2 * Copyright 2019 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
23 #include "amdgpu_amdkfd.h"
24 #include "gc/gc_10_1_0_offset.h"
25 #include "gc/gc_10_1_0_sh_mask.h"
26 #include "athub/athub_2_0_0_offset.h"
27 #include "athub/athub_2_0_0_sh_mask.h"
28 #include "oss/osssys_5_0_0_offset.h"
29 #include "oss/osssys_5_0_0_sh_mask.h"
30 #include "soc15_common.h"
31 #include "v10_structs.h"
35 enum hqd_dequeue_request_type {
42 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
44 return (struct amdgpu_device *)kgd;
47 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
48 uint32_t queue, uint32_t vmid)
50 struct amdgpu_device *adev = get_amdgpu_device(kgd);
52 mutex_lock(&adev->srbm_mutex);
53 nv_grbm_select(adev, mec, pipe, queue, vmid);
56 static void unlock_srbm(struct kgd_dev *kgd)
58 struct amdgpu_device *adev = get_amdgpu_device(kgd);
60 nv_grbm_select(adev, 0, 0, 0, 0);
61 mutex_unlock(&adev->srbm_mutex);
64 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
67 struct amdgpu_device *adev = get_amdgpu_device(kgd);
69 uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
70 uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
72 lock_srbm(kgd, mec, pipe, queue_id, 0);
75 static uint64_t get_queue_mask(struct amdgpu_device *adev,
76 uint32_t pipe_id, uint32_t queue_id)
78 unsigned int bit = pipe_id * adev->gfx.mec.num_queue_per_pipe +
84 static void release_queue(struct kgd_dev *kgd)
89 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
90 uint32_t sh_mem_config,
91 uint32_t sh_mem_ape1_base,
92 uint32_t sh_mem_ape1_limit,
93 uint32_t sh_mem_bases)
95 struct amdgpu_device *adev = get_amdgpu_device(kgd);
97 lock_srbm(kgd, 0, 0, 0, vmid);
99 WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
100 WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
101 /* APE1 no longer exists on GFX9 */
106 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, u32 pasid,
109 struct amdgpu_device *adev = get_amdgpu_device(kgd);
112 * We have to assume that there is no outstanding mapping.
113 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
114 * a mapping is in progress or because a mapping finished
115 * and the SW cleared it.
116 * So the protocol is to always wait & clear.
118 uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
119 ATC_VMID0_PASID_MAPPING__VALID_MASK;
121 pr_debug("pasid 0x%x vmid %d, reg value %x\n", pasid, vmid, pasid_mapping);
123 pr_debug("ATHUB, reg %x\n", SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid);
124 WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
128 /* TODO: uncomment this code when the hardware support is ready. */
129 while (!(RREG32(SOC15_REG_OFFSET(
131 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
135 pr_debug("ATHUB mapping update finished\n");
136 WREG32(SOC15_REG_OFFSET(ATHUB, 0,
137 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
141 /* Mapping vmid to pasid also for IH block */
142 pr_debug("update mapping for IH block and mmhub");
143 WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
149 /* TODO - RING0 form of field is obsolete, seems to date back to SI
153 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
155 struct amdgpu_device *adev = get_amdgpu_device(kgd);
159 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
160 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
162 lock_srbm(kgd, mec, pipe, 0, 0);
164 WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
165 CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
166 CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
173 static uint32_t get_sdma_rlc_reg_offset(struct amdgpu_device *adev,
174 unsigned int engine_id,
175 unsigned int queue_id)
177 uint32_t sdma_engine_reg_base[2] = {
178 SOC15_REG_OFFSET(SDMA0, 0,
179 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
180 /* On gfx10, mmSDMA1_xxx registers are defined NOT based
181 * on SDMA1 base address (dw 0x1860) but based on SDMA0
182 * base address (dw 0x1260). Therefore use mmSDMA0_RLC0_RB_CNTL
183 * instead of mmSDMA1_RLC0_RB_CNTL for the base address calc
186 SOC15_REG_OFFSET(SDMA1, 0,
187 mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL
190 uint32_t retval = sdma_engine_reg_base[engine_id]
191 + queue_id * (mmSDMA0_RLC1_RB_CNTL - mmSDMA0_RLC0_RB_CNTL);
193 pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n", engine_id,
200 static uint32_t get_watch_base_addr(struct amdgpu_device *adev)
202 uint32_t retval = SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) -
205 pr_debug("kfd: reg watch base address: 0x%x\n", retval);
211 static inline struct v10_compute_mqd *get_mqd(void *mqd)
213 return (struct v10_compute_mqd *)mqd;
216 static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
218 return (struct v10_sdma_mqd *)mqd;
221 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
222 uint32_t queue_id, uint32_t __user *wptr,
223 uint32_t wptr_shift, uint32_t wptr_mask,
224 struct mm_struct *mm)
226 struct amdgpu_device *adev = get_amdgpu_device(kgd);
227 struct v10_compute_mqd *m;
229 uint32_t reg, hqd_base, data;
233 pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id);
234 acquire_queue(kgd, pipe_id, queue_id);
236 /* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
237 mqd_hqd = &m->cp_mqd_base_addr_lo;
238 hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
241 reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
242 WREG32(reg, mqd_hqd[reg - hqd_base]);
245 /* Activate doorbell logic before triggering WPTR poll. */
246 data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
247 CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
248 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data);
251 /* Don't read wptr with get_user because the user
252 * context may not be accessible (if this function
253 * runs in a work queue). Instead trigger a one-shot
254 * polling read from memory in the CP. This assumes
255 * that wptr is GPU-accessible in the queue's VMID via
256 * ATC or SVM. WPTR==RPTR before starting the poll so
257 * the CP starts fetching new commands from the right
260 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
261 * tricky. Assume that the queue didn't overflow. The
262 * number of valid bits in the 32-bit RPTR depends on
263 * the queue size. The remaining bits are taken from
264 * the saved 64-bit WPTR. If the WPTR wrapped, add the
267 uint32_t queue_size =
268 2 << REG_GET_FIELD(m->cp_hqd_pq_control,
269 CP_HQD_PQ_CONTROL, QUEUE_SIZE);
270 uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
272 if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
273 guessed_wptr += queue_size;
274 guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
275 guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
277 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
278 lower_32_bits(guessed_wptr));
279 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
280 upper_32_bits(guessed_wptr));
281 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
282 lower_32_bits((uint64_t)wptr));
283 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
284 upper_32_bits((uint64_t)wptr));
285 pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__,
286 (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
287 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
288 (uint32_t)get_queue_mask(adev, pipe_id, queue_id));
291 /* Start the EOP fetcher */
292 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
293 REG_SET_FIELD(m->cp_hqd_eop_rptr,
294 CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
296 data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
297 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
304 static int kgd_hiq_mqd_load(struct kgd_dev *kgd, void *mqd,
305 uint32_t pipe_id, uint32_t queue_id,
306 uint32_t doorbell_off)
308 struct amdgpu_device *adev = get_amdgpu_device(kgd);
309 struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
310 struct v10_compute_mqd *m;
316 acquire_queue(kgd, pipe_id, queue_id);
318 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
319 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
321 pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
322 mec, pipe, queue_id);
324 spin_lock(&adev->gfx.kiq.ring_lock);
325 r = amdgpu_ring_alloc(kiq_ring, 7);
327 pr_err("Failed to alloc KIQ (%d).\n", r);
331 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
332 amdgpu_ring_write(kiq_ring,
333 PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */
334 PACKET3_MAP_QUEUES_VMID(m->cp_hqd_vmid) | /* VMID */
335 PACKET3_MAP_QUEUES_QUEUE(queue_id) |
336 PACKET3_MAP_QUEUES_PIPE(pipe) |
337 PACKET3_MAP_QUEUES_ME((mec - 1)) |
338 PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */
339 PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */
340 PACKET3_MAP_QUEUES_ENGINE_SEL(1) | /* engine_sel: hiq */
341 PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */
342 amdgpu_ring_write(kiq_ring,
343 PACKET3_MAP_QUEUES_DOORBELL_OFFSET(doorbell_off));
344 amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_lo);
345 amdgpu_ring_write(kiq_ring, m->cp_mqd_base_addr_hi);
346 amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_lo);
347 amdgpu_ring_write(kiq_ring, m->cp_hqd_pq_wptr_poll_addr_hi);
348 amdgpu_ring_commit(kiq_ring);
351 spin_unlock(&adev->gfx.kiq.ring_lock);
357 static int kgd_hqd_dump(struct kgd_dev *kgd,
358 uint32_t pipe_id, uint32_t queue_id,
359 uint32_t (**dump)[2], uint32_t *n_regs)
361 struct amdgpu_device *adev = get_amdgpu_device(kgd);
363 #define HQD_N_REGS 56
364 #define DUMP_REG(addr) do { \
365 if (WARN_ON_ONCE(i >= HQD_N_REGS)) \
367 (*dump)[i][0] = (addr) << 2; \
368 (*dump)[i++][1] = RREG32(addr); \
371 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
375 acquire_queue(kgd, pipe_id, queue_id);
377 for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
378 reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
383 WARN_ON_ONCE(i != HQD_N_REGS);
389 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
390 uint32_t __user *wptr, struct mm_struct *mm)
392 struct amdgpu_device *adev = get_amdgpu_device(kgd);
393 struct v10_sdma_mqd *m;
394 uint32_t sdma_rlc_reg_offset;
395 unsigned long end_jiffies;
398 uint64_t __user *wptr64 = (uint64_t __user *)wptr;
400 m = get_sdma_mqd(mqd);
401 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
404 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
405 m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
407 end_jiffies = msecs_to_jiffies(2000) + jiffies;
409 data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
410 if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
412 if (time_after(jiffies, end_jiffies)) {
413 pr_err("SDMA RLC not idle in %s\n", __func__);
416 usleep_range(500, 1000);
419 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL_OFFSET,
420 m->sdmax_rlcx_doorbell_offset);
422 data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
424 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
425 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
426 m->sdmax_rlcx_rb_rptr);
427 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI,
428 m->sdmax_rlcx_rb_rptr_hi);
430 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
431 if (read_user_wptr(mm, wptr64, data64)) {
432 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
433 lower_32_bits(data64));
434 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
435 upper_32_bits(data64));
437 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
438 m->sdmax_rlcx_rb_rptr);
439 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR_HI,
440 m->sdmax_rlcx_rb_rptr_hi);
442 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
444 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
445 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
446 m->sdmax_rlcx_rb_base_hi);
447 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
448 m->sdmax_rlcx_rb_rptr_addr_lo);
449 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
450 m->sdmax_rlcx_rb_rptr_addr_hi);
452 data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
454 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
459 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
460 uint32_t engine_id, uint32_t queue_id,
461 uint32_t (**dump)[2], uint32_t *n_regs)
463 struct amdgpu_device *adev = get_amdgpu_device(kgd);
464 uint32_t sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev,
465 engine_id, queue_id);
468 #define HQD_N_REGS (19+6+7+10)
470 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
474 for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
475 DUMP_REG(sdma_rlc_reg_offset + reg);
476 for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
477 DUMP_REG(sdma_rlc_reg_offset + reg);
478 for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
479 reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
480 DUMP_REG(sdma_rlc_reg_offset + reg);
481 for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
482 reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
483 DUMP_REG(sdma_rlc_reg_offset + reg);
485 WARN_ON_ONCE(i != HQD_N_REGS);
491 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
492 uint32_t pipe_id, uint32_t queue_id)
494 struct amdgpu_device *adev = get_amdgpu_device(kgd);
499 acquire_queue(kgd, pipe_id, queue_id);
500 act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
502 low = lower_32_bits(queue_address >> 8);
503 high = upper_32_bits(queue_address >> 8);
505 if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
506 high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
513 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
515 struct amdgpu_device *adev = get_amdgpu_device(kgd);
516 struct v10_sdma_mqd *m;
517 uint32_t sdma_rlc_reg_offset;
518 uint32_t sdma_rlc_rb_cntl;
520 m = get_sdma_mqd(mqd);
521 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
524 sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
526 if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
532 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
533 enum kfd_preempt_type reset_type,
534 unsigned int utimeout, uint32_t pipe_id,
537 struct amdgpu_device *adev = get_amdgpu_device(kgd);
538 enum hqd_dequeue_request_type type;
539 unsigned long end_jiffies;
541 struct v10_compute_mqd *m = get_mqd(mqd);
543 if (amdgpu_in_reset(adev))
551 acquire_queue(kgd, pipe_id, queue_id);
553 if (m->cp_hqd_vmid == 0)
554 WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
556 switch (reset_type) {
557 case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
560 case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
568 #if 0 /* Is this still needed? */
569 /* Workaround: If IQ timer is active and the wait time is close to or
570 * equal to 0, dequeueing is not safe. Wait until either the wait time
571 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
572 * cleared before continuing. Also, ensure wait times are set to at
575 local_irq_save(flags);
577 retry = 5000; /* wait for 500 usecs at maximum */
579 temp = RREG32(mmCP_HQD_IQ_TIMER);
580 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
581 pr_debug("HW is processing IQ\n");
584 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
585 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
586 == 3) /* SEM-rearm is safe */
588 /* Wait time 3 is safe for CP, but our MMIO read/write
589 * time is close to 1 microsecond, so check for 10 to
590 * leave more buffer room
592 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
595 pr_debug("IQ timer is active\n");
600 pr_err("CP HQD IQ timer status time out\n");
608 temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
609 if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
611 pr_debug("Dequeue request is pending\n");
614 pr_err("CP HQD dequeue request time out\n");
620 local_irq_restore(flags);
624 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
626 end_jiffies = (utimeout * HZ / 1000) + jiffies;
628 temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
629 if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
631 if (time_after(jiffies, end_jiffies)) {
632 pr_err("cp queue preemption time out.\n");
636 usleep_range(500, 1000);
643 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
644 unsigned int utimeout)
646 struct amdgpu_device *adev = get_amdgpu_device(kgd);
647 struct v10_sdma_mqd *m;
648 uint32_t sdma_rlc_reg_offset;
650 unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
652 m = get_sdma_mqd(mqd);
653 sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(adev, m->sdma_engine_id,
656 temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
657 temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
658 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
661 temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
662 if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
664 if (time_after(jiffies, end_jiffies)) {
665 pr_err("SDMA RLC not idle in %s\n", __func__);
668 usleep_range(500, 1000);
671 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
672 WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
673 RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
674 SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
676 m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
677 m->sdmax_rlcx_rb_rptr_hi =
678 RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_HI);
683 static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
684 uint8_t vmid, uint16_t *p_pasid)
687 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
689 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
691 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
693 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
696 static int kgd_address_watch_disable(struct kgd_dev *kgd)
701 static int kgd_address_watch_execute(struct kgd_dev *kgd,
702 unsigned int watch_point_id,
710 static int kgd_wave_control_execute(struct kgd_dev *kgd,
711 uint32_t gfx_index_val,
714 struct amdgpu_device *adev = get_amdgpu_device(kgd);
717 mutex_lock(&adev->grbm_idx_mutex);
719 WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), gfx_index_val);
720 WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
722 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
723 INSTANCE_BROADCAST_WRITES, 1);
724 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
725 SA_BROADCAST_WRITES, 1);
726 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
727 SE_BROADCAST_WRITES, 1);
729 WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), data);
730 mutex_unlock(&adev->grbm_idx_mutex);
735 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
736 unsigned int watch_point_id,
737 unsigned int reg_offset)
742 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
743 uint64_t page_table_base)
745 struct amdgpu_device *adev = get_amdgpu_device(kgd);
747 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
748 pr_err("trying to set page table base for wrong VMID %u\n",
753 /* SDMA is on gfxhub as well for Navi1* series */
754 adev->gfxhub.funcs->setup_vm_pt_regs(adev, vmid, page_table_base);
757 const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
758 .program_sh_mem_settings = kgd_program_sh_mem_settings,
759 .set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
760 .init_interrupts = kgd_init_interrupts,
761 .hqd_load = kgd_hqd_load,
762 .hiq_mqd_load = kgd_hiq_mqd_load,
763 .hqd_sdma_load = kgd_hqd_sdma_load,
764 .hqd_dump = kgd_hqd_dump,
765 .hqd_sdma_dump = kgd_hqd_sdma_dump,
766 .hqd_is_occupied = kgd_hqd_is_occupied,
767 .hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
768 .hqd_destroy = kgd_hqd_destroy,
769 .hqd_sdma_destroy = kgd_hqd_sdma_destroy,
770 .address_watch_disable = kgd_address_watch_disable,
771 .address_watch_execute = kgd_address_watch_execute,
772 .wave_control_execute = kgd_wave_control_execute,
773 .address_watch_get_offset = kgd_address_watch_get_offset,
774 .get_atc_vmid_pasid_mapping_info =
775 get_atc_vmid_pasid_mapping_info,
776 .set_vm_context_page_table_base = set_vm_context_page_table_base,