2 * Copyright 2014-2018 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 #define pr_fmt(fmt) "kfd2kgd: " fmt
25 #include <linux/module.h>
26 #include <linux/fdtable.h>
27 #include <linux/uaccess.h>
28 #include <linux/firmware.h>
31 #include "amdgpu_amdkfd.h"
32 #include "amdgpu_ucode.h"
33 #include "soc15_hw_ip.h"
34 #include "gc/gc_9_0_offset.h"
35 #include "gc/gc_9_0_sh_mask.h"
36 #include "vega10_enum.h"
37 #include "sdma0/sdma0_4_0_offset.h"
38 #include "sdma0/sdma0_4_0_sh_mask.h"
39 #include "sdma1/sdma1_4_0_offset.h"
40 #include "sdma1/sdma1_4_0_sh_mask.h"
41 #include "athub/athub_1_0_offset.h"
42 #include "athub/athub_1_0_sh_mask.h"
43 #include "oss/osssys_4_0_offset.h"
44 #include "oss/osssys_4_0_sh_mask.h"
45 #include "soc15_common.h"
46 #include "v9_structs.h"
50 /* HACK: MMHUB and GC both have VM-related register with the same
51 * names but different offsets. Define the MMHUB register we need here
52 * with a prefix. A proper solution would be to move the functions
53 * programming these registers into gfx_v9_0.c and mmhub_v1_0.c
56 #define mmMMHUB_VM_INVALIDATE_ENG16_REQ 0x06f3
57 #define mmMMHUB_VM_INVALIDATE_ENG16_REQ_BASE_IDX 0
59 #define mmMMHUB_VM_INVALIDATE_ENG16_ACK 0x0705
60 #define mmMMHUB_VM_INVALIDATE_ENG16_ACK_BASE_IDX 0
62 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32 0x072b
63 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32_BASE_IDX 0
64 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR_HI32 0x072c
65 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR_HI32_BASE_IDX 0
67 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32 0x074b
68 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32_BASE_IDX 0
69 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32 0x074c
70 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32_BASE_IDX 0
72 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32 0x076b
73 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32_BASE_IDX 0
74 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32 0x076c
75 #define mmMMHUB_VM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32_BASE_IDX 0
77 #define mmMMHUB_VM_INVALIDATE_ENG16_ADDR_RANGE_LO32 0x0727
78 #define mmMMHUB_VM_INVALIDATE_ENG16_ADDR_RANGE_LO32_BASE_IDX 0
79 #define mmMMHUB_VM_INVALIDATE_ENG16_ADDR_RANGE_HI32 0x0728
80 #define mmMMHUB_VM_INVALIDATE_ENG16_ADDR_RANGE_HI32_BASE_IDX 0
82 #define V9_PIPE_PER_MEC (4)
83 #define V9_QUEUES_PER_PIPE_MEC (8)
85 enum hqd_dequeue_request_type {
92 * Register access functions
95 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
96 uint32_t sh_mem_config,
97 uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit,
98 uint32_t sh_mem_bases);
99 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
101 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
102 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
103 uint32_t queue_id, uint32_t __user *wptr,
104 uint32_t wptr_shift, uint32_t wptr_mask,
105 struct mm_struct *mm);
106 static int kgd_hqd_dump(struct kgd_dev *kgd,
107 uint32_t pipe_id, uint32_t queue_id,
108 uint32_t (**dump)[2], uint32_t *n_regs);
109 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
110 uint32_t __user *wptr, struct mm_struct *mm);
111 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
112 uint32_t engine_id, uint32_t queue_id,
113 uint32_t (**dump)[2], uint32_t *n_regs);
114 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
115 uint32_t pipe_id, uint32_t queue_id);
116 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
117 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
118 enum kfd_preempt_type reset_type,
119 unsigned int utimeout, uint32_t pipe_id,
121 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
122 unsigned int utimeout);
123 static int kgd_address_watch_disable(struct kgd_dev *kgd);
124 static int kgd_address_watch_execute(struct kgd_dev *kgd,
125 unsigned int watch_point_id,
129 static int kgd_wave_control_execute(struct kgd_dev *kgd,
130 uint32_t gfx_index_val,
132 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
133 unsigned int watch_point_id,
134 unsigned int reg_offset);
136 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
138 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
140 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
141 uint32_t page_table_base);
142 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type);
143 static void set_scratch_backing_va(struct kgd_dev *kgd,
144 uint64_t va, uint32_t vmid);
145 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
146 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
148 /* Because of REG_GET_FIELD() being used, we put this function in the
149 * asic specific file.
151 static int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
152 struct tile_config *config)
154 struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
156 config->gb_addr_config = adev->gfx.config.gb_addr_config;
158 config->tile_config_ptr = adev->gfx.config.tile_mode_array;
159 config->num_tile_configs =
160 ARRAY_SIZE(adev->gfx.config.tile_mode_array);
161 config->macro_tile_config_ptr =
162 adev->gfx.config.macrotile_mode_array;
163 config->num_macro_tile_configs =
164 ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
169 static const struct kfd2kgd_calls kfd2kgd = {
170 .init_gtt_mem_allocation = alloc_gtt_mem,
171 .free_gtt_mem = free_gtt_mem,
172 .get_local_mem_info = get_local_mem_info,
173 .get_gpu_clock_counter = get_gpu_clock_counter,
174 .get_max_engine_clock_in_mhz = get_max_engine_clock_in_mhz,
175 .alloc_pasid = amdgpu_pasid_alloc,
176 .free_pasid = amdgpu_pasid_free,
177 .program_sh_mem_settings = kgd_program_sh_mem_settings,
178 .set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
179 .init_interrupts = kgd_init_interrupts,
180 .hqd_load = kgd_hqd_load,
181 .hqd_sdma_load = kgd_hqd_sdma_load,
182 .hqd_dump = kgd_hqd_dump,
183 .hqd_sdma_dump = kgd_hqd_sdma_dump,
184 .hqd_is_occupied = kgd_hqd_is_occupied,
185 .hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
186 .hqd_destroy = kgd_hqd_destroy,
187 .hqd_sdma_destroy = kgd_hqd_sdma_destroy,
188 .address_watch_disable = kgd_address_watch_disable,
189 .address_watch_execute = kgd_address_watch_execute,
190 .wave_control_execute = kgd_wave_control_execute,
191 .address_watch_get_offset = kgd_address_watch_get_offset,
192 .get_atc_vmid_pasid_mapping_pasid =
193 get_atc_vmid_pasid_mapping_pasid,
194 .get_atc_vmid_pasid_mapping_valid =
195 get_atc_vmid_pasid_mapping_valid,
196 .get_fw_version = get_fw_version,
197 .set_scratch_backing_va = set_scratch_backing_va,
198 .get_tile_config = amdgpu_amdkfd_get_tile_config,
199 .get_cu_info = get_cu_info,
200 .get_vram_usage = amdgpu_amdkfd_get_vram_usage,
201 .create_process_vm = amdgpu_amdkfd_gpuvm_create_process_vm,
202 .acquire_process_vm = amdgpu_amdkfd_gpuvm_acquire_process_vm,
203 .destroy_process_vm = amdgpu_amdkfd_gpuvm_destroy_process_vm,
204 .get_process_page_dir = amdgpu_amdkfd_gpuvm_get_process_page_dir,
205 .set_vm_context_page_table_base = set_vm_context_page_table_base,
206 .alloc_memory_of_gpu = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu,
207 .free_memory_of_gpu = amdgpu_amdkfd_gpuvm_free_memory_of_gpu,
208 .map_memory_to_gpu = amdgpu_amdkfd_gpuvm_map_memory_to_gpu,
209 .unmap_memory_to_gpu = amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu,
210 .sync_memory = amdgpu_amdkfd_gpuvm_sync_memory,
211 .map_gtt_bo_to_kernel = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel,
212 .restore_process_bos = amdgpu_amdkfd_gpuvm_restore_process_bos,
213 .invalidate_tlbs = invalidate_tlbs,
214 .invalidate_tlbs_vmid = invalidate_tlbs_vmid,
215 .submit_ib = amdgpu_amdkfd_submit_ib,
218 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_9_0_get_functions(void)
220 return (struct kfd2kgd_calls *)&kfd2kgd;
223 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
225 return (struct amdgpu_device *)kgd;
228 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
229 uint32_t queue, uint32_t vmid)
231 struct amdgpu_device *adev = get_amdgpu_device(kgd);
233 mutex_lock(&adev->srbm_mutex);
234 soc15_grbm_select(adev, mec, pipe, queue, vmid);
237 static void unlock_srbm(struct kgd_dev *kgd)
239 struct amdgpu_device *adev = get_amdgpu_device(kgd);
241 soc15_grbm_select(adev, 0, 0, 0, 0);
242 mutex_unlock(&adev->srbm_mutex);
245 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
248 struct amdgpu_device *adev = get_amdgpu_device(kgd);
250 uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
251 uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
253 lock_srbm(kgd, mec, pipe, queue_id, 0);
256 static uint32_t get_queue_mask(struct amdgpu_device *adev,
257 uint32_t pipe_id, uint32_t queue_id)
259 unsigned int bit = (pipe_id * adev->gfx.mec.num_queue_per_pipe +
262 return ((uint32_t)1) << bit;
265 static void release_queue(struct kgd_dev *kgd)
270 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
271 uint32_t sh_mem_config,
272 uint32_t sh_mem_ape1_base,
273 uint32_t sh_mem_ape1_limit,
274 uint32_t sh_mem_bases)
276 struct amdgpu_device *adev = get_amdgpu_device(kgd);
278 lock_srbm(kgd, 0, 0, 0, vmid);
280 WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
281 WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
282 /* APE1 no longer exists on GFX9 */
287 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
290 struct amdgpu_device *adev = get_amdgpu_device(kgd);
293 * We have to assume that there is no outstanding mapping.
294 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
295 * a mapping is in progress or because a mapping finished
296 * and the SW cleared it.
297 * So the protocol is to always wait & clear.
299 uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
300 ATC_VMID0_PASID_MAPPING__VALID_MASK;
303 * need to do this twice, once for gfx and once for mmhub
304 * for ATC add 16 to VMID for mmhub, for IH different registers.
305 * ATC_VMID0..15 registers are separate from ATC_VMID16..31.
308 WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
311 while (!(RREG32(SOC15_REG_OFFSET(
313 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
317 WREG32(SOC15_REG_OFFSET(ATHUB, 0,
318 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
321 /* Mapping vmid to pasid also for IH block */
322 WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
325 WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID16_PASID_MAPPING) + vmid,
328 while (!(RREG32(SOC15_REG_OFFSET(
330 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
331 (1U << (vmid + 16))))
334 WREG32(SOC15_REG_OFFSET(ATHUB, 0,
335 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
338 /* Mapping vmid to pasid also for IH block */
339 WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid,
344 /* TODO - RING0 form of field is obsolete, seems to date back to SI
348 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
350 struct amdgpu_device *adev = get_amdgpu_device(kgd);
354 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
355 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
357 lock_srbm(kgd, mec, pipe, 0, 0);
359 WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
360 CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
361 CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
368 static uint32_t get_sdma_base_addr(struct amdgpu_device *adev,
369 unsigned int engine_id,
370 unsigned int queue_id)
373 SOC15_REG_OFFSET(SDMA0, 0,
374 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
375 SOC15_REG_OFFSET(SDMA1, 0,
376 mmSDMA1_RLC0_RB_CNTL) - mmSDMA1_RLC0_RB_CNTL
380 retval = base[engine_id] + queue_id * (mmSDMA0_RLC1_RB_CNTL -
381 mmSDMA0_RLC0_RB_CNTL);
383 pr_debug("sdma base address: 0x%x\n", retval);
388 static inline struct v9_mqd *get_mqd(void *mqd)
390 return (struct v9_mqd *)mqd;
393 static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
395 return (struct v9_sdma_mqd *)mqd;
398 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
399 uint32_t queue_id, uint32_t __user *wptr,
400 uint32_t wptr_shift, uint32_t wptr_mask,
401 struct mm_struct *mm)
403 struct amdgpu_device *adev = get_amdgpu_device(kgd);
406 uint32_t reg, hqd_base, data;
410 acquire_queue(kgd, pipe_id, queue_id);
412 /* HIQ is set during driver init period with vmid set to 0*/
413 if (m->cp_hqd_vmid == 0) {
414 uint32_t value, mec, pipe;
416 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
417 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
419 pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
420 mec, pipe, queue_id);
421 value = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS));
422 value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
423 ((mec << 5) | (pipe << 3) | queue_id | 0x80));
424 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS), value);
427 /* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
428 mqd_hqd = &m->cp_mqd_base_addr_lo;
429 hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
432 reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
433 WREG32(reg, mqd_hqd[reg - hqd_base]);
436 /* Activate doorbell logic before triggering WPTR poll. */
437 data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
438 CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
439 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data);
442 /* Don't read wptr with get_user because the user
443 * context may not be accessible (if this function
444 * runs in a work queue). Instead trigger a one-shot
445 * polling read from memory in the CP. This assumes
446 * that wptr is GPU-accessible in the queue's VMID via
447 * ATC or SVM. WPTR==RPTR before starting the poll so
448 * the CP starts fetching new commands from the right
451 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
452 * tricky. Assume that the queue didn't overflow. The
453 * number of valid bits in the 32-bit RPTR depends on
454 * the queue size. The remaining bits are taken from
455 * the saved 64-bit WPTR. If the WPTR wrapped, add the
458 uint32_t queue_size =
459 2 << REG_GET_FIELD(m->cp_hqd_pq_control,
460 CP_HQD_PQ_CONTROL, QUEUE_SIZE);
461 uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
463 if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
464 guessed_wptr += queue_size;
465 guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
466 guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
468 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
469 lower_32_bits(guessed_wptr));
470 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
471 upper_32_bits(guessed_wptr));
472 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
473 lower_32_bits((uint64_t)wptr));
474 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
475 upper_32_bits((uint64_t)wptr));
476 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
477 get_queue_mask(adev, pipe_id, queue_id));
480 /* Start the EOP fetcher */
481 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
482 REG_SET_FIELD(m->cp_hqd_eop_rptr,
483 CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
485 data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
486 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
493 static int kgd_hqd_dump(struct kgd_dev *kgd,
494 uint32_t pipe_id, uint32_t queue_id,
495 uint32_t (**dump)[2], uint32_t *n_regs)
497 struct amdgpu_device *adev = get_amdgpu_device(kgd);
499 #define HQD_N_REGS 56
500 #define DUMP_REG(addr) do { \
501 if (WARN_ON_ONCE(i >= HQD_N_REGS)) \
503 (*dump)[i][0] = (addr) << 2; \
504 (*dump)[i++][1] = RREG32(addr); \
507 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
511 acquire_queue(kgd, pipe_id, queue_id);
513 for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
514 reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
519 WARN_ON_ONCE(i != HQD_N_REGS);
525 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
526 uint32_t __user *wptr, struct mm_struct *mm)
528 struct amdgpu_device *adev = get_amdgpu_device(kgd);
529 struct v9_sdma_mqd *m;
530 uint32_t sdma_base_addr, sdmax_gfx_context_cntl;
531 unsigned long end_jiffies;
534 uint64_t __user *wptr64 = (uint64_t __user *)wptr;
536 m = get_sdma_mqd(mqd);
537 sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
539 sdmax_gfx_context_cntl = m->sdma_engine_id ?
540 SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GFX_CONTEXT_CNTL) :
541 SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GFX_CONTEXT_CNTL);
543 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
544 m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
546 end_jiffies = msecs_to_jiffies(2000) + jiffies;
548 data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
549 if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
551 if (time_after(jiffies, end_jiffies))
553 usleep_range(500, 1000);
555 data = RREG32(sdmax_gfx_context_cntl);
556 data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
558 WREG32(sdmax_gfx_context_cntl, data);
560 WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL_OFFSET,
561 m->sdmax_rlcx_doorbell_offset);
563 data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
565 WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data);
566 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdmax_rlcx_rb_rptr);
567 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI,
568 m->sdmax_rlcx_rb_rptr_hi);
570 WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
571 if (read_user_wptr(mm, wptr64, data64)) {
572 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
573 lower_32_bits(data64));
574 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
575 upper_32_bits(data64));
577 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
578 m->sdmax_rlcx_rb_rptr);
579 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
580 m->sdmax_rlcx_rb_rptr_hi);
582 WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
584 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
585 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI,
586 m->sdmax_rlcx_rb_base_hi);
587 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
588 m->sdmax_rlcx_rb_rptr_addr_lo);
589 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
590 m->sdmax_rlcx_rb_rptr_addr_hi);
592 data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
594 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
599 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
600 uint32_t engine_id, uint32_t queue_id,
601 uint32_t (**dump)[2], uint32_t *n_regs)
603 struct amdgpu_device *adev = get_amdgpu_device(kgd);
604 uint32_t sdma_base_addr = get_sdma_base_addr(adev, engine_id, queue_id);
607 #define HQD_N_REGS (19+6+7+10)
609 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
613 for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
614 DUMP_REG(sdma_base_addr + reg);
615 for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
616 DUMP_REG(sdma_base_addr + reg);
617 for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
618 reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
619 DUMP_REG(sdma_base_addr + reg);
620 for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
621 reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
622 DUMP_REG(sdma_base_addr + reg);
624 WARN_ON_ONCE(i != HQD_N_REGS);
630 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
631 uint32_t pipe_id, uint32_t queue_id)
633 struct amdgpu_device *adev = get_amdgpu_device(kgd);
638 acquire_queue(kgd, pipe_id, queue_id);
639 act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
641 low = lower_32_bits(queue_address >> 8);
642 high = upper_32_bits(queue_address >> 8);
644 if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
645 high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
652 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
654 struct amdgpu_device *adev = get_amdgpu_device(kgd);
655 struct v9_sdma_mqd *m;
656 uint32_t sdma_base_addr;
657 uint32_t sdma_rlc_rb_cntl;
659 m = get_sdma_mqd(mqd);
660 sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
663 sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
665 if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
671 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
672 enum kfd_preempt_type reset_type,
673 unsigned int utimeout, uint32_t pipe_id,
676 struct amdgpu_device *adev = get_amdgpu_device(kgd);
677 enum hqd_dequeue_request_type type;
678 unsigned long end_jiffies;
680 struct v9_mqd *m = get_mqd(mqd);
682 acquire_queue(kgd, pipe_id, queue_id);
684 if (m->cp_hqd_vmid == 0)
685 WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
687 switch (reset_type) {
688 case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
691 case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
699 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
701 end_jiffies = (utimeout * HZ / 1000) + jiffies;
703 temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
704 if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
706 if (time_after(jiffies, end_jiffies)) {
707 pr_err("cp queue preemption time out.\n");
711 usleep_range(500, 1000);
718 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
719 unsigned int utimeout)
721 struct amdgpu_device *adev = get_amdgpu_device(kgd);
722 struct v9_sdma_mqd *m;
723 uint32_t sdma_base_addr;
725 unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
727 m = get_sdma_mqd(mqd);
728 sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
731 temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
732 temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
733 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, temp);
736 temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
737 if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
739 if (time_after(jiffies, end_jiffies))
741 usleep_range(500, 1000);
744 WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, 0);
745 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
746 RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL) |
747 SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
749 m->sdmax_rlcx_rb_rptr = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR);
750 m->sdmax_rlcx_rb_rptr_hi =
751 RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI);
756 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
760 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
762 reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
764 return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
767 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
771 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
773 reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
775 return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
778 static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid)
780 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
781 uint32_t req = (1 << vmid) |
782 (0 << VM_INVALIDATE_ENG16_REQ__FLUSH_TYPE__SHIFT) | /* legacy */
783 VM_INVALIDATE_ENG16_REQ__INVALIDATE_L2_PTES_MASK |
784 VM_INVALIDATE_ENG16_REQ__INVALIDATE_L2_PDE0_MASK |
785 VM_INVALIDATE_ENG16_REQ__INVALIDATE_L2_PDE1_MASK |
786 VM_INVALIDATE_ENG16_REQ__INVALIDATE_L2_PDE2_MASK |
787 VM_INVALIDATE_ENG16_REQ__INVALIDATE_L1_PTES_MASK;
789 mutex_lock(&adev->srbm_mutex);
791 /* Use legacy mode tlb invalidation.
793 * Currently on Raven the code below is broken for anything but
794 * legacy mode due to a MMHUB power gating problem. A workaround
795 * is for MMHUB to wait until the condition PER_VMID_INVALIDATE_REQ
796 * == PER_VMID_INVALIDATE_ACK instead of simply waiting for the ack
799 * TODO 1: agree on the right set of invalidation registers for
800 * KFD use. Use the last one for now. Invalidate both GC and
803 * TODO 2: support range-based invalidation, requires kfg2kgd
806 WREG32(SOC15_REG_OFFSET(GC, 0, mmVM_INVALIDATE_ENG16_ADDR_RANGE_LO32),
808 WREG32(SOC15_REG_OFFSET(GC, 0, mmVM_INVALIDATE_ENG16_ADDR_RANGE_HI32),
811 WREG32(SOC15_REG_OFFSET(MMHUB, 0,
812 mmMMHUB_VM_INVALIDATE_ENG16_ADDR_RANGE_LO32),
814 WREG32(SOC15_REG_OFFSET(MMHUB, 0,
815 mmMMHUB_VM_INVALIDATE_ENG16_ADDR_RANGE_HI32),
818 WREG32(SOC15_REG_OFFSET(GC, 0, mmVM_INVALIDATE_ENG16_REQ), req);
820 WREG32(SOC15_REG_OFFSET(MMHUB, 0, mmMMHUB_VM_INVALIDATE_ENG16_REQ),
823 while (!(RREG32(SOC15_REG_OFFSET(GC, 0, mmVM_INVALIDATE_ENG16_ACK)) &
827 while (!(RREG32(SOC15_REG_OFFSET(MMHUB, 0,
828 mmMMHUB_VM_INVALIDATE_ENG16_ACK)) &
832 mutex_unlock(&adev->srbm_mutex);
836 static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid)
840 struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
842 spin_lock(&adev->gfx.kiq.ring_lock);
843 amdgpu_ring_alloc(ring, 12); /* fence + invalidate_tlbs package*/
844 amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0));
845 amdgpu_ring_write(ring,
846 PACKET3_INVALIDATE_TLBS_DST_SEL(1) |
847 PACKET3_INVALIDATE_TLBS_ALL_HUB(1) |
848 PACKET3_INVALIDATE_TLBS_PASID(pasid) |
849 PACKET3_INVALIDATE_TLBS_FLUSH_TYPE(0)); /* legacy */
850 amdgpu_fence_emit_polling(ring, &seq);
851 amdgpu_ring_commit(ring);
852 spin_unlock(&adev->gfx.kiq.ring_lock);
854 r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
856 DRM_ERROR("wait for kiq fence error: %ld.\n", r);
863 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
865 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
867 struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
870 return invalidate_tlbs_with_kiq(adev, pasid);
872 for (vmid = 0; vmid < 16; vmid++) {
873 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
875 if (get_atc_vmid_pasid_mapping_valid(kgd, vmid)) {
876 if (get_atc_vmid_pasid_mapping_pasid(kgd, vmid)
878 write_vmid_invalidate_request(kgd, vmid);
887 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
889 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
891 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
892 pr_err("non kfd vmid %d\n", vmid);
896 write_vmid_invalidate_request(kgd, vmid);
900 static int kgd_address_watch_disable(struct kgd_dev *kgd)
905 static int kgd_address_watch_execute(struct kgd_dev *kgd,
906 unsigned int watch_point_id,
914 static int kgd_wave_control_execute(struct kgd_dev *kgd,
915 uint32_t gfx_index_val,
918 struct amdgpu_device *adev = get_amdgpu_device(kgd);
921 mutex_lock(&adev->grbm_idx_mutex);
923 WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), gfx_index_val);
924 WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
926 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
927 INSTANCE_BROADCAST_WRITES, 1);
928 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
929 SH_BROADCAST_WRITES, 1);
930 data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
931 SE_BROADCAST_WRITES, 1);
933 WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), data);
934 mutex_unlock(&adev->grbm_idx_mutex);
939 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
940 unsigned int watch_point_id,
941 unsigned int reg_offset)
946 static void set_scratch_backing_va(struct kgd_dev *kgd,
947 uint64_t va, uint32_t vmid)
949 /* No longer needed on GFXv9. The scratch base address is
950 * passed to the shader by the CP. It's the user mode driver's
955 /* FIXME: Does this need to be ASIC-specific code? */
956 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type)
958 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
959 const union amdgpu_firmware_header *hdr;
963 hdr = (const union amdgpu_firmware_header *)adev->gfx.pfp_fw->data;
967 hdr = (const union amdgpu_firmware_header *)adev->gfx.me_fw->data;
971 hdr = (const union amdgpu_firmware_header *)adev->gfx.ce_fw->data;
974 case KGD_ENGINE_MEC1:
975 hdr = (const union amdgpu_firmware_header *)adev->gfx.mec_fw->data;
978 case KGD_ENGINE_MEC2:
979 hdr = (const union amdgpu_firmware_header *)adev->gfx.mec2_fw->data;
983 hdr = (const union amdgpu_firmware_header *)adev->gfx.rlc_fw->data;
986 case KGD_ENGINE_SDMA1:
987 hdr = (const union amdgpu_firmware_header *)adev->sdma.instance[0].fw->data;
990 case KGD_ENGINE_SDMA2:
991 hdr = (const union amdgpu_firmware_header *)adev->sdma.instance[1].fw->data;
1001 /* Only 12 bit in use*/
1002 return hdr->common.ucode_version;
1005 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
1006 uint32_t page_table_base)
1008 struct amdgpu_device *adev = get_amdgpu_device(kgd);
1009 uint64_t base = (uint64_t)page_table_base << PAGE_SHIFT |
1012 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
1013 pr_err("trying to set page table base for wrong VMID %u\n",
1018 /* TODO: take advantage of per-process address space size. For
1019 * now, all processes share the same address space size, like
1020 * on GFX8 and older.
1022 WREG32(SOC15_REG_OFFSET(MMHUB, 0, mmMMHUB_VM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32) + (vmid*2), 0);
1023 WREG32(SOC15_REG_OFFSET(MMHUB, 0, mmMMHUB_VM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32) + (vmid*2), 0);
1025 WREG32(SOC15_REG_OFFSET(MMHUB, 0, mmMMHUB_VM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32) + (vmid*2),
1026 lower_32_bits(adev->vm_manager.max_pfn - 1));
1027 WREG32(SOC15_REG_OFFSET(MMHUB, 0, mmMMHUB_VM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32) + (vmid*2),
1028 upper_32_bits(adev->vm_manager.max_pfn - 1));
1030 WREG32(SOC15_REG_OFFSET(MMHUB, 0, mmMMHUB_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32) + (vmid*2), lower_32_bits(base));
1031 WREG32(SOC15_REG_OFFSET(MMHUB, 0, mmMMHUB_VM_CONTEXT0_PAGE_TABLE_BASE_ADDR_HI32) + (vmid*2), upper_32_bits(base));
1033 WREG32(SOC15_REG_OFFSET(GC, 0, mmVM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32) + (vmid*2), 0);
1034 WREG32(SOC15_REG_OFFSET(GC, 0, mmVM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32) + (vmid*2), 0);
1036 WREG32(SOC15_REG_OFFSET(GC, 0, mmVM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32) + (vmid*2),
1037 lower_32_bits(adev->vm_manager.max_pfn - 1));
1038 WREG32(SOC15_REG_OFFSET(GC, 0, mmVM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32) + (vmid*2),
1039 upper_32_bits(adev->vm_manager.max_pfn - 1));
1041 WREG32(SOC15_REG_OFFSET(GC, 0, mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32) + (vmid*2), lower_32_bits(base));
1042 WREG32(SOC15_REG_OFFSET(GC, 0, mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_HI32) + (vmid*2), upper_32_bits(base));