Commit | Line | Data |
---|---|---|
d87f36a0 | 1 | // SPDX-License-Identifier: GPL-2.0 OR MIT |
64c7f8cf | 2 | /* |
d87f36a0 | 3 | * Copyright 2014-2022 Advanced Micro Devices, Inc. |
64c7f8cf BG |
4 | * |
5 | * Permission is hereby granted, free of charge, to any person obtaining a | |
6 | * copy of this software and associated documentation files (the "Software"), | |
7 | * to deal in the Software without restriction, including without limitation | |
8 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, | |
9 | * and/or sell copies of the Software, and to permit persons to whom the | |
10 | * Software is furnished to do so, subject to the following conditions: | |
11 | * | |
12 | * The above copyright notice and this permission notice shall be included in | |
13 | * all copies or substantial portions of the Software. | |
14 | * | |
15 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
16 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
17 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
18 | * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR | |
19 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, | |
20 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR | |
21 | * OTHER DEALINGS IN THE SOFTWARE. | |
22 | * | |
23 | */ | |
24 | ||
26103436 FK |
25 | #include <linux/ratelimit.h> |
26 | #include <linux/printk.h> | |
64c7f8cf BG |
27 | #include <linux/slab.h> |
28 | #include <linux/list.h> | |
29 | #include <linux/types.h> | |
64c7f8cf | 30 | #include <linux/bitops.h> |
99331a51 | 31 | #include <linux/sched.h> |
64c7f8cf BG |
32 | #include "kfd_priv.h" |
33 | #include "kfd_device_queue_manager.h" | |
34 | #include "kfd_mqd_manager.h" | |
35 | #include "cik_regs.h" | |
36 | #include "kfd_kernel_queue.h" | |
5b87245f | 37 | #include "amdgpu_amdkfd.h" |
cc009e61 | 38 | #include "mes_api_def.h" |
64c7f8cf BG |
39 | |
40 | /* Size of the per-pipe EOP queue */ | |
41 | #define CIK_HPD_EOP_BYTES_LOG2 11 | |
42 | #define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2) | |
43 | ||
64c7f8cf | 44 | static int set_pasid_vmid_mapping(struct device_queue_manager *dqm, |
c7b6bac9 | 45 | u32 pasid, unsigned int vmid); |
64c7f8cf | 46 | |
c4744e24 YZ |
47 | static int execute_queues_cpsch(struct device_queue_manager *dqm, |
48 | enum kfd_unmap_queues_filter filter, | |
7cee6a68 JK |
49 | uint32_t filter_param, |
50 | uint32_t grace_period); | |
7da2bcf8 | 51 | static int unmap_queues_cpsch(struct device_queue_manager *dqm, |
4465f466 | 52 | enum kfd_unmap_queues_filter filter, |
7cee6a68 JK |
53 | uint32_t filter_param, |
54 | uint32_t grace_period, | |
55 | bool reset); | |
64c7f8cf | 56 | |
60a00956 FK |
57 | static int map_queues_cpsch(struct device_queue_manager *dqm); |
58 | ||
bcea3081 | 59 | static void deallocate_sdma_queue(struct device_queue_manager *dqm, |
1b4670f6 | 60 | struct queue *q); |
64c7f8cf | 61 | |
d39b7737 OZ |
62 | static inline void deallocate_hqd(struct device_queue_manager *dqm, |
63 | struct queue *q); | |
64 | static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q); | |
65 | static int allocate_sdma_queue(struct device_queue_manager *dqm, | |
2485c12c | 66 | struct queue *q, const uint32_t *restore_sdma_id); |
73ea648d SL |
67 | static void kfd_process_hw_exception(struct work_struct *work); |
68 | ||
bcea3081 BG |
69 | static inline |
70 | enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type) | |
64c7f8cf | 71 | { |
1b4670f6 | 72 | if (type == KFD_QUEUE_TYPE_SDMA || type == KFD_QUEUE_TYPE_SDMA_XGMI) |
85d258f9 BG |
73 | return KFD_MQD_TYPE_SDMA; |
74 | return KFD_MQD_TYPE_CP; | |
64c7f8cf BG |
75 | } |
76 | ||
d0b63bb3 AR |
77 | static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe) |
78 | { | |
79 | int i; | |
8dc1db31 MJ |
80 | int pipe_offset = (mec * dqm->dev->kfd->shared_resources.num_pipe_per_mec |
81 | + pipe) * dqm->dev->kfd->shared_resources.num_queue_per_pipe; | |
d0b63bb3 AR |
82 | |
83 | /* queue is available for KFD usage if bit is 1 */ | |
8dc1db31 | 84 | for (i = 0; i < dqm->dev->kfd->shared_resources.num_queue_per_pipe; ++i) |
d0b63bb3 | 85 | if (test_bit(pipe_offset + i, |
8dc1db31 | 86 | dqm->dev->kfd->shared_resources.cp_queue_bitmap)) |
d0b63bb3 AR |
87 | return true; |
88 | return false; | |
89 | } | |
90 | ||
e6945304 | 91 | unsigned int get_cp_queues_num(struct device_queue_manager *dqm) |
64ea8f4a | 92 | { |
8dc1db31 | 93 | return bitmap_weight(dqm->dev->kfd->shared_resources.cp_queue_bitmap, |
d0b63bb3 | 94 | KGD_MAX_QUEUES); |
64ea8f4a OG |
95 | } |
96 | ||
d0b63bb3 | 97 | unsigned int get_queues_per_pipe(struct device_queue_manager *dqm) |
64c7f8cf | 98 | { |
8dc1db31 | 99 | return dqm->dev->kfd->shared_resources.num_queue_per_pipe; |
d0b63bb3 AR |
100 | } |
101 | ||
102 | unsigned int get_pipes_per_mec(struct device_queue_manager *dqm) | |
103 | { | |
8dc1db31 | 104 | return dqm->dev->kfd->shared_resources.num_pipe_per_mec; |
64c7f8cf BG |
105 | } |
106 | ||
c7637c95 YZ |
107 | static unsigned int get_num_all_sdma_engines(struct device_queue_manager *dqm) |
108 | { | |
ee2f17f4 AL |
109 | return kfd_get_num_sdma_engines(dqm->dev) + |
110 | kfd_get_num_xgmi_sdma_engines(dqm->dev); | |
c7637c95 YZ |
111 | } |
112 | ||
98bb9222 YZ |
113 | unsigned int get_num_sdma_queues(struct device_queue_manager *dqm) |
114 | { | |
ee2f17f4 | 115 | return kfd_get_num_sdma_engines(dqm->dev) * |
8dc1db31 | 116 | dqm->dev->kfd->device_info.num_sdma_queues_per_engine; |
98bb9222 YZ |
117 | } |
118 | ||
1b4670f6 OZ |
119 | unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm) |
120 | { | |
ee2f17f4 | 121 | return kfd_get_num_xgmi_sdma_engines(dqm->dev) * |
8dc1db31 | 122 | dqm->dev->kfd->device_info.num_sdma_queues_per_engine; |
1b4670f6 OZ |
123 | } |
124 | ||
cc009e61 MJ |
125 | static inline uint64_t get_reserved_sdma_queues_bitmap(struct device_queue_manager *dqm) |
126 | { | |
8dc1db31 | 127 | return dqm->dev->kfd->device_info.reserved_sdma_queues_bitmap; |
cc009e61 MJ |
128 | } |
129 | ||
a805889a MJ |
130 | static void init_sdma_bitmaps(struct device_queue_manager *dqm) |
131 | { | |
132 | bitmap_zero(dqm->sdma_bitmap, KFD_MAX_SDMA_QUEUES); | |
133 | bitmap_set(dqm->sdma_bitmap, 0, get_num_sdma_queues(dqm)); | |
134 | ||
135 | bitmap_zero(dqm->xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES); | |
136 | bitmap_set(dqm->xgmi_sdma_bitmap, 0, get_num_xgmi_sdma_queues(dqm)); | |
137 | } | |
138 | ||
a22fc854 | 139 | void program_sh_mem_settings(struct device_queue_manager *dqm, |
64c7f8cf BG |
140 | struct qcm_process_device *qpd) |
141 | { | |
c4050ff1 LL |
142 | uint32_t xcc_mask = dqm->dev->xcc_mask; |
143 | int xcc_id; | |
e2069a7b | 144 | |
c4050ff1 | 145 | for_each_inst(xcc_id, xcc_mask) |
e2069a7b | 146 | dqm->dev->kfd2kgd->program_sh_mem_settings( |
c4050ff1 LL |
147 | dqm->dev->adev, qpd->vmid, qpd->sh_mem_config, |
148 | qpd->sh_mem_ape1_base, qpd->sh_mem_ape1_limit, | |
149 | qpd->sh_mem_bases, xcc_id); | |
64c7f8cf BG |
150 | } |
151 | ||
cc009e61 MJ |
152 | static void kfd_hws_hang(struct device_queue_manager *dqm) |
153 | { | |
154 | /* | |
155 | * Issue a GPU reset if HWS is unresponsive | |
156 | */ | |
157 | dqm->is_hws_hang = true; | |
158 | ||
159 | /* It's possible we're detecting a HWS hang in the | |
160 | * middle of a GPU reset. No need to schedule another | |
161 | * reset in this case. | |
162 | */ | |
163 | if (!dqm->is_resetting) | |
164 | schedule_work(&dqm->hw_exception_work); | |
165 | } | |
166 | ||
167 | static int convert_to_mes_queue_type(int queue_type) | |
168 | { | |
169 | int mes_queue_type; | |
170 | ||
171 | switch (queue_type) { | |
172 | case KFD_QUEUE_TYPE_COMPUTE: | |
173 | mes_queue_type = MES_QUEUE_TYPE_COMPUTE; | |
174 | break; | |
175 | case KFD_QUEUE_TYPE_SDMA: | |
176 | mes_queue_type = MES_QUEUE_TYPE_SDMA; | |
177 | break; | |
178 | default: | |
179 | WARN(1, "Invalid queue type %d", queue_type); | |
180 | mes_queue_type = -EINVAL; | |
181 | break; | |
182 | } | |
183 | ||
184 | return mes_queue_type; | |
185 | } | |
186 | ||
187 | static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q, | |
188 | struct qcm_process_device *qpd) | |
189 | { | |
190 | struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev; | |
191 | struct kfd_process_device *pdd = qpd_to_pdd(qpd); | |
192 | struct mes_add_queue_input queue_input; | |
04fd0739 | 193 | int r, queue_type; |
e77a541f | 194 | uint64_t wptr_addr_off; |
cc009e61 MJ |
195 | |
196 | if (dqm->is_hws_hang) | |
197 | return -EIO; | |
198 | ||
199 | memset(&queue_input, 0x0, sizeof(struct mes_add_queue_input)); | |
200 | queue_input.process_id = qpd->pqm->process->pasid; | |
201 | queue_input.page_table_base_addr = qpd->page_table_base; | |
202 | queue_input.process_va_start = 0; | |
203 | queue_input.process_va_end = adev->vm_manager.max_pfn - 1; | |
204 | /* MES unit for quantum is 100ns */ | |
205 | queue_input.process_quantum = KFD_MES_PROCESS_QUANTUM; /* Equivalent to 10ms. */ | |
206 | queue_input.process_context_addr = pdd->proc_ctx_gpu_addr; | |
207 | queue_input.gang_quantum = KFD_MES_GANG_QUANTUM; /* Equivalent to 1ms */ | |
208 | queue_input.gang_context_addr = q->gang_ctx_gpu_addr; | |
209 | queue_input.inprocess_gang_priority = q->properties.priority; | |
210 | queue_input.gang_global_priority_level = | |
211 | AMDGPU_MES_PRIORITY_LEVEL_NORMAL; | |
212 | queue_input.doorbell_offset = q->properties.doorbell_off; | |
213 | queue_input.mqd_addr = q->gart_mqd_addr; | |
fe4e9ff9 | 214 | queue_input.wptr_addr = (uint64_t)q->properties.write_ptr; |
e77a541f GS |
215 | |
216 | if (q->wptr_bo) { | |
7f347e3f | 217 | wptr_addr_off = (uint64_t)q->properties.write_ptr & (PAGE_SIZE - 1); |
fe4e9ff9 JX |
218 | queue_input.wptr_mc_addr = ((uint64_t)q->wptr_bo->tbo.resource->start << PAGE_SHIFT) + wptr_addr_off; |
219 | } | |
e77a541f | 220 | |
a9579956 | 221 | queue_input.is_kfd_process = 1; |
3e9cf234 GS |
222 | queue_input.is_aql_queue = (q->properties.format == KFD_QUEUE_FORMAT_AQL); |
223 | queue_input.queue_size = q->properties.queue_size >> 2; | |
a9579956 | 224 | |
cc009e61 MJ |
225 | queue_input.paging = false; |
226 | queue_input.tba_addr = qpd->tba_addr; | |
227 | queue_input.tma_addr = qpd->tma_addr; | |
228 | ||
04fd0739 GS |
229 | queue_type = convert_to_mes_queue_type(q->properties.type); |
230 | if (queue_type < 0) { | |
cc009e61 MJ |
231 | pr_err("Queue type not supported with MES, queue:%d\n", |
232 | q->properties.type); | |
233 | return -EINVAL; | |
234 | } | |
04fd0739 | 235 | queue_input.queue_type = (uint32_t)queue_type; |
cc009e61 MJ |
236 | |
237 | if (q->gws) { | |
238 | queue_input.gws_base = 0; | |
239 | queue_input.gws_size = qpd->num_gws; | |
240 | } | |
241 | ||
242 | amdgpu_mes_lock(&adev->mes); | |
243 | r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input); | |
244 | amdgpu_mes_unlock(&adev->mes); | |
245 | if (r) { | |
246 | pr_err("failed to add hardware queue to MES, doorbell=0x%x\n", | |
247 | q->properties.doorbell_off); | |
248 | pr_err("MES might be in unrecoverable state, issue a GPU reset\n"); | |
249 | kfd_hws_hang(dqm); | |
250 | } | |
251 | ||
252 | return r; | |
253 | } | |
254 | ||
255 | static int remove_queue_mes(struct device_queue_manager *dqm, struct queue *q, | |
256 | struct qcm_process_device *qpd) | |
257 | { | |
258 | struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev; | |
259 | int r; | |
260 | struct mes_remove_queue_input queue_input; | |
261 | ||
262 | if (dqm->is_hws_hang) | |
263 | return -EIO; | |
264 | ||
265 | memset(&queue_input, 0x0, sizeof(struct mes_remove_queue_input)); | |
266 | queue_input.doorbell_offset = q->properties.doorbell_off; | |
267 | queue_input.gang_context_addr = q->gang_ctx_gpu_addr; | |
268 | ||
269 | amdgpu_mes_lock(&adev->mes); | |
270 | r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input); | |
271 | amdgpu_mes_unlock(&adev->mes); | |
272 | ||
273 | if (r) { | |
274 | pr_err("failed to remove hardware queue from MES, doorbell=0x%x\n", | |
275 | q->properties.doorbell_off); | |
276 | pr_err("MES might be in unrecoverable state, issue a GPU reset\n"); | |
277 | kfd_hws_hang(dqm); | |
278 | } | |
279 | ||
280 | return r; | |
281 | } | |
282 | ||
283 | static int remove_all_queues_mes(struct device_queue_manager *dqm) | |
284 | { | |
285 | struct device_process_node *cur; | |
286 | struct qcm_process_device *qpd; | |
287 | struct queue *q; | |
288 | int retval = 0; | |
289 | ||
290 | list_for_each_entry(cur, &dqm->queues, list) { | |
291 | qpd = cur->qpd; | |
292 | list_for_each_entry(q, &qpd->queues_list, list) { | |
293 | if (q->properties.is_active) { | |
294 | retval = remove_queue_mes(dqm, q, qpd); | |
295 | if (retval) { | |
296 | pr_err("%s: Failed to remove queue %d for dev %d", | |
297 | __func__, | |
298 | q->properties.queue_id, | |
299 | dqm->dev->id); | |
300 | return retval; | |
301 | } | |
302 | } | |
303 | } | |
304 | } | |
305 | ||
306 | return retval; | |
307 | } | |
308 | ||
204d8998 | 309 | static void increment_queue_count(struct device_queue_manager *dqm, |
ab4d51d4 DYS |
310 | struct qcm_process_device *qpd, |
311 | struct queue *q) | |
b42902f4 YZ |
312 | { |
313 | dqm->active_queue_count++; | |
ab4d51d4 DYS |
314 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || |
315 | q->properties.type == KFD_QUEUE_TYPE_DIQ) | |
b42902f4 | 316 | dqm->active_cp_queue_count++; |
ab4d51d4 DYS |
317 | |
318 | if (q->properties.is_gws) { | |
319 | dqm->gws_queue_count++; | |
320 | qpd->mapped_gws_queue = true; | |
321 | } | |
b42902f4 YZ |
322 | } |
323 | ||
204d8998 | 324 | static void decrement_queue_count(struct device_queue_manager *dqm, |
ab4d51d4 DYS |
325 | struct qcm_process_device *qpd, |
326 | struct queue *q) | |
b42902f4 YZ |
327 | { |
328 | dqm->active_queue_count--; | |
ab4d51d4 DYS |
329 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || |
330 | q->properties.type == KFD_QUEUE_TYPE_DIQ) | |
b42902f4 | 331 | dqm->active_cp_queue_count--; |
ab4d51d4 DYS |
332 | |
333 | if (q->properties.is_gws) { | |
334 | dqm->gws_queue_count--; | |
335 | qpd->mapped_gws_queue = false; | |
336 | } | |
b42902f4 YZ |
337 | } |
338 | ||
5bb6a8fa DYS |
339 | /* |
340 | * Allocate a doorbell ID to this queue. | |
341 | * If doorbell_id is passed in, make sure requested ID is valid then allocate it. | |
342 | */ | |
343 | static int allocate_doorbell(struct qcm_process_device *qpd, | |
344 | struct queue *q, | |
345 | uint32_t const *restore_id) | |
ef568db7 | 346 | { |
8dc1db31 | 347 | struct kfd_node *dev = qpd->dqm->dev; |
ef568db7 | 348 | |
dd0ae064 | 349 | if (!KFD_IS_SOC15(dev)) { |
ef568db7 FK |
350 | /* On pre-SOC15 chips we need to use the queue ID to |
351 | * preserve the user mode ABI. | |
352 | */ | |
5bb6a8fa DYS |
353 | |
354 | if (restore_id && *restore_id != q->properties.queue_id) | |
355 | return -EINVAL; | |
356 | ||
ef568db7 | 357 | q->doorbell_id = q->properties.queue_id; |
1b4670f6 OZ |
358 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
359 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
234441dd YZ |
360 | /* For SDMA queues on SOC15 with 8-byte doorbell, use static |
361 | * doorbell assignments based on the engine and queue id. | |
362 | * The doobell index distance between RLC (2*i) and (2*i+1) | |
363 | * for a SDMA engine is 512. | |
ef568db7 | 364 | */ |
234441dd | 365 | |
8dc1db31 | 366 | uint32_t *idx_offset = dev->kfd->shared_resources.sdma_doorbell_idx; |
643e40d4 MJ |
367 | |
368 | /* | |
369 | * q->properties.sdma_engine_id corresponds to the virtual | |
370 | * sdma engine number. However, for doorbell allocation, | |
371 | * we need the physical sdma engine id in order to get the | |
372 | * correct doorbell offset. | |
373 | */ | |
374 | uint32_t valid_id = idx_offset[qpd->dqm->dev->node_id * | |
375 | get_num_all_sdma_engines(qpd->dqm) + | |
376 | q->properties.sdma_engine_id] | |
5bb6a8fa DYS |
377 | + (q->properties.sdma_queue_id & 1) |
378 | * KFD_QUEUE_DOORBELL_MIRROR_OFFSET | |
379 | + (q->properties.sdma_queue_id >> 1); | |
380 | ||
381 | if (restore_id && *restore_id != valid_id) | |
382 | return -EINVAL; | |
383 | q->doorbell_id = valid_id; | |
ef568db7 | 384 | } else { |
5bb6a8fa DYS |
385 | /* For CP queues on SOC15 */ |
386 | if (restore_id) { | |
387 | /* make sure that ID is free */ | |
388 | if (__test_and_set_bit(*restore_id, qpd->doorbell_bitmap)) | |
389 | return -EINVAL; | |
390 | ||
391 | q->doorbell_id = *restore_id; | |
392 | } else { | |
393 | /* or reserve a free doorbell ID */ | |
394 | unsigned int found; | |
395 | ||
396 | found = find_first_zero_bit(qpd->doorbell_bitmap, | |
397 | KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); | |
398 | if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) { | |
399 | pr_debug("No doorbells available"); | |
400 | return -EBUSY; | |
401 | } | |
402 | set_bit(found, qpd->doorbell_bitmap); | |
403 | q->doorbell_id = found; | |
ef568db7 | 404 | } |
ef568db7 FK |
405 | } |
406 | ||
407 | q->properties.doorbell_off = | |
8dc1db31 | 408 | kfd_get_doorbell_dw_offset_in_bar(dev->kfd, qpd_to_pdd(qpd), |
ef568db7 | 409 | q->doorbell_id); |
ef568db7 FK |
410 | return 0; |
411 | } | |
412 | ||
413 | static void deallocate_doorbell(struct qcm_process_device *qpd, | |
414 | struct queue *q) | |
415 | { | |
416 | unsigned int old; | |
8dc1db31 | 417 | struct kfd_node *dev = qpd->dqm->dev; |
ef568db7 | 418 | |
dd0ae064 | 419 | if (!KFD_IS_SOC15(dev) || |
1b4670f6 OZ |
420 | q->properties.type == KFD_QUEUE_TYPE_SDMA || |
421 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) | |
ef568db7 FK |
422 | return; |
423 | ||
424 | old = test_and_clear_bit(q->doorbell_id, qpd->doorbell_bitmap); | |
425 | WARN_ON(!old); | |
426 | } | |
427 | ||
b53ef0df MJ |
428 | static void program_trap_handler_settings(struct device_queue_manager *dqm, |
429 | struct qcm_process_device *qpd) | |
430 | { | |
c4050ff1 LL |
431 | uint32_t xcc_mask = dqm->dev->xcc_mask; |
432 | int xcc_id; | |
e2069a7b | 433 | |
b53ef0df | 434 | if (dqm->dev->kfd2kgd->program_trap_handler_settings) |
c4050ff1 | 435 | for_each_inst(xcc_id, xcc_mask) |
e2069a7b | 436 | dqm->dev->kfd2kgd->program_trap_handler_settings( |
c4050ff1 LL |
437 | dqm->dev->adev, qpd->vmid, qpd->tba_addr, |
438 | qpd->tma_addr, xcc_id); | |
b53ef0df MJ |
439 | } |
440 | ||
64c7f8cf BG |
441 | static int allocate_vmid(struct device_queue_manager *dqm, |
442 | struct qcm_process_device *qpd, | |
443 | struct queue *q) | |
444 | { | |
d9d4623c | 445 | int allocated_vmid = -1, i; |
64c7f8cf | 446 | |
d9d4623c YZ |
447 | for (i = dqm->dev->vm_info.first_vmid_kfd; |
448 | i <= dqm->dev->vm_info.last_vmid_kfd; i++) { | |
449 | if (!dqm->vmid_pasid[i]) { | |
450 | allocated_vmid = i; | |
451 | break; | |
452 | } | |
453 | } | |
454 | ||
455 | if (allocated_vmid < 0) { | |
456 | pr_err("no more vmid to allocate\n"); | |
457 | return -ENOSPC; | |
458 | } | |
459 | ||
460 | pr_debug("vmid allocated: %d\n", allocated_vmid); | |
461 | ||
462 | dqm->vmid_pasid[allocated_vmid] = q->process->pasid; | |
64c7f8cf | 463 | |
d9d4623c | 464 | set_pasid_vmid_mapping(dqm, q->process->pasid, allocated_vmid); |
64c7f8cf | 465 | |
64c7f8cf BG |
466 | qpd->vmid = allocated_vmid; |
467 | q->properties.vmid = allocated_vmid; | |
468 | ||
64c7f8cf BG |
469 | program_sh_mem_settings(dqm, qpd); |
470 | ||
8dc1db31 | 471 | if (KFD_IS_SOC15(dqm->dev) && dqm->dev->kfd->cwsr_enabled) |
b53ef0df MJ |
472 | program_trap_handler_settings(dqm, qpd); |
473 | ||
403575c4 FK |
474 | /* qpd->page_table_base is set earlier when register_process() |
475 | * is called, i.e. when the first queue is created. | |
476 | */ | |
3356c38d | 477 | dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->adev, |
403575c4 FK |
478 | qpd->vmid, |
479 | qpd->page_table_base); | |
480 | /* invalidate the VM context after pasid and vmid mapping is set up */ | |
3543b055 | 481 | kfd_flush_tlb(qpd_to_pdd(qpd), TLB_FLUSH_LEGACY); |
403575c4 | 482 | |
c637b36a | 483 | if (dqm->dev->kfd2kgd->set_scratch_backing_va) |
3356c38d | 484 | dqm->dev->kfd2kgd->set_scratch_backing_va(dqm->dev->adev, |
c637b36a | 485 | qpd->sh_hidden_private_base, qpd->vmid); |
d39b7737 | 486 | |
64c7f8cf BG |
487 | return 0; |
488 | } | |
489 | ||
8dc1db31 | 490 | static int flush_texture_cache_nocpsch(struct kfd_node *kdev, |
552764b6 FK |
491 | struct qcm_process_device *qpd) |
492 | { | |
9af5379c | 493 | const struct packet_manager_funcs *pmf = qpd->dqm->packet_mgr.pmf; |
f6e27ff1 | 494 | int ret; |
552764b6 FK |
495 | |
496 | if (!qpd->ib_kaddr) | |
497 | return -ENOMEM; | |
498 | ||
f6e27ff1 FK |
499 | ret = pmf->release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr); |
500 | if (ret) | |
501 | return ret; | |
552764b6 | 502 | |
6bfc7c7e | 503 | return amdgpu_amdkfd_submit_ib(kdev->adev, KGD_ENGINE_MEC1, qpd->vmid, |
f6e27ff1 FK |
504 | qpd->ib_base, (uint32_t *)qpd->ib_kaddr, |
505 | pmf->release_mem_size / sizeof(uint32_t)); | |
552764b6 FK |
506 | } |
507 | ||
64c7f8cf BG |
508 | static void deallocate_vmid(struct device_queue_manager *dqm, |
509 | struct qcm_process_device *qpd, | |
510 | struct queue *q) | |
511 | { | |
552764b6 | 512 | /* On GFX v7, CP doesn't flush TC at dequeue */ |
7eb0502a | 513 | if (q->device->adev->asic_type == CHIP_HAWAII) |
552764b6 FK |
514 | if (flush_texture_cache_nocpsch(q->device, qpd)) |
515 | pr_err("Failed to flush TC\n"); | |
516 | ||
3543b055 | 517 | kfd_flush_tlb(qpd_to_pdd(qpd), TLB_FLUSH_LEGACY); |
403575c4 | 518 | |
2030664b BG |
519 | /* Release the vmid mapping */ |
520 | set_pasid_vmid_mapping(dqm, 0, qpd->vmid); | |
d9d4623c | 521 | dqm->vmid_pasid[qpd->vmid] = 0; |
2030664b | 522 | |
64c7f8cf BG |
523 | qpd->vmid = 0; |
524 | q->properties.vmid = 0; | |
525 | } | |
526 | ||
527 | static int create_queue_nocpsch(struct device_queue_manager *dqm, | |
528 | struct queue *q, | |
2485c12c | 529 | struct qcm_process_device *qpd, |
42c6c482 | 530 | const struct kfd_criu_queue_priv_data *qd, |
3a9822d7 | 531 | const void *restore_mqd, const void *restore_ctl_stack) |
64c7f8cf | 532 | { |
d39b7737 | 533 | struct mqd_manager *mqd_mgr; |
64c7f8cf BG |
534 | int retval; |
535 | ||
efeaed4d | 536 | dqm_lock(dqm); |
64c7f8cf | 537 | |
b8cbab04 | 538 | if (dqm->total_queue_count >= max_num_of_queues_per_device) { |
79775b62 | 539 | pr_warn("Can't create new usermode queue because %d queues were already created\n", |
b8cbab04 | 540 | dqm->total_queue_count); |
ab7c1648 KR |
541 | retval = -EPERM; |
542 | goto out_unlock; | |
b8cbab04 OG |
543 | } |
544 | ||
64c7f8cf BG |
545 | if (list_empty(&qpd->queues_list)) { |
546 | retval = allocate_vmid(dqm, qpd, q); | |
ab7c1648 KR |
547 | if (retval) |
548 | goto out_unlock; | |
64c7f8cf | 549 | } |
64c7f8cf | 550 | q->properties.vmid = qpd->vmid; |
26103436 | 551 | /* |
bb2d2128 FK |
552 | * Eviction state logic: mark all queues as evicted, even ones |
553 | * not currently active. Restoring inactive queues later only | |
554 | * updates the is_evicted flag but is a no-op otherwise. | |
26103436 | 555 | */ |
bb2d2128 | 556 | q->properties.is_evicted = !!qpd->evicted; |
64c7f8cf | 557 | |
373d7080 FK |
558 | q->properties.tba_addr = qpd->tba_addr; |
559 | q->properties.tma_addr = qpd->tma_addr; | |
560 | ||
d091bc0a OZ |
561 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
562 | q->properties.type)]; | |
d39b7737 OZ |
563 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) { |
564 | retval = allocate_hqd(dqm, q); | |
565 | if (retval) | |
566 | goto deallocate_vmid; | |
567 | pr_debug("Loading mqd to hqd on pipe %d, queue %d\n", | |
568 | q->pipe, q->queue); | |
569 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || | |
570 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
2485c12c | 571 | retval = allocate_sdma_queue(dqm, q, qd ? &qd->sdma_id : NULL); |
d39b7737 OZ |
572 | if (retval) |
573 | goto deallocate_vmid; | |
574 | dqm->asic_ops.init_sdma_vm(dqm, q, qpd); | |
575 | } | |
576 | ||
5bb6a8fa | 577 | retval = allocate_doorbell(qpd, q, qd ? &qd->doorbell_id : NULL); |
d39b7737 OZ |
578 | if (retval) |
579 | goto out_deallocate_hqd; | |
580 | ||
6a6ef5ee OZ |
581 | /* Temporarily release dqm lock to avoid a circular lock dependency */ |
582 | dqm_unlock(dqm); | |
d091bc0a | 583 | q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties); |
6a6ef5ee OZ |
584 | dqm_lock(dqm); |
585 | ||
d091bc0a OZ |
586 | if (!q->mqd_mem_obj) { |
587 | retval = -ENOMEM; | |
588 | goto out_deallocate_doorbell; | |
589 | } | |
42c6c482 DYS |
590 | |
591 | if (qd) | |
592 | mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr, | |
3a9822d7 DYS |
593 | &q->properties, restore_mqd, restore_ctl_stack, |
594 | qd->ctl_stack_size); | |
42c6c482 DYS |
595 | else |
596 | mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, | |
597 | &q->gart_mqd_addr, &q->properties); | |
598 | ||
d39b7737 | 599 | if (q->properties.is_active) { |
2c99a547 PY |
600 | if (!dqm->sched_running) { |
601 | WARN_ONCE(1, "Load non-HWS mqd while stopped\n"); | |
602 | goto add_queue_to_list; | |
603 | } | |
d39b7737 OZ |
604 | |
605 | if (WARN(q->process->mm != current->mm, | |
606 | "should only run in user thread")) | |
607 | retval = -EFAULT; | |
608 | else | |
609 | retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe, | |
610 | q->queue, &q->properties, current->mm); | |
611 | if (retval) | |
d091bc0a | 612 | goto out_free_mqd; |
64c7f8cf BG |
613 | } |
614 | ||
2c99a547 | 615 | add_queue_to_list: |
64c7f8cf | 616 | list_add(&q->list, &qpd->queues_list); |
bc920fd4 | 617 | qpd->queue_count++; |
b6819cec | 618 | if (q->properties.is_active) |
ab4d51d4 | 619 | increment_queue_count(dqm, qpd, q); |
64c7f8cf | 620 | |
b8cbab04 OG |
621 | /* |
622 | * Unconditionally increment this counter, regardless of the queue's | |
623 | * type or whether the queue is active. | |
624 | */ | |
625 | dqm->total_queue_count++; | |
626 | pr_debug("Total of %d queues are accountable so far\n", | |
627 | dqm->total_queue_count); | |
d091bc0a | 628 | goto out_unlock; |
b8cbab04 | 629 | |
d091bc0a OZ |
630 | out_free_mqd: |
631 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); | |
d39b7737 OZ |
632 | out_deallocate_doorbell: |
633 | deallocate_doorbell(qpd, q); | |
634 | out_deallocate_hqd: | |
635 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) | |
636 | deallocate_hqd(dqm, q); | |
637 | else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || | |
638 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) | |
639 | deallocate_sdma_queue(dqm, q); | |
640 | deallocate_vmid: | |
641 | if (list_empty(&qpd->queues_list)) | |
642 | deallocate_vmid(dqm, qpd, q); | |
ab7c1648 | 643 | out_unlock: |
efeaed4d | 644 | dqm_unlock(dqm); |
ab7c1648 | 645 | return retval; |
64c7f8cf BG |
646 | } |
647 | ||
648 | static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q) | |
649 | { | |
650 | bool set; | |
f0ec5b99 | 651 | int pipe, bit, i; |
64c7f8cf BG |
652 | |
653 | set = false; | |
654 | ||
8eabaf54 KR |
655 | for (pipe = dqm->next_pipe_to_allocate, i = 0; |
656 | i < get_pipes_per_mec(dqm); | |
d0b63bb3 AR |
657 | pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) { |
658 | ||
659 | if (!is_pipe_enabled(dqm, 0, pipe)) | |
660 | continue; | |
661 | ||
64c7f8cf | 662 | if (dqm->allocated_queues[pipe] != 0) { |
4252bf68 HK |
663 | bit = ffs(dqm->allocated_queues[pipe]) - 1; |
664 | dqm->allocated_queues[pipe] &= ~(1 << bit); | |
64c7f8cf BG |
665 | q->pipe = pipe; |
666 | q->queue = bit; | |
667 | set = true; | |
668 | break; | |
669 | } | |
670 | } | |
671 | ||
991ca8ee | 672 | if (!set) |
64c7f8cf BG |
673 | return -EBUSY; |
674 | ||
79775b62 | 675 | pr_debug("hqd slot - pipe %d, queue %d\n", q->pipe, q->queue); |
64c7f8cf | 676 | /* horizontal hqd allocation */ |
d0b63bb3 | 677 | dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm); |
64c7f8cf BG |
678 | |
679 | return 0; | |
680 | } | |
681 | ||
682 | static inline void deallocate_hqd(struct device_queue_manager *dqm, | |
683 | struct queue *q) | |
684 | { | |
4252bf68 | 685 | dqm->allocated_queues[q->pipe] |= (1 << q->queue); |
64c7f8cf BG |
686 | } |
687 | ||
5bdd3eb2 MJ |
688 | #define SQ_IND_CMD_CMD_KILL 0x00000003 |
689 | #define SQ_IND_CMD_MODE_BROADCAST 0x00000001 | |
690 | ||
8dc1db31 | 691 | static int dbgdev_wave_reset_wavefronts(struct kfd_node *dev, struct kfd_process *p) |
5bdd3eb2 MJ |
692 | { |
693 | int status = 0; | |
694 | unsigned int vmid; | |
695 | uint16_t queried_pasid; | |
696 | union SQ_CMD_BITS reg_sq_cmd; | |
697 | union GRBM_GFX_INDEX_BITS reg_gfx_index; | |
698 | struct kfd_process_device *pdd; | |
699 | int first_vmid_to_scan = dev->vm_info.first_vmid_kfd; | |
700 | int last_vmid_to_scan = dev->vm_info.last_vmid_kfd; | |
c4050ff1 LL |
701 | uint32_t xcc_mask = dev->xcc_mask; |
702 | int xcc_id; | |
5bdd3eb2 MJ |
703 | |
704 | reg_sq_cmd.u32All = 0; | |
705 | reg_gfx_index.u32All = 0; | |
706 | ||
707 | pr_debug("Killing all process wavefronts\n"); | |
708 | ||
d55957fb YZ |
709 | if (!dev->kfd2kgd->get_atc_vmid_pasid_mapping_info) { |
710 | pr_err("no vmid pasid mapping supported \n"); | |
711 | return -EOPNOTSUPP; | |
712 | } | |
713 | ||
5bdd3eb2 MJ |
714 | /* Scan all registers in the range ATC_VMID8_PASID_MAPPING .. |
715 | * ATC_VMID15_PASID_MAPPING | |
716 | * to check which VMID the current process is mapped to. | |
717 | */ | |
718 | ||
d55957fb YZ |
719 | for (vmid = first_vmid_to_scan; vmid <= last_vmid_to_scan; vmid++) { |
720 | status = dev->kfd2kgd->get_atc_vmid_pasid_mapping_info | |
721 | (dev->adev, vmid, &queried_pasid); | |
c8b0507f | 722 | |
d55957fb YZ |
723 | if (status && queried_pasid == p->pasid) { |
724 | pr_debug("Killing wave fronts of vmid %d and pasid 0x%x\n", | |
725 | vmid, p->pasid); | |
726 | break; | |
5bdd3eb2 MJ |
727 | } |
728 | } | |
729 | ||
730 | if (vmid > last_vmid_to_scan) { | |
731 | pr_err("Didn't find vmid for pasid 0x%x\n", p->pasid); | |
732 | return -EFAULT; | |
733 | } | |
734 | ||
735 | /* taking the VMID for that process on the safe way using PDD */ | |
736 | pdd = kfd_get_process_device_data(dev, p); | |
737 | if (!pdd) | |
738 | return -EFAULT; | |
739 | ||
740 | reg_gfx_index.bits.sh_broadcast_writes = 1; | |
741 | reg_gfx_index.bits.se_broadcast_writes = 1; | |
742 | reg_gfx_index.bits.instance_broadcast_writes = 1; | |
743 | reg_sq_cmd.bits.mode = SQ_IND_CMD_MODE_BROADCAST; | |
744 | reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_KILL; | |
745 | reg_sq_cmd.bits.vm_id = vmid; | |
746 | ||
c4050ff1 LL |
747 | for_each_inst(xcc_id, xcc_mask) |
748 | dev->kfd2kgd->wave_control_execute( | |
749 | dev->adev, reg_gfx_index.u32All, | |
750 | reg_sq_cmd.u32All, xcc_id); | |
5bdd3eb2 MJ |
751 | |
752 | return 0; | |
753 | } | |
754 | ||
9fd3f1bf FK |
755 | /* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked |
756 | * to avoid asynchronized access | |
757 | */ | |
758 | static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm, | |
64c7f8cf BG |
759 | struct qcm_process_device *qpd, |
760 | struct queue *q) | |
761 | { | |
762 | int retval; | |
8d5f3552 | 763 | struct mqd_manager *mqd_mgr; |
64c7f8cf | 764 | |
fdfa090b OZ |
765 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
766 | q->properties.type)]; | |
64c7f8cf | 767 | |
c7637c95 | 768 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) |
c2e1b3a4 | 769 | deallocate_hqd(dqm, q); |
c7637c95 | 770 | else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) |
1b4670f6 | 771 | deallocate_sdma_queue(dqm, q); |
c7637c95 | 772 | else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) |
1b4670f6 | 773 | deallocate_sdma_queue(dqm, q); |
c7637c95 | 774 | else { |
79775b62 | 775 | pr_debug("q->properties.type %d is invalid\n", |
7113cd65 | 776 | q->properties.type); |
9fd3f1bf | 777 | return -EINVAL; |
64c7f8cf | 778 | } |
9fd3f1bf | 779 | dqm->total_queue_count--; |
64c7f8cf | 780 | |
ef568db7 FK |
781 | deallocate_doorbell(qpd, q); |
782 | ||
2c99a547 PY |
783 | if (!dqm->sched_running) { |
784 | WARN_ONCE(1, "Destroy non-HWS queue while stopped\n"); | |
785 | return 0; | |
786 | } | |
787 | ||
8d5f3552 | 788 | retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, |
c2e1b3a4 | 789 | KFD_PREEMPT_TYPE_WAVEFRONT_RESET, |
b90e3fbe | 790 | KFD_UNMAP_LATENCY_MS, |
64c7f8cf | 791 | q->pipe, q->queue); |
9fd3f1bf FK |
792 | if (retval == -ETIME) |
793 | qpd->reset_wavefronts = true; | |
64c7f8cf | 794 | |
64c7f8cf | 795 | list_del(&q->list); |
9fd3f1bf FK |
796 | if (list_empty(&qpd->queues_list)) { |
797 | if (qpd->reset_wavefronts) { | |
798 | pr_warn("Resetting wave fronts (nocpsch) on dev %p\n", | |
799 | dqm->dev); | |
800 | /* dbgdev_wave_reset_wavefronts has to be called before | |
801 | * deallocate_vmid(), i.e. when vmid is still in use. | |
802 | */ | |
803 | dbgdev_wave_reset_wavefronts(dqm->dev, | |
804 | qpd->pqm->process); | |
805 | qpd->reset_wavefronts = false; | |
806 | } | |
807 | ||
64c7f8cf | 808 | deallocate_vmid(dqm, qpd, q); |
9fd3f1bf | 809 | } |
bc920fd4 | 810 | qpd->queue_count--; |
ab4d51d4 DYS |
811 | if (q->properties.is_active) |
812 | decrement_queue_count(dqm, qpd, q); | |
b8cbab04 | 813 | |
9fd3f1bf FK |
814 | return retval; |
815 | } | |
b8cbab04 | 816 | |
9fd3f1bf FK |
817 | static int destroy_queue_nocpsch(struct device_queue_manager *dqm, |
818 | struct qcm_process_device *qpd, | |
819 | struct queue *q) | |
820 | { | |
821 | int retval; | |
d69fd951 MJ |
822 | uint64_t sdma_val = 0; |
823 | struct kfd_process_device *pdd = qpd_to_pdd(qpd); | |
a7b2451d AL |
824 | struct mqd_manager *mqd_mgr = |
825 | dqm->mqd_mgrs[get_mqd_type_from_queue_type(q->properties.type)]; | |
d69fd951 MJ |
826 | |
827 | /* Get the SDMA queue stats */ | |
828 | if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) || | |
829 | (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { | |
818b0324 | 830 | retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr, |
d69fd951 MJ |
831 | &sdma_val); |
832 | if (retval) | |
833 | pr_err("Failed to read SDMA queue counter for queue: %d\n", | |
834 | q->properties.queue_id); | |
835 | } | |
9fd3f1bf | 836 | |
efeaed4d | 837 | dqm_lock(dqm); |
9fd3f1bf | 838 | retval = destroy_queue_nocpsch_locked(dqm, qpd, q); |
d69fd951 MJ |
839 | if (!retval) |
840 | pdd->sdma_past_activity_counter += sdma_val; | |
efeaed4d | 841 | dqm_unlock(dqm); |
9fd3f1bf | 842 | |
a7b2451d AL |
843 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); |
844 | ||
64c7f8cf BG |
845 | return retval; |
846 | } | |
847 | ||
c6e559eb LY |
848 | static int update_queue(struct device_queue_manager *dqm, struct queue *q, |
849 | struct mqd_update_info *minfo) | |
64c7f8cf | 850 | { |
8636e53c | 851 | int retval = 0; |
8d5f3552 | 852 | struct mqd_manager *mqd_mgr; |
26103436 | 853 | struct kfd_process_device *pdd; |
b6ffbab8 | 854 | bool prev_active = false; |
64c7f8cf | 855 | |
efeaed4d | 856 | dqm_lock(dqm); |
26103436 FK |
857 | pdd = kfd_get_process_device_data(q->device, q->process); |
858 | if (!pdd) { | |
859 | retval = -ENODEV; | |
860 | goto out_unlock; | |
861 | } | |
fdfa090b OZ |
862 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
863 | q->properties.type)]; | |
64c7f8cf | 864 | |
60a00956 FK |
865 | /* Save previous activity state for counters */ |
866 | prev_active = q->properties.is_active; | |
867 | ||
868 | /* Make sure the queue is unmapped before updating the MQD */ | |
d146c5a7 | 869 | if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { |
8dc1db31 | 870 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
cc009e61 | 871 | retval = unmap_queues_cpsch(dqm, |
7cee6a68 | 872 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD, false); |
cc009e61 MJ |
873 | else if (prev_active) |
874 | retval = remove_queue_mes(dqm, q, &pdd->qpd); | |
875 | ||
894a8293 | 876 | if (retval) { |
60a00956 FK |
877 | pr_err("unmap queue failed\n"); |
878 | goto out_unlock; | |
879 | } | |
894a8293 | 880 | } else if (prev_active && |
60a00956 | 881 | (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || |
1b4670f6 OZ |
882 | q->properties.type == KFD_QUEUE_TYPE_SDMA || |
883 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { | |
2c99a547 PY |
884 | |
885 | if (!dqm->sched_running) { | |
886 | WARN_ONCE(1, "Update non-HWS queue while stopped\n"); | |
887 | goto out_unlock; | |
888 | } | |
889 | ||
8d5f3552 | 890 | retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, |
8dc1db31 | 891 | (dqm->dev->kfd->cwsr_enabled ? |
2243f493 RB |
892 | KFD_PREEMPT_TYPE_WAVEFRONT_SAVE : |
893 | KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN), | |
60a00956 FK |
894 | KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); |
895 | if (retval) { | |
896 | pr_err("destroy mqd failed\n"); | |
897 | goto out_unlock; | |
898 | } | |
899 | } | |
900 | ||
c6e559eb | 901 | mqd_mgr->update_mqd(mqd_mgr, q->mqd, &q->properties, minfo); |
60a00956 | 902 | |
096d1a3e FK |
903 | /* |
904 | * check active state vs. the previous state and modify | |
905 | * counter accordingly. map_queues_cpsch uses the | |
81b820b3 | 906 | * dqm->active_queue_count to determine whether a new runlist must be |
096d1a3e FK |
907 | * uploaded. |
908 | */ | |
ab4d51d4 DYS |
909 | if (q->properties.is_active && !prev_active) { |
910 | increment_queue_count(dqm, &pdd->qpd, q); | |
911 | } else if (!q->properties.is_active && prev_active) { | |
912 | decrement_queue_count(dqm, &pdd->qpd, q); | |
913 | } else if (q->gws && !q->properties.is_gws) { | |
b8020b03 JG |
914 | if (q->properties.is_active) { |
915 | dqm->gws_queue_count++; | |
916 | pdd->qpd.mapped_gws_queue = true; | |
917 | } | |
918 | q->properties.is_gws = true; | |
919 | } else if (!q->gws && q->properties.is_gws) { | |
920 | if (q->properties.is_active) { | |
921 | dqm->gws_queue_count--; | |
922 | pdd->qpd.mapped_gws_queue = false; | |
923 | } | |
924 | q->properties.is_gws = false; | |
925 | } | |
926 | ||
cc009e61 | 927 | if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { |
8dc1db31 | 928 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
cc009e61 | 929 | retval = map_queues_cpsch(dqm); |
f4f9b827 | 930 | else if (q->properties.is_active) |
cc009e61 MJ |
931 | retval = add_queue_mes(dqm, q, &pdd->qpd); |
932 | } else if (q->properties.is_active && | |
60a00956 | 933 | (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || |
1b4670f6 OZ |
934 | q->properties.type == KFD_QUEUE_TYPE_SDMA || |
935 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { | |
1b19aa5a FK |
936 | if (WARN(q->process->mm != current->mm, |
937 | "should only run in user thread")) | |
938 | retval = -EFAULT; | |
939 | else | |
940 | retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, | |
941 | q->pipe, q->queue, | |
942 | &q->properties, current->mm); | |
943 | } | |
b6ffbab8 | 944 | |
ab7c1648 | 945 | out_unlock: |
efeaed4d | 946 | dqm_unlock(dqm); |
64c7f8cf BG |
947 | return retval; |
948 | } | |
949 | ||
26103436 FK |
950 | static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, |
951 | struct qcm_process_device *qpd) | |
952 | { | |
953 | struct queue *q; | |
8d5f3552 | 954 | struct mqd_manager *mqd_mgr; |
26103436 | 955 | struct kfd_process_device *pdd; |
bb2d2128 | 956 | int retval, ret = 0; |
26103436 | 957 | |
efeaed4d | 958 | dqm_lock(dqm); |
26103436 FK |
959 | if (qpd->evicted++ > 0) /* already evicted, do nothing */ |
960 | goto out; | |
961 | ||
962 | pdd = qpd_to_pdd(qpd); | |
783a25f4 | 963 | pr_debug_ratelimited("Evicting PASID 0x%x queues\n", |
26103436 FK |
964 | pdd->process->pasid); |
965 | ||
4327bed2 | 966 | pdd->last_evict_timestamp = get_jiffies_64(); |
bb2d2128 FK |
967 | /* Mark all queues as evicted. Deactivate all active queues on |
968 | * the qpd. | |
969 | */ | |
26103436 | 970 | list_for_each_entry(q, &qpd->queues_list, list) { |
bb2d2128 | 971 | q->properties.is_evicted = true; |
26103436 FK |
972 | if (!q->properties.is_active) |
973 | continue; | |
bb2d2128 | 974 | |
fdfa090b OZ |
975 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
976 | q->properties.type)]; | |
26103436 | 977 | q->properties.is_active = false; |
ab4d51d4 | 978 | decrement_queue_count(dqm, qpd, q); |
2c99a547 PY |
979 | |
980 | if (WARN_ONCE(!dqm->sched_running, "Evict when stopped\n")) | |
981 | continue; | |
982 | ||
8d5f3552 | 983 | retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, |
8dc1db31 | 984 | (dqm->dev->kfd->cwsr_enabled ? |
2243f493 RB |
985 | KFD_PREEMPT_TYPE_WAVEFRONT_SAVE : |
986 | KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN), | |
26103436 | 987 | KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); |
bb2d2128 FK |
988 | if (retval && !ret) |
989 | /* Return the first error, but keep going to | |
990 | * maintain a consistent eviction state | |
991 | */ | |
992 | ret = retval; | |
26103436 FK |
993 | } |
994 | ||
995 | out: | |
efeaed4d | 996 | dqm_unlock(dqm); |
bb2d2128 | 997 | return ret; |
26103436 FK |
998 | } |
999 | ||
1000 | static int evict_process_queues_cpsch(struct device_queue_manager *dqm, | |
1001 | struct qcm_process_device *qpd) | |
1002 | { | |
1003 | struct queue *q; | |
1004 | struct kfd_process_device *pdd; | |
1005 | int retval = 0; | |
1006 | ||
efeaed4d | 1007 | dqm_lock(dqm); |
26103436 FK |
1008 | if (qpd->evicted++ > 0) /* already evicted, do nothing */ |
1009 | goto out; | |
1010 | ||
1011 | pdd = qpd_to_pdd(qpd); | |
0ab2d753 JK |
1012 | |
1013 | /* The debugger creates processes that temporarily have not acquired | |
1014 | * all VMs for all devices and has no VMs itself. | |
1015 | * Skip queue eviction on process eviction. | |
1016 | */ | |
1017 | if (!pdd->drm_priv) | |
1018 | goto out; | |
1019 | ||
783a25f4 | 1020 | pr_debug_ratelimited("Evicting PASID 0x%x queues\n", |
26103436 FK |
1021 | pdd->process->pasid); |
1022 | ||
bb2d2128 FK |
1023 | /* Mark all queues as evicted. Deactivate all active queues on |
1024 | * the qpd. | |
1025 | */ | |
26103436 | 1026 | list_for_each_entry(q, &qpd->queues_list, list) { |
bb2d2128 | 1027 | q->properties.is_evicted = true; |
26103436 FK |
1028 | if (!q->properties.is_active) |
1029 | continue; | |
bb2d2128 | 1030 | |
26103436 | 1031 | q->properties.is_active = false; |
ab4d51d4 | 1032 | decrement_queue_count(dqm, qpd, q); |
cc009e61 | 1033 | |
8dc1db31 | 1034 | if (dqm->dev->kfd->shared_resources.enable_mes) { |
cc009e61 MJ |
1035 | retval = remove_queue_mes(dqm, q, qpd); |
1036 | if (retval) { | |
1037 | pr_err("Failed to evict queue %d\n", | |
1038 | q->properties.queue_id); | |
1039 | goto out; | |
1040 | } | |
1041 | } | |
26103436 | 1042 | } |
4327bed2 | 1043 | pdd->last_evict_timestamp = get_jiffies_64(); |
8dc1db31 | 1044 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
cc009e61 MJ |
1045 | retval = execute_queues_cpsch(dqm, |
1046 | qpd->is_debug ? | |
1047 | KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES : | |
7cee6a68 JK |
1048 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, |
1049 | USE_DEFAULT_GRACE_PERIOD); | |
26103436 FK |
1050 | |
1051 | out: | |
efeaed4d | 1052 | dqm_unlock(dqm); |
26103436 FK |
1053 | return retval; |
1054 | } | |
1055 | ||
1056 | static int restore_process_queues_nocpsch(struct device_queue_manager *dqm, | |
1057 | struct qcm_process_device *qpd) | |
1058 | { | |
1b19aa5a | 1059 | struct mm_struct *mm = NULL; |
26103436 | 1060 | struct queue *q; |
8d5f3552 | 1061 | struct mqd_manager *mqd_mgr; |
26103436 | 1062 | struct kfd_process_device *pdd; |
e715c6d0 | 1063 | uint64_t pd_base; |
4327bed2 | 1064 | uint64_t eviction_duration; |
bb2d2128 | 1065 | int retval, ret = 0; |
26103436 FK |
1066 | |
1067 | pdd = qpd_to_pdd(qpd); | |
1068 | /* Retrieve PD base */ | |
b40a6ab2 | 1069 | pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv); |
26103436 | 1070 | |
efeaed4d | 1071 | dqm_lock(dqm); |
26103436 FK |
1072 | if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ |
1073 | goto out; | |
1074 | if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */ | |
1075 | qpd->evicted--; | |
1076 | goto out; | |
1077 | } | |
1078 | ||
783a25f4 | 1079 | pr_debug_ratelimited("Restoring PASID 0x%x queues\n", |
26103436 FK |
1080 | pdd->process->pasid); |
1081 | ||
1082 | /* Update PD Base in QPD */ | |
1083 | qpd->page_table_base = pd_base; | |
e715c6d0 | 1084 | pr_debug("Updated PD address to 0x%llx\n", pd_base); |
26103436 FK |
1085 | |
1086 | if (!list_empty(&qpd->queues_list)) { | |
1087 | dqm->dev->kfd2kgd->set_vm_context_page_table_base( | |
3356c38d | 1088 | dqm->dev->adev, |
26103436 FK |
1089 | qpd->vmid, |
1090 | qpd->page_table_base); | |
3543b055 | 1091 | kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY); |
26103436 FK |
1092 | } |
1093 | ||
1b19aa5a FK |
1094 | /* Take a safe reference to the mm_struct, which may otherwise |
1095 | * disappear even while the kfd_process is still referenced. | |
1096 | */ | |
1097 | mm = get_task_mm(pdd->process->lead_thread); | |
1098 | if (!mm) { | |
bb2d2128 | 1099 | ret = -EFAULT; |
1b19aa5a FK |
1100 | goto out; |
1101 | } | |
1102 | ||
bb2d2128 FK |
1103 | /* Remove the eviction flags. Activate queues that are not |
1104 | * inactive for other reasons. | |
1105 | */ | |
26103436 | 1106 | list_for_each_entry(q, &qpd->queues_list, list) { |
bb2d2128 FK |
1107 | q->properties.is_evicted = false; |
1108 | if (!QUEUE_IS_ACTIVE(q->properties)) | |
26103436 | 1109 | continue; |
bb2d2128 | 1110 | |
fdfa090b OZ |
1111 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
1112 | q->properties.type)]; | |
26103436 | 1113 | q->properties.is_active = true; |
ab4d51d4 | 1114 | increment_queue_count(dqm, qpd, q); |
2c99a547 PY |
1115 | |
1116 | if (WARN_ONCE(!dqm->sched_running, "Restore when stopped\n")) | |
1117 | continue; | |
1118 | ||
8d5f3552 | 1119 | retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe, |
1b19aa5a | 1120 | q->queue, &q->properties, mm); |
bb2d2128 FK |
1121 | if (retval && !ret) |
1122 | /* Return the first error, but keep going to | |
1123 | * maintain a consistent eviction state | |
1124 | */ | |
1125 | ret = retval; | |
26103436 FK |
1126 | } |
1127 | qpd->evicted = 0; | |
4327bed2 PC |
1128 | eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp; |
1129 | atomic64_add(eviction_duration, &pdd->evict_duration_counter); | |
26103436 | 1130 | out: |
1b19aa5a FK |
1131 | if (mm) |
1132 | mmput(mm); | |
efeaed4d | 1133 | dqm_unlock(dqm); |
bb2d2128 | 1134 | return ret; |
26103436 FK |
1135 | } |
1136 | ||
1137 | static int restore_process_queues_cpsch(struct device_queue_manager *dqm, | |
1138 | struct qcm_process_device *qpd) | |
1139 | { | |
1140 | struct queue *q; | |
1141 | struct kfd_process_device *pdd; | |
4327bed2 | 1142 | uint64_t eviction_duration; |
26103436 FK |
1143 | int retval = 0; |
1144 | ||
1145 | pdd = qpd_to_pdd(qpd); | |
26103436 | 1146 | |
efeaed4d | 1147 | dqm_lock(dqm); |
26103436 FK |
1148 | if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ |
1149 | goto out; | |
1150 | if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */ | |
1151 | qpd->evicted--; | |
1152 | goto out; | |
1153 | } | |
1154 | ||
0ab2d753 JK |
1155 | /* The debugger creates processes that temporarily have not acquired |
1156 | * all VMs for all devices and has no VMs itself. | |
1157 | * Skip queue restore on process restore. | |
1158 | */ | |
1159 | if (!pdd->drm_priv) | |
1160 | goto vm_not_acquired; | |
1161 | ||
783a25f4 | 1162 | pr_debug_ratelimited("Restoring PASID 0x%x queues\n", |
26103436 FK |
1163 | pdd->process->pasid); |
1164 | ||
1165 | /* Update PD Base in QPD */ | |
0ab2d753 JK |
1166 | qpd->page_table_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv); |
1167 | pr_debug("Updated PD address to 0x%llx\n", qpd->page_table_base); | |
26103436 FK |
1168 | |
1169 | /* activate all active queues on the qpd */ | |
1170 | list_for_each_entry(q, &qpd->queues_list, list) { | |
26103436 | 1171 | q->properties.is_evicted = false; |
bb2d2128 FK |
1172 | if (!QUEUE_IS_ACTIVE(q->properties)) |
1173 | continue; | |
1174 | ||
26103436 | 1175 | q->properties.is_active = true; |
ab4d51d4 | 1176 | increment_queue_count(dqm, &pdd->qpd, q); |
cc009e61 | 1177 | |
8dc1db31 | 1178 | if (dqm->dev->kfd->shared_resources.enable_mes) { |
cc009e61 MJ |
1179 | retval = add_queue_mes(dqm, q, qpd); |
1180 | if (retval) { | |
1181 | pr_err("Failed to restore queue %d\n", | |
1182 | q->properties.queue_id); | |
1183 | goto out; | |
1184 | } | |
1185 | } | |
26103436 | 1186 | } |
8dc1db31 | 1187 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
cc009e61 | 1188 | retval = execute_queues_cpsch(dqm, |
7cee6a68 | 1189 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); |
4327bed2 PC |
1190 | eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp; |
1191 | atomic64_add(eviction_duration, &pdd->evict_duration_counter); | |
0ab2d753 JK |
1192 | vm_not_acquired: |
1193 | qpd->evicted = 0; | |
26103436 | 1194 | out: |
efeaed4d | 1195 | dqm_unlock(dqm); |
26103436 FK |
1196 | return retval; |
1197 | } | |
1198 | ||
58dcd5bf | 1199 | static int register_process(struct device_queue_manager *dqm, |
64c7f8cf BG |
1200 | struct qcm_process_device *qpd) |
1201 | { | |
1202 | struct device_process_node *n; | |
403575c4 | 1203 | struct kfd_process_device *pdd; |
e715c6d0 | 1204 | uint64_t pd_base; |
a22fc854 | 1205 | int retval; |
64c7f8cf | 1206 | |
dbf56ab1 | 1207 | n = kzalloc(sizeof(*n), GFP_KERNEL); |
64c7f8cf BG |
1208 | if (!n) |
1209 | return -ENOMEM; | |
1210 | ||
1211 | n->qpd = qpd; | |
1212 | ||
403575c4 FK |
1213 | pdd = qpd_to_pdd(qpd); |
1214 | /* Retrieve PD base */ | |
b40a6ab2 | 1215 | pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv); |
403575c4 | 1216 | |
efeaed4d | 1217 | dqm_lock(dqm); |
64c7f8cf BG |
1218 | list_add(&n->list, &dqm->queues); |
1219 | ||
403575c4 FK |
1220 | /* Update PD Base in QPD */ |
1221 | qpd->page_table_base = pd_base; | |
e715c6d0 | 1222 | pr_debug("Updated PD address to 0x%llx\n", pd_base); |
403575c4 | 1223 | |
bfd5e378 | 1224 | retval = dqm->asic_ops.update_qpd(dqm, qpd); |
a22fc854 | 1225 | |
f756e631 | 1226 | dqm->processes_count++; |
64c7f8cf | 1227 | |
efeaed4d | 1228 | dqm_unlock(dqm); |
64c7f8cf | 1229 | |
32cce8bc FK |
1230 | /* Outside the DQM lock because under the DQM lock we can't do |
1231 | * reclaim or take other locks that others hold while reclaiming. | |
1232 | */ | |
1233 | kfd_inc_compute_active(dqm->dev); | |
1234 | ||
a22fc854 | 1235 | return retval; |
64c7f8cf BG |
1236 | } |
1237 | ||
58dcd5bf | 1238 | static int unregister_process(struct device_queue_manager *dqm, |
64c7f8cf BG |
1239 | struct qcm_process_device *qpd) |
1240 | { | |
1241 | int retval; | |
1242 | struct device_process_node *cur, *next; | |
1243 | ||
1e5ec956 OG |
1244 | pr_debug("qpd->queues_list is %s\n", |
1245 | list_empty(&qpd->queues_list) ? "empty" : "not empty"); | |
64c7f8cf BG |
1246 | |
1247 | retval = 0; | |
efeaed4d | 1248 | dqm_lock(dqm); |
64c7f8cf BG |
1249 | |
1250 | list_for_each_entry_safe(cur, next, &dqm->queues, list) { | |
1251 | if (qpd == cur->qpd) { | |
1252 | list_del(&cur->list); | |
f5d896bb | 1253 | kfree(cur); |
f756e631 | 1254 | dqm->processes_count--; |
64c7f8cf BG |
1255 | goto out; |
1256 | } | |
1257 | } | |
1258 | /* qpd not found in dqm list */ | |
1259 | retval = 1; | |
1260 | out: | |
efeaed4d | 1261 | dqm_unlock(dqm); |
32cce8bc FK |
1262 | |
1263 | /* Outside the DQM lock because under the DQM lock we can't do | |
1264 | * reclaim or take other locks that others hold while reclaiming. | |
1265 | */ | |
1266 | if (!retval) | |
1267 | kfd_dec_compute_active(dqm->dev); | |
1268 | ||
64c7f8cf BG |
1269 | return retval; |
1270 | } | |
1271 | ||
1272 | static int | |
c7b6bac9 | 1273 | set_pasid_vmid_mapping(struct device_queue_manager *dqm, u32 pasid, |
64c7f8cf BG |
1274 | unsigned int vmid) |
1275 | { | |
c4050ff1 LL |
1276 | uint32_t xcc_mask = dqm->dev->xcc_mask; |
1277 | int xcc_id, ret; | |
e2069a7b | 1278 | |
c4050ff1 | 1279 | for_each_inst(xcc_id, xcc_mask) { |
e2069a7b | 1280 | ret = dqm->dev->kfd2kgd->set_pasid_vmid_mapping( |
c4050ff1 | 1281 | dqm->dev->adev, pasid, vmid, xcc_id); |
e2069a7b MJ |
1282 | if (ret) |
1283 | break; | |
1284 | } | |
1285 | ||
1286 | return ret; | |
64c7f8cf BG |
1287 | } |
1288 | ||
2249d558 AL |
1289 | static void init_interrupts(struct device_queue_manager *dqm) |
1290 | { | |
c4050ff1 LL |
1291 | uint32_t xcc_mask = dqm->dev->xcc_mask; |
1292 | unsigned int i, xcc_id; | |
2249d558 | 1293 | |
b695c97b LL |
1294 | for_each_inst(xcc_id, xcc_mask) { |
1295 | for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++) { | |
1296 | if (is_pipe_enabled(dqm, 0, i)) { | |
e2069a7b | 1297 | dqm->dev->kfd2kgd->init_interrupts( |
c4050ff1 | 1298 | dqm->dev->adev, i, xcc_id); |
b695c97b | 1299 | } |
e2069a7b MJ |
1300 | } |
1301 | } | |
2249d558 AL |
1302 | } |
1303 | ||
64c7f8cf BG |
1304 | static int initialize_nocpsch(struct device_queue_manager *dqm) |
1305 | { | |
86194cf8 | 1306 | int pipe, queue; |
64c7f8cf | 1307 | |
79775b62 | 1308 | pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm)); |
64c7f8cf | 1309 | |
ab7c1648 KR |
1310 | dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm), |
1311 | sizeof(unsigned int), GFP_KERNEL); | |
1312 | if (!dqm->allocated_queues) | |
1313 | return -ENOMEM; | |
1314 | ||
efeaed4d | 1315 | mutex_init(&dqm->lock_hidden); |
64c7f8cf | 1316 | INIT_LIST_HEAD(&dqm->queues); |
81b820b3 | 1317 | dqm->active_queue_count = dqm->next_pipe_to_allocate = 0; |
b42902f4 | 1318 | dqm->active_cp_queue_count = 0; |
b8020b03 | 1319 | dqm->gws_queue_count = 0; |
64c7f8cf | 1320 | |
86194cf8 FK |
1321 | for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) { |
1322 | int pipe_offset = pipe * get_queues_per_pipe(dqm); | |
1323 | ||
1324 | for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) | |
1325 | if (test_bit(pipe_offset + queue, | |
8dc1db31 | 1326 | dqm->dev->kfd->shared_resources.cp_queue_bitmap)) |
86194cf8 FK |
1327 | dqm->allocated_queues[pipe] |= 1 << queue; |
1328 | } | |
64c7f8cf | 1329 | |
d9d4623c YZ |
1330 | memset(dqm->vmid_pasid, 0, sizeof(dqm->vmid_pasid)); |
1331 | ||
b292cafe | 1332 | init_sdma_bitmaps(dqm); |
64c7f8cf | 1333 | |
64c7f8cf BG |
1334 | return 0; |
1335 | } | |
1336 | ||
58dcd5bf | 1337 | static void uninitialize(struct device_queue_manager *dqm) |
64c7f8cf | 1338 | { |
6f9d54fd OG |
1339 | int i; |
1340 | ||
81b820b3 | 1341 | WARN_ON(dqm->active_queue_count > 0 || dqm->processes_count > 0); |
64c7f8cf BG |
1342 | |
1343 | kfree(dqm->allocated_queues); | |
6f9d54fd | 1344 | for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++) |
8d5f3552 | 1345 | kfree(dqm->mqd_mgrs[i]); |
efeaed4d | 1346 | mutex_destroy(&dqm->lock_hidden); |
64c7f8cf BG |
1347 | } |
1348 | ||
1349 | static int start_nocpsch(struct device_queue_manager *dqm) | |
1350 | { | |
6f4cb84a FK |
1351 | int r = 0; |
1352 | ||
52055039 | 1353 | pr_info("SW scheduler is used"); |
2249d558 | 1354 | init_interrupts(dqm); |
2243f493 | 1355 | |
7eb0502a | 1356 | if (dqm->dev->adev->asic_type == CHIP_HAWAII) |
6f4cb84a FK |
1357 | r = pm_init(&dqm->packet_mgr, dqm); |
1358 | if (!r) | |
1359 | dqm->sched_running = true; | |
2c99a547 | 1360 | |
6f4cb84a | 1361 | return r; |
64c7f8cf BG |
1362 | } |
1363 | ||
1364 | static int stop_nocpsch(struct device_queue_manager *dqm) | |
1365 | { | |
5e406012 MJ |
1366 | dqm_lock(dqm); |
1367 | if (!dqm->sched_running) { | |
1368 | dqm_unlock(dqm); | |
1369 | return 0; | |
1370 | } | |
1371 | ||
7eb0502a | 1372 | if (dqm->dev->adev->asic_type == CHIP_HAWAII) |
9af5379c | 1373 | pm_uninit(&dqm->packet_mgr, false); |
2c99a547 | 1374 | dqm->sched_running = false; |
5e406012 | 1375 | dqm_unlock(dqm); |
2c99a547 | 1376 | |
64c7f8cf BG |
1377 | return 0; |
1378 | } | |
1379 | ||
09c34e8d FK |
1380 | static void pre_reset(struct device_queue_manager *dqm) |
1381 | { | |
1382 | dqm_lock(dqm); | |
1383 | dqm->is_resetting = true; | |
1384 | dqm_unlock(dqm); | |
1385 | } | |
1386 | ||
bcea3081 | 1387 | static int allocate_sdma_queue(struct device_queue_manager *dqm, |
2485c12c | 1388 | struct queue *q, const uint32_t *restore_sdma_id) |
bcea3081 BG |
1389 | { |
1390 | int bit; | |
1391 | ||
1b4670f6 | 1392 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { |
a805889a | 1393 | if (bitmap_empty(dqm->sdma_bitmap, KFD_MAX_SDMA_QUEUES)) { |
c7637c95 | 1394 | pr_err("No more SDMA queue to allocate\n"); |
1b4670f6 | 1395 | return -ENOMEM; |
c7637c95 YZ |
1396 | } |
1397 | ||
2485c12c DYS |
1398 | if (restore_sdma_id) { |
1399 | /* Re-use existing sdma_id */ | |
a805889a | 1400 | if (!test_bit(*restore_sdma_id, dqm->sdma_bitmap)) { |
2485c12c DYS |
1401 | pr_err("SDMA queue already in use\n"); |
1402 | return -EBUSY; | |
1403 | } | |
a805889a | 1404 | clear_bit(*restore_sdma_id, dqm->sdma_bitmap); |
2485c12c DYS |
1405 | q->sdma_id = *restore_sdma_id; |
1406 | } else { | |
1407 | /* Find first available sdma_id */ | |
a805889a MJ |
1408 | bit = find_first_bit(dqm->sdma_bitmap, |
1409 | get_num_sdma_queues(dqm)); | |
1410 | clear_bit(bit, dqm->sdma_bitmap); | |
2485c12c DYS |
1411 | q->sdma_id = bit; |
1412 | } | |
1413 | ||
a805889a | 1414 | q->properties.sdma_engine_id = |
a805889a | 1415 | q->sdma_id % kfd_get_num_sdma_engines(dqm->dev); |
1b4670f6 | 1416 | q->properties.sdma_queue_id = q->sdma_id / |
ee2f17f4 | 1417 | kfd_get_num_sdma_engines(dqm->dev); |
1b4670f6 | 1418 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { |
a805889a | 1419 | if (bitmap_empty(dqm->xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES)) { |
c7637c95 | 1420 | pr_err("No more XGMI SDMA queue to allocate\n"); |
1b4670f6 | 1421 | return -ENOMEM; |
c7637c95 | 1422 | } |
2485c12c DYS |
1423 | if (restore_sdma_id) { |
1424 | /* Re-use existing sdma_id */ | |
a805889a | 1425 | if (!test_bit(*restore_sdma_id, dqm->xgmi_sdma_bitmap)) { |
2485c12c DYS |
1426 | pr_err("SDMA queue already in use\n"); |
1427 | return -EBUSY; | |
1428 | } | |
a805889a | 1429 | clear_bit(*restore_sdma_id, dqm->xgmi_sdma_bitmap); |
2485c12c DYS |
1430 | q->sdma_id = *restore_sdma_id; |
1431 | } else { | |
a805889a MJ |
1432 | bit = find_first_bit(dqm->xgmi_sdma_bitmap, |
1433 | get_num_xgmi_sdma_queues(dqm)); | |
1434 | clear_bit(bit, dqm->xgmi_sdma_bitmap); | |
2485c12c DYS |
1435 | q->sdma_id = bit; |
1436 | } | |
1b4670f6 OZ |
1437 | /* sdma_engine_id is sdma id including |
1438 | * both PCIe-optimized SDMAs and XGMI- | |
1439 | * optimized SDMAs. The calculation below | |
1440 | * assumes the first N engines are always | |
1441 | * PCIe-optimized ones | |
1442 | */ | |
ee2f17f4 AL |
1443 | q->properties.sdma_engine_id = |
1444 | kfd_get_num_sdma_engines(dqm->dev) + | |
1445 | q->sdma_id % kfd_get_num_xgmi_sdma_engines(dqm->dev); | |
1b4670f6 | 1446 | q->properties.sdma_queue_id = q->sdma_id / |
ee2f17f4 | 1447 | kfd_get_num_xgmi_sdma_engines(dqm->dev); |
1b4670f6 | 1448 | } |
e78579aa | 1449 | |
e78579aa YZ |
1450 | pr_debug("SDMA engine id: %d\n", q->properties.sdma_engine_id); |
1451 | pr_debug("SDMA queue id: %d\n", q->properties.sdma_queue_id); | |
bcea3081 BG |
1452 | |
1453 | return 0; | |
1454 | } | |
1455 | ||
1456 | static void deallocate_sdma_queue(struct device_queue_manager *dqm, | |
1b4670f6 | 1457 | struct queue *q) |
bcea3081 | 1458 | { |
1b4670f6 OZ |
1459 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { |
1460 | if (q->sdma_id >= get_num_sdma_queues(dqm)) | |
1461 | return; | |
a805889a | 1462 | set_bit(q->sdma_id, dqm->sdma_bitmap); |
1b4670f6 OZ |
1463 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { |
1464 | if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm)) | |
1465 | return; | |
a805889a | 1466 | set_bit(q->sdma_id, dqm->xgmi_sdma_bitmap); |
1b4670f6 | 1467 | } |
bcea3081 BG |
1468 | } |
1469 | ||
64c7f8cf BG |
1470 | /* |
1471 | * Device Queue Manager implementation for cp scheduler | |
1472 | */ | |
1473 | ||
1474 | static int set_sched_resources(struct device_queue_manager *dqm) | |
1475 | { | |
d0b63bb3 | 1476 | int i, mec; |
64c7f8cf | 1477 | struct scheduling_resources res; |
64c7f8cf | 1478 | |
74c5b85d | 1479 | res.vmid_mask = dqm->dev->compute_vmid_bitmap; |
d0b63bb3 AR |
1480 | |
1481 | res.queue_mask = 0; | |
1482 | for (i = 0; i < KGD_MAX_QUEUES; ++i) { | |
8dc1db31 MJ |
1483 | mec = (i / dqm->dev->kfd->shared_resources.num_queue_per_pipe) |
1484 | / dqm->dev->kfd->shared_resources.num_pipe_per_mec; | |
d0b63bb3 | 1485 | |
8dc1db31 | 1486 | if (!test_bit(i, dqm->dev->kfd->shared_resources.cp_queue_bitmap)) |
d0b63bb3 AR |
1487 | continue; |
1488 | ||
1489 | /* only acquire queues from the first MEC */ | |
1490 | if (mec > 0) | |
1491 | continue; | |
1492 | ||
1493 | /* This situation may be hit in the future if a new HW | |
1494 | * generation exposes more than 64 queues. If so, the | |
8eabaf54 KR |
1495 | * definition of res.queue_mask needs updating |
1496 | */ | |
1d11ee89 | 1497 | if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) { |
d0b63bb3 AR |
1498 | pr_err("Invalid queue enabled by amdgpu: %d\n", i); |
1499 | break; | |
1500 | } | |
1501 | ||
d09f85d5 YZ |
1502 | res.queue_mask |= 1ull |
1503 | << amdgpu_queue_mask_bit_to_set_resource_bit( | |
56c5977e | 1504 | dqm->dev->adev, i); |
d0b63bb3 | 1505 | } |
d9848e14 OZ |
1506 | res.gws_mask = ~0ull; |
1507 | res.oac_mask = res.gds_heap_base = res.gds_heap_size = 0; | |
64c7f8cf | 1508 | |
79775b62 KR |
1509 | pr_debug("Scheduling resources:\n" |
1510 | "vmid mask: 0x%8X\n" | |
1511 | "queue mask: 0x%8llX\n", | |
64c7f8cf BG |
1512 | res.vmid_mask, res.queue_mask); |
1513 | ||
9af5379c | 1514 | return pm_send_set_resources(&dqm->packet_mgr, &res); |
64c7f8cf BG |
1515 | } |
1516 | ||
1517 | static int initialize_cpsch(struct device_queue_manager *dqm) | |
1518 | { | |
79775b62 | 1519 | pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm)); |
64c7f8cf | 1520 | |
efeaed4d | 1521 | mutex_init(&dqm->lock_hidden); |
64c7f8cf | 1522 | INIT_LIST_HEAD(&dqm->queues); |
81b820b3 | 1523 | dqm->active_queue_count = dqm->processes_count = 0; |
b42902f4 | 1524 | dqm->active_cp_queue_count = 0; |
b8020b03 | 1525 | dqm->gws_queue_count = 0; |
64c7f8cf | 1526 | dqm->active_runlist = false; |
73ea648d SL |
1527 | INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception); |
1528 | ||
b292cafe FK |
1529 | init_sdma_bitmaps(dqm); |
1530 | ||
7cee6a68 JK |
1531 | if (dqm->dev->kfd2kgd->get_iq_wait_times) |
1532 | dqm->dev->kfd2kgd->get_iq_wait_times(dqm->dev->adev, | |
1533 | &dqm->wait_times); | |
bfd5e378 | 1534 | return 0; |
64c7f8cf BG |
1535 | } |
1536 | ||
1537 | static int start_cpsch(struct device_queue_manager *dqm) | |
1538 | { | |
64c7f8cf BG |
1539 | int retval; |
1540 | ||
64c7f8cf BG |
1541 | retval = 0; |
1542 | ||
4f942aae | 1543 | dqm_lock(dqm); |
64c7f8cf | 1544 | |
8dc1db31 | 1545 | if (!dqm->dev->kfd->shared_resources.enable_mes) { |
cc009e61 MJ |
1546 | retval = pm_init(&dqm->packet_mgr, dqm); |
1547 | if (retval) | |
1548 | goto fail_packet_manager_init; | |
64c7f8cf | 1549 | |
cc009e61 MJ |
1550 | retval = set_sched_resources(dqm); |
1551 | if (retval) | |
1552 | goto fail_set_sched_resources; | |
1553 | } | |
79775b62 | 1554 | pr_debug("Allocating fence memory\n"); |
64c7f8cf BG |
1555 | |
1556 | /* allocate fence memory on the gart */ | |
a86aa3ca OG |
1557 | retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr), |
1558 | &dqm->fence_mem); | |
64c7f8cf | 1559 | |
4eacc26b | 1560 | if (retval) |
64c7f8cf BG |
1561 | goto fail_allocate_vidmem; |
1562 | ||
b010affe | 1563 | dqm->fence_addr = (uint64_t *)dqm->fence_mem->cpu_ptr; |
64c7f8cf | 1564 | dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr; |
2249d558 AL |
1565 | |
1566 | init_interrupts(dqm); | |
1567 | ||
73ea648d SL |
1568 | /* clear hang status when driver try to start the hw scheduler */ |
1569 | dqm->is_hws_hang = false; | |
09c34e8d | 1570 | dqm->is_resetting = false; |
2c99a547 | 1571 | dqm->sched_running = true; |
7cee6a68 | 1572 | |
8dc1db31 | 1573 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
7cee6a68 | 1574 | execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); |
efeaed4d | 1575 | dqm_unlock(dqm); |
64c7f8cf BG |
1576 | |
1577 | return 0; | |
1578 | fail_allocate_vidmem: | |
1579 | fail_set_sched_resources: | |
8dc1db31 | 1580 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
cc009e61 | 1581 | pm_uninit(&dqm->packet_mgr, false); |
64c7f8cf | 1582 | fail_packet_manager_init: |
4f942aae | 1583 | dqm_unlock(dqm); |
64c7f8cf BG |
1584 | return retval; |
1585 | } | |
1586 | ||
1587 | static int stop_cpsch(struct device_queue_manager *dqm) | |
1588 | { | |
c2a77fde FK |
1589 | bool hanging; |
1590 | ||
efeaed4d | 1591 | dqm_lock(dqm); |
c96cb659 | 1592 | if (!dqm->sched_running) { |
1593 | dqm_unlock(dqm); | |
1594 | return 0; | |
1595 | } | |
1596 | ||
cc009e61 | 1597 | if (!dqm->is_hws_hang) { |
8dc1db31 | 1598 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
7cee6a68 | 1599 | unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD, false); |
cc009e61 MJ |
1600 | else |
1601 | remove_all_queues_mes(dqm); | |
1602 | } | |
1603 | ||
c2a77fde | 1604 | hanging = dqm->is_hws_hang || dqm->is_resetting; |
2c99a547 | 1605 | dqm->sched_running = false; |
64c7f8cf | 1606 | |
8dc1db31 | 1607 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
cc009e61 | 1608 | pm_release_ib(&dqm->packet_mgr); |
087d7641 | 1609 | |
a86aa3ca | 1610 | kfd_gtt_sa_free(dqm->dev, dqm->fence_mem); |
8dc1db31 | 1611 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
cc009e61 | 1612 | pm_uninit(&dqm->packet_mgr, hanging); |
4f942aae | 1613 | dqm_unlock(dqm); |
64c7f8cf BG |
1614 | |
1615 | return 0; | |
1616 | } | |
1617 | ||
1618 | static int create_kernel_queue_cpsch(struct device_queue_manager *dqm, | |
1619 | struct kernel_queue *kq, | |
1620 | struct qcm_process_device *qpd) | |
1621 | { | |
efeaed4d | 1622 | dqm_lock(dqm); |
b8cbab04 | 1623 | if (dqm->total_queue_count >= max_num_of_queues_per_device) { |
79775b62 | 1624 | pr_warn("Can't create new kernel queue because %d queues were already created\n", |
b8cbab04 | 1625 | dqm->total_queue_count); |
efeaed4d | 1626 | dqm_unlock(dqm); |
b8cbab04 OG |
1627 | return -EPERM; |
1628 | } | |
1629 | ||
1630 | /* | |
1631 | * Unconditionally increment this counter, regardless of the queue's | |
1632 | * type or whether the queue is active. | |
1633 | */ | |
1634 | dqm->total_queue_count++; | |
1635 | pr_debug("Total of %d queues are accountable so far\n", | |
1636 | dqm->total_queue_count); | |
1637 | ||
64c7f8cf | 1638 | list_add(&kq->list, &qpd->priv_queue_list); |
ab4d51d4 | 1639 | increment_queue_count(dqm, qpd, kq->queue); |
64c7f8cf | 1640 | qpd->is_debug = true; |
7cee6a68 JK |
1641 | execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, |
1642 | USE_DEFAULT_GRACE_PERIOD); | |
efeaed4d | 1643 | dqm_unlock(dqm); |
64c7f8cf BG |
1644 | |
1645 | return 0; | |
1646 | } | |
1647 | ||
1648 | static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm, | |
1649 | struct kernel_queue *kq, | |
1650 | struct qcm_process_device *qpd) | |
1651 | { | |
efeaed4d | 1652 | dqm_lock(dqm); |
64c7f8cf | 1653 | list_del(&kq->list); |
ab4d51d4 | 1654 | decrement_queue_count(dqm, qpd, kq->queue); |
64c7f8cf | 1655 | qpd->is_debug = false; |
7cee6a68 JK |
1656 | execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, |
1657 | USE_DEFAULT_GRACE_PERIOD); | |
b8cbab04 OG |
1658 | /* |
1659 | * Unconditionally decrement this counter, regardless of the queue's | |
1660 | * type. | |
1661 | */ | |
8b58f261 | 1662 | dqm->total_queue_count--; |
b8cbab04 OG |
1663 | pr_debug("Total of %d queues are accountable so far\n", |
1664 | dqm->total_queue_count); | |
efeaed4d | 1665 | dqm_unlock(dqm); |
64c7f8cf BG |
1666 | } |
1667 | ||
1668 | static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, | |
2485c12c | 1669 | struct qcm_process_device *qpd, |
42c6c482 | 1670 | const struct kfd_criu_queue_priv_data *qd, |
3a9822d7 | 1671 | const void *restore_mqd, const void *restore_ctl_stack) |
64c7f8cf BG |
1672 | { |
1673 | int retval; | |
8d5f3552 | 1674 | struct mqd_manager *mqd_mgr; |
64c7f8cf | 1675 | |
b8cbab04 | 1676 | if (dqm->total_queue_count >= max_num_of_queues_per_device) { |
79775b62 | 1677 | pr_warn("Can't create new usermode queue because %d queues were already created\n", |
b8cbab04 | 1678 | dqm->total_queue_count); |
70d488fb OZ |
1679 | retval = -EPERM; |
1680 | goto out; | |
b8cbab04 OG |
1681 | } |
1682 | ||
1b4670f6 OZ |
1683 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
1684 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
38bb4226 | 1685 | dqm_lock(dqm); |
2485c12c | 1686 | retval = allocate_sdma_queue(dqm, q, qd ? &qd->sdma_id : NULL); |
38bb4226 | 1687 | dqm_unlock(dqm); |
894a8293 | 1688 | if (retval) |
70d488fb | 1689 | goto out; |
e139cd2a | 1690 | } |
ef568db7 | 1691 | |
5bb6a8fa | 1692 | retval = allocate_doorbell(qpd, q, qd ? &qd->doorbell_id : NULL); |
ef568db7 FK |
1693 | if (retval) |
1694 | goto out_deallocate_sdma_queue; | |
1695 | ||
70d488fb OZ |
1696 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
1697 | q->properties.type)]; | |
70df8273 | 1698 | |
eec0b4cf OZ |
1699 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
1700 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) | |
1701 | dqm->asic_ops.init_sdma_vm(dqm, q, qpd); | |
373d7080 FK |
1702 | q->properties.tba_addr = qpd->tba_addr; |
1703 | q->properties.tma_addr = qpd->tma_addr; | |
70d488fb OZ |
1704 | q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties); |
1705 | if (!q->mqd_mem_obj) { | |
1706 | retval = -ENOMEM; | |
1707 | goto out_deallocate_doorbell; | |
1708 | } | |
70df8273 EH |
1709 | |
1710 | dqm_lock(dqm); | |
1711 | /* | |
1712 | * Eviction state logic: mark all queues as evicted, even ones | |
1713 | * not currently active. Restoring inactive queues later only | |
1714 | * updates the is_evicted flag but is a no-op otherwise. | |
1715 | */ | |
1716 | q->properties.is_evicted = !!qpd->evicted; | |
42c6c482 DYS |
1717 | |
1718 | if (qd) | |
1719 | mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr, | |
3a9822d7 DYS |
1720 | &q->properties, restore_mqd, restore_ctl_stack, |
1721 | qd->ctl_stack_size); | |
42c6c482 DYS |
1722 | else |
1723 | mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, | |
1724 | &q->gart_mqd_addr, &q->properties); | |
89cd9d23 | 1725 | |
64c7f8cf | 1726 | list_add(&q->list, &qpd->queues_list); |
bc920fd4 | 1727 | qpd->queue_count++; |
f38abc15 | 1728 | |
64c7f8cf | 1729 | if (q->properties.is_active) { |
ab4d51d4 | 1730 | increment_queue_count(dqm, qpd, q); |
b42902f4 | 1731 | |
8dc1db31 | 1732 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
cc009e61 | 1733 | retval = execute_queues_cpsch(dqm, |
7cee6a68 | 1734 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); |
cc3cb791 | 1735 | else |
cc009e61 | 1736 | retval = add_queue_mes(dqm, q, qpd); |
cc3cb791 | 1737 | if (retval) |
1738 | goto cleanup_queue; | |
64c7f8cf BG |
1739 | } |
1740 | ||
b8cbab04 OG |
1741 | /* |
1742 | * Unconditionally increment this counter, regardless of the queue's | |
1743 | * type or whether the queue is active. | |
1744 | */ | |
1745 | dqm->total_queue_count++; | |
1746 | ||
1747 | pr_debug("Total of %d queues are accountable so far\n", | |
1748 | dqm->total_queue_count); | |
1749 | ||
efeaed4d | 1750 | dqm_unlock(dqm); |
72a01d23 FK |
1751 | return retval; |
1752 | ||
cc009e61 MJ |
1753 | cleanup_queue: |
1754 | qpd->queue_count--; | |
1755 | list_del(&q->list); | |
1756 | if (q->properties.is_active) | |
1757 | decrement_queue_count(dqm, qpd, q); | |
1758 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); | |
1759 | dqm_unlock(dqm); | |
70d488fb OZ |
1760 | out_deallocate_doorbell: |
1761 | deallocate_doorbell(qpd, q); | |
72a01d23 | 1762 | out_deallocate_sdma_queue: |
1b4670f6 | 1763 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
38bb4226 OZ |
1764 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { |
1765 | dqm_lock(dqm); | |
1b4670f6 | 1766 | deallocate_sdma_queue(dqm, q); |
38bb4226 OZ |
1767 | dqm_unlock(dqm); |
1768 | } | |
70d488fb | 1769 | out: |
64c7f8cf BG |
1770 | return retval; |
1771 | } | |
1772 | ||
b010affe QH |
1773 | int amdkfd_fence_wait_timeout(uint64_t *fence_addr, |
1774 | uint64_t fence_value, | |
8c72c3d7 | 1775 | unsigned int timeout_ms) |
64c7f8cf | 1776 | { |
8c72c3d7 | 1777 | unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies; |
64c7f8cf BG |
1778 | |
1779 | while (*fence_addr != fence_value) { | |
8c72c3d7 | 1780 | if (time_after(jiffies, end_jiffies)) { |
79775b62 | 1781 | pr_err("qcm fence wait loop timeout expired\n"); |
0e9a860c YZ |
1782 | /* In HWS case, this is used to halt the driver thread |
1783 | * in order not to mess up CP states before doing | |
1784 | * scandumps for FW debugging. | |
1785 | */ | |
1786 | while (halt_if_hws_hang) | |
1787 | schedule(); | |
1788 | ||
64c7f8cf BG |
1789 | return -ETIME; |
1790 | } | |
99331a51 | 1791 | schedule(); |
64c7f8cf BG |
1792 | } |
1793 | ||
1794 | return 0; | |
1795 | } | |
1796 | ||
60a00956 FK |
1797 | /* dqm->lock mutex has to be locked before calling this function */ |
1798 | static int map_queues_cpsch(struct device_queue_manager *dqm) | |
1799 | { | |
1800 | int retval; | |
1801 | ||
2c99a547 PY |
1802 | if (!dqm->sched_running) |
1803 | return 0; | |
81b820b3 | 1804 | if (dqm->active_queue_count <= 0 || dqm->processes_count <= 0) |
60a00956 | 1805 | return 0; |
60a00956 FK |
1806 | if (dqm->active_runlist) |
1807 | return 0; | |
1808 | ||
9af5379c | 1809 | retval = pm_send_runlist(&dqm->packet_mgr, &dqm->queues); |
14328aa5 | 1810 | pr_debug("%s sent runlist\n", __func__); |
60a00956 FK |
1811 | if (retval) { |
1812 | pr_err("failed to execute runlist\n"); | |
1813 | return retval; | |
1814 | } | |
1815 | dqm->active_runlist = true; | |
1816 | ||
1817 | return retval; | |
1818 | } | |
1819 | ||
ac30c783 | 1820 | /* dqm->lock mutex has to be locked before calling this function */ |
7da2bcf8 | 1821 | static int unmap_queues_cpsch(struct device_queue_manager *dqm, |
4465f466 | 1822 | enum kfd_unmap_queues_filter filter, |
7cee6a68 JK |
1823 | uint32_t filter_param, |
1824 | uint32_t grace_period, | |
1825 | bool reset) | |
64c7f8cf | 1826 | { |
9fd3f1bf | 1827 | int retval = 0; |
51a0f459 | 1828 | struct mqd_manager *mqd_mgr; |
64c7f8cf | 1829 | |
2c99a547 PY |
1830 | if (!dqm->sched_running) |
1831 | return 0; | |
b8c20c74 | 1832 | if (dqm->is_hws_hang || dqm->is_resetting) |
73ea648d | 1833 | return -EIO; |
991ca8ee | 1834 | if (!dqm->active_runlist) |
ac30c783 | 1835 | return retval; |
bcea3081 | 1836 | |
7cee6a68 JK |
1837 | if (grace_period != USE_DEFAULT_GRACE_PERIOD) { |
1838 | retval = pm_update_grace_period(&dqm->packet_mgr, grace_period); | |
1839 | if (retval) | |
1840 | return retval; | |
1841 | } | |
1842 | ||
d2cb0b21 | 1843 | retval = pm_send_unmap_queue(&dqm->packet_mgr, filter, filter_param, reset); |
4eacc26b | 1844 | if (retval) |
ac30c783 | 1845 | return retval; |
64c7f8cf BG |
1846 | |
1847 | *dqm->fence_addr = KFD_FENCE_INIT; | |
9af5379c | 1848 | pm_send_query_status(&dqm->packet_mgr, dqm->fence_gpu_addr, |
64c7f8cf BG |
1849 | KFD_FENCE_COMPLETED); |
1850 | /* should be timed out */ | |
c3447e81 | 1851 | retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED, |
14328aa5 | 1852 | queue_preemption_timeout_ms); |
09c34e8d FK |
1853 | if (retval) { |
1854 | pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n"); | |
cc009e61 | 1855 | kfd_hws_hang(dqm); |
ac30c783 | 1856 | return retval; |
09c34e8d | 1857 | } |
9fd3f1bf | 1858 | |
51a0f459 OZ |
1859 | /* In the current MEC firmware implementation, if compute queue |
1860 | * doesn't response to the preemption request in time, HIQ will | |
1861 | * abandon the unmap request without returning any timeout error | |
1862 | * to driver. Instead, MEC firmware will log the doorbell of the | |
1863 | * unresponding compute queue to HIQ.MQD.queue_doorbell_id fields. | |
1864 | * To make sure the queue unmap was successful, driver need to | |
1865 | * check those fields | |
1866 | */ | |
1867 | mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]; | |
9af5379c | 1868 | if (mqd_mgr->read_doorbell_id(dqm->packet_mgr.priv_queue->queue->mqd)) { |
51a0f459 OZ |
1869 | pr_err("HIQ MQD's queue_doorbell_id0 is not 0, Queue preemption time out\n"); |
1870 | while (halt_if_hws_hang) | |
1871 | schedule(); | |
1872 | return -ETIME; | |
1873 | } | |
1874 | ||
7cee6a68 JK |
1875 | /* We need to reset the grace period value for this device */ |
1876 | if (grace_period != USE_DEFAULT_GRACE_PERIOD) { | |
1877 | if (pm_update_grace_period(&dqm->packet_mgr, | |
1878 | USE_DEFAULT_GRACE_PERIOD)) | |
1879 | pr_err("Failed to reset grace period\n"); | |
1880 | } | |
1881 | ||
9af5379c | 1882 | pm_release_ib(&dqm->packet_mgr); |
64c7f8cf BG |
1883 | dqm->active_runlist = false; |
1884 | ||
64c7f8cf BG |
1885 | return retval; |
1886 | } | |
1887 | ||
dec63443 TZ |
1888 | /* only for compute queue */ |
1889 | static int reset_queues_cpsch(struct device_queue_manager *dqm, | |
1890 | uint16_t pasid) | |
1891 | { | |
1892 | int retval; | |
1893 | ||
1894 | dqm_lock(dqm); | |
1895 | ||
1896 | retval = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_BY_PASID, | |
7cee6a68 | 1897 | pasid, USE_DEFAULT_GRACE_PERIOD, true); |
dec63443 TZ |
1898 | |
1899 | dqm_unlock(dqm); | |
1900 | return retval; | |
1901 | } | |
1902 | ||
ac30c783 | 1903 | /* dqm->lock mutex has to be locked before calling this function */ |
c4744e24 YZ |
1904 | static int execute_queues_cpsch(struct device_queue_manager *dqm, |
1905 | enum kfd_unmap_queues_filter filter, | |
7cee6a68 JK |
1906 | uint32_t filter_param, |
1907 | uint32_t grace_period) | |
64c7f8cf BG |
1908 | { |
1909 | int retval; | |
1910 | ||
73ea648d SL |
1911 | if (dqm->is_hws_hang) |
1912 | return -EIO; | |
7cee6a68 | 1913 | retval = unmap_queues_cpsch(dqm, filter, filter_param, grace_period, false); |
09c34e8d | 1914 | if (retval) |
ac30c783 | 1915 | return retval; |
64c7f8cf | 1916 | |
60a00956 | 1917 | return map_queues_cpsch(dqm); |
64c7f8cf BG |
1918 | } |
1919 | ||
1920 | static int destroy_queue_cpsch(struct device_queue_manager *dqm, | |
1921 | struct qcm_process_device *qpd, | |
1922 | struct queue *q) | |
1923 | { | |
1924 | int retval; | |
8d5f3552 | 1925 | struct mqd_manager *mqd_mgr; |
d69fd951 MJ |
1926 | uint64_t sdma_val = 0; |
1927 | struct kfd_process_device *pdd = qpd_to_pdd(qpd); | |
1928 | ||
1929 | /* Get the SDMA queue stats */ | |
1930 | if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) || | |
1931 | (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { | |
818b0324 | 1932 | retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr, |
d69fd951 MJ |
1933 | &sdma_val); |
1934 | if (retval) | |
1935 | pr_err("Failed to read SDMA queue counter for queue: %d\n", | |
1936 | q->properties.queue_id); | |
1937 | } | |
992839ad | 1938 | |
64c7f8cf BG |
1939 | retval = 0; |
1940 | ||
1941 | /* remove queue from list to prevent rescheduling after preemption */ | |
efeaed4d | 1942 | dqm_lock(dqm); |
992839ad YS |
1943 | |
1944 | if (qpd->is_debug) { | |
1945 | /* | |
1946 | * error, currently we do not allow to destroy a queue | |
1947 | * of a currently debugged process | |
1948 | */ | |
1949 | retval = -EBUSY; | |
1950 | goto failed_try_destroy_debugged_queue; | |
1951 | ||
1952 | } | |
1953 | ||
fdfa090b OZ |
1954 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
1955 | q->properties.type)]; | |
64c7f8cf | 1956 | |
ef568db7 FK |
1957 | deallocate_doorbell(qpd, q); |
1958 | ||
d69fd951 MJ |
1959 | if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) || |
1960 | (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { | |
1b4670f6 | 1961 | deallocate_sdma_queue(dqm, q); |
d69fd951 MJ |
1962 | pdd->sdma_past_activity_counter += sdma_val; |
1963 | } | |
bcea3081 | 1964 | |
64c7f8cf | 1965 | list_del(&q->list); |
bc920fd4 | 1966 | qpd->queue_count--; |
8c07f33e | 1967 | if (q->properties.is_active) { |
8dc1db31 | 1968 | if (!dqm->dev->kfd->shared_resources.enable_mes) { |
8c07f33e PY |
1969 | decrement_queue_count(dqm, qpd, q); |
1970 | retval = execute_queues_cpsch(dqm, | |
7cee6a68 JK |
1971 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, |
1972 | USE_DEFAULT_GRACE_PERIOD); | |
8c07f33e PY |
1973 | if (retval == -ETIME) |
1974 | qpd->reset_wavefronts = true; | |
1975 | } else { | |
1976 | retval = remove_queue_mes(dqm, q, qpd); | |
1977 | } | |
1978 | } | |
64c7f8cf | 1979 | |
b8cbab04 OG |
1980 | /* |
1981 | * Unconditionally decrement this counter, regardless of the queue's | |
1982 | * type | |
1983 | */ | |
1984 | dqm->total_queue_count--; | |
1985 | pr_debug("Total of %d queues are accountable so far\n", | |
1986 | dqm->total_queue_count); | |
64c7f8cf | 1987 | |
efeaed4d | 1988 | dqm_unlock(dqm); |
64c7f8cf | 1989 | |
8636e53c OZ |
1990 | /* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */ |
1991 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); | |
89cd9d23 | 1992 | |
9e827224 | 1993 | return retval; |
64c7f8cf | 1994 | |
992839ad YS |
1995 | failed_try_destroy_debugged_queue: |
1996 | ||
efeaed4d | 1997 | dqm_unlock(dqm); |
64c7f8cf BG |
1998 | return retval; |
1999 | } | |
2000 | ||
2001 | /* | |
2002 | * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to | |
2003 | * stay in user mode. | |
2004 | */ | |
2005 | #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL | |
2006 | /* APE1 limit is inclusive and 64K aligned. */ | |
2007 | #define APE1_LIMIT_ALIGNMENT 0xFFFF | |
2008 | ||
2009 | static bool set_cache_memory_policy(struct device_queue_manager *dqm, | |
2010 | struct qcm_process_device *qpd, | |
2011 | enum cache_policy default_policy, | |
2012 | enum cache_policy alternate_policy, | |
2013 | void __user *alternate_aperture_base, | |
2014 | uint64_t alternate_aperture_size) | |
2015 | { | |
bed4f110 FK |
2016 | bool retval = true; |
2017 | ||
2018 | if (!dqm->asic_ops.set_cache_memory_policy) | |
2019 | return retval; | |
64c7f8cf | 2020 | |
efeaed4d | 2021 | dqm_lock(dqm); |
64c7f8cf BG |
2022 | |
2023 | if (alternate_aperture_size == 0) { | |
2024 | /* base > limit disables APE1 */ | |
2025 | qpd->sh_mem_ape1_base = 1; | |
2026 | qpd->sh_mem_ape1_limit = 0; | |
2027 | } else { | |
2028 | /* | |
2029 | * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]}, | |
2030 | * SH_MEM_APE1_BASE[31:0], 0x0000 } | |
2031 | * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]}, | |
2032 | * SH_MEM_APE1_LIMIT[31:0], 0xFFFF } | |
2033 | * Verify that the base and size parameters can be | |
2034 | * represented in this format and convert them. | |
2035 | * Additionally restrict APE1 to user-mode addresses. | |
2036 | */ | |
2037 | ||
2038 | uint64_t base = (uintptr_t)alternate_aperture_base; | |
2039 | uint64_t limit = base + alternate_aperture_size - 1; | |
2040 | ||
ab7c1648 KR |
2041 | if (limit <= base || (base & APE1_FIXED_BITS_MASK) != 0 || |
2042 | (limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT) { | |
2043 | retval = false; | |
64c7f8cf | 2044 | goto out; |
ab7c1648 | 2045 | } |
64c7f8cf BG |
2046 | |
2047 | qpd->sh_mem_ape1_base = base >> 16; | |
2048 | qpd->sh_mem_ape1_limit = limit >> 16; | |
2049 | } | |
2050 | ||
bfd5e378 | 2051 | retval = dqm->asic_ops.set_cache_memory_policy( |
a22fc854 BG |
2052 | dqm, |
2053 | qpd, | |
2054 | default_policy, | |
2055 | alternate_policy, | |
2056 | alternate_aperture_base, | |
2057 | alternate_aperture_size); | |
64c7f8cf | 2058 | |
d146c5a7 | 2059 | if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0)) |
64c7f8cf BG |
2060 | program_sh_mem_settings(dqm, qpd); |
2061 | ||
79775b62 | 2062 | pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n", |
64c7f8cf BG |
2063 | qpd->sh_mem_config, qpd->sh_mem_ape1_base, |
2064 | qpd->sh_mem_ape1_limit); | |
2065 | ||
64c7f8cf | 2066 | out: |
efeaed4d | 2067 | dqm_unlock(dqm); |
ab7c1648 | 2068 | return retval; |
64c7f8cf BG |
2069 | } |
2070 | ||
9fd3f1bf FK |
2071 | static int process_termination_nocpsch(struct device_queue_manager *dqm, |
2072 | struct qcm_process_device *qpd) | |
2073 | { | |
a7b2451d | 2074 | struct queue *q; |
9fd3f1bf FK |
2075 | struct device_process_node *cur, *next_dpn; |
2076 | int retval = 0; | |
32cce8bc | 2077 | bool found = false; |
9fd3f1bf | 2078 | |
efeaed4d | 2079 | dqm_lock(dqm); |
9fd3f1bf FK |
2080 | |
2081 | /* Clear all user mode queues */ | |
a7b2451d AL |
2082 | while (!list_empty(&qpd->queues_list)) { |
2083 | struct mqd_manager *mqd_mgr; | |
9fd3f1bf FK |
2084 | int ret; |
2085 | ||
a7b2451d AL |
2086 | q = list_first_entry(&qpd->queues_list, struct queue, list); |
2087 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( | |
2088 | q->properties.type)]; | |
9fd3f1bf FK |
2089 | ret = destroy_queue_nocpsch_locked(dqm, qpd, q); |
2090 | if (ret) | |
2091 | retval = ret; | |
a7b2451d AL |
2092 | dqm_unlock(dqm); |
2093 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); | |
2094 | dqm_lock(dqm); | |
9fd3f1bf FK |
2095 | } |
2096 | ||
2097 | /* Unregister process */ | |
2098 | list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) { | |
2099 | if (qpd == cur->qpd) { | |
2100 | list_del(&cur->list); | |
2101 | kfree(cur); | |
2102 | dqm->processes_count--; | |
32cce8bc | 2103 | found = true; |
9fd3f1bf FK |
2104 | break; |
2105 | } | |
2106 | } | |
2107 | ||
efeaed4d | 2108 | dqm_unlock(dqm); |
32cce8bc FK |
2109 | |
2110 | /* Outside the DQM lock because under the DQM lock we can't do | |
2111 | * reclaim or take other locks that others hold while reclaiming. | |
2112 | */ | |
2113 | if (found) | |
2114 | kfd_dec_compute_active(dqm->dev); | |
2115 | ||
9fd3f1bf FK |
2116 | return retval; |
2117 | } | |
2118 | ||
5df099e8 JC |
2119 | static int get_wave_state(struct device_queue_manager *dqm, |
2120 | struct queue *q, | |
2121 | void __user *ctl_stack, | |
2122 | u32 *ctl_stack_used_size, | |
2123 | u32 *save_area_used_size) | |
2124 | { | |
4e6c6fc1 | 2125 | struct mqd_manager *mqd_mgr; |
5df099e8 JC |
2126 | |
2127 | dqm_lock(dqm); | |
2128 | ||
d7c0b047 | 2129 | mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP]; |
5df099e8 | 2130 | |
63f6e012 | 2131 | if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE || |
8dc1db31 | 2132 | q->properties.is_active || !q->device->kfd->cwsr_enabled || |
63f6e012 JK |
2133 | !mqd_mgr->get_wave_state) { |
2134 | dqm_unlock(dqm); | |
2135 | return -EINVAL; | |
5df099e8 JC |
2136 | } |
2137 | ||
5df099e8 | 2138 | dqm_unlock(dqm); |
63f6e012 JK |
2139 | |
2140 | /* | |
2141 | * get_wave_state is outside the dqm lock to prevent circular locking | |
2142 | * and the queue should be protected against destruction by the process | |
2143 | * lock. | |
2144 | */ | |
7fe51e6f MJ |
2145 | return mqd_mgr->get_wave_state(mqd_mgr, q->mqd, &q->properties, |
2146 | ctl_stack, ctl_stack_used_size, save_area_used_size); | |
5df099e8 | 2147 | } |
9fd3f1bf | 2148 | |
42c6c482 DYS |
2149 | static void get_queue_checkpoint_info(struct device_queue_manager *dqm, |
2150 | const struct queue *q, | |
3a9822d7 DYS |
2151 | u32 *mqd_size, |
2152 | u32 *ctl_stack_size) | |
42c6c482 DYS |
2153 | { |
2154 | struct mqd_manager *mqd_mgr; | |
2155 | enum KFD_MQD_TYPE mqd_type = | |
2156 | get_mqd_type_from_queue_type(q->properties.type); | |
2157 | ||
2158 | dqm_lock(dqm); | |
2159 | mqd_mgr = dqm->mqd_mgrs[mqd_type]; | |
2160 | *mqd_size = mqd_mgr->mqd_size; | |
3a9822d7 DYS |
2161 | *ctl_stack_size = 0; |
2162 | ||
2163 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE && mqd_mgr->get_checkpoint_info) | |
2164 | mqd_mgr->get_checkpoint_info(mqd_mgr, q->mqd, ctl_stack_size); | |
42c6c482 DYS |
2165 | |
2166 | dqm_unlock(dqm); | |
2167 | } | |
2168 | ||
2169 | static int checkpoint_mqd(struct device_queue_manager *dqm, | |
2170 | const struct queue *q, | |
3a9822d7 DYS |
2171 | void *mqd, |
2172 | void *ctl_stack) | |
42c6c482 DYS |
2173 | { |
2174 | struct mqd_manager *mqd_mgr; | |
2175 | int r = 0; | |
2176 | enum KFD_MQD_TYPE mqd_type = | |
2177 | get_mqd_type_from_queue_type(q->properties.type); | |
2178 | ||
2179 | dqm_lock(dqm); | |
2180 | ||
8dc1db31 | 2181 | if (q->properties.is_active || !q->device->kfd->cwsr_enabled) { |
42c6c482 DYS |
2182 | r = -EINVAL; |
2183 | goto dqm_unlock; | |
2184 | } | |
2185 | ||
2186 | mqd_mgr = dqm->mqd_mgrs[mqd_type]; | |
2187 | if (!mqd_mgr->checkpoint_mqd) { | |
2188 | r = -EOPNOTSUPP; | |
2189 | goto dqm_unlock; | |
2190 | } | |
2191 | ||
3a9822d7 | 2192 | mqd_mgr->checkpoint_mqd(mqd_mgr, q->mqd, mqd, ctl_stack); |
42c6c482 DYS |
2193 | |
2194 | dqm_unlock: | |
2195 | dqm_unlock(dqm); | |
2196 | return r; | |
2197 | } | |
2198 | ||
9fd3f1bf FK |
2199 | static int process_termination_cpsch(struct device_queue_manager *dqm, |
2200 | struct qcm_process_device *qpd) | |
2201 | { | |
2202 | int retval; | |
56f221b6 | 2203 | struct queue *q; |
9fd3f1bf | 2204 | struct kernel_queue *kq, *kq_next; |
8d5f3552 | 2205 | struct mqd_manager *mqd_mgr; |
9fd3f1bf FK |
2206 | struct device_process_node *cur, *next_dpn; |
2207 | enum kfd_unmap_queues_filter filter = | |
2208 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES; | |
32cce8bc | 2209 | bool found = false; |
9fd3f1bf FK |
2210 | |
2211 | retval = 0; | |
2212 | ||
efeaed4d | 2213 | dqm_lock(dqm); |
9fd3f1bf FK |
2214 | |
2215 | /* Clean all kernel queues */ | |
2216 | list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) { | |
2217 | list_del(&kq->list); | |
ab4d51d4 | 2218 | decrement_queue_count(dqm, qpd, kq->queue); |
9fd3f1bf FK |
2219 | qpd->is_debug = false; |
2220 | dqm->total_queue_count--; | |
2221 | filter = KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES; | |
2222 | } | |
2223 | ||
2224 | /* Clear all user mode queues */ | |
2225 | list_for_each_entry(q, &qpd->queues_list, list) { | |
c7637c95 | 2226 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) |
1b4670f6 | 2227 | deallocate_sdma_queue(dqm, q); |
c7637c95 | 2228 | else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) |
1b4670f6 | 2229 | deallocate_sdma_queue(dqm, q); |
9fd3f1bf | 2230 | |
cc009e61 | 2231 | if (q->properties.is_active) { |
ab4d51d4 | 2232 | decrement_queue_count(dqm, qpd, q); |
9fd3f1bf | 2233 | |
8dc1db31 | 2234 | if (dqm->dev->kfd->shared_resources.enable_mes) { |
cc009e61 MJ |
2235 | retval = remove_queue_mes(dqm, q, qpd); |
2236 | if (retval) | |
2237 | pr_err("Failed to remove queue %d\n", | |
2238 | q->properties.queue_id); | |
2239 | } | |
2240 | } | |
2241 | ||
9fd3f1bf FK |
2242 | dqm->total_queue_count--; |
2243 | } | |
2244 | ||
2245 | /* Unregister process */ | |
2246 | list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) { | |
2247 | if (qpd == cur->qpd) { | |
2248 | list_del(&cur->list); | |
2249 | kfree(cur); | |
2250 | dqm->processes_count--; | |
32cce8bc | 2251 | found = true; |
9fd3f1bf FK |
2252 | break; |
2253 | } | |
2254 | } | |
2255 | ||
8dc1db31 | 2256 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
7cee6a68 | 2257 | retval = execute_queues_cpsch(dqm, filter, 0, USE_DEFAULT_GRACE_PERIOD); |
cc009e61 | 2258 | |
73ea648d | 2259 | if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) { |
9fd3f1bf FK |
2260 | pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev); |
2261 | dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process); | |
2262 | qpd->reset_wavefronts = false; | |
2263 | } | |
2264 | ||
89cd9d23 | 2265 | /* Lastly, free mqd resources. |
8636e53c | 2266 | * Do free_mqd() after dqm_unlock to avoid circular locking. |
89cd9d23 | 2267 | */ |
56f221b6 | 2268 | while (!list_empty(&qpd->queues_list)) { |
2269 | q = list_first_entry(&qpd->queues_list, struct queue, list); | |
fdfa090b OZ |
2270 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
2271 | q->properties.type)]; | |
9fd3f1bf | 2272 | list_del(&q->list); |
bc920fd4 | 2273 | qpd->queue_count--; |
56f221b6 | 2274 | dqm_unlock(dqm); |
8636e53c | 2275 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); |
56f221b6 | 2276 | dqm_lock(dqm); |
9fd3f1bf | 2277 | } |
56f221b6 | 2278 | dqm_unlock(dqm); |
2279 | ||
2280 | /* Outside the DQM lock because under the DQM lock we can't do | |
2281 | * reclaim or take other locks that others hold while reclaiming. | |
2282 | */ | |
2283 | if (found) | |
2284 | kfd_dec_compute_active(dqm->dev); | |
9fd3f1bf | 2285 | |
9fd3f1bf FK |
2286 | return retval; |
2287 | } | |
2288 | ||
fdfa090b OZ |
2289 | static int init_mqd_managers(struct device_queue_manager *dqm) |
2290 | { | |
2291 | int i, j; | |
2292 | struct mqd_manager *mqd_mgr; | |
2293 | ||
2294 | for (i = 0; i < KFD_MQD_TYPE_MAX; i++) { | |
2295 | mqd_mgr = dqm->asic_ops.mqd_manager_init(i, dqm->dev); | |
2296 | if (!mqd_mgr) { | |
2297 | pr_err("mqd manager [%d] initialization failed\n", i); | |
2298 | goto out_free; | |
2299 | } | |
2300 | dqm->mqd_mgrs[i] = mqd_mgr; | |
2301 | } | |
2302 | ||
2303 | return 0; | |
2304 | ||
2305 | out_free: | |
2306 | for (j = 0; j < i; j++) { | |
2307 | kfree(dqm->mqd_mgrs[j]); | |
2308 | dqm->mqd_mgrs[j] = NULL; | |
2309 | } | |
2310 | ||
2311 | return -ENOMEM; | |
2312 | } | |
11614c36 OZ |
2313 | |
2314 | /* Allocate one hiq mqd (HWS) and all SDMA mqd in a continuous trunk*/ | |
2315 | static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm) | |
2316 | { | |
2317 | int retval; | |
8dc1db31 | 2318 | struct kfd_node *dev = dqm->dev; |
11614c36 OZ |
2319 | struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd; |
2320 | uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size * | |
c7637c95 | 2321 | get_num_all_sdma_engines(dqm) * |
8dc1db31 | 2322 | dev->kfd->device_info.num_sdma_queues_per_engine + |
2f77b9a2 | 2323 | (dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size * |
c4050ff1 | 2324 | NUM_XCC(dqm->dev->xcc_mask)); |
11614c36 | 2325 | |
6bfc7c7e | 2326 | retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, size, |
11614c36 | 2327 | &(mem_obj->gtt_mem), &(mem_obj->gpu_addr), |
f2cc50ce | 2328 | (void *)&(mem_obj->cpu_ptr), false); |
11614c36 OZ |
2329 | |
2330 | return retval; | |
2331 | } | |
2332 | ||
8dc1db31 | 2333 | struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev) |
64c7f8cf BG |
2334 | { |
2335 | struct device_queue_manager *dqm; | |
2336 | ||
79775b62 | 2337 | pr_debug("Loading device queue manager\n"); |
a22fc854 | 2338 | |
dbf56ab1 | 2339 | dqm = kzalloc(sizeof(*dqm), GFP_KERNEL); |
64c7f8cf BG |
2340 | if (!dqm) |
2341 | return NULL; | |
2342 | ||
7eb0502a | 2343 | switch (dev->adev->asic_type) { |
d146c5a7 FK |
2344 | /* HWS is not available on Hawaii. */ |
2345 | case CHIP_HAWAII: | |
2346 | /* HWS depends on CWSR for timely dequeue. CWSR is not | |
2347 | * available on Tonga. | |
2348 | * | |
2349 | * FIXME: This argument also applies to Kaveri. | |
2350 | */ | |
2351 | case CHIP_TONGA: | |
2352 | dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS; | |
2353 | break; | |
2354 | default: | |
2355 | dqm->sched_policy = sched_policy; | |
2356 | break; | |
2357 | } | |
2358 | ||
64c7f8cf | 2359 | dqm->dev = dev; |
d146c5a7 | 2360 | switch (dqm->sched_policy) { |
64c7f8cf BG |
2361 | case KFD_SCHED_POLICY_HWS: |
2362 | case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: | |
2363 | /* initialize dqm for cp scheduling */ | |
45c9a5e4 OG |
2364 | dqm->ops.create_queue = create_queue_cpsch; |
2365 | dqm->ops.initialize = initialize_cpsch; | |
2366 | dqm->ops.start = start_cpsch; | |
2367 | dqm->ops.stop = stop_cpsch; | |
09c34e8d | 2368 | dqm->ops.pre_reset = pre_reset; |
45c9a5e4 OG |
2369 | dqm->ops.destroy_queue = destroy_queue_cpsch; |
2370 | dqm->ops.update_queue = update_queue; | |
58dcd5bf YZ |
2371 | dqm->ops.register_process = register_process; |
2372 | dqm->ops.unregister_process = unregister_process; | |
2373 | dqm->ops.uninitialize = uninitialize; | |
45c9a5e4 OG |
2374 | dqm->ops.create_kernel_queue = create_kernel_queue_cpsch; |
2375 | dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch; | |
2376 | dqm->ops.set_cache_memory_policy = set_cache_memory_policy; | |
9fd3f1bf | 2377 | dqm->ops.process_termination = process_termination_cpsch; |
26103436 FK |
2378 | dqm->ops.evict_process_queues = evict_process_queues_cpsch; |
2379 | dqm->ops.restore_process_queues = restore_process_queues_cpsch; | |
5df099e8 | 2380 | dqm->ops.get_wave_state = get_wave_state; |
dec63443 | 2381 | dqm->ops.reset_queues = reset_queues_cpsch; |
42c6c482 DYS |
2382 | dqm->ops.get_queue_checkpoint_info = get_queue_checkpoint_info; |
2383 | dqm->ops.checkpoint_mqd = checkpoint_mqd; | |
64c7f8cf BG |
2384 | break; |
2385 | case KFD_SCHED_POLICY_NO_HWS: | |
2386 | /* initialize dqm for no cp scheduling */ | |
45c9a5e4 OG |
2387 | dqm->ops.start = start_nocpsch; |
2388 | dqm->ops.stop = stop_nocpsch; | |
09c34e8d | 2389 | dqm->ops.pre_reset = pre_reset; |
45c9a5e4 OG |
2390 | dqm->ops.create_queue = create_queue_nocpsch; |
2391 | dqm->ops.destroy_queue = destroy_queue_nocpsch; | |
2392 | dqm->ops.update_queue = update_queue; | |
58dcd5bf YZ |
2393 | dqm->ops.register_process = register_process; |
2394 | dqm->ops.unregister_process = unregister_process; | |
45c9a5e4 | 2395 | dqm->ops.initialize = initialize_nocpsch; |
58dcd5bf | 2396 | dqm->ops.uninitialize = uninitialize; |
45c9a5e4 | 2397 | dqm->ops.set_cache_memory_policy = set_cache_memory_policy; |
9fd3f1bf | 2398 | dqm->ops.process_termination = process_termination_nocpsch; |
26103436 FK |
2399 | dqm->ops.evict_process_queues = evict_process_queues_nocpsch; |
2400 | dqm->ops.restore_process_queues = | |
2401 | restore_process_queues_nocpsch; | |
5df099e8 | 2402 | dqm->ops.get_wave_state = get_wave_state; |
42c6c482 DYS |
2403 | dqm->ops.get_queue_checkpoint_info = get_queue_checkpoint_info; |
2404 | dqm->ops.checkpoint_mqd = checkpoint_mqd; | |
64c7f8cf BG |
2405 | break; |
2406 | default: | |
d146c5a7 | 2407 | pr_err("Invalid scheduling policy %d\n", dqm->sched_policy); |
32fa8219 | 2408 | goto out_free; |
64c7f8cf BG |
2409 | } |
2410 | ||
7eb0502a | 2411 | switch (dev->adev->asic_type) { |
a22fc854 | 2412 | case CHIP_CARRIZO: |
bfd5e378 | 2413 | device_queue_manager_init_vi(&dqm->asic_ops); |
300dec95 OG |
2414 | break; |
2415 | ||
a22fc854 | 2416 | case CHIP_KAVERI: |
bfd5e378 | 2417 | device_queue_manager_init_cik(&dqm->asic_ops); |
300dec95 | 2418 | break; |
97672cbe FK |
2419 | |
2420 | case CHIP_HAWAII: | |
2421 | device_queue_manager_init_cik_hawaii(&dqm->asic_ops); | |
2422 | break; | |
2423 | ||
2424 | case CHIP_TONGA: | |
2425 | case CHIP_FIJI: | |
2426 | case CHIP_POLARIS10: | |
2427 | case CHIP_POLARIS11: | |
846a44d7 | 2428 | case CHIP_POLARIS12: |
ed81cd6e | 2429 | case CHIP_VEGAM: |
97672cbe FK |
2430 | device_queue_manager_init_vi_tonga(&dqm->asic_ops); |
2431 | break; | |
bed4f110 | 2432 | |
e596b903 | 2433 | default: |
cc009e61 MJ |
2434 | if (KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0)) |
2435 | device_queue_manager_init_v11(&dqm->asic_ops); | |
2436 | else if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1)) | |
e4804a39 GS |
2437 | device_queue_manager_init_v10_navi10(&dqm->asic_ops); |
2438 | else if (KFD_GC_VERSION(dev) >= IP_VERSION(9, 0, 1)) | |
2439 | device_queue_manager_init_v9(&dqm->asic_ops); | |
2440 | else { | |
2441 | WARN(1, "Unexpected ASIC family %u", | |
7eb0502a | 2442 | dev->adev->asic_type); |
e4804a39 GS |
2443 | goto out_free; |
2444 | } | |
a22fc854 BG |
2445 | } |
2446 | ||
fdfa090b OZ |
2447 | if (init_mqd_managers(dqm)) |
2448 | goto out_free; | |
2449 | ||
8dc1db31 | 2450 | if (!dev->kfd->shared_resources.enable_mes && allocate_hiq_sdma_mqd(dqm)) { |
11614c36 OZ |
2451 | pr_err("Failed to allocate hiq sdma mqd trunk buffer\n"); |
2452 | goto out_free; | |
2453 | } | |
2454 | ||
32fa8219 FK |
2455 | if (!dqm->ops.initialize(dqm)) |
2456 | return dqm; | |
64c7f8cf | 2457 | |
32fa8219 FK |
2458 | out_free: |
2459 | kfree(dqm); | |
2460 | return NULL; | |
64c7f8cf BG |
2461 | } |
2462 | ||
8dc1db31 | 2463 | static void deallocate_hiq_sdma_mqd(struct kfd_node *dev, |
7fd5a6fb | 2464 | struct kfd_mem_obj *mqd) |
11614c36 OZ |
2465 | { |
2466 | WARN(!mqd, "No hiq sdma mqd trunk to free"); | |
2467 | ||
6bfc7c7e | 2468 | amdgpu_amdkfd_free_gtt_mem(dev->adev, mqd->gtt_mem); |
11614c36 OZ |
2469 | } |
2470 | ||
64c7f8cf BG |
2471 | void device_queue_manager_uninit(struct device_queue_manager *dqm) |
2472 | { | |
5e406012 | 2473 | dqm->ops.stop(dqm); |
45c9a5e4 | 2474 | dqm->ops.uninitialize(dqm); |
8dc1db31 | 2475 | if (!dqm->dev->kfd->shared_resources.enable_mes) |
2e2b9baf | 2476 | deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd); |
64c7f8cf BG |
2477 | kfree(dqm); |
2478 | } | |
851a645e | 2479 | |
03e5b167 | 2480 | int kfd_dqm_evict_pasid(struct device_queue_manager *dqm, u32 pasid) |
2640c3fa | 2481 | { |
2482 | struct kfd_process_device *pdd; | |
2483 | struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); | |
2484 | int ret = 0; | |
2485 | ||
2486 | if (!p) | |
2487 | return -EINVAL; | |
8a491bb3 | 2488 | WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid); |
2640c3fa | 2489 | pdd = kfd_get_process_device_data(dqm->dev, p); |
2490 | if (pdd) | |
2491 | ret = dqm->ops.evict_process_queues(dqm, &pdd->qpd); | |
2492 | kfd_unref_process(p); | |
2493 | ||
2494 | return ret; | |
2495 | } | |
2496 | ||
73ea648d SL |
2497 | static void kfd_process_hw_exception(struct work_struct *work) |
2498 | { | |
2499 | struct device_queue_manager *dqm = container_of(work, | |
2500 | struct device_queue_manager, hw_exception_work); | |
6bfc7c7e | 2501 | amdgpu_amdkfd_gpu_reset(dqm->dev->adev); |
73ea648d SL |
2502 | } |
2503 | ||
851a645e FK |
2504 | #if defined(CONFIG_DEBUG_FS) |
2505 | ||
2506 | static void seq_reg_dump(struct seq_file *m, | |
2507 | uint32_t (*dump)[2], uint32_t n_regs) | |
2508 | { | |
2509 | uint32_t i, count; | |
2510 | ||
2511 | for (i = 0, count = 0; i < n_regs; i++) { | |
2512 | if (count == 0 || | |
2513 | dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) { | |
2514 | seq_printf(m, "%s %08x: %08x", | |
2515 | i ? "\n" : "", | |
2516 | dump[i][0], dump[i][1]); | |
2517 | count = 7; | |
2518 | } else { | |
2519 | seq_printf(m, " %08x", dump[i][1]); | |
2520 | count--; | |
2521 | } | |
2522 | } | |
2523 | ||
2524 | seq_puts(m, "\n"); | |
2525 | } | |
2526 | ||
2527 | int dqm_debugfs_hqds(struct seq_file *m, void *data) | |
2528 | { | |
2529 | struct device_queue_manager *dqm = data; | |
c4050ff1 | 2530 | uint32_t xcc_mask = dqm->dev->xcc_mask; |
851a645e FK |
2531 | uint32_t (*dump)[2], n_regs; |
2532 | int pipe, queue; | |
c4050ff1 | 2533 | int r = 0, xcc_id; |
643e40d4 | 2534 | uint32_t sdma_engine_start; |
851a645e | 2535 | |
2c99a547 | 2536 | if (!dqm->sched_running) { |
2243f493 | 2537 | seq_puts(m, " Device is stopped\n"); |
2c99a547 PY |
2538 | return 0; |
2539 | } | |
2540 | ||
c4050ff1 | 2541 | for_each_inst(xcc_id, xcc_mask) { |
e2069a7b | 2542 | r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev, |
c4050ff1 LL |
2543 | KFD_CIK_HIQ_PIPE, |
2544 | KFD_CIK_HIQ_QUEUE, &dump, | |
2545 | &n_regs, xcc_id); | |
e2069a7b | 2546 | if (!r) { |
c4050ff1 LL |
2547 | seq_printf( |
2548 | m, | |
e2069a7b | 2549 | " Inst %d, HIQ on MEC %d Pipe %d Queue %d\n", |
c4050ff1 LL |
2550 | xcc_id, |
2551 | KFD_CIK_HIQ_PIPE / get_pipes_per_mec(dqm) + 1, | |
2552 | KFD_CIK_HIQ_PIPE % get_pipes_per_mec(dqm), | |
e2069a7b MJ |
2553 | KFD_CIK_HIQ_QUEUE); |
2554 | seq_reg_dump(m, dump, n_regs); | |
24f48a42 | 2555 | |
e2069a7b MJ |
2556 | kfree(dump); |
2557 | } | |
24f48a42 | 2558 | |
e2069a7b MJ |
2559 | for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) { |
2560 | int pipe_offset = pipe * get_queues_per_pipe(dqm); | |
851a645e | 2561 | |
e2069a7b MJ |
2562 | for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) { |
2563 | if (!test_bit(pipe_offset + queue, | |
8dc1db31 | 2564 | dqm->dev->kfd->shared_resources.cp_queue_bitmap)) |
e2069a7b | 2565 | continue; |
851a645e | 2566 | |
c4050ff1 LL |
2567 | r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev, |
2568 | pipe, queue, | |
2569 | &dump, &n_regs, | |
2570 | xcc_id); | |
e2069a7b MJ |
2571 | if (r) |
2572 | break; | |
851a645e | 2573 | |
c4050ff1 LL |
2574 | seq_printf(m, |
2575 | " Inst %d, CP Pipe %d, Queue %d\n", | |
2576 | xcc_id, pipe, queue); | |
e2069a7b | 2577 | seq_reg_dump(m, dump, n_regs); |
851a645e | 2578 | |
e2069a7b MJ |
2579 | kfree(dump); |
2580 | } | |
851a645e FK |
2581 | } |
2582 | } | |
2583 | ||
643e40d4 MJ |
2584 | sdma_engine_start = dqm->dev->node_id * get_num_all_sdma_engines(dqm); |
2585 | for (pipe = sdma_engine_start; | |
2586 | pipe < (sdma_engine_start + get_num_all_sdma_engines(dqm)); | |
2587 | pipe++) { | |
d5094189 | 2588 | for (queue = 0; |
8dc1db31 | 2589 | queue < dqm->dev->kfd->device_info.num_sdma_queues_per_engine; |
d5094189 | 2590 | queue++) { |
851a645e | 2591 | r = dqm->dev->kfd2kgd->hqd_sdma_dump( |
420185fd | 2592 | dqm->dev->adev, pipe, queue, &dump, &n_regs); |
851a645e FK |
2593 | if (r) |
2594 | break; | |
2595 | ||
2596 | seq_printf(m, " SDMA Engine %d, RLC %d\n", | |
2597 | pipe, queue); | |
2598 | seq_reg_dump(m, dump, n_regs); | |
2599 | ||
2600 | kfree(dump); | |
2601 | } | |
2602 | } | |
2603 | ||
2604 | return r; | |
2605 | } | |
2606 | ||
4f942aae | 2607 | int dqm_debugfs_hang_hws(struct device_queue_manager *dqm) |
a29ec470 SL |
2608 | { |
2609 | int r = 0; | |
2610 | ||
2611 | dqm_lock(dqm); | |
4f942aae OZ |
2612 | r = pm_debugfs_hang_hws(&dqm->packet_mgr); |
2613 | if (r) { | |
2614 | dqm_unlock(dqm); | |
2615 | return r; | |
2616 | } | |
a29ec470 | 2617 | dqm->active_runlist = true; |
7cee6a68 JK |
2618 | r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, |
2619 | 0, USE_DEFAULT_GRACE_PERIOD); | |
a29ec470 SL |
2620 | dqm_unlock(dqm); |
2621 | ||
2622 | return r; | |
2623 | } | |
2624 | ||
851a645e | 2625 | #endif |