {
534 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
535 lower_32_bits(data64));
536 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
537 upper_32_bits(data64));
538 } else {
539 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
540 m->sdmax_rlcx_rb_rptr);
541 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
542 m->sdmax_rlcx_rb_rptr_hi);
543 }
544 WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
545
546 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
547 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI,
548 m->sdmax_rlcx_rb_base_hi);
549 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
550 m->sdmax_rlcx_rb_rptr_addr_lo);
551 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
552 m->sdmax_rlcx_rb_rptr_addr_hi);
553
554 data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
555 RB_ENABLE, 1);
556 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
557
558 return 0;
559}
560
561static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
562 uint32_t engine_id, uint32_t queue_id,
563 uint32_t (**dump)[2], uint32_t *n_regs)
564{
565 struct amdgpu_device *adev = get_amdgpu_device(kgd);
566 uint32_t sdma_base_addr = get_sdma_base_addr(adev, engine_id, queue_id);
567 uint32_t i = 0, reg;
568#undef HQD_N_REGS
569#define HQD_N_REGS (19+6+7+10)
570
571 pr_debug("sdma dump engine id %d queue_id %d\n", engine_id, queue_id);
572 pr_debug("sdma base addr %x\n", sdma_base_addr);
573
574 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
575 if (*dump == NULL)
576 return -ENOMEM;
577
578 for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
579 DUMP_REG(sdma_base_addr + reg);
580 for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
581 DUMP_REG(sdma_base_addr + reg);
582 for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
583 reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
584 DUMP_REG(sdma_base_addr + reg);
585 for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
586 reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
587 DUMP_REG(sdma_base_addr + reg);
588
589 WARN_ON_ONCE(i != HQD_N_REGS);
590 *n_regs = i;
591
592 return 0;
593}
594
595static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
596 uint32_t pipe_id, uint32_t queue_id)
597{
598 struct amdgpu_device *adev = get_amdgpu_device(kgd);
599 uint32_t act;
600 bool retval = false;
601 uint32_t low, high;
602
603 acquire_queue(kgd, pipe_id, queue_id);
604 act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
605 if (act) {
606 low = lower_32_bits(queue_address >> 8);
607 high = upper_32_bits(queue_address >> 8);
608
609 if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
610 high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
611 retval = true;
612 }
613 release_queue(kgd);
614 return retval;
615}
616
617static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
618{
619 struct amdgpu_device *adev = get_amdgpu_device(kgd);
620 struct v10_sdma_mqd *m;
621 uint32_t sdma_base_addr;
622 uint32_t sdma_rlc_rb_cntl;
623
624 m = get_sdma_mqd(mqd);
625 sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
626 m->sdma_queue_id);
627
628 sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
629
630 if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
631 return true;
632
633 return false;
634}
635
636static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
637 enum kfd_preempt_type reset_type,
638 unsigned int utimeout, uint32_t pipe_id,
639 uint32_t queue_id)
640{
641 struct amdgpu_device *adev = get_amdgpu_device(kgd);
642 enum hqd_dequeue_request_type type;
643 unsigned long end_jiffies;
644 uint32_t temp;
645 struct v10_compute_mqd *m = get_mqd(mqd);
646
647#if 0
648 unsigned long flags;
649 int retry;
650#endif
651
652 acquire_queue(kgd, pipe_id, queue_id);
653
654 if (m->cp_hqd_vmid == 0)
655 WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
656
657 switch (reset_type) {
658 case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
659 type = DRAIN_PIPE;
660 break;
661 case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
662 type = RESET_WAVES;
663 break;
664 default:
665 type = DRAIN_PIPE;
666 break;
667 }
668
669#if 0 /* Is this still needed? */
670 /* Workaround: If IQ timer is active and the wait time is close to or
671 * equal to 0, dequeueing is not safe. Wait until either the wait time
672 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
673 * cleared before continuing. Also, ensure wait times are set to at
674 * least 0x3.
675 */
676 local_irq_save(flags);
677 preempt_disable();
678 retry = 5000; /* wait for 500 usecs at maximum */
679 while (true) {
680 temp = RREG32(mmCP_HQD_IQ_TIMER);
681 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
682 pr_debug("HW is processing IQ\n");
683 goto loop;
684 }
685 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
686 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
687 == 3) /* SEM-rearm is safe */
688 break;
689 /* Wait time 3 is safe for CP, but our MMIO read/write
690 * time is close to 1 microsecond, so check for 10 to
691 * leave more buffer room
692 */
693 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
694 >= 10)
695 break;
696 pr_debug("IQ timer is active\n");
697 } else
698 break;
699loop:
700 if (!retry) {
701 pr_err("CP HQD IQ timer status time out\n");
702 break;
703 }
704 ndelay(100);
705 --retry;
706 }
707 retry = 1000;
708 while (true) {
709 temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
710 if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
711 break;
712 pr_debug("Dequeue request is pending\n");
713
714 if (!retry) {
715 pr_err("CP HQD dequeue request time out\n");
716 break;
717 }
718 ndelay(100);
719 --retry;
720 }
721 local_irq_restore(flags);
722 preempt_enable();
723#endif
724
725 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
726
727 end_jiffies = (utimeout * HZ / 1000) + jiffies;
728 while (true) {
729 temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
730 if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
731 break;
732 if (time_after(jiffies, end_jiffies)) {
733 pr_err("cp queue preemption time out.\n");
734 release_queue(kgd);
735 return -ETIME;
736 }
737 usleep_range(500, 1000);
738 }
739
740 release_queue(kgd);
741 return 0;
742}
743
744static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
745 unsigned int utimeout)
746{
747 struct amdgpu_device *adev = get_amdgpu_device(kgd);
748 struct v10_sdma_mqd *m;
749 uint32_t sdma_base_addr;
750 uint32_t temp;
751 unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
752
753 m = get_sdma_mqd(mqd);
754 sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
755 m->sdma_queue_id);
756
757 temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
758 temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
759 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, temp);
760
761 while (true) {
762 temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
763 if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
764 break;
765 if (time_after(jiffies, end_jiffies))
766 return -ETIME;
767 usleep_range(500, 1000);
768 }
769
770 WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, 0);
771 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
772 RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL) |
773 SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
774
775 m->sdmax_rlcx_rb_rptr = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR);
776 m->sdmax_rlcx_rb_rptr_hi =
777 RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI);
778
779 return 0;
780}
781
782static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
783 uint8_t vmid)
784{
785 uint32_t reg;
786 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
787
788 reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
789 + vmid);
790 return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
791}
792
793static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
794 uint8_t vmid)
795{
796 uint32_t reg;
797 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
798
799 reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
800 + vmid);
801 return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
802}
803
804static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid)
805{
806 signed long r;
807 uint32_t seq;
808 struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
809
810 spin_lock(&adev->gfx.kiq.ring_lock);
811 amdgpu_ring_alloc(ring, 12); /* fence + invalidate_tlbs package*/
812 amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0));
813 amdgpu_ring_write(ring,
814 PACKET3_INVALIDATE_TLBS_DST_SEL(1) |
815 PACKET3_INVALIDATE_TLBS_PASID(pasid));
816 amdgpu_fence_emit_polling(ring, &seq);
817 amdgpu_ring_commit(ring);
818 spin_unlock(&adev->gfx.kiq.ring_lock);
819
820 r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
821 if (r < 1) {
822 DRM_ERROR("wait for kiq fence error: %ld.\n", r);
823 return -ETIME;
824 }
825
826 return 0;
827}
828
829static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
830{
831 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
832 int vmid;
833 struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
834
835 if (amdgpu_emu_mode == 0 && ring->sched.ready)
836 return invalidate_tlbs_with_kiq(adev, pasid);
837
838 for (vmid = 0; vmid < 16; vmid++) {
839 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
840 continue;
841 if (get_atc_vmid_pasid_mapping_valid(kgd, vmid)) {
842 if (get_atc_vmid_pasid_mapping_pasid(kgd, vmid)
843 == pasid) {
844 amdgpu_gmc_flush_gpu_tlb(adev, vmid,
845 AMDGPU_GFXHUB_0, 0);
846 break;
847 }
848 }
849 }
850
851 return 0;
852}
853
854static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
855{
856 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
857
858 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
859 pr_err("non kfd vmid %d\n", vmid);
860 return 0;
861 }
862
863 amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB_0, 0);
864 return 0;
865}
866
867static int kgd_address_watch_disable(struct kgd_dev *kgd)
868{
869 return 0;
870}
871
872static int kgd_address_watch_execute(struct kgd_dev *kgd,
873 unsigned int watch_point_id,
874 uint32_t cntl_val,
875 uint32_t addr_hi,
876 uint32_t addr_lo)
877{
878 return 0;
879}
880
881static int kgd_wave_control_execute(struct kgd_dev *kgd,
882 uint32_t gfx_index_val,
883 uint32_t sq_cmd)
884{
885 struct amdgpu_device *adev = get_amdgpu_device(kgd);
886 uint32_t data = 0;
887
888 mutex_lock(&adev->grbm_idx_mutex);
889
890 WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), gfx_index_val);
891 WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
892
893 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
894 INSTANCE_BROADCAST_WRITES, 1);
895 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
896 SA_BROADCAST_WRITES, 1);
897 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
898 SE_BROADCAST_WRITES, 1);
899
900 WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), data);
901 mutex_unlock(&adev->grbm_idx_mutex);
902
903 return 0;
904}
905
906static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
907 unsigned int watch_point_id,
908 unsigned int reg_offset)
909{
910 return 0;
911}
912
913static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
914 uint64_t page_table_base)
915{
916 struct amdgpu_device *adev = get_amdgpu_device(kgd);
917 uint64_t base = page_table_base | AMDGPU_PTE_VALID;
918
919 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
920 pr_err("trying to set page table base for wrong VMID %u\n",
921 vmid);
922 return;
923 }
924
925 /* TODO: take advantage of per-process address space size. For
926 * now, all processes share the same address space size, like
927 * on GFX8 and older.
928 */
929 WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32) + (vmid*2), 0);
930 WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32) + (vmid*2), 0);
931
932 WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32) + (vmid*2),
933 lower_32_bits(adev->vm_manager.max_pfn - 1));
934 WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32) + (vmid*2),
935 upper_32_bits(adev->vm_manager.max_pfn - 1));
936
937 WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32) + (vmid*2), lower_32_bits(base));
938 WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_HI32) + (vmid*2), upper_32_bits(base));
939}