Linux Audio

Check our new training course

Loading...
v5.4
  1/*
  2 * Copyright 2014 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
 23#include <linux/module.h>
 24#include <linux/fdtable.h>
 25#include <linux/uaccess.h>
 26#include <linux/mmu_context.h>
 27
 28#include "amdgpu.h"
 29#include "amdgpu_amdkfd.h"
 30#include "gfx_v8_0.h"
 31#include "gca/gfx_8_0_sh_mask.h"
 32#include "gca/gfx_8_0_d.h"
 33#include "gca/gfx_8_0_enum.h"
 34#include "oss/oss_3_0_sh_mask.h"
 35#include "oss/oss_3_0_d.h"
 36#include "gmc/gmc_8_1_sh_mask.h"
 37#include "gmc/gmc_8_1_d.h"
 38#include "vi_structs.h"
 39#include "vid.h"
 40
 41enum hqd_dequeue_request_type {
 42	NO_ACTION = 0,
 43	DRAIN_PIPE,
 44	RESET_WAVES
 45};
 46
 47/*
 48 * Register access functions
 49 */
 50
 51static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
 52		uint32_t sh_mem_config,
 53		uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit,
 54		uint32_t sh_mem_bases);
 55static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
 56		unsigned int vmid);
 57static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
 58static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
 59			uint32_t queue_id, uint32_t __user *wptr,
 60			uint32_t wptr_shift, uint32_t wptr_mask,
 61			struct mm_struct *mm);
 62static int kgd_hqd_dump(struct kgd_dev *kgd,
 63			uint32_t pipe_id, uint32_t queue_id,
 64			uint32_t (**dump)[2], uint32_t *n_regs);
 65static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
 66			     uint32_t __user *wptr, struct mm_struct *mm);
 67static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
 68			     uint32_t engine_id, uint32_t queue_id,
 69			     uint32_t (**dump)[2], uint32_t *n_regs);
 70static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
 71		uint32_t pipe_id, uint32_t queue_id);
 72static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
 73static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
 74				enum kfd_preempt_type reset_type,
 75				unsigned int utimeout, uint32_t pipe_id,
 76				uint32_t queue_id);
 77static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
 78				unsigned int utimeout);
 79static int kgd_address_watch_disable(struct kgd_dev *kgd);
 80static int kgd_address_watch_execute(struct kgd_dev *kgd,
 81					unsigned int watch_point_id,
 82					uint32_t cntl_val,
 83					uint32_t addr_hi,
 84					uint32_t addr_lo);
 85static int kgd_wave_control_execute(struct kgd_dev *kgd,
 86					uint32_t gfx_index_val,
 87					uint32_t sq_cmd);
 88static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
 89					unsigned int watch_point_id,
 90					unsigned int reg_offset);
 91
 92static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
 93		uint8_t vmid);
 94static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
 95		uint8_t vmid);
 96static void set_scratch_backing_va(struct kgd_dev *kgd,
 97					uint64_t va, uint32_t vmid);
 98static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
 99		uint64_t page_table_base);
100static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid);
101static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid);
102
103/* Because of REG_GET_FIELD() being used, we put this function in the
104 * asic specific file.
105 */
106static int get_tile_config(struct kgd_dev *kgd,
107		struct tile_config *config)
108{
109	struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
110
111	config->gb_addr_config = adev->gfx.config.gb_addr_config;
112	config->num_banks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
113				MC_ARB_RAMCFG, NOOFBANK);
114	config->num_ranks = REG_GET_FIELD(adev->gfx.config.mc_arb_ramcfg,
115				MC_ARB_RAMCFG, NOOFRANKS);
116
117	config->tile_config_ptr = adev->gfx.config.tile_mode_array;
118	config->num_tile_configs =
119			ARRAY_SIZE(adev->gfx.config.tile_mode_array);
120	config->macro_tile_config_ptr =
121			adev->gfx.config.macrotile_mode_array;
122	config->num_macro_tile_configs =
123			ARRAY_SIZE(adev->gfx.config.macrotile_mode_array);
124
125	return 0;
126}
127
128static const struct kfd2kgd_calls kfd2kgd = {
129	.program_sh_mem_settings = kgd_program_sh_mem_settings,
130	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
131	.init_interrupts = kgd_init_interrupts,
132	.hqd_load = kgd_hqd_load,
133	.hqd_sdma_load = kgd_hqd_sdma_load,
134	.hqd_dump = kgd_hqd_dump,
135	.hqd_sdma_dump = kgd_hqd_sdma_dump,
136	.hqd_is_occupied = kgd_hqd_is_occupied,
137	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
138	.hqd_destroy = kgd_hqd_destroy,
139	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
140	.address_watch_disable = kgd_address_watch_disable,
141	.address_watch_execute = kgd_address_watch_execute,
142	.wave_control_execute = kgd_wave_control_execute,
143	.address_watch_get_offset = kgd_address_watch_get_offset,
144	.get_atc_vmid_pasid_mapping_pasid =
145			get_atc_vmid_pasid_mapping_pasid,
146	.get_atc_vmid_pasid_mapping_valid =
147			get_atc_vmid_pasid_mapping_valid,
148	.set_scratch_backing_va = set_scratch_backing_va,
149	.get_tile_config = get_tile_config,
150	.set_vm_context_page_table_base = set_vm_context_page_table_base,
151	.invalidate_tlbs = invalidate_tlbs,
152	.invalidate_tlbs_vmid = invalidate_tlbs_vmid,
153};
154
155struct kfd2kgd_calls *amdgpu_amdkfd_gfx_8_0_get_functions(void)
156{
157	return (struct kfd2kgd_calls *)&kfd2kgd;
158}
159
160static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
161{
162	return (struct amdgpu_device *)kgd;
163}
164
165static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
166			uint32_t queue, uint32_t vmid)
167{
168	struct amdgpu_device *adev = get_amdgpu_device(kgd);
169	uint32_t value = PIPEID(pipe) | MEID(mec) | VMID(vmid) | QUEUEID(queue);
170
171	mutex_lock(&adev->srbm_mutex);
172	WREG32(mmSRBM_GFX_CNTL, value);
173}
174
175static void unlock_srbm(struct kgd_dev *kgd)
176{
177	struct amdgpu_device *adev = get_amdgpu_device(kgd);
178
179	WREG32(mmSRBM_GFX_CNTL, 0);
180	mutex_unlock(&adev->srbm_mutex);
181}
182
183static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
184				uint32_t queue_id)
185{
186	struct amdgpu_device *adev = get_amdgpu_device(kgd);
187
188	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
189	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
190
191	lock_srbm(kgd, mec, pipe, queue_id, 0);
192}
193
194static void release_queue(struct kgd_dev *kgd)
195{
196	unlock_srbm(kgd);
197}
198
199static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
200					uint32_t sh_mem_config,
201					uint32_t sh_mem_ape1_base,
202					uint32_t sh_mem_ape1_limit,
203					uint32_t sh_mem_bases)
204{
205	struct amdgpu_device *adev = get_amdgpu_device(kgd);
206
207	lock_srbm(kgd, 0, 0, 0, vmid);
208
209	WREG32(mmSH_MEM_CONFIG, sh_mem_config);
210	WREG32(mmSH_MEM_APE1_BASE, sh_mem_ape1_base);
211	WREG32(mmSH_MEM_APE1_LIMIT, sh_mem_ape1_limit);
212	WREG32(mmSH_MEM_BASES, sh_mem_bases);
213
214	unlock_srbm(kgd);
215}
216
217static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
218					unsigned int vmid)
219{
220	struct amdgpu_device *adev = get_amdgpu_device(kgd);
221
222	/*
223	 * We have to assume that there is no outstanding mapping.
224	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
225	 * a mapping is in progress or because a mapping finished
226	 * and the SW cleared it.
227	 * So the protocol is to always wait & clear.
228	 */
229	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
230			ATC_VMID0_PASID_MAPPING__VALID_MASK;
231
232	WREG32(mmATC_VMID0_PASID_MAPPING + vmid, pasid_mapping);
233
234	while (!(RREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS) & (1U << vmid)))
235		cpu_relax();
236	WREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
237
238	/* Mapping vmid to pasid also for IH block */
239	WREG32(mmIH_VMID_0_LUT + vmid, pasid_mapping);
240
241	return 0;
242}
243
244static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
245{
246	struct amdgpu_device *adev = get_amdgpu_device(kgd);
247	uint32_t mec;
248	uint32_t pipe;
249
250	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
251	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
252
253	lock_srbm(kgd, mec, pipe, 0, 0);
254
255	WREG32(mmCPC_INT_CNTL, CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
256			CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
257
258	unlock_srbm(kgd);
259
260	return 0;
261}
262
263static inline uint32_t get_sdma_base_addr(struct vi_sdma_mqd *m)
264{
265	uint32_t retval;
266
267	retval = m->sdma_engine_id * SDMA1_REGISTER_OFFSET +
268		m->sdma_queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
269	pr_debug("sdma base address: 0x%x\n", retval);
 
 
270
271	return retval;
272}
273
274static inline struct vi_mqd *get_mqd(void *mqd)
275{
276	return (struct vi_mqd *)mqd;
277}
278
279static inline struct vi_sdma_mqd *get_sdma_mqd(void *mqd)
280{
281	return (struct vi_sdma_mqd *)mqd;
282}
283
284static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
285			uint32_t queue_id, uint32_t __user *wptr,
286			uint32_t wptr_shift, uint32_t wptr_mask,
287			struct mm_struct *mm)
288{
289	struct amdgpu_device *adev = get_amdgpu_device(kgd);
290	struct vi_mqd *m;
291	uint32_t *mqd_hqd;
292	uint32_t reg, wptr_val, data;
293	bool valid_wptr = false;
294
295	m = get_mqd(mqd);
296
297	acquire_queue(kgd, pipe_id, queue_id);
298
299	/* HIQ is set during driver init period with vmid set to 0*/
300	if (m->cp_hqd_vmid == 0) {
301		uint32_t value, mec, pipe;
302
303		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
304		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
305
306		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
307			mec, pipe, queue_id);
308		value = RREG32(mmRLC_CP_SCHEDULERS);
309		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
310			((mec << 5) | (pipe << 3) | queue_id | 0x80));
311		WREG32(mmRLC_CP_SCHEDULERS, value);
312	}
313
314	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
315	mqd_hqd = &m->cp_mqd_base_addr_lo;
316
317	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_CONTROL; reg++)
318		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
319
320	/* Tonga errata: EOP RPTR/WPTR should be left unmodified.
321	 * This is safe since EOP RPTR==WPTR for any inactive HQD
322	 * on ASICs that do not support context-save.
323	 * EOP writes/reads can start anywhere in the ring.
324	 */
325	if (get_amdgpu_device(kgd)->asic_type != CHIP_TONGA) {
326		WREG32(mmCP_HQD_EOP_RPTR, m->cp_hqd_eop_rptr);
327		WREG32(mmCP_HQD_EOP_WPTR, m->cp_hqd_eop_wptr);
328		WREG32(mmCP_HQD_EOP_WPTR_MEM, m->cp_hqd_eop_wptr_mem);
329	}
330
331	for (reg = mmCP_HQD_EOP_EVENTS; reg <= mmCP_HQD_ERROR; reg++)
332		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
333
334	/* Copy userspace write pointer value to register.
335	 * Activate doorbell logic to monitor subsequent changes.
336	 */
337	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
338			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
339	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, data);
340
341	/* read_user_ptr may take the mm->mmap_sem.
342	 * release srbm_mutex to avoid circular dependency between
343	 * srbm_mutex->mm_sem->reservation_ww_class_mutex->srbm_mutex.
344	 */
345	release_queue(kgd);
346	valid_wptr = read_user_wptr(mm, wptr, wptr_val);
347	acquire_queue(kgd, pipe_id, queue_id);
348	if (valid_wptr)
349		WREG32(mmCP_HQD_PQ_WPTR, (wptr_val << wptr_shift) & wptr_mask);
350
351	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
352	WREG32(mmCP_HQD_ACTIVE, data);
353
354	release_queue(kgd);
355
356	return 0;
357}
358
359static int kgd_hqd_dump(struct kgd_dev *kgd,
360			uint32_t pipe_id, uint32_t queue_id,
361			uint32_t (**dump)[2], uint32_t *n_regs)
362{
363	struct amdgpu_device *adev = get_amdgpu_device(kgd);
364	uint32_t i = 0, reg;
365#define HQD_N_REGS (54+4)
366#define DUMP_REG(addr) do {				\
367		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
368			break;				\
369		(*dump)[i][0] = (addr) << 2;		\
370		(*dump)[i++][1] = RREG32(addr);		\
371	} while (0)
372
373	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
374	if (*dump == NULL)
375		return -ENOMEM;
376
377	acquire_queue(kgd, pipe_id, queue_id);
378
379	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE0);
380	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE1);
381	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE2);
382	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE3);
383
384	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_DONES; reg++)
385		DUMP_REG(reg);
386
387	release_queue(kgd);
388
389	WARN_ON_ONCE(i != HQD_N_REGS);
390	*n_regs = i;
391
392	return 0;
393}
394
395static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
396			     uint32_t __user *wptr, struct mm_struct *mm)
397{
398	struct amdgpu_device *adev = get_amdgpu_device(kgd);
399	struct vi_sdma_mqd *m;
400	unsigned long end_jiffies;
401	uint32_t sdma_base_addr;
402	uint32_t data;
403
404	m = get_sdma_mqd(mqd);
405	sdma_base_addr = get_sdma_base_addr(m);
406	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
407		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
408
409	end_jiffies = msecs_to_jiffies(2000) + jiffies;
410	while (true) {
411		data = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
412		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
413			break;
414		if (time_after(jiffies, end_jiffies))
 
415			return -ETIME;
 
416		usleep_range(500, 1000);
417	}
418	if (m->sdma_engine_id) {
419		data = RREG32(mmSDMA1_GFX_CONTEXT_CNTL);
420		data = REG_SET_FIELD(data, SDMA1_GFX_CONTEXT_CNTL,
421				RESUME_CTX, 0);
422		WREG32(mmSDMA1_GFX_CONTEXT_CNTL, data);
423	} else {
424		data = RREG32(mmSDMA0_GFX_CONTEXT_CNTL);
425		data = REG_SET_FIELD(data, SDMA0_GFX_CONTEXT_CNTL,
426				RESUME_CTX, 0);
427		WREG32(mmSDMA0_GFX_CONTEXT_CNTL, data);
428	}
429
430	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
431			     ENABLE, 1);
432	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, data);
433	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR, m->sdmax_rlcx_rb_rptr);
 
434
435	if (read_user_wptr(mm, wptr, data))
436		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR, data);
437	else
438		WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_WPTR,
439		       m->sdmax_rlcx_rb_rptr);
440
441	WREG32(sdma_base_addr + mmSDMA0_RLC0_VIRTUAL_ADDR,
442				m->sdmax_rlcx_virtual_addr);
443	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
444	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_BASE_HI,
445			m->sdmax_rlcx_rb_base_hi);
446	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
447			m->sdmax_rlcx_rb_rptr_addr_lo);
448	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
449			m->sdmax_rlcx_rb_rptr_addr_hi);
450
451	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
452			     RB_ENABLE, 1);
453	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, data);
454
455	return 0;
456}
457
458static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
459			     uint32_t engine_id, uint32_t queue_id,
460			     uint32_t (**dump)[2], uint32_t *n_regs)
461{
462	struct amdgpu_device *adev = get_amdgpu_device(kgd);
463	uint32_t sdma_offset = engine_id * SDMA1_REGISTER_OFFSET +
464		queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
465	uint32_t i = 0, reg;
466#undef HQD_N_REGS
467#define HQD_N_REGS (19+4+2+3+7)
468
469	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
470	if (*dump == NULL)
471		return -ENOMEM;
472
473	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
474		DUMP_REG(sdma_offset + reg);
475	for (reg = mmSDMA0_RLC0_VIRTUAL_ADDR; reg <= mmSDMA0_RLC0_WATERMARK;
476	     reg++)
477		DUMP_REG(sdma_offset + reg);
478	for (reg = mmSDMA0_RLC0_CSA_ADDR_LO; reg <= mmSDMA0_RLC0_CSA_ADDR_HI;
479	     reg++)
480		DUMP_REG(sdma_offset + reg);
481	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN; reg <= mmSDMA0_RLC0_DUMMY_REG;
482	     reg++)
483		DUMP_REG(sdma_offset + reg);
484	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0; reg <= mmSDMA0_RLC0_MIDCMD_CNTL;
485	     reg++)
486		DUMP_REG(sdma_offset + reg);
487
488	WARN_ON_ONCE(i != HQD_N_REGS);
489	*n_regs = i;
490
491	return 0;
492}
493
494static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
495				uint32_t pipe_id, uint32_t queue_id)
496{
497	struct amdgpu_device *adev = get_amdgpu_device(kgd);
498	uint32_t act;
499	bool retval = false;
500	uint32_t low, high;
501
502	acquire_queue(kgd, pipe_id, queue_id);
503	act = RREG32(mmCP_HQD_ACTIVE);
504	if (act) {
505		low = lower_32_bits(queue_address >> 8);
506		high = upper_32_bits(queue_address >> 8);
507
508		if (low == RREG32(mmCP_HQD_PQ_BASE) &&
509				high == RREG32(mmCP_HQD_PQ_BASE_HI))
510			retval = true;
511	}
512	release_queue(kgd);
513	return retval;
514}
515
516static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
517{
518	struct amdgpu_device *adev = get_amdgpu_device(kgd);
519	struct vi_sdma_mqd *m;
520	uint32_t sdma_base_addr;
521	uint32_t sdma_rlc_rb_cntl;
522
523	m = get_sdma_mqd(mqd);
524	sdma_base_addr = get_sdma_base_addr(m);
525
526	sdma_rlc_rb_cntl = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
527
528	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
529		return true;
530
531	return false;
532}
533
534static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
535				enum kfd_preempt_type reset_type,
536				unsigned int utimeout, uint32_t pipe_id,
537				uint32_t queue_id)
538{
539	struct amdgpu_device *adev = get_amdgpu_device(kgd);
540	uint32_t temp;
541	enum hqd_dequeue_request_type type;
542	unsigned long flags, end_jiffies;
543	int retry;
544	struct vi_mqd *m = get_mqd(mqd);
545
546	if (adev->in_gpu_reset)
547		return -EIO;
548
549	acquire_queue(kgd, pipe_id, queue_id);
550
551	if (m->cp_hqd_vmid == 0)
552		WREG32_FIELD(RLC_CP_SCHEDULERS, scheduler1, 0);
553
554	switch (reset_type) {
555	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
556		type = DRAIN_PIPE;
557		break;
558	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
559		type = RESET_WAVES;
560		break;
561	default:
562		type = DRAIN_PIPE;
563		break;
564	}
565
566	/* Workaround: If IQ timer is active and the wait time is close to or
567	 * equal to 0, dequeueing is not safe. Wait until either the wait time
568	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
569	 * cleared before continuing. Also, ensure wait times are set to at
570	 * least 0x3.
571	 */
572	local_irq_save(flags);
573	preempt_disable();
574	retry = 5000; /* wait for 500 usecs at maximum */
575	while (true) {
576		temp = RREG32(mmCP_HQD_IQ_TIMER);
577		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
578			pr_debug("HW is processing IQ\n");
579			goto loop;
580		}
581		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
582			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
583					== 3) /* SEM-rearm is safe */
584				break;
585			/* Wait time 3 is safe for CP, but our MMIO read/write
586			 * time is close to 1 microsecond, so check for 10 to
587			 * leave more buffer room
588			 */
589			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
590					>= 10)
591				break;
592			pr_debug("IQ timer is active\n");
593		} else
594			break;
595loop:
596		if (!retry) {
597			pr_err("CP HQD IQ timer status time out\n");
598			break;
599		}
600		ndelay(100);
601		--retry;
602	}
603	retry = 1000;
604	while (true) {
605		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
606		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
607			break;
608		pr_debug("Dequeue request is pending\n");
609
610		if (!retry) {
611			pr_err("CP HQD dequeue request time out\n");
612			break;
613		}
614		ndelay(100);
615		--retry;
616	}
617	local_irq_restore(flags);
618	preempt_enable();
619
620	WREG32(mmCP_HQD_DEQUEUE_REQUEST, type);
621
622	end_jiffies = (utimeout * HZ / 1000) + jiffies;
623	while (true) {
624		temp = RREG32(mmCP_HQD_ACTIVE);
625		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
626			break;
627		if (time_after(jiffies, end_jiffies)) {
628			pr_err("cp queue preemption time out.\n");
629			release_queue(kgd);
630			return -ETIME;
631		}
632		usleep_range(500, 1000);
633	}
634
635	release_queue(kgd);
636	return 0;
637}
638
639static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
640				unsigned int utimeout)
641{
642	struct amdgpu_device *adev = get_amdgpu_device(kgd);
643	struct vi_sdma_mqd *m;
644	uint32_t sdma_base_addr;
645	uint32_t temp;
646	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
647
648	m = get_sdma_mqd(mqd);
649	sdma_base_addr = get_sdma_base_addr(m);
650
651	temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL);
652	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
653	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL, temp);
654
655	while (true) {
656		temp = RREG32(sdma_base_addr + mmSDMA0_RLC0_CONTEXT_STATUS);
657		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
658			break;
659		if (time_after(jiffies, end_jiffies))
 
660			return -ETIME;
 
661		usleep_range(500, 1000);
662	}
663
664	WREG32(sdma_base_addr + mmSDMA0_RLC0_DOORBELL, 0);
665	WREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL,
666		RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_CNTL) |
667		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
668
669	m->sdmax_rlcx_rb_rptr = RREG32(sdma_base_addr + mmSDMA0_RLC0_RB_RPTR);
670
671	return 0;
672}
673
674static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd,
675							uint8_t vmid)
676{
677	uint32_t reg;
678	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
679
680	reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
681	return reg & ATC_VMID0_PASID_MAPPING__VALID_MASK;
682}
683
684static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
685								uint8_t vmid)
686{
687	uint32_t reg;
688	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
689
690	reg = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
691	return reg & ATC_VMID0_PASID_MAPPING__PASID_MASK;
692}
693
694static int kgd_address_watch_disable(struct kgd_dev *kgd)
695{
696	return 0;
697}
698
699static int kgd_address_watch_execute(struct kgd_dev *kgd,
700					unsigned int watch_point_id,
701					uint32_t cntl_val,
702					uint32_t addr_hi,
703					uint32_t addr_lo)
704{
705	return 0;
706}
707
708static int kgd_wave_control_execute(struct kgd_dev *kgd,
709					uint32_t gfx_index_val,
710					uint32_t sq_cmd)
711{
712	struct amdgpu_device *adev = get_amdgpu_device(kgd);
713	uint32_t data = 0;
714
715	mutex_lock(&adev->grbm_idx_mutex);
716
717	WREG32(mmGRBM_GFX_INDEX, gfx_index_val);
718	WREG32(mmSQ_CMD, sq_cmd);
719
720	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
721		INSTANCE_BROADCAST_WRITES, 1);
722	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
723		SH_BROADCAST_WRITES, 1);
724	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
725		SE_BROADCAST_WRITES, 1);
726
727	WREG32(mmGRBM_GFX_INDEX, data);
728	mutex_unlock(&adev->grbm_idx_mutex);
729
730	return 0;
731}
732
733static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
734					unsigned int watch_point_id,
735					unsigned int reg_offset)
736{
737	return 0;
738}
739
740static void set_scratch_backing_va(struct kgd_dev *kgd,
741					uint64_t va, uint32_t vmid)
742{
743	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
744
745	lock_srbm(kgd, 0, 0, 0, vmid);
746	WREG32(mmSH_HIDDEN_PRIVATE_BASE_VMID, va);
747	unlock_srbm(kgd);
748}
749
750static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
751		uint64_t page_table_base)
752{
753	struct amdgpu_device *adev = get_amdgpu_device(kgd);
754
755	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
756		pr_err("trying to set page table base for wrong VMID\n");
757		return;
758	}
759	WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8,
760			lower_32_bits(page_table_base));
761}
762
763static int invalidate_tlbs(struct kgd_dev *kgd, uint16_t pasid)
764{
765	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
766	int vmid;
767	unsigned int tmp;
768
769	if (adev->in_gpu_reset)
770		return -EIO;
771
772	for (vmid = 0; vmid < 16; vmid++) {
773		if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid))
774			continue;
775
776		tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
777		if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
778			(tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
779			WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
780			RREG32(mmVM_INVALIDATE_RESPONSE);
781			break;
782		}
783	}
784
785	return 0;
786}
787
788static int invalidate_tlbs_vmid(struct kgd_dev *kgd, uint16_t vmid)
789{
790	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
791
792	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
793		pr_err("non kfd vmid %d\n", vmid);
794		return -EINVAL;
795	}
796
797	WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
798	RREG32(mmVM_INVALIDATE_RESPONSE);
799	return 0;
800}
v5.9
  1/*
  2 * Copyright 2014 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
 
 
 
 
 
 23#include "amdgpu.h"
 24#include "amdgpu_amdkfd.h"
 25#include "gfx_v8_0.h"
 26#include "gca/gfx_8_0_sh_mask.h"
 27#include "gca/gfx_8_0_d.h"
 28#include "gca/gfx_8_0_enum.h"
 29#include "oss/oss_3_0_sh_mask.h"
 30#include "oss/oss_3_0_d.h"
 31#include "gmc/gmc_8_1_sh_mask.h"
 32#include "gmc/gmc_8_1_d.h"
 33#include "vi_structs.h"
 34#include "vid.h"
 35
 36enum hqd_dequeue_request_type {
 37	NO_ACTION = 0,
 38	DRAIN_PIPE,
 39	RESET_WAVES
 40};
 41
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 42static inline struct amdgpu_device *get_amdgpu_device(struct kgd_dev *kgd)
 43{
 44	return (struct amdgpu_device *)kgd;
 45}
 46
 47static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
 48			uint32_t queue, uint32_t vmid)
 49{
 50	struct amdgpu_device *adev = get_amdgpu_device(kgd);
 51	uint32_t value = PIPEID(pipe) | MEID(mec) | VMID(vmid) | QUEUEID(queue);
 52
 53	mutex_lock(&adev->srbm_mutex);
 54	WREG32(mmSRBM_GFX_CNTL, value);
 55}
 56
 57static void unlock_srbm(struct kgd_dev *kgd)
 58{
 59	struct amdgpu_device *adev = get_amdgpu_device(kgd);
 60
 61	WREG32(mmSRBM_GFX_CNTL, 0);
 62	mutex_unlock(&adev->srbm_mutex);
 63}
 64
 65static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
 66				uint32_t queue_id)
 67{
 68	struct amdgpu_device *adev = get_amdgpu_device(kgd);
 69
 70	uint32_t mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
 71	uint32_t pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
 72
 73	lock_srbm(kgd, mec, pipe, queue_id, 0);
 74}
 75
 76static void release_queue(struct kgd_dev *kgd)
 77{
 78	unlock_srbm(kgd);
 79}
 80
 81static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
 82					uint32_t sh_mem_config,
 83					uint32_t sh_mem_ape1_base,
 84					uint32_t sh_mem_ape1_limit,
 85					uint32_t sh_mem_bases)
 86{
 87	struct amdgpu_device *adev = get_amdgpu_device(kgd);
 88
 89	lock_srbm(kgd, 0, 0, 0, vmid);
 90
 91	WREG32(mmSH_MEM_CONFIG, sh_mem_config);
 92	WREG32(mmSH_MEM_APE1_BASE, sh_mem_ape1_base);
 93	WREG32(mmSH_MEM_APE1_LIMIT, sh_mem_ape1_limit);
 94	WREG32(mmSH_MEM_BASES, sh_mem_bases);
 95
 96	unlock_srbm(kgd);
 97}
 98
 99static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
100					unsigned int vmid)
101{
102	struct amdgpu_device *adev = get_amdgpu_device(kgd);
103
104	/*
105	 * We have to assume that there is no outstanding mapping.
106	 * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0 because
107	 * a mapping is in progress or because a mapping finished
108	 * and the SW cleared it.
109	 * So the protocol is to always wait & clear.
110	 */
111	uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
112			ATC_VMID0_PASID_MAPPING__VALID_MASK;
113
114	WREG32(mmATC_VMID0_PASID_MAPPING + vmid, pasid_mapping);
115
116	while (!(RREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS) & (1U << vmid)))
117		cpu_relax();
118	WREG32(mmATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
119
120	/* Mapping vmid to pasid also for IH block */
121	WREG32(mmIH_VMID_0_LUT + vmid, pasid_mapping);
122
123	return 0;
124}
125
126static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
127{
128	struct amdgpu_device *adev = get_amdgpu_device(kgd);
129	uint32_t mec;
130	uint32_t pipe;
131
132	mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
133	pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
134
135	lock_srbm(kgd, mec, pipe, 0, 0);
136
137	WREG32(mmCPC_INT_CNTL, CP_INT_CNTL_RING0__TIME_STAMP_INT_ENABLE_MASK |
138			CP_INT_CNTL_RING0__OPCODE_ERROR_INT_ENABLE_MASK);
139
140	unlock_srbm(kgd);
141
142	return 0;
143}
144
145static inline uint32_t get_sdma_rlc_reg_offset(struct vi_sdma_mqd *m)
146{
147	uint32_t retval;
148
149	retval = m->sdma_engine_id * SDMA1_REGISTER_OFFSET +
150		m->sdma_queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
151
152	pr_debug("RLC register offset for SDMA%d RLC%d: 0x%x\n",
153			m->sdma_engine_id, m->sdma_queue_id, retval);
154
155	return retval;
156}
157
158static inline struct vi_mqd *get_mqd(void *mqd)
159{
160	return (struct vi_mqd *)mqd;
161}
162
163static inline struct vi_sdma_mqd *get_sdma_mqd(void *mqd)
164{
165	return (struct vi_sdma_mqd *)mqd;
166}
167
168static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
169			uint32_t queue_id, uint32_t __user *wptr,
170			uint32_t wptr_shift, uint32_t wptr_mask,
171			struct mm_struct *mm)
172{
173	struct amdgpu_device *adev = get_amdgpu_device(kgd);
174	struct vi_mqd *m;
175	uint32_t *mqd_hqd;
176	uint32_t reg, wptr_val, data;
177	bool valid_wptr = false;
178
179	m = get_mqd(mqd);
180
181	acquire_queue(kgd, pipe_id, queue_id);
182
183	/* HIQ is set during driver init period with vmid set to 0*/
184	if (m->cp_hqd_vmid == 0) {
185		uint32_t value, mec, pipe;
186
187		mec = (pipe_id / adev->gfx.mec.num_pipe_per_mec) + 1;
188		pipe = (pipe_id % adev->gfx.mec.num_pipe_per_mec);
189
190		pr_debug("kfd: set HIQ, mec:%d, pipe:%d, queue:%d.\n",
191			mec, pipe, queue_id);
192		value = RREG32(mmRLC_CP_SCHEDULERS);
193		value = REG_SET_FIELD(value, RLC_CP_SCHEDULERS, scheduler1,
194			((mec << 5) | (pipe << 3) | queue_id | 0x80));
195		WREG32(mmRLC_CP_SCHEDULERS, value);
196	}
197
198	/* HQD registers extend from CP_MQD_BASE_ADDR to CP_HQD_EOP_WPTR_MEM. */
199	mqd_hqd = &m->cp_mqd_base_addr_lo;
200
201	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_CONTROL; reg++)
202		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
203
204	/* Tonga errata: EOP RPTR/WPTR should be left unmodified.
205	 * This is safe since EOP RPTR==WPTR for any inactive HQD
206	 * on ASICs that do not support context-save.
207	 * EOP writes/reads can start anywhere in the ring.
208	 */
209	if (get_amdgpu_device(kgd)->asic_type != CHIP_TONGA) {
210		WREG32(mmCP_HQD_EOP_RPTR, m->cp_hqd_eop_rptr);
211		WREG32(mmCP_HQD_EOP_WPTR, m->cp_hqd_eop_wptr);
212		WREG32(mmCP_HQD_EOP_WPTR_MEM, m->cp_hqd_eop_wptr_mem);
213	}
214
215	for (reg = mmCP_HQD_EOP_EVENTS; reg <= mmCP_HQD_ERROR; reg++)
216		WREG32(reg, mqd_hqd[reg - mmCP_MQD_BASE_ADDR]);
217
218	/* Copy userspace write pointer value to register.
219	 * Activate doorbell logic to monitor subsequent changes.
220	 */
221	data = REG_SET_FIELD(m->cp_hqd_pq_doorbell_control,
222			     CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
223	WREG32(mmCP_HQD_PQ_DOORBELL_CONTROL, data);
224
225	/* read_user_ptr may take the mm->mmap_lock.
226	 * release srbm_mutex to avoid circular dependency between
227	 * srbm_mutex->mm_sem->reservation_ww_class_mutex->srbm_mutex.
228	 */
229	release_queue(kgd);
230	valid_wptr = read_user_wptr(mm, wptr, wptr_val);
231	acquire_queue(kgd, pipe_id, queue_id);
232	if (valid_wptr)
233		WREG32(mmCP_HQD_PQ_WPTR, (wptr_val << wptr_shift) & wptr_mask);
234
235	data = REG_SET_FIELD(m->cp_hqd_active, CP_HQD_ACTIVE, ACTIVE, 1);
236	WREG32(mmCP_HQD_ACTIVE, data);
237
238	release_queue(kgd);
239
240	return 0;
241}
242
243static int kgd_hqd_dump(struct kgd_dev *kgd,
244			uint32_t pipe_id, uint32_t queue_id,
245			uint32_t (**dump)[2], uint32_t *n_regs)
246{
247	struct amdgpu_device *adev = get_amdgpu_device(kgd);
248	uint32_t i = 0, reg;
249#define HQD_N_REGS (54+4)
250#define DUMP_REG(addr) do {				\
251		if (WARN_ON_ONCE(i >= HQD_N_REGS))	\
252			break;				\
253		(*dump)[i][0] = (addr) << 2;		\
254		(*dump)[i++][1] = RREG32(addr);		\
255	} while (0)
256
257	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
258	if (*dump == NULL)
259		return -ENOMEM;
260
261	acquire_queue(kgd, pipe_id, queue_id);
262
263	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE0);
264	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE1);
265	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE2);
266	DUMP_REG(mmCOMPUTE_STATIC_THREAD_MGMT_SE3);
267
268	for (reg = mmCP_MQD_BASE_ADDR; reg <= mmCP_HQD_EOP_DONES; reg++)
269		DUMP_REG(reg);
270
271	release_queue(kgd);
272
273	WARN_ON_ONCE(i != HQD_N_REGS);
274	*n_regs = i;
275
276	return 0;
277}
278
279static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd,
280			     uint32_t __user *wptr, struct mm_struct *mm)
281{
282	struct amdgpu_device *adev = get_amdgpu_device(kgd);
283	struct vi_sdma_mqd *m;
284	unsigned long end_jiffies;
285	uint32_t sdma_rlc_reg_offset;
286	uint32_t data;
287
288	m = get_sdma_mqd(mqd);
289	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
290	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
291		m->sdmax_rlcx_rb_cntl & (~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK));
292
293	end_jiffies = msecs_to_jiffies(2000) + jiffies;
294	while (true) {
295		data = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
296		if (data & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
297			break;
298		if (time_after(jiffies, end_jiffies)) {
299			pr_err("SDMA RLC not idle in %s\n", __func__);
300			return -ETIME;
301		}
302		usleep_range(500, 1000);
303	}
 
 
 
 
 
 
 
 
 
 
 
304
305	data = REG_SET_FIELD(m->sdmax_rlcx_doorbell, SDMA0_RLC0_DOORBELL,
306			     ENABLE, 1);
307	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, data);
308	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR,
309				m->sdmax_rlcx_rb_rptr);
310
311	if (read_user_wptr(mm, wptr, data))
312		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR, data);
313	else
314		WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_WPTR,
315		       m->sdmax_rlcx_rb_rptr);
316
317	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_VIRTUAL_ADDR,
318				m->sdmax_rlcx_virtual_addr);
319	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE, m->sdmax_rlcx_rb_base);
320	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_BASE_HI,
321			m->sdmax_rlcx_rb_base_hi);
322	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_LO,
323			m->sdmax_rlcx_rb_rptr_addr_lo);
324	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR_ADDR_HI,
325			m->sdmax_rlcx_rb_rptr_addr_hi);
326
327	data = REG_SET_FIELD(m->sdmax_rlcx_rb_cntl, SDMA0_RLC0_RB_CNTL,
328			     RB_ENABLE, 1);
329	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, data);
330
331	return 0;
332}
333
334static int kgd_hqd_sdma_dump(struct kgd_dev *kgd,
335			     uint32_t engine_id, uint32_t queue_id,
336			     uint32_t (**dump)[2], uint32_t *n_regs)
337{
338	struct amdgpu_device *adev = get_amdgpu_device(kgd);
339	uint32_t sdma_offset = engine_id * SDMA1_REGISTER_OFFSET +
340		queue_id * KFD_VI_SDMA_QUEUE_OFFSET;
341	uint32_t i = 0, reg;
342#undef HQD_N_REGS
343#define HQD_N_REGS (19+4+2+3+7)
344
345	*dump = kmalloc_array(HQD_N_REGS * 2, sizeof(uint32_t), GFP_KERNEL);
346	if (*dump == NULL)
347		return -ENOMEM;
348
349	for (reg = mmSDMA0_RLC0_RB_CNTL; reg <= mmSDMA0_RLC0_DOORBELL; reg++)
350		DUMP_REG(sdma_offset + reg);
351	for (reg = mmSDMA0_RLC0_VIRTUAL_ADDR; reg <= mmSDMA0_RLC0_WATERMARK;
352	     reg++)
353		DUMP_REG(sdma_offset + reg);
354	for (reg = mmSDMA0_RLC0_CSA_ADDR_LO; reg <= mmSDMA0_RLC0_CSA_ADDR_HI;
355	     reg++)
356		DUMP_REG(sdma_offset + reg);
357	for (reg = mmSDMA0_RLC0_IB_SUB_REMAIN; reg <= mmSDMA0_RLC0_DUMMY_REG;
358	     reg++)
359		DUMP_REG(sdma_offset + reg);
360	for (reg = mmSDMA0_RLC0_MIDCMD_DATA0; reg <= mmSDMA0_RLC0_MIDCMD_CNTL;
361	     reg++)
362		DUMP_REG(sdma_offset + reg);
363
364	WARN_ON_ONCE(i != HQD_N_REGS);
365	*n_regs = i;
366
367	return 0;
368}
369
370static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
371				uint32_t pipe_id, uint32_t queue_id)
372{
373	struct amdgpu_device *adev = get_amdgpu_device(kgd);
374	uint32_t act;
375	bool retval = false;
376	uint32_t low, high;
377
378	acquire_queue(kgd, pipe_id, queue_id);
379	act = RREG32(mmCP_HQD_ACTIVE);
380	if (act) {
381		low = lower_32_bits(queue_address >> 8);
382		high = upper_32_bits(queue_address >> 8);
383
384		if (low == RREG32(mmCP_HQD_PQ_BASE) &&
385				high == RREG32(mmCP_HQD_PQ_BASE_HI))
386			retval = true;
387	}
388	release_queue(kgd);
389	return retval;
390}
391
392static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
393{
394	struct amdgpu_device *adev = get_amdgpu_device(kgd);
395	struct vi_sdma_mqd *m;
396	uint32_t sdma_rlc_reg_offset;
397	uint32_t sdma_rlc_rb_cntl;
398
399	m = get_sdma_mqd(mqd);
400	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
401
402	sdma_rlc_rb_cntl = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
403
404	if (sdma_rlc_rb_cntl & SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK)
405		return true;
406
407	return false;
408}
409
410static int kgd_hqd_destroy(struct kgd_dev *kgd, void *mqd,
411				enum kfd_preempt_type reset_type,
412				unsigned int utimeout, uint32_t pipe_id,
413				uint32_t queue_id)
414{
415	struct amdgpu_device *adev = get_amdgpu_device(kgd);
416	uint32_t temp;
417	enum hqd_dequeue_request_type type;
418	unsigned long flags, end_jiffies;
419	int retry;
420	struct vi_mqd *m = get_mqd(mqd);
421
422	if (adev->in_gpu_reset)
423		return -EIO;
424
425	acquire_queue(kgd, pipe_id, queue_id);
426
427	if (m->cp_hqd_vmid == 0)
428		WREG32_FIELD(RLC_CP_SCHEDULERS, scheduler1, 0);
429
430	switch (reset_type) {
431	case KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN:
432		type = DRAIN_PIPE;
433		break;
434	case KFD_PREEMPT_TYPE_WAVEFRONT_RESET:
435		type = RESET_WAVES;
436		break;
437	default:
438		type = DRAIN_PIPE;
439		break;
440	}
441
442	/* Workaround: If IQ timer is active and the wait time is close to or
443	 * equal to 0, dequeueing is not safe. Wait until either the wait time
444	 * is larger or timer is cleared. Also, ensure that IQ_REQ_PEND is
445	 * cleared before continuing. Also, ensure wait times are set to at
446	 * least 0x3.
447	 */
448	local_irq_save(flags);
449	preempt_disable();
450	retry = 5000; /* wait for 500 usecs at maximum */
451	while (true) {
452		temp = RREG32(mmCP_HQD_IQ_TIMER);
453		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, PROCESSING_IQ)) {
454			pr_debug("HW is processing IQ\n");
455			goto loop;
456		}
457		if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, ACTIVE)) {
458			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, RETRY_TYPE)
459					== 3) /* SEM-rearm is safe */
460				break;
461			/* Wait time 3 is safe for CP, but our MMIO read/write
462			 * time is close to 1 microsecond, so check for 10 to
463			 * leave more buffer room
464			 */
465			if (REG_GET_FIELD(temp, CP_HQD_IQ_TIMER, WAIT_TIME)
466					>= 10)
467				break;
468			pr_debug("IQ timer is active\n");
469		} else
470			break;
471loop:
472		if (!retry) {
473			pr_err("CP HQD IQ timer status time out\n");
474			break;
475		}
476		ndelay(100);
477		--retry;
478	}
479	retry = 1000;
480	while (true) {
481		temp = RREG32(mmCP_HQD_DEQUEUE_REQUEST);
482		if (!(temp & CP_HQD_DEQUEUE_REQUEST__IQ_REQ_PEND_MASK))
483			break;
484		pr_debug("Dequeue request is pending\n");
485
486		if (!retry) {
487			pr_err("CP HQD dequeue request time out\n");
488			break;
489		}
490		ndelay(100);
491		--retry;
492	}
493	local_irq_restore(flags);
494	preempt_enable();
495
496	WREG32(mmCP_HQD_DEQUEUE_REQUEST, type);
497
498	end_jiffies = (utimeout * HZ / 1000) + jiffies;
499	while (true) {
500		temp = RREG32(mmCP_HQD_ACTIVE);
501		if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
502			break;
503		if (time_after(jiffies, end_jiffies)) {
504			pr_err("cp queue preemption time out.\n");
505			release_queue(kgd);
506			return -ETIME;
507		}
508		usleep_range(500, 1000);
509	}
510
511	release_queue(kgd);
512	return 0;
513}
514
515static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
516				unsigned int utimeout)
517{
518	struct amdgpu_device *adev = get_amdgpu_device(kgd);
519	struct vi_sdma_mqd *m;
520	uint32_t sdma_rlc_reg_offset;
521	uint32_t temp;
522	unsigned long end_jiffies = (utimeout * HZ / 1000) + jiffies;
523
524	m = get_sdma_mqd(mqd);
525	sdma_rlc_reg_offset = get_sdma_rlc_reg_offset(m);
526
527	temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL);
528	temp = temp & ~SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK;
529	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL, temp);
530
531	while (true) {
532		temp = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_CONTEXT_STATUS);
533		if (temp & SDMA0_RLC0_CONTEXT_STATUS__IDLE_MASK)
534			break;
535		if (time_after(jiffies, end_jiffies)) {
536			pr_err("SDMA RLC not idle in %s\n", __func__);
537			return -ETIME;
538		}
539		usleep_range(500, 1000);
540	}
541
542	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_DOORBELL, 0);
543	WREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL,
544		RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_CNTL) |
545		SDMA0_RLC0_RB_CNTL__RB_ENABLE_MASK);
546
547	m->sdmax_rlcx_rb_rptr = RREG32(sdma_rlc_reg_offset + mmSDMA0_RLC0_RB_RPTR);
548
549	return 0;
550}
551
552static bool get_atc_vmid_pasid_mapping_info(struct kgd_dev *kgd,
553					uint8_t vmid, uint16_t *p_pasid)
554{
555	uint32_t value;
556	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
557
558	value = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
559	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
 
 
 
 
 
 
 
560
561	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
 
562}
563
564static int kgd_address_watch_disable(struct kgd_dev *kgd)
565{
566	return 0;
567}
568
569static int kgd_address_watch_execute(struct kgd_dev *kgd,
570					unsigned int watch_point_id,
571					uint32_t cntl_val,
572					uint32_t addr_hi,
573					uint32_t addr_lo)
574{
575	return 0;
576}
577
578static int kgd_wave_control_execute(struct kgd_dev *kgd,
579					uint32_t gfx_index_val,
580					uint32_t sq_cmd)
581{
582	struct amdgpu_device *adev = get_amdgpu_device(kgd);
583	uint32_t data = 0;
584
585	mutex_lock(&adev->grbm_idx_mutex);
586
587	WREG32(mmGRBM_GFX_INDEX, gfx_index_val);
588	WREG32(mmSQ_CMD, sq_cmd);
589
590	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
591		INSTANCE_BROADCAST_WRITES, 1);
592	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
593		SH_BROADCAST_WRITES, 1);
594	data = REG_SET_FIELD(data, GRBM_GFX_INDEX,
595		SE_BROADCAST_WRITES, 1);
596
597	WREG32(mmGRBM_GFX_INDEX, data);
598	mutex_unlock(&adev->grbm_idx_mutex);
599
600	return 0;
601}
602
603static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
604					unsigned int watch_point_id,
605					unsigned int reg_offset)
606{
607	return 0;
608}
609
610static void set_scratch_backing_va(struct kgd_dev *kgd,
611					uint64_t va, uint32_t vmid)
612{
613	struct amdgpu_device *adev = (struct amdgpu_device *) kgd;
614
615	lock_srbm(kgd, 0, 0, 0, vmid);
616	WREG32(mmSH_HIDDEN_PRIVATE_BASE_VMID, va);
617	unlock_srbm(kgd);
618}
619
620static void set_vm_context_page_table_base(struct kgd_dev *kgd, uint32_t vmid,
621		uint64_t page_table_base)
622{
623	struct amdgpu_device *adev = get_amdgpu_device(kgd);
624
625	if (!amdgpu_amdkfd_is_kfd_vmid(adev, vmid)) {
626		pr_err("trying to set page table base for wrong VMID\n");
627		return;
628	}
629	WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8,
630			lower_32_bits(page_table_base));
631}
632
633const struct kfd2kgd_calls gfx_v8_kfd2kgd = {
634	.program_sh_mem_settings = kgd_program_sh_mem_settings,
635	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
636	.init_interrupts = kgd_init_interrupts,
637	.hqd_load = kgd_hqd_load,
638	.hqd_sdma_load = kgd_hqd_sdma_load,
639	.hqd_dump = kgd_hqd_dump,
640	.hqd_sdma_dump = kgd_hqd_sdma_dump,
641	.hqd_is_occupied = kgd_hqd_is_occupied,
642	.hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
643	.hqd_destroy = kgd_hqd_destroy,
644	.hqd_sdma_destroy = kgd_hqd_sdma_destroy,
645	.address_watch_disable = kgd_address_watch_disable,
646	.address_watch_execute = kgd_address_watch_execute,
647	.wave_control_execute = kgd_wave_control_execute,
648	.address_watch_get_offset = kgd_address_watch_get_offset,
649	.get_atc_vmid_pasid_mapping_info =
650			get_atc_vmid_pasid_mapping_info,
651	.set_scratch_backing_va = set_scratch_backing_va,
652	.set_vm_context_page_table_base = set_vm_context_page_table_base,
653};