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 <drm/drmP.h>
31 #include "amdgpu.h"
32 #include "amdgpu_amdkfd.h"
33 #include "amdgpu_ucode.h"
34 #include "soc15_hw_ip.h"
35 #include "gc/gc_10_1_0_offset.h"
36 #include "gc/gc_10_1_0_sh_mask.h"
37 #include "navi10_enum.h"
38 #include "athub/athub_2_0_0_offset.h"
39 #include "athub/athub_2_0_0_sh_mask.h"
40 #include "oss/osssys_5_0_0_offset.h"
41 #include "oss/osssys_5_0_0_sh_mask.h"
42 #include "soc15_common.h"
43 #include "v10_structs.h"
44 #include "nv.h"
45 #include "nvd.h"
46 
47 enum hqd_dequeue_request_type {
48 	NO_ACTION = 0,
49 	DRAIN_PIPE,
50 	RESET_WAVES,
51 	SAVE_WAVES
52 };
53 
54 /*
55  * Register access functions
56  */
57 
58 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
59 		uint32_t sh_mem_config,
60 		uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit,
61 		uint32_t sh_mem_bases);
62 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
63 		unsigned int vmid);
64 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
65 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
66 			uint32_t queue_id, uint32_t __user *wptr,
67 			uint32_t wptr_shift, uint32_t wptr_mask,
68 			struct mm_struct *mm);
69 static int kgd_hqd_dump(struct kgd_dev *kgd,
70 			uint32_t pipe_id, uint32_t queue_id,
71 			uint32_t (**dump)[2], uint32_t *n_regs);
72 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
73 			     uint32_t __user *wptr, struct mm_struct *mm);
74 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
75 			     uint32_t engine_id, uint32_t queue_id,
76 			     uint32_t (**dump)[2], uint32_t *n_regs);
77 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
78 		uint32_t pipe_id, uint32_t queue_id);
79 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
80 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
81 				enum kfd_preempt_type reset_type,
82 				unsigned int utimeout, uint32_t pipe_id,
83 				uint32_t queue_id);
84 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
85 				unsigned int utimeout);
86 #if 0
87 static uint32_t get_watch_base_addr(struct amdgpu_device *adev);
88 #endif
89 static int kgd_address_watch_disable(struct kgd_dev *kgd);
90 static int kgd_address_watch_execute(struct kgd_dev *kgd,
91 					unsigned int watch_point_id,
92 					uint32_t cntl_val,
93 					uint32_t addr_hi,
94 					uint32_t addr_lo);
95 static int kgd_wave_control_execute(struct kgd_dev *kgd,
96 					uint32_t gfx_index_val,
97 					uint32_t sq_cmd);
98 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
99 					unsigned int watch_point_id,
100 					unsigned int reg_offset);
101 
102 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
103 		uint8_t vmid);
104 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
105 		uint8_t vmid);
106 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
107 		uint64_t page_table_base);
108 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
109 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
110 
111 /* Because of REG_GET_FIELD() being used, we put this function in the
112  * asic specific file.
113  */
114 static int amdgpu_amdkfd_get_tile_config(struct kgd_dev *kgd,
115 		struct tile_config *config)
116 {
117 	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
118 
119 	config->gb_addr_config = adev->gfx.config.gb_addr_config;
120 #if 0
121 /* TODO - confirm REG_GET_FIELD x2, should be OK as is... but
122  * MC_ARB_RAMCFG register doesn't exist on Vega10 - initial amdgpu
123  * changes commented out related code, doing the same here for now but
124  * need to sync with Ken et al
125  */
126 	config->num_banks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
127 				MC_ARB_RAMCFG, NOOFBANK);
128 	config->num_ranks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
129 				MC_ARB_RAMCFG, NOOFRANKS);
130 #endif
131 
132 	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
133 	config->num_tile_configs =
134 			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
135 	config->macro_tile_config_ptr =
136 			adev->gfx.config.macrotile_mode_array;
137 	config->num_macro_tile_configs =
138 			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
139 
140 	return 0;
141 }
142 
143 static const struct kfd2kgd_calls kfd2kgd = {
144 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
145 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
146 	.init_interrupts = kgd_init_interrupts,
147 	.hqd_load = kgd_hqd_load,
148 	.hqd_sdma_load = kgd_hqd_sdma_load,
149 	.hqd_dump = kgd_hqd_dump,
150 	.hqd_sdma_dump = kgd_hqd_sdma_dump,
151 	.hqd_is_occupied = kgd_hqd_is_occupied,
152 	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
153 	.hqd_destroy = kgd_hqd_destroy,
154 	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
155 	.address_watch_disable = kgd_address_watch_disable,
156 	.address_watch_execute = kgd_address_watch_execute,
157 	.wave_control_execute = kgd_wave_control_execute,
158 	.address_watch_get_offset = kgd_address_watch_get_offset,
159 	.get_atc_vmid_pasid_mapping_pasid =
160 			get_atc_vmid_pasid_mapping_pasid,
161 	.get_atc_vmid_pasid_mapping_valid =
162 			get_atc_vmid_pasid_mapping_valid,
163 	.invalidate_tlbs = invalidate_tlbs,
164 	.invalidate_tlbs_vmid = invalidate_tlbs_vmid,
165 	.set_vm_context_page_table_base = set_vm_context_page_table_base,
166 	.get_tile_config = amdgpu_amdkfd_get_tile_config,
167 };
168 
169 struct kfd2kgd_calls *amdgpu_amdkfd_gfx_10_0_get_functions()
170 {
171 	return (struct kfd2kgd_calls *)&kfd2kgd;
172 }
173 
174 static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
175 {
176 	return (struct amdgpu_device *)kgd;
177 }
178 
179 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
180 			uint32_t queue, uint32_t vmid)
181 {
182 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
183 
184 	mutex_lock(&adev->srbm_mutex);
185 	nv_grbm_select(adev, mec, pipe, queue, vmid);
186 }
187 
188 static void unlock_srbm(struct kgd_dev *kgd)
189 {
190 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
191 
192 	nv_grbm_select(adev, 0, 0, 0, 0);
193 	mutex_unlock(&adev->srbm_mutex);
194 }
195 
196 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
197 				uint32_t queue_id)
198 {
199 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
200 
201 	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
202 	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
203 
204 	lock_srbm(kgd, mec, pipe, queue_id, 0);
205 }
206 
207 static uint32_t get_queue_mask(struct amdgpu_device *adev,
208 			       uint32_t pipe_id, uint32_t queue_id)
209 {
210 	unsigned int bit = (pipe_id * adev->gfx.mec.num_queue_per_pipe +
211 			    queue_id) & 31;
212 
213 	return ((uint32_t)1) << bit;
214 }
215 
216 static void release_queue(struct kgd_dev *kgd)
217 {
218 	unlock_srbm(kgd);
219 }
220 
221 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
222 					uint32_t sh_mem_config,
223 					uint32_t sh_mem_ape1_base,
224 					uint32_t sh_mem_ape1_limit,
225 					uint32_t sh_mem_bases)
226 {
227 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
228 
229 	lock_srbm(kgd, 0, 0, 0, vmid);
230 
231 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_CONFIG), sh_mem_config);
232 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSH_MEM_BASES), sh_mem_bases);
233 	/* APE1 no longer exists on GFX9 */
234 
235 	unlock_srbm(kgd);
236 }
237 
238 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
239 					unsigned int vmid)
240 {
241 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
242 
243 	/*
244 	 * We have to assume that there is no outstanding mapping.
245 	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
246 	 * a mapping is in progress or because a mapping finished
247 	 * and the SW cleared it.
248 	 * So the protocol is to always wait & clear.
249 	 */
250 	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
251 			ATC_VMID0_PASID_MAPPING__VALID_MASK;
252 
253 	pr_debug("pasid 0x%x vmid %d, reg value %x\n", pasid, vmid, pasid_mapping);
254 	/*
255 	 * need to do this twice, once for gfx and once for mmhub
256 	 * for ATC add 16 to VMID for mmhub, for IH different registers.
257 	 * ATC_VMID0..15 registers are separate from ATC_VMID16..31.
258 	 */
259 
260 	pr_debug("ATHUB, reg %x\n", SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid);
261 	WREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) + vmid,
262 	       pasid_mapping);
263 
264 #if 0
265 	/* TODO: uncomment this code when the hardware support is ready. */
266 	while (!(RREG32(SOC15_REG_OFFSET(
267 				ATHUB, 0,
268 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS)) &
269 		 (1U << vmid)))
270 		cpu_relax();
271 
272 	pr_debug("ATHUB mapping update finished\n");
273 	WREG32(SOC15_REG_OFFSET(ATHUB, 0,
274 				mmATC_VMID_PASID_MAPPING_UPDATE_STATUS),
275 	       1U << vmid);
276 #endif
277 
278 	/* Mapping vmid to pasid also for IH block */
279 	pr_debug("update mapping for IH block and mmhub");
280 	WREG32(SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid,
281 	       pasid_mapping);
282 
283 	return 0;
284 }
285 
286 /* TODO - RING0 form of field is obsolete, seems to date back to SI
287  * but still works
288  */
289 
290 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
291 {
292 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
293 	uint32_t mec;
294 	uint32_t pipe;
295 
296 	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
297 	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
298 
299 	lock_srbm(kgd, mec, pipe, 0, 0);
300 
301 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCPC_INT_CNTL),
302 		CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
303 		CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
304 
305 	unlock_srbm(kgd);
306 
307 	return 0;
308 }
309 
310 static uint32_t get_sdma_base_addr(struct amdgpu_device *adev,
311 				unsigned int engine_id,
312 				unsigned int queue_id)
313 {
314 	uint32_t base[2] = {
315 		SOC15_REG_OFFSET(SDMA0, 0,
316 				 mmSDMA0_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL,
317 		/* On gfx10, mmSDMA1_xxx registers are defined NOT based
318 		 * on SDMA1 base address (dw 0x1860) but based on SDMA0
319 		 * base address (dw 0x1260). Therefore use mmSDMA0_RLC0_RB_CNTL
320 		 * instead of mmSDMA1_RLC0_RB_CNTL for the base address calc
321 		 * below
322 		 */
323 		SOC15_REG_OFFSET(SDMA1, 0,
324 				 mmSDMA1_RLC0_RB_CNTL) - mmSDMA0_RLC0_RB_CNTL
325 	};
326 	uint32_t retval;
327 
328 	retval = base[engine_id] + queue_id * (mmSDMA0_RLC1_RB_CNTL -
329 					       mmSDMA0_RLC0_RB_CNTL);
330 
331 	pr_debug("sdma base address: 0x%x\n", retval);
332 
333 	return retval;
334 }
335 
336 #if 0
337 static uint32_t get_watch_base_addr(struct amdgpu_device *adev)
338 {
339 	uint32_t retval = SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) -
340 			mmTCP_WATCH0_ADDR_H;
341 
342 	pr_debug("kfd: reg watch base address: 0x%x\n", retval);
343 
344 	return retval;
345 }
346 #endif
347 
348 static inline struct v10_compute_mqd *get_mqd(void *mqd)
349 {
350 	return (struct v10_compute_mqd *)mqd;
351 }
352 
353 static inline struct v10_sdma_mqd *get_sdma_mqd(void *mqd)
354 {
355 	return (struct v10_sdma_mqd *)mqd;
356 }
357 
358 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
359 			uint32_t queue_id, uint32_t __user *wptr,
360 			uint32_t wptr_shift, uint32_t wptr_mask,
361 			struct mm_struct *mm)
362 {
363 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
364 	struct v10_compute_mqd *m;
365 	uint32_t *mqd_hqd;
366 	uint32_t reg, hqd_base, data;
367 
368 	m = get_mqd(mqd);
369 
370 	pr_debug("Load hqd of pipe %d queue %d\n", pipe_id, queue_id);
371 	acquire_queue(kgd, pipe_id, queue_id);
372 
373 	/* HIQ is set during driver init period with vmid set to 0*/
374 	if (m->cp_hqd_vmid == 0) {
375 		uint32_t value, mec, pipe;
376 
377 		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
378 		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
379 
380 		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
381 			mec, pipe, queue_id);
382 		value = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS));
383 		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
384 			((mec << 5) | (pipe << 3) | queue_id | 0x80));
385 		WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CP_SCHEDULERS), value);
386 	}
387 
388 	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
389 	mqd_hqd = &m->cp_mqd_base_addr_lo;
390 	hqd_base = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
391 
392 	for (reg = hqd_base;
393 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
394 		WREG32(reg, mqd_hqd[reg - hqd_base]);
395 
396 
397 	/* Activate doorbell logic before triggering WPTR poll. */
398 	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
399 			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
400 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL), data);
401 
402 	if (wptr) {
403 		/* Don't read wptr with get_user because the user
404 		 * context may not be accessible (if this function
405 		 * runs in a work queue). Instead trigger a one-shot
406 		 * polling read from memory in the CP. This assumes
407 		 * that wptr is GPU-accessible in the queue's VMID via
408 		 * ATC or SVM. WPTR==RPTR before starting the poll so
409 		 * the CP starts fetching new commands from the right
410 		 * place.
411 		 *
412 		 * Guessing a 64-bit WPTR from a 32-bit RPTR is a bit
413 		 * tricky. Assume that the queue didn't overflow. The
414 		 * number of valid bits in the 32-bit RPTR depends on
415 		 * the queue size. The remaining bits are taken from
416 		 * the saved 64-bit WPTR. If the WPTR wrapped, add the
417 		 * queue size.
418 		 */
419 		uint32_t queue_size =
420 			2 << REG_GET_FIELD(m->cp_hqd_pq_control,
421 					   CP_HQD_PQ_CONTROL, QUEUE_SIZE);
422 		uint64_t guessed_wptr = m->cp_hqd_pq_rptr & (queue_size - 1);
423 
424 		if ((m->cp_hqd_pq_wptr_lo & (queue_size - 1)) < guessed_wptr)
425 			guessed_wptr += queue_size;
426 		guessed_wptr += m->cp_hqd_pq_wptr_lo & ~(queue_size - 1);
427 		guessed_wptr += (uint64_t)m->cp_hqd_pq_wptr_hi << 32;
428 
429 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_LO),
430 		       lower_32_bits(guessed_wptr));
431 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI),
432 		       upper_32_bits(guessed_wptr));
433 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR),
434 		       lower_32_bits((uint64_t)wptr));
435 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI),
436 		       upper_32_bits((uint64_t)wptr));
437 		pr_debug("%s setting CP_PQ_WPTR_POLL_CNTL1 to %x\n", __func__, get_queue_mask(adev, pipe_id, queue_id));
438 		WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_PQ_WPTR_POLL_CNTL1),
439 		       get_queue_mask(adev, pipe_id, queue_id));
440 	}
441 
442 	/* Start the EOP fetcher */
443 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_EOP_RPTR),
444 	       REG_SET_FIELD(m->cp_hqd_eop_rptr,
445 			     CP_HQD_EOP_RPTR, INIT_FETCHER, 1));
446 
447 	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
448 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE), data);
449 
450 	release_queue(kgd);
451 
452 	return 0;
453 }
454 
455 static int kgd_hqd_dump(struct kgd_dev *kgd,
456 			uint32_t pipe_id, uint32_t queue_id,
457 			uint32_t (**dump)[2], uint32_t *n_regs)
458 {
459 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
460 	uint32_t i = 0, reg;
461 #define HQD_N_REGS 56
462 #define DUMP_REG(addr) do {				\
463 		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
464 			break;				\
465 		(*dump)[i][0] = (addr) << 2;		\
466 		(*dump)[i++][1] = RREG32(addr);		\
467 	} while (0)
468 
469 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
470 	if (*dump == NULL)
471 		return -ENOMEM;
472 
473 	acquire_queue(kgd, pipe_id, queue_id);
474 
475 	for (reg = SOC15_REG_OFFSET(GC, 0, mmCP_MQD_BASE_ADDR);
476 	     reg <= SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_WPTR_HI); reg++)
477 		DUMP_REG(reg);
478 
479 	release_queue(kgd);
480 
481 	WARN_ON_ONCE(i != HQD_N_REGS);
482 	*n_regs = i;
483 
484 	return 0;
485 }
486 
487 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
488 			     uint32_t __user *wptr, struct mm_struct *mm)
489 {
490 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
491 	struct v10_sdma_mqd *m;
492 	uint32_t sdma_base_addr, sdmax_gfx_context_cntl;
493 	unsigned long end_jiffies;
494 	uint32_t data;
495 	uint64_t data64;
496 	uint64_t __user *wptr64 = (uint64_t __user *)wptr;
497 
498 	m = get_sdma_mqd(mqd);
499 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
500 					    m->sdma_queue_id);
501 	pr_debug("sdma load base addr %x for engine %d, queue %d\n", sdma_base_addr, m->sdma_engine_id, m->sdma_queue_id);
502 	sdmax_gfx_context_cntl = m->sdma_engine_id ?
503 		SOC15_REG_OFFSET(SDMA1, 0, mmSDMA1_GFX_CONTEXT_CNTL) :
504 		SOC15_REG_OFFSET(SDMA0, 0, mmSDMA0_GFX_CONTEXT_CNTL);
505 
506 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
507 		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
508 
509 	end_jiffies = msecs_to_jiffies(2000) + jiffies;
510 	while (true) {
511 		data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
512 		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
513 			break;
514 		if (time_after(jiffies, end_jiffies))
515 			return -ETIME;
516 		usleep_range(500, 1000);
517 	}
518 	data = RREG32(sdmax_gfx_context_cntl);
519 	data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
520 			     RESUME_CTX, 0);
521 	WREG32(sdmax_gfx_context_cntl, data);
522 
523 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL_OFFSET,
524 	       m->sdmax_rlcx_doorbell_offset);
525 
526 	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
527 			     ENABLE, 1);
528 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data);
529 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdmax_rlcx_rb_rptr);
530 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI,
531 				m->sdmax_rlcx_rb_rptr_hi);
532 
533 	WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 1);
534 	if (read_user_wptr(mm, wptr64, data64)) {
535 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
536 		       lower_32_bits(data64));
537 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
538 		       upper_32_bits(data64));
539 	} else {
540 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
541 		       m->sdmax_rlcx_rb_rptr);
542 		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR_HI,
543 		       m->sdmax_rlcx_rb_rptr_hi);
544 	}
545 	WREG32(sdma_base_addr + mmSDMA0_RLC0_MINOR_PTR_UPDATE, 0);
546 
547 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
548 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI,
549 			m->sdmax_rlcx_rb_base_hi);
550 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
551 			m->sdmax_rlcx_rb_rptr_addr_lo);
552 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
553 			m->sdmax_rlcx_rb_rptr_addr_hi);
554 
555 	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
556 			     RB_ENABLE, 1);
557 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
558 
559 	return 0;
560 }
561 
562 static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
563 			     uint32_t engine_id, uint32_t queue_id,
564 			     uint32_t (**dump)[2], uint32_t *n_regs)
565 {
566 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
567 	uint32_t sdma_base_addr = get_sdma_base_addr(adev, engine_id, queue_id);
568 	uint32_t i = 0, reg;
569 #undef HQD_N_REGS
570 #define HQD_N_REGS (19+6+7+10)
571 
572 	pr_debug("sdma dump engine id %d queue_id %d\n", engine_id, queue_id);
573 	pr_debug("sdma base addr %x\n", sdma_base_addr);
574 
575 	*dump = kmalloc(HQD_N_REGS*2*sizeof(uint32_t), GFP_KERNEL);
576 	if (*dump == NULL)
577 		return -ENOMEM;
578 
579 	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
580 		DUMP_REG(sdma_base_addr + reg);
581 	for (reg = mmSDMA0_RLC0_STATUS; reg <= mmSDMA0_RLC0_CSA_ADDR_HI; reg++)
582 		DUMP_REG(sdma_base_addr + reg);
583 	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN;
584 	     reg <= mmSDMA0_RLC0_MINOR_PTR_UPDATE; reg++)
585 		DUMP_REG(sdma_base_addr + reg);
586 	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0;
587 	     reg <= mmSDMA0_RLC0_MIDCMD_CNTL; reg++)
588 		DUMP_REG(sdma_base_addr + reg);
589 
590 	WARN_ON_ONCE(i != HQD_N_REGS);
591 	*n_regs = i;
592 
593 	return 0;
594 }
595 
596 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
597 				uint32_t pipe_id, uint32_t queue_id)
598 {
599 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
600 	uint32_t act;
601 	bool retval = false;
602 	uint32_t low, high;
603 
604 	acquire_queue(kgd, pipe_id, queue_id);
605 	act = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
606 	if (act) {
607 		low = lower_32_bits(queue_address >> 8);
608 		high = upper_32_bits(queue_address >> 8);
609 
610 		if (low == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE)) &&
611 		   high == RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_PQ_BASE_HI)))
612 			retval = true;
613 	}
614 	release_queue(kgd);
615 	return retval;
616 }
617 
618 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
619 {
620 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
621 	struct v10_sdma_mqd *m;
622 	uint32_t sdma_base_addr;
623 	uint32_t sdma_rlc_rb_cntl;
624 
625 	m = get_sdma_mqd(mqd);
626 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
627 					    m->sdma_queue_id);
628 
629 	sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
630 
631 	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
632 		return true;
633 
634 	return false;
635 }
636 
637 static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
638 				enum kfd_preempt_type reset_type,
639 				unsigned int utimeout, uint32_t pipe_id,
640 				uint32_t queue_id)
641 {
642 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
643 	enum hqd_dequeue_request_type type;
644 	unsigned long end_jiffies;
645 	uint32_t temp;
646 	struct v10_compute_mqd *m = get_mqd(mqd);
647 
648 #if 0
649 	unsigned long flags;
650 	int retry;
651 #endif
652 
653 	acquire_queue(kgd, pipe_id, queue_id);
654 
655 	if (m->cp_hqd_vmid == 0)
656 		WREG32_FIELD15(GC, 0, RLC_CP_SCHEDULERS, scheduler1, 0);
657 
658 	switch (reset_type) {
659 	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
660 		type = DRAIN_PIPE;
661 		break;
662 	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
663 		type = RESET_WAVES;
664 		break;
665 	default:
666 		type = DRAIN_PIPE;
667 		break;
668 	}
669 
670 #if 0 /* Is this still needed? */
671 	/* Workaround: If IQ timer is active and the wait time is close to or
672 	 * equal to 0, dequeueing is not safe. Wait until either the wait time
673 	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
674 	 * cleared before continuing. Also, ensure wait times are set to at
675 	 * least 0x3.
676 	 */
677 	local_irq_save(flags);
678 	preempt_disable();
679 	retry = 5000; /* wait for 500 usecs at maximum */
680 	while (true) {
681 		temp = RREG32(mmCP_HQD_IQ_TIMER);
682 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
683 			pr_debug("HW is processing IQ\n");
684 			goto loop;
685 		}
686 		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
687 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
688 					== 3) /* SEM-rearm is safe */
689 				break;
690 			/* Wait time 3 is safe for CP, but our MMIO read/write
691 			 * time is close to 1 microsecond, so check for 10 to
692 			 * leave more buffer room
693 			 */
694 			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
695 					>= 10)
696 				break;
697 			pr_debug("IQ timer is active\n");
698 		} else
699 			break;
700 loop:
701 		if (!retry) {
702 			pr_err("CP HQD IQ timer status time out\n");
703 			break;
704 		}
705 		ndelay(100);
706 		--retry;
707 	}
708 	retry = 1000;
709 	while (true) {
710 		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
711 		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
712 			break;
713 		pr_debug("Dequeue request is pending\n");
714 
715 		if (!retry) {
716 			pr_err("CP HQD dequeue request time out\n");
717 			break;
718 		}
719 		ndelay(100);
720 		--retry;
721 	}
722 	local_irq_restore(flags);
723 	preempt_enable();
724 #endif
725 
726 	WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_DEQUEUE_REQUEST), type);
727 
728 	end_jiffies = (utimeout * HZ / 1000) + jiffies;
729 	while (true) {
730 		temp = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_HQD_ACTIVE));
731 		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
732 			break;
733 		if (time_after(jiffies, end_jiffies)) {
734 			pr_err("cp queue preemption time out.\n");
735 			release_queue(kgd);
736 			return -ETIME;
737 		}
738 		usleep_range(500, 1000);
739 	}
740 
741 	release_queue(kgd);
742 	return 0;
743 }
744 
745 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
746 				unsigned int utimeout)
747 {
748 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
749 	struct v10_sdma_mqd *m;
750 	uint32_t sdma_base_addr;
751 	uint32_t temp;
752 	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
753 
754 	m = get_sdma_mqd(mqd);
755 	sdma_base_addr = get_sdma_base_addr(adev, m->sdma_engine_id,
756 					    m->sdma_queue_id);
757 
758 	temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
759 	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
760 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, temp);
761 
762 	while (true) {
763 		temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
764 		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
765 			break;
766 		if (time_after(jiffies, end_jiffies))
767 			return -ETIME;
768 		usleep_range(500, 1000);
769 	}
770 
771 	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, 0);
772 	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
773 		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL) |
774 		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
775 
776 	m->sdmax_rlcx_rb_rptr = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR);
777 	m->sdmax_rlcx_rb_rptr_hi =
778 		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_HI);
779 
780 	return 0;
781 }
782 
783 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
784 							uint8_t vmid)
785 {
786 	uint32_t reg;
787 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
788 
789 	reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
790 		     + vmid);
791 	return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
792 }
793 
794 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
795 								uint8_t vmid)
796 {
797 	uint32_t reg;
798 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
799 
800 	reg = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
801 		     + vmid);
802 	return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
803 }
804 
805 static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid)
806 {
807 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
808 	uint32_t req = (1 << vmid) |
809 		(0 << GCVM_INVALIDATE_ENG0_REQ__FLUSH_TYPE__SHIFT) |/* legacy */
810 		GCVM_INVALIDATE_ENG0_REQ__INVALIDATE_L2_PTES_MASK |
811 		GCVM_INVALIDATE_ENG0_REQ__INVALIDATE_L2_PDE0_MASK |
812 		GCVM_INVALIDATE_ENG0_REQ__INVALIDATE_L2_PDE1_MASK |
813 		GCVM_INVALIDATE_ENG0_REQ__INVALIDATE_L2_PDE2_MASK |
814 		GCVM_INVALIDATE_ENG0_REQ__INVALIDATE_L1_PTES_MASK;
815 
816 	mutex_lock(&adev->srbm_mutex);
817 
818 	/* Use light weight invalidation.
819 	 *
820 	 * TODO 1: agree on the right set of invalidation registers for
821 	 * KFD use. Use the last one for now. Invalidate only GCHUB as
822 	 * SDMA is now moved to GCHUB
823 	 *
824 	 * TODO 2: support range-based invalidation, requires kfg2kgd
825 	 * interface change
826 	 */
827 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_INVALIDATE_ENG0_ADDR_RANGE_LO32),
828 				0xffffffff);
829 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_INVALIDATE_ENG0_ADDR_RANGE_HI32),
830 				0x0000001f);
831 
832 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_INVALIDATE_ENG0_REQ), req);
833 
834 	while (!(RREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_INVALIDATE_ENG0_ACK)) &
835 					(1 << vmid)))
836 		cpu_relax();
837 
838 	mutex_unlock(&adev->srbm_mutex);
839 }
840 
841 static int invalidate_tlbs_with_kiq(struct amdgpu_device *adev, uint16_t pasid)
842 {
843 	signed long r;
844 	uint32_t seq;
845 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
846 
847 	spin_lock(&adev->gfx.kiq.ring_lock);
848 	amdgpu_ring_alloc(ring, 12); /* fence + invalidate_tlbs package*/
849 	amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0));
850 	amdgpu_ring_write(ring,
851 			PACKET3_INVALIDATE_TLBS_DST_SEL(1) |
852 			PACKET3_INVALIDATE_TLBS_PASID(pasid));
853 	amdgpu_fence_emit_polling(ring, &seq);
854 	amdgpu_ring_commit(ring);
855 	spin_unlock(&adev->gfx.kiq.ring_lock);
856 
857 	r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
858 	if (r < 1) {
859 		DRM_ERROR("wait for kiq fence error: %ld.\n", r);
860 		return -ETIME;
861 	}
862 
863 	return 0;
864 }
865 
866 static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
867 {
868 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
869 	int vmid;
870 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
871 
872 	if (amdgpu_emu_mode == 0 && ring->sched.ready)
873 		return invalidate_tlbs_with_kiq(adev, pasid);
874 
875 	for (vmid = 0; vmid < 16; vmid++) {
876 		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
877 			continue;
878 		if (get_atc_vmid_pasid_mapping_valid(kgd, vmid)) {
879 			if (get_atc_vmid_pasid_mapping_pasid(kgd, vmid)
880 				== pasid) {
881 				write_vmid_invalidate_request(kgd, vmid);
882 				break;
883 			}
884 		}
885 	}
886 
887 	return 0;
888 }
889 
890 static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
891 {
892 	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
893 
894 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
895 		pr_err("non kfd vmid %d\n", vmid);
896 		return 0;
897 	}
898 
899 	write_vmid_invalidate_request(kgd, vmid);
900 	return 0;
901 }
902 
903 static int kgd_address_watch_disable(struct kgd_dev *kgd)
904 {
905 	return 0;
906 }
907 
908 static int kgd_address_watch_execute(struct kgd_dev *kgd,
909 					unsigned int watch_point_id,
910 					uint32_t cntl_val,
911 					uint32_t addr_hi,
912 					uint32_t addr_lo)
913 {
914 	return 0;
915 }
916 
917 static int kgd_wave_control_execute(struct kgd_dev *kgd,
918 					uint32_t gfx_index_val,
919 					uint32_t sq_cmd)
920 {
921 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
922 	uint32_t data = 0;
923 
924 	mutex_lock(&adev->grbm_idx_mutex);
925 
926 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), gfx_index_val);
927 	WREG32(SOC15_REG_OFFSET(GC, 0, mmSQ_CMD), sq_cmd);
928 
929 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
930 		INSTANCE_BROADCAST_WRITES, 1);
931 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
932 		SA_BROADCAST_WRITES, 1);
933 	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
934 		SE_BROADCAST_WRITES, 1);
935 
936 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGRBM_GFX_INDEX), data);
937 	mutex_unlock(&adev->grbm_idx_mutex);
938 
939 	return 0;
940 }
941 
942 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
943 					unsigned int watch_point_id,
944 					unsigned int reg_offset)
945 {
946 	return 0;
947 }
948 
949 static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
950 		uint64_t page_table_base)
951 {
952 	struct amdgpu_device *adev = get_amdgpu_device(kgd);
953 	uint64_t base = page_table_base | AMDGPU_PTE_VALID;
954 
955 	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
956 		pr_err("trying to set page table base for wrong VMID %u\n",
957 		       vmid);
958 		return;
959 	}
960 
961 	/* TODO: take advantage of per-process address space size. For
962 	 * now, all processes share the same address space size, like
963 	 * on GFX8 and older.
964 	 */
965 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_START_ADDR_LO32) + (vmid*2), 0);
966 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_START_ADDR_HI32) + (vmid*2), 0);
967 
968 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_END_ADDR_LO32) + (vmid*2),
969 			lower_32_bits(adev->vm_manager.max_pfn - 1));
970 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_END_ADDR_HI32) + (vmid*2),
971 			upper_32_bits(adev->vm_manager.max_pfn - 1));
972 
973 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_LO32) + (vmid*2), lower_32_bits(base));
974 	WREG32(SOC15_REG_OFFSET(GC, 0, mmGCVM_CONTEXT0_PAGE_TABLE_BASE_ADDR_HI32) + (vmid*2), upper_32_bits(base));
975 }
976