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 #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>
29 #include <linux/mmu_context.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"
46 enum hqd_dequeue_request_type {
54 * Register access functions
57 static 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);
61 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
63 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
64 static 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);
68 static 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);
71 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
72 uint32_t __user *wptr, struct mm_struct *mm);
73 static 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);
76 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
77 uint32_t pipe_id, uint32_t queue_id);
78 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
79 static 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,
83 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
84 unsigned int utimeout);
86 static uint32_t get_watch_base_addr(struct amdgpu_device *adev);
88 static int kgd_address_watch_disable(struct kgd_dev *kgd);
89 static int kgd_address_watch_execute(struct kgd_dev *kgd,
90 unsigned int watch_point_id,
94 static int kgd_wave_control_execute(struct kgd_dev *kgd,
95 uint32_t gfx_index_val,
97 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
98 unsigned int watch_point_id,
99 unsigned int reg_offset);
101 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
103 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
105 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
106 uint64_t page_table_base);
107 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
108 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
110 /* Because of REG_GET_FIELD() being used, we put this function in the
111 * asic specific file.
113 static int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
114 struct tile_config *config)
116 struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
118 config->gb_addr_config = adev->gfx.config.gb_addr_config;
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
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);
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);
142 static 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,
168 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_10_0_get_functions()
170 return (struct kfd2kgd_calls *)&kfd2kgd;
173 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
175 return (struct amdgpu_device *)kgd;
178 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
179 uint32_t queue, uint32_t vmid)
181 struct amdgpu_device *adev = get_amdgpu_device(kgd);
183 mutex_lock(&adev->srbm_mutex);
184 nv_grbm_select(adev, mec, pipe, queue, vmid);
187 static void unlock_srbm(struct kgd_dev *kgd)
189 struct amdgpu_device *adev = get_amdgpu_device(kgd);
191 nv_grbm_select(adev, 0, 0, 0, 0);
192 mutex_unlock(&adev->srbm_mutex);
195 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
198 struct amdgpu_device *adev = get_amdgpu_device(kgd);
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);
203 lock_srbm(kgd, mec, pipe, queue_id, 0);
206 static uint32_t get_queue_mask(struct amdgpu_device *adev,
207 uint32_t pipe_id, uint32_t queue_id)
209 unsigned int bit = (pipe_id * adev->gfx.mec.num_queue_per_pipe +
212 return ((uint32_t)1) << bit;
215 static void release_queue(struct kgd_dev *kgd)
220 static 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)
226 struct amdgpu_device *adev = get_amdgpu_device(kgd);
228 lock_srbm(kgd, 0, 0, 0, vmid);
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 */
237 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
240 struct amdgpu_device *adev = get_amdgpu_device(kgd);
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.
249 uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
250 ATC_VMID0_PASID_MAPPING__VALID_MASK;
252 pr_debug("pasid 0x%x vmid %d, reg value %x\n", pasid, vmid, pasid_mapping);
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.
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,
264 /* TODO: uncomment this code when the hardware support is ready. */
265 while (!(RREG32(SOC15_REG_OFFSET(
267 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
271 pr_debug("ATHUB mapping update finished\n");
272 WREG32(SOC15_REG_OFFSET(ATHUB, 0,
273 mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
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,
285 /* TODO - RING0 form of field is obsolete, seems to date back to SI
289 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
291 struct amdgpu_device *adev = get_amdgpu_device(kgd);
295 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
296 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
298 lock_srbm(kgd, mec, pipe, 0, 0);
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);
309 static uint32_t get_sdma_base_addr(struct amdgpu_device *adev,
310 unsigned int engine_id,
311 unsigned int queue_id)
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
322 SOC15_REG_OFFSET(SDMA1, 0,
323 mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL
327 retval = base[engine_id] + queue_id * (mmSDMA0_RLC1_RB_CNTL -
328 mmSDMA0_RLC0_RB_CNTL);
330 pr_debug("sdma base address: 0x%x\n", retval);
336 static uint32_t get_watch_base_addr(struct amdgpu_device *adev)
338 uint32_t retval = SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) -
341 pr_debug("kfd: reg watch base address: 0x%x\n", retval);
347 static inline struct v10_compute_mqd *get_mqd(void *mqd)
349 return (struct v10_compute_mqd *)mqd;
352 static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
354 return (struct v10_sdma_mqd *)mqd;
357 static 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)
362 struct amdgpu_device *adev = get_amdgpu_device(kgd);
363 struct v10_compute_mqd *m;
365 uint32_t reg, hqd_base, data;
369 pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id);
370 acquire_queue(kgd, pipe_id, queue_id);
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;
376 mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
377 pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
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);
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);
392 reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
393 WREG32(reg, mqd_hqd[reg - hqd_base]);
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);
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
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
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);
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;
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));
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));
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);
454 static 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)
458 struct amdgpu_device *adev = get_amdgpu_device(kgd);
460 #define HQD_N_REGS 56
461 #define DUMP_REG(addr) do { \
462 if (WARN_ON_ONCE(i >= HQD_N_REGS)) \
464 (*dump)[i][0] = (addr) << 2; \
465 (*dump)[i++][1] = RREG32(addr); \
468 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
472 acquire_queue(kgd, pipe_id, queue_id);
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++)
480 WARN_ON_ONCE(i != HQD_N_REGS);
486 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
487 uint32_t __user *wptr, struct mm_struct *mm)
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;
495 uint64_t __user *wptr64 = (uint64_t __user *)wptr;
497 m = get_sdma_mqd(mqd);
498 sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_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);
505 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
506 m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
508 end_jiffies = msecs_to_jiffies(2000) + jiffies;
510 data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
511 if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
513 if (time_after(jiffies, end_jiffies))
515 usleep_range(500, 1000);
517 data = RREG32(sdmax_gfx_context_cntl);
518 data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
520 WREG32(sdmax_gfx_context_cntl, data);
522 WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL_OFFSET,
523 m->sdmax_rlcx_doorbell_offset);
525 data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
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);
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));
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);
544 WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
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);
554 data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
556 WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
561 static 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)
565 struct amdgpu_device *adev = get_amdgpu_device(kgd);
566 uint32_t sdma_base_addr = get_sdma_base_addr(adev, engine_id, queue_id);
569 #define HQD_N_REGS (19+6+7+10)
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);
574 *dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
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);
589 WARN_ON_ONCE(i != HQD_N_REGS);
595 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
596 uint32_t pipe_id, uint32_t queue_id)
598 struct amdgpu_device *adev = get_amdgpu_device(kgd);
603 acquire_queue(kgd, pipe_id, queue_id);
604 act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
606 low = lower_32_bits(queue_address >> 8);
607 high = upper_32_bits(queue_address >> 8);
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)))
617 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
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;
624 m = get_sdma_mqd(mqd);
625 sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
628 sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
630 if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
636 static 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,
641 struct amdgpu_device *adev = get_amdgpu_device(kgd);
642 enum hqd_dequeue_request_type type;
643 unsigned long end_jiffies;
645 struct v10_compute_mqd *m = get_mqd(mqd);
652 acquire_queue(kgd, pipe_id, queue_id);
654 if (m->cp_hqd_vmid == 0)
655 WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
657 switch (reset_type) {
658 case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
661 case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
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
676 local_irq_save(flags);
678 retry = 5000; /* wait for 500 usecs at maximum */
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");
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 */
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
693 if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
696 pr_debug("IQ timer is active\n");
701 pr_err("CP HQD IQ timer status time out\n");
709 temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
710 if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
712 pr_debug("Dequeue request is pending\n");
715 pr_err("CP HQD dequeue request time out\n");
721 local_irq_restore(flags);
725 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
727 end_jiffies = (utimeout * HZ / 1000) + jiffies;
729 temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
730 if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
732 if (time_after(jiffies, end_jiffies)) {
733 pr_err("cp queue preemption time out.\n");
737 usleep_range(500, 1000);
744 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
745 unsigned int utimeout)
747 struct amdgpu_device *adev = get_amdgpu_device(kgd);
748 struct v10_sdma_mqd *m;
749 uint32_t sdma_base_addr;
751 unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
753 m = get_sdma_mqd(mqd);
754 sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
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);
762 temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
763 if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
765 if (time_after(jiffies, end_jiffies))
767 usleep_range(500, 1000);
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);
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);
782 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
786 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
788 reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
790 return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
793 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
797 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
799 reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
801 return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
804 static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid)
808 struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
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);
820 r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
822 DRM_ERROR("wait for kiq fence error: %ld.\n", r);
829 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
831 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
833 struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
835 if (amdgpu_emu_mode == 0 && ring->sched.ready)
836 return invalidate_tlbs_with_kiq(adev, pasid);
838 for (vmid = 0; vmid < 16; vmid++) {
839 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
841 if (get_atc_vmid_pasid_mapping_valid(kgd, vmid)) {
842 if (get_atc_vmid_pasid_mapping_pasid(kgd, vmid)
844 amdgpu_gmc_flush_gpu_tlb(adev, vmid,
854 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
856 struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
858 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
859 pr_err("non kfd vmid %d\n", vmid);
863 amdgpu_gmc_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB_0, 0);
867 static int kgd_address_watch_disable(struct kgd_dev *kgd)
872 static int kgd_address_watch_execute(struct kgd_dev *kgd,
873 unsigned int watch_point_id,
881 static int kgd_wave_control_execute(struct kgd_dev *kgd,
882 uint32_t gfx_index_val,
885 struct amdgpu_device *adev = get_amdgpu_device(kgd);
888 mutex_lock(&adev->grbm_idx_mutex);
890 WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), gfx_index_val);
891 WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
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);
900 WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), data);
901 mutex_unlock(&adev->grbm_idx_mutex);
906 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
907 unsigned int watch_point_id,
908 unsigned int reg_offset)
913 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
914 uint64_t page_table_base)
916 struct amdgpu_device *adev = get_amdgpu_device(kgd);
917 uint64_t base = page_table_base | AMDGPU_PTE_VALID;
919 if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
920 pr_err("trying to set page table base for wrong VMID %u\n",
925 /* TODO: take advantage of per-process address space size. For
926 * now, all processes share the same address space size, like
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);
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));
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));