Commit | Line | Data |
---|---|---|
64c7f8cf BG |
1 | /* |
2 | * Copyright 2014 Advanced Micro Devices, Inc. | |
3 | * | |
4 | * Permission is hereby granted, free of charge, to any person obtaining a | |
5 | * copy of this software and associated documentation files (the "Software"), | |
6 | * to deal in the Software without restriction, including without limitation | |
7 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, | |
8 | * and/or sell copies of the Software, and to permit persons to whom the | |
9 | * Software is furnished to do so, subject to the following conditions: | |
10 | * | |
11 | * The above copyright notice and this permission notice shall be included in | |
12 | * all copies or substantial portions of the Software. | |
13 | * | |
14 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
15 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
16 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
17 | * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR | |
18 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, | |
19 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR | |
20 | * OTHER DEALINGS IN THE SOFTWARE. | |
21 | * | |
22 | */ | |
23 | ||
26103436 FK |
24 | #include <linux/ratelimit.h> |
25 | #include <linux/printk.h> | |
64c7f8cf BG |
26 | #include <linux/slab.h> |
27 | #include <linux/list.h> | |
28 | #include <linux/types.h> | |
64c7f8cf | 29 | #include <linux/bitops.h> |
99331a51 | 30 | #include <linux/sched.h> |
64c7f8cf BG |
31 | #include "kfd_priv.h" |
32 | #include "kfd_device_queue_manager.h" | |
33 | #include "kfd_mqd_manager.h" | |
34 | #include "cik_regs.h" | |
35 | #include "kfd_kernel_queue.h" | |
5b87245f | 36 | #include "amdgpu_amdkfd.h" |
64c7f8cf BG |
37 | |
38 | /* Size of the per-pipe EOP queue */ | |
39 | #define CIK_HPD_EOP_BYTES_LOG2 11 | |
40 | #define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2) | |
41 | ||
64c7f8cf BG |
42 | static int set_pasid_vmid_mapping(struct device_queue_manager *dqm, |
43 | unsigned int pasid, unsigned int vmid); | |
44 | ||
c4744e24 YZ |
45 | static int execute_queues_cpsch(struct device_queue_manager *dqm, |
46 | enum kfd_unmap_queues_filter filter, | |
47 | uint32_t filter_param); | |
7da2bcf8 | 48 | static int unmap_queues_cpsch(struct device_queue_manager *dqm, |
4465f466 YZ |
49 | enum kfd_unmap_queues_filter filter, |
50 | uint32_t filter_param); | |
64c7f8cf | 51 | |
60a00956 FK |
52 | static int map_queues_cpsch(struct device_queue_manager *dqm); |
53 | ||
bcea3081 | 54 | static void deallocate_sdma_queue(struct device_queue_manager *dqm, |
1b4670f6 | 55 | struct queue *q); |
64c7f8cf | 56 | |
d39b7737 OZ |
57 | static inline void deallocate_hqd(struct device_queue_manager *dqm, |
58 | struct queue *q); | |
59 | static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q); | |
60 | static int allocate_sdma_queue(struct device_queue_manager *dqm, | |
61 | struct queue *q); | |
73ea648d SL |
62 | static void kfd_process_hw_exception(struct work_struct *work); |
63 | ||
bcea3081 BG |
64 | static inline |
65 | enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type) | |
64c7f8cf | 66 | { |
1b4670f6 | 67 | if (type == KFD_QUEUE_TYPE_SDMA || type == KFD_QUEUE_TYPE_SDMA_XGMI) |
85d258f9 BG |
68 | return KFD_MQD_TYPE_SDMA; |
69 | return KFD_MQD_TYPE_CP; | |
64c7f8cf BG |
70 | } |
71 | ||
d0b63bb3 AR |
72 | static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe) |
73 | { | |
74 | int i; | |
75 | int pipe_offset = mec * dqm->dev->shared_resources.num_pipe_per_mec | |
76 | + pipe * dqm->dev->shared_resources.num_queue_per_pipe; | |
77 | ||
78 | /* queue is available for KFD usage if bit is 1 */ | |
79 | for (i = 0; i < dqm->dev->shared_resources.num_queue_per_pipe; ++i) | |
80 | if (test_bit(pipe_offset + i, | |
81 | dqm->dev->shared_resources.queue_bitmap)) | |
82 | return true; | |
83 | return false; | |
84 | } | |
85 | ||
d0b63bb3 | 86 | unsigned int get_queues_num(struct device_queue_manager *dqm) |
64ea8f4a | 87 | { |
d0b63bb3 AR |
88 | return bitmap_weight(dqm->dev->shared_resources.queue_bitmap, |
89 | KGD_MAX_QUEUES); | |
64ea8f4a OG |
90 | } |
91 | ||
d0b63bb3 | 92 | unsigned int get_queues_per_pipe(struct device_queue_manager *dqm) |
64c7f8cf | 93 | { |
d0b63bb3 AR |
94 | return dqm->dev->shared_resources.num_queue_per_pipe; |
95 | } | |
96 | ||
97 | unsigned int get_pipes_per_mec(struct device_queue_manager *dqm) | |
98 | { | |
d0b63bb3 | 99 | return dqm->dev->shared_resources.num_pipe_per_mec; |
64c7f8cf BG |
100 | } |
101 | ||
98bb9222 YZ |
102 | static unsigned int get_num_sdma_engines(struct device_queue_manager *dqm) |
103 | { | |
104 | return dqm->dev->device_info->num_sdma_engines; | |
105 | } | |
106 | ||
1b4670f6 OZ |
107 | static unsigned int get_num_xgmi_sdma_engines(struct device_queue_manager *dqm) |
108 | { | |
109 | return dqm->dev->device_info->num_xgmi_sdma_engines; | |
110 | } | |
111 | ||
98bb9222 YZ |
112 | unsigned int get_num_sdma_queues(struct device_queue_manager *dqm) |
113 | { | |
114 | return dqm->dev->device_info->num_sdma_engines | |
d5094189 | 115 | * dqm->dev->device_info->num_sdma_queues_per_engine; |
98bb9222 YZ |
116 | } |
117 | ||
1b4670f6 OZ |
118 | unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm) |
119 | { | |
120 | return dqm->dev->device_info->num_xgmi_sdma_engines | |
121 | * dqm->dev->device_info->num_sdma_queues_per_engine; | |
122 | } | |
123 | ||
a22fc854 | 124 | void program_sh_mem_settings(struct device_queue_manager *dqm, |
64c7f8cf BG |
125 | struct qcm_process_device *qpd) |
126 | { | |
cea405b1 XZ |
127 | return dqm->dev->kfd2kgd->program_sh_mem_settings( |
128 | dqm->dev->kgd, qpd->vmid, | |
64c7f8cf BG |
129 | qpd->sh_mem_config, |
130 | qpd->sh_mem_ape1_base, | |
131 | qpd->sh_mem_ape1_limit, | |
132 | qpd->sh_mem_bases); | |
133 | } | |
134 | ||
ef568db7 FK |
135 | static int allocate_doorbell(struct qcm_process_device *qpd, struct queue *q) |
136 | { | |
137 | struct kfd_dev *dev = qpd->dqm->dev; | |
138 | ||
139 | if (!KFD_IS_SOC15(dev->device_info->asic_family)) { | |
140 | /* On pre-SOC15 chips we need to use the queue ID to | |
141 | * preserve the user mode ABI. | |
142 | */ | |
143 | q->doorbell_id = q->properties.queue_id; | |
1b4670f6 OZ |
144 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
145 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
234441dd YZ |
146 | /* For SDMA queues on SOC15 with 8-byte doorbell, use static |
147 | * doorbell assignments based on the engine and queue id. | |
148 | * The doobell index distance between RLC (2*i) and (2*i+1) | |
149 | * for a SDMA engine is 512. | |
ef568db7 | 150 | */ |
234441dd YZ |
151 | uint32_t *idx_offset = |
152 | dev->shared_resources.sdma_doorbell_idx; | |
153 | ||
154 | q->doorbell_id = idx_offset[q->properties.sdma_engine_id] | |
155 | + (q->properties.sdma_queue_id & 1) | |
156 | * KFD_QUEUE_DOORBELL_MIRROR_OFFSET | |
157 | + (q->properties.sdma_queue_id >> 1); | |
ef568db7 FK |
158 | } else { |
159 | /* For CP queues on SOC15 reserve a free doorbell ID */ | |
160 | unsigned int found; | |
161 | ||
162 | found = find_first_zero_bit(qpd->doorbell_bitmap, | |
163 | KFD_MAX_NUM_OF_QUEUES_PER_PROCESS); | |
164 | if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) { | |
165 | pr_debug("No doorbells available"); | |
166 | return -EBUSY; | |
167 | } | |
168 | set_bit(found, qpd->doorbell_bitmap); | |
169 | q->doorbell_id = found; | |
170 | } | |
171 | ||
172 | q->properties.doorbell_off = | |
173 | kfd_doorbell_id_to_offset(dev, q->process, | |
174 | q->doorbell_id); | |
175 | ||
176 | return 0; | |
177 | } | |
178 | ||
179 | static void deallocate_doorbell(struct qcm_process_device *qpd, | |
180 | struct queue *q) | |
181 | { | |
182 | unsigned int old; | |
183 | struct kfd_dev *dev = qpd->dqm->dev; | |
184 | ||
185 | if (!KFD_IS_SOC15(dev->device_info->asic_family) || | |
1b4670f6 OZ |
186 | q->properties.type == KFD_QUEUE_TYPE_SDMA || |
187 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) | |
ef568db7 FK |
188 | return; |
189 | ||
190 | old = test_and_clear_bit(q->doorbell_id, qpd->doorbell_bitmap); | |
191 | WARN_ON(!old); | |
192 | } | |
193 | ||
64c7f8cf BG |
194 | static int allocate_vmid(struct device_queue_manager *dqm, |
195 | struct qcm_process_device *qpd, | |
196 | struct queue *q) | |
197 | { | |
198 | int bit, allocated_vmid; | |
199 | ||
200 | if (dqm->vmid_bitmap == 0) | |
201 | return -ENOMEM; | |
202 | ||
4252bf68 HK |
203 | bit = ffs(dqm->vmid_bitmap) - 1; |
204 | dqm->vmid_bitmap &= ~(1 << bit); | |
64c7f8cf | 205 | |
44008d7a | 206 | allocated_vmid = bit + dqm->dev->vm_info.first_vmid_kfd; |
79775b62 | 207 | pr_debug("vmid allocation %d\n", allocated_vmid); |
64c7f8cf BG |
208 | qpd->vmid = allocated_vmid; |
209 | q->properties.vmid = allocated_vmid; | |
210 | ||
211 | set_pasid_vmid_mapping(dqm, q->process->pasid, q->properties.vmid); | |
212 | program_sh_mem_settings(dqm, qpd); | |
213 | ||
403575c4 FK |
214 | /* qpd->page_table_base is set earlier when register_process() |
215 | * is called, i.e. when the first queue is created. | |
216 | */ | |
217 | dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->kgd, | |
218 | qpd->vmid, | |
219 | qpd->page_table_base); | |
220 | /* invalidate the VM context after pasid and vmid mapping is set up */ | |
221 | kfd_flush_tlb(qpd_to_pdd(qpd)); | |
222 | ||
d39b7737 OZ |
223 | dqm->dev->kfd2kgd->set_scratch_backing_va( |
224 | dqm->dev->kgd, qpd->sh_hidden_private_base, qpd->vmid); | |
225 | ||
64c7f8cf BG |
226 | return 0; |
227 | } | |
228 | ||
552764b6 FK |
229 | static int flush_texture_cache_nocpsch(struct kfd_dev *kdev, |
230 | struct qcm_process_device *qpd) | |
231 | { | |
f6e27ff1 FK |
232 | const struct packet_manager_funcs *pmf = qpd->dqm->packets.pmf; |
233 | int ret; | |
552764b6 FK |
234 | |
235 | if (!qpd->ib_kaddr) | |
236 | return -ENOMEM; | |
237 | ||
f6e27ff1 FK |
238 | ret = pmf->release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr); |
239 | if (ret) | |
240 | return ret; | |
552764b6 | 241 | |
5b87245f | 242 | return amdgpu_amdkfd_submit_ib(kdev->kgd, KGD_ENGINE_MEC1, qpd->vmid, |
f6e27ff1 FK |
243 | qpd->ib_base, (uint32_t *)qpd->ib_kaddr, |
244 | pmf->release_mem_size / sizeof(uint32_t)); | |
552764b6 FK |
245 | } |
246 | ||
64c7f8cf BG |
247 | static void deallocate_vmid(struct device_queue_manager *dqm, |
248 | struct qcm_process_device *qpd, | |
249 | struct queue *q) | |
250 | { | |
44008d7a | 251 | int bit = qpd->vmid - dqm->dev->vm_info.first_vmid_kfd; |
64c7f8cf | 252 | |
552764b6 FK |
253 | /* On GFX v7, CP doesn't flush TC at dequeue */ |
254 | if (q->device->device_info->asic_family == CHIP_HAWAII) | |
255 | if (flush_texture_cache_nocpsch(q->device, qpd)) | |
256 | pr_err("Failed to flush TC\n"); | |
257 | ||
403575c4 FK |
258 | kfd_flush_tlb(qpd_to_pdd(qpd)); |
259 | ||
2030664b BG |
260 | /* Release the vmid mapping */ |
261 | set_pasid_vmid_mapping(dqm, 0, qpd->vmid); | |
262 | ||
4252bf68 | 263 | dqm->vmid_bitmap |= (1 << bit); |
64c7f8cf BG |
264 | qpd->vmid = 0; |
265 | q->properties.vmid = 0; | |
266 | } | |
267 | ||
268 | static int create_queue_nocpsch(struct device_queue_manager *dqm, | |
269 | struct queue *q, | |
b46cb7d7 | 270 | struct qcm_process_device *qpd) |
64c7f8cf | 271 | { |
d39b7737 | 272 | struct mqd_manager *mqd_mgr; |
64c7f8cf BG |
273 | int retval; |
274 | ||
64c7f8cf BG |
275 | print_queue(q); |
276 | ||
efeaed4d | 277 | dqm_lock(dqm); |
64c7f8cf | 278 | |
b8cbab04 | 279 | if (dqm->total_queue_count >= max_num_of_queues_per_device) { |
79775b62 | 280 | pr_warn("Can't create new usermode queue because %d queues were already created\n", |
b8cbab04 | 281 | dqm->total_queue_count); |
ab7c1648 KR |
282 | retval = -EPERM; |
283 | goto out_unlock; | |
b8cbab04 OG |
284 | } |
285 | ||
64c7f8cf BG |
286 | if (list_empty(&qpd->queues_list)) { |
287 | retval = allocate_vmid(dqm, qpd, q); | |
ab7c1648 KR |
288 | if (retval) |
289 | goto out_unlock; | |
64c7f8cf | 290 | } |
64c7f8cf | 291 | q->properties.vmid = qpd->vmid; |
26103436 | 292 | /* |
bb2d2128 FK |
293 | * Eviction state logic: mark all queues as evicted, even ones |
294 | * not currently active. Restoring inactive queues later only | |
295 | * updates the is_evicted flag but is a no-op otherwise. | |
26103436 | 296 | */ |
bb2d2128 | 297 | q->properties.is_evicted = !!qpd->evicted; |
64c7f8cf | 298 | |
373d7080 FK |
299 | q->properties.tba_addr = qpd->tba_addr; |
300 | q->properties.tma_addr = qpd->tma_addr; | |
301 | ||
d091bc0a OZ |
302 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
303 | q->properties.type)]; | |
d39b7737 OZ |
304 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) { |
305 | retval = allocate_hqd(dqm, q); | |
306 | if (retval) | |
307 | goto deallocate_vmid; | |
308 | pr_debug("Loading mqd to hqd on pipe %d, queue %d\n", | |
309 | q->pipe, q->queue); | |
310 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || | |
311 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
312 | retval = allocate_sdma_queue(dqm, q); | |
313 | if (retval) | |
314 | goto deallocate_vmid; | |
315 | dqm->asic_ops.init_sdma_vm(dqm, q, qpd); | |
316 | } | |
317 | ||
318 | retval = allocate_doorbell(qpd, q); | |
319 | if (retval) | |
320 | goto out_deallocate_hqd; | |
321 | ||
6a6ef5ee OZ |
322 | /* Temporarily release dqm lock to avoid a circular lock dependency */ |
323 | dqm_unlock(dqm); | |
d091bc0a | 324 | q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties); |
6a6ef5ee OZ |
325 | dqm_lock(dqm); |
326 | ||
d091bc0a OZ |
327 | if (!q->mqd_mem_obj) { |
328 | retval = -ENOMEM; | |
329 | goto out_deallocate_doorbell; | |
330 | } | |
8636e53c OZ |
331 | mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, |
332 | &q->gart_mqd_addr, &q->properties); | |
d39b7737 OZ |
333 | if (q->properties.is_active) { |
334 | ||
335 | if (WARN(q->process->mm != current->mm, | |
336 | "should only run in user thread")) | |
337 | retval = -EFAULT; | |
338 | else | |
339 | retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe, | |
340 | q->queue, &q->properties, current->mm); | |
341 | if (retval) | |
d091bc0a | 342 | goto out_free_mqd; |
64c7f8cf BG |
343 | } |
344 | ||
345 | list_add(&q->list, &qpd->queues_list); | |
bc920fd4 | 346 | qpd->queue_count++; |
b6819cec JC |
347 | if (q->properties.is_active) |
348 | dqm->queue_count++; | |
64c7f8cf | 349 | |
bcea3081 BG |
350 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) |
351 | dqm->sdma_queue_count++; | |
1b4670f6 OZ |
352 | else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) |
353 | dqm->xgmi_sdma_queue_count++; | |
64c7f8cf | 354 | |
b8cbab04 OG |
355 | /* |
356 | * Unconditionally increment this counter, regardless of the queue's | |
357 | * type or whether the queue is active. | |
358 | */ | |
359 | dqm->total_queue_count++; | |
360 | pr_debug("Total of %d queues are accountable so far\n", | |
361 | dqm->total_queue_count); | |
d091bc0a | 362 | goto out_unlock; |
b8cbab04 | 363 | |
d091bc0a OZ |
364 | out_free_mqd: |
365 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); | |
d39b7737 OZ |
366 | out_deallocate_doorbell: |
367 | deallocate_doorbell(qpd, q); | |
368 | out_deallocate_hqd: | |
369 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) | |
370 | deallocate_hqd(dqm, q); | |
371 | else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || | |
372 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) | |
373 | deallocate_sdma_queue(dqm, q); | |
374 | deallocate_vmid: | |
375 | if (list_empty(&qpd->queues_list)) | |
376 | deallocate_vmid(dqm, qpd, q); | |
ab7c1648 | 377 | out_unlock: |
efeaed4d | 378 | dqm_unlock(dqm); |
ab7c1648 | 379 | return retval; |
64c7f8cf BG |
380 | } |
381 | ||
382 | static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q) | |
383 | { | |
384 | bool set; | |
f0ec5b99 | 385 | int pipe, bit, i; |
64c7f8cf BG |
386 | |
387 | set = false; | |
388 | ||
8eabaf54 KR |
389 | for (pipe = dqm->next_pipe_to_allocate, i = 0; |
390 | i < get_pipes_per_mec(dqm); | |
d0b63bb3 AR |
391 | pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) { |
392 | ||
393 | if (!is_pipe_enabled(dqm, 0, pipe)) | |
394 | continue; | |
395 | ||
64c7f8cf | 396 | if (dqm->allocated_queues[pipe] != 0) { |
4252bf68 HK |
397 | bit = ffs(dqm->allocated_queues[pipe]) - 1; |
398 | dqm->allocated_queues[pipe] &= ~(1 << bit); | |
64c7f8cf BG |
399 | q->pipe = pipe; |
400 | q->queue = bit; | |
401 | set = true; | |
402 | break; | |
403 | } | |
404 | } | |
405 | ||
991ca8ee | 406 | if (!set) |
64c7f8cf BG |
407 | return -EBUSY; |
408 | ||
79775b62 | 409 | pr_debug("hqd slot - pipe %d, queue %d\n", q->pipe, q->queue); |
64c7f8cf | 410 | /* horizontal hqd allocation */ |
d0b63bb3 | 411 | dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm); |
64c7f8cf BG |
412 | |
413 | return 0; | |
414 | } | |
415 | ||
416 | static inline void deallocate_hqd(struct device_queue_manager *dqm, | |
417 | struct queue *q) | |
418 | { | |
4252bf68 | 419 | dqm->allocated_queues[q->pipe] |= (1 << q->queue); |
64c7f8cf BG |
420 | } |
421 | ||
9fd3f1bf FK |
422 | /* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked |
423 | * to avoid asynchronized access | |
424 | */ | |
425 | static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm, | |
64c7f8cf BG |
426 | struct qcm_process_device *qpd, |
427 | struct queue *q) | |
428 | { | |
429 | int retval; | |
8d5f3552 | 430 | struct mqd_manager *mqd_mgr; |
64c7f8cf | 431 | |
fdfa090b OZ |
432 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
433 | q->properties.type)]; | |
64c7f8cf | 434 | |
c2e1b3a4 | 435 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) { |
c2e1b3a4 BG |
436 | deallocate_hqd(dqm, q); |
437 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { | |
c2e1b3a4 | 438 | dqm->sdma_queue_count--; |
1b4670f6 OZ |
439 | deallocate_sdma_queue(dqm, q); |
440 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
441 | dqm->xgmi_sdma_queue_count--; | |
442 | deallocate_sdma_queue(dqm, q); | |
7113cd65 | 443 | } else { |
79775b62 | 444 | pr_debug("q->properties.type %d is invalid\n", |
7113cd65 | 445 | q->properties.type); |
9fd3f1bf | 446 | return -EINVAL; |
64c7f8cf | 447 | } |
9fd3f1bf | 448 | dqm->total_queue_count--; |
64c7f8cf | 449 | |
ef568db7 FK |
450 | deallocate_doorbell(qpd, q); |
451 | ||
8d5f3552 | 452 | retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, |
c2e1b3a4 | 453 | KFD_PREEMPT_TYPE_WAVEFRONT_RESET, |
b90e3fbe | 454 | KFD_UNMAP_LATENCY_MS, |
64c7f8cf | 455 | q->pipe, q->queue); |
9fd3f1bf FK |
456 | if (retval == -ETIME) |
457 | qpd->reset_wavefronts = true; | |
64c7f8cf | 458 | |
8636e53c | 459 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); |
64c7f8cf BG |
460 | |
461 | list_del(&q->list); | |
9fd3f1bf FK |
462 | if (list_empty(&qpd->queues_list)) { |
463 | if (qpd->reset_wavefronts) { | |
464 | pr_warn("Resetting wave fronts (nocpsch) on dev %p\n", | |
465 | dqm->dev); | |
466 | /* dbgdev_wave_reset_wavefronts has to be called before | |
467 | * deallocate_vmid(), i.e. when vmid is still in use. | |
468 | */ | |
469 | dbgdev_wave_reset_wavefronts(dqm->dev, | |
470 | qpd->pqm->process); | |
471 | qpd->reset_wavefronts = false; | |
472 | } | |
473 | ||
64c7f8cf | 474 | deallocate_vmid(dqm, qpd, q); |
9fd3f1bf | 475 | } |
bc920fd4 | 476 | qpd->queue_count--; |
b6819cec JC |
477 | if (q->properties.is_active) |
478 | dqm->queue_count--; | |
b8cbab04 | 479 | |
9fd3f1bf FK |
480 | return retval; |
481 | } | |
b8cbab04 | 482 | |
9fd3f1bf FK |
483 | static int destroy_queue_nocpsch(struct device_queue_manager *dqm, |
484 | struct qcm_process_device *qpd, | |
485 | struct queue *q) | |
486 | { | |
487 | int retval; | |
488 | ||
efeaed4d | 489 | dqm_lock(dqm); |
9fd3f1bf | 490 | retval = destroy_queue_nocpsch_locked(dqm, qpd, q); |
efeaed4d | 491 | dqm_unlock(dqm); |
9fd3f1bf | 492 | |
64c7f8cf BG |
493 | return retval; |
494 | } | |
495 | ||
496 | static int update_queue(struct device_queue_manager *dqm, struct queue *q) | |
497 | { | |
8636e53c | 498 | int retval = 0; |
8d5f3552 | 499 | struct mqd_manager *mqd_mgr; |
26103436 | 500 | struct kfd_process_device *pdd; |
b6ffbab8 | 501 | bool prev_active = false; |
64c7f8cf | 502 | |
efeaed4d | 503 | dqm_lock(dqm); |
26103436 FK |
504 | pdd = kfd_get_process_device_data(q->device, q->process); |
505 | if (!pdd) { | |
506 | retval = -ENODEV; | |
507 | goto out_unlock; | |
508 | } | |
fdfa090b OZ |
509 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
510 | q->properties.type)]; | |
64c7f8cf | 511 | |
60a00956 FK |
512 | /* Save previous activity state for counters */ |
513 | prev_active = q->properties.is_active; | |
514 | ||
515 | /* Make sure the queue is unmapped before updating the MQD */ | |
d146c5a7 | 516 | if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { |
60a00956 FK |
517 | retval = unmap_queues_cpsch(dqm, |
518 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); | |
894a8293 | 519 | if (retval) { |
60a00956 FK |
520 | pr_err("unmap queue failed\n"); |
521 | goto out_unlock; | |
522 | } | |
894a8293 | 523 | } else if (prev_active && |
60a00956 | 524 | (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || |
1b4670f6 OZ |
525 | q->properties.type == KFD_QUEUE_TYPE_SDMA || |
526 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { | |
8d5f3552 | 527 | retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, |
60a00956 FK |
528 | KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN, |
529 | KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); | |
530 | if (retval) { | |
531 | pr_err("destroy mqd failed\n"); | |
532 | goto out_unlock; | |
533 | } | |
534 | } | |
535 | ||
8636e53c | 536 | mqd_mgr->update_mqd(mqd_mgr, q->mqd, &q->properties); |
60a00956 | 537 | |
096d1a3e FK |
538 | /* |
539 | * check active state vs. the previous state and modify | |
540 | * counter accordingly. map_queues_cpsch uses the | |
541 | * dqm->queue_count to determine whether a new runlist must be | |
542 | * uploaded. | |
543 | */ | |
544 | if (q->properties.is_active && !prev_active) | |
545 | dqm->queue_count++; | |
546 | else if (!q->properties.is_active && prev_active) | |
547 | dqm->queue_count--; | |
548 | ||
d146c5a7 | 549 | if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) |
60a00956 | 550 | retval = map_queues_cpsch(dqm); |
894a8293 | 551 | else if (q->properties.is_active && |
60a00956 | 552 | (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || |
1b4670f6 OZ |
553 | q->properties.type == KFD_QUEUE_TYPE_SDMA || |
554 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { | |
1b19aa5a FK |
555 | if (WARN(q->process->mm != current->mm, |
556 | "should only run in user thread")) | |
557 | retval = -EFAULT; | |
558 | else | |
559 | retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, | |
560 | q->pipe, q->queue, | |
561 | &q->properties, current->mm); | |
562 | } | |
b6ffbab8 | 563 | |
ab7c1648 | 564 | out_unlock: |
efeaed4d | 565 | dqm_unlock(dqm); |
64c7f8cf BG |
566 | return retval; |
567 | } | |
568 | ||
26103436 FK |
569 | static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, |
570 | struct qcm_process_device *qpd) | |
571 | { | |
572 | struct queue *q; | |
8d5f3552 | 573 | struct mqd_manager *mqd_mgr; |
26103436 | 574 | struct kfd_process_device *pdd; |
bb2d2128 | 575 | int retval, ret = 0; |
26103436 | 576 | |
efeaed4d | 577 | dqm_lock(dqm); |
26103436 FK |
578 | if (qpd->evicted++ > 0) /* already evicted, do nothing */ |
579 | goto out; | |
580 | ||
581 | pdd = qpd_to_pdd(qpd); | |
582 | pr_info_ratelimited("Evicting PASID %u queues\n", | |
583 | pdd->process->pasid); | |
584 | ||
bb2d2128 FK |
585 | /* Mark all queues as evicted. Deactivate all active queues on |
586 | * the qpd. | |
587 | */ | |
26103436 | 588 | list_for_each_entry(q, &qpd->queues_list, list) { |
bb2d2128 | 589 | q->properties.is_evicted = true; |
26103436 FK |
590 | if (!q->properties.is_active) |
591 | continue; | |
bb2d2128 | 592 | |
fdfa090b OZ |
593 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
594 | q->properties.type)]; | |
26103436 | 595 | q->properties.is_active = false; |
8d5f3552 | 596 | retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, |
26103436 FK |
597 | KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN, |
598 | KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); | |
bb2d2128 FK |
599 | if (retval && !ret) |
600 | /* Return the first error, but keep going to | |
601 | * maintain a consistent eviction state | |
602 | */ | |
603 | ret = retval; | |
26103436 FK |
604 | dqm->queue_count--; |
605 | } | |
606 | ||
607 | out: | |
efeaed4d | 608 | dqm_unlock(dqm); |
bb2d2128 | 609 | return ret; |
26103436 FK |
610 | } |
611 | ||
612 | static int evict_process_queues_cpsch(struct device_queue_manager *dqm, | |
613 | struct qcm_process_device *qpd) | |
614 | { | |
615 | struct queue *q; | |
616 | struct kfd_process_device *pdd; | |
617 | int retval = 0; | |
618 | ||
efeaed4d | 619 | dqm_lock(dqm); |
26103436 FK |
620 | if (qpd->evicted++ > 0) /* already evicted, do nothing */ |
621 | goto out; | |
622 | ||
623 | pdd = qpd_to_pdd(qpd); | |
624 | pr_info_ratelimited("Evicting PASID %u queues\n", | |
625 | pdd->process->pasid); | |
626 | ||
bb2d2128 FK |
627 | /* Mark all queues as evicted. Deactivate all active queues on |
628 | * the qpd. | |
629 | */ | |
26103436 | 630 | list_for_each_entry(q, &qpd->queues_list, list) { |
bb2d2128 | 631 | q->properties.is_evicted = true; |
26103436 FK |
632 | if (!q->properties.is_active) |
633 | continue; | |
bb2d2128 | 634 | |
26103436 FK |
635 | q->properties.is_active = false; |
636 | dqm->queue_count--; | |
637 | } | |
638 | retval = execute_queues_cpsch(dqm, | |
639 | qpd->is_debug ? | |
640 | KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES : | |
641 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); | |
642 | ||
643 | out: | |
efeaed4d | 644 | dqm_unlock(dqm); |
26103436 FK |
645 | return retval; |
646 | } | |
647 | ||
648 | static int restore_process_queues_nocpsch(struct device_queue_manager *dqm, | |
649 | struct qcm_process_device *qpd) | |
650 | { | |
1b19aa5a | 651 | struct mm_struct *mm = NULL; |
26103436 | 652 | struct queue *q; |
8d5f3552 | 653 | struct mqd_manager *mqd_mgr; |
26103436 | 654 | struct kfd_process_device *pdd; |
e715c6d0 | 655 | uint64_t pd_base; |
bb2d2128 | 656 | int retval, ret = 0; |
26103436 FK |
657 | |
658 | pdd = qpd_to_pdd(qpd); | |
659 | /* Retrieve PD base */ | |
5b87245f | 660 | pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm); |
26103436 | 661 | |
efeaed4d | 662 | dqm_lock(dqm); |
26103436 FK |
663 | if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ |
664 | goto out; | |
665 | if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */ | |
666 | qpd->evicted--; | |
667 | goto out; | |
668 | } | |
669 | ||
670 | pr_info_ratelimited("Restoring PASID %u queues\n", | |
671 | pdd->process->pasid); | |
672 | ||
673 | /* Update PD Base in QPD */ | |
674 | qpd->page_table_base = pd_base; | |
e715c6d0 | 675 | pr_debug("Updated PD address to 0x%llx\n", pd_base); |
26103436 FK |
676 | |
677 | if (!list_empty(&qpd->queues_list)) { | |
678 | dqm->dev->kfd2kgd->set_vm_context_page_table_base( | |
679 | dqm->dev->kgd, | |
680 | qpd->vmid, | |
681 | qpd->page_table_base); | |
682 | kfd_flush_tlb(pdd); | |
683 | } | |
684 | ||
1b19aa5a FK |
685 | /* Take a safe reference to the mm_struct, which may otherwise |
686 | * disappear even while the kfd_process is still referenced. | |
687 | */ | |
688 | mm = get_task_mm(pdd->process->lead_thread); | |
689 | if (!mm) { | |
bb2d2128 | 690 | ret = -EFAULT; |
1b19aa5a FK |
691 | goto out; |
692 | } | |
693 | ||
bb2d2128 FK |
694 | /* Remove the eviction flags. Activate queues that are not |
695 | * inactive for other reasons. | |
696 | */ | |
26103436 | 697 | list_for_each_entry(q, &qpd->queues_list, list) { |
bb2d2128 FK |
698 | q->properties.is_evicted = false; |
699 | if (!QUEUE_IS_ACTIVE(q->properties)) | |
26103436 | 700 | continue; |
bb2d2128 | 701 | |
fdfa090b OZ |
702 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
703 | q->properties.type)]; | |
26103436 | 704 | q->properties.is_active = true; |
8d5f3552 | 705 | retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe, |
1b19aa5a | 706 | q->queue, &q->properties, mm); |
bb2d2128 FK |
707 | if (retval && !ret) |
708 | /* Return the first error, but keep going to | |
709 | * maintain a consistent eviction state | |
710 | */ | |
711 | ret = retval; | |
26103436 FK |
712 | dqm->queue_count++; |
713 | } | |
714 | qpd->evicted = 0; | |
715 | out: | |
1b19aa5a FK |
716 | if (mm) |
717 | mmput(mm); | |
efeaed4d | 718 | dqm_unlock(dqm); |
bb2d2128 | 719 | return ret; |
26103436 FK |
720 | } |
721 | ||
722 | static int restore_process_queues_cpsch(struct device_queue_manager *dqm, | |
723 | struct qcm_process_device *qpd) | |
724 | { | |
725 | struct queue *q; | |
726 | struct kfd_process_device *pdd; | |
e715c6d0 | 727 | uint64_t pd_base; |
26103436 FK |
728 | int retval = 0; |
729 | ||
730 | pdd = qpd_to_pdd(qpd); | |
731 | /* Retrieve PD base */ | |
5b87245f | 732 | pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm); |
26103436 | 733 | |
efeaed4d | 734 | dqm_lock(dqm); |
26103436 FK |
735 | if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ |
736 | goto out; | |
737 | if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */ | |
738 | qpd->evicted--; | |
739 | goto out; | |
740 | } | |
741 | ||
742 | pr_info_ratelimited("Restoring PASID %u queues\n", | |
743 | pdd->process->pasid); | |
744 | ||
745 | /* Update PD Base in QPD */ | |
746 | qpd->page_table_base = pd_base; | |
e715c6d0 | 747 | pr_debug("Updated PD address to 0x%llx\n", pd_base); |
26103436 FK |
748 | |
749 | /* activate all active queues on the qpd */ | |
750 | list_for_each_entry(q, &qpd->queues_list, list) { | |
26103436 | 751 | q->properties.is_evicted = false; |
bb2d2128 FK |
752 | if (!QUEUE_IS_ACTIVE(q->properties)) |
753 | continue; | |
754 | ||
26103436 FK |
755 | q->properties.is_active = true; |
756 | dqm->queue_count++; | |
757 | } | |
758 | retval = execute_queues_cpsch(dqm, | |
759 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); | |
bb2d2128 | 760 | qpd->evicted = 0; |
26103436 | 761 | out: |
efeaed4d | 762 | dqm_unlock(dqm); |
26103436 FK |
763 | return retval; |
764 | } | |
765 | ||
58dcd5bf | 766 | static int register_process(struct device_queue_manager *dqm, |
64c7f8cf BG |
767 | struct qcm_process_device *qpd) |
768 | { | |
769 | struct device_process_node *n; | |
403575c4 | 770 | struct kfd_process_device *pdd; |
e715c6d0 | 771 | uint64_t pd_base; |
a22fc854 | 772 | int retval; |
64c7f8cf | 773 | |
dbf56ab1 | 774 | n = kzalloc(sizeof(*n), GFP_KERNEL); |
64c7f8cf BG |
775 | if (!n) |
776 | return -ENOMEM; | |
777 | ||
778 | n->qpd = qpd; | |
779 | ||
403575c4 FK |
780 | pdd = qpd_to_pdd(qpd); |
781 | /* Retrieve PD base */ | |
5b87245f | 782 | pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm); |
403575c4 | 783 | |
efeaed4d | 784 | dqm_lock(dqm); |
64c7f8cf BG |
785 | list_add(&n->list, &dqm->queues); |
786 | ||
403575c4 FK |
787 | /* Update PD Base in QPD */ |
788 | qpd->page_table_base = pd_base; | |
e715c6d0 | 789 | pr_debug("Updated PD address to 0x%llx\n", pd_base); |
403575c4 | 790 | |
bfd5e378 | 791 | retval = dqm->asic_ops.update_qpd(dqm, qpd); |
a22fc854 | 792 | |
f756e631 | 793 | dqm->processes_count++; |
64c7f8cf | 794 | |
efeaed4d | 795 | dqm_unlock(dqm); |
64c7f8cf | 796 | |
32cce8bc FK |
797 | /* Outside the DQM lock because under the DQM lock we can't do |
798 | * reclaim or take other locks that others hold while reclaiming. | |
799 | */ | |
800 | kfd_inc_compute_active(dqm->dev); | |
801 | ||
a22fc854 | 802 | return retval; |
64c7f8cf BG |
803 | } |
804 | ||
58dcd5bf | 805 | static int unregister_process(struct device_queue_manager *dqm, |
64c7f8cf BG |
806 | struct qcm_process_device *qpd) |
807 | { | |
808 | int retval; | |
809 | struct device_process_node *cur, *next; | |
810 | ||
1e5ec956 OG |
811 | pr_debug("qpd->queues_list is %s\n", |
812 | list_empty(&qpd->queues_list) ? "empty" : "not empty"); | |
64c7f8cf BG |
813 | |
814 | retval = 0; | |
efeaed4d | 815 | dqm_lock(dqm); |
64c7f8cf BG |
816 | |
817 | list_for_each_entry_safe(cur, next, &dqm->queues, list) { | |
818 | if (qpd == cur->qpd) { | |
819 | list_del(&cur->list); | |
f5d896bb | 820 | kfree(cur); |
f756e631 | 821 | dqm->processes_count--; |
64c7f8cf BG |
822 | goto out; |
823 | } | |
824 | } | |
825 | /* qpd not found in dqm list */ | |
826 | retval = 1; | |
827 | out: | |
efeaed4d | 828 | dqm_unlock(dqm); |
32cce8bc FK |
829 | |
830 | /* Outside the DQM lock because under the DQM lock we can't do | |
831 | * reclaim or take other locks that others hold while reclaiming. | |
832 | */ | |
833 | if (!retval) | |
834 | kfd_dec_compute_active(dqm->dev); | |
835 | ||
64c7f8cf BG |
836 | return retval; |
837 | } | |
838 | ||
839 | static int | |
840 | set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid, | |
841 | unsigned int vmid) | |
842 | { | |
cea405b1 | 843 | return dqm->dev->kfd2kgd->set_pasid_vmid_mapping( |
deb99d7c | 844 | dqm->dev->kgd, pasid, vmid); |
64c7f8cf BG |
845 | } |
846 | ||
2249d558 AL |
847 | static void init_interrupts(struct device_queue_manager *dqm) |
848 | { | |
849 | unsigned int i; | |
850 | ||
d0b63bb3 AR |
851 | for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++) |
852 | if (is_pipe_enabled(dqm, 0, i)) | |
853 | dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd, i); | |
2249d558 AL |
854 | } |
855 | ||
64c7f8cf BG |
856 | static int initialize_nocpsch(struct device_queue_manager *dqm) |
857 | { | |
86194cf8 | 858 | int pipe, queue; |
64c7f8cf | 859 | |
79775b62 | 860 | pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm)); |
64c7f8cf | 861 | |
ab7c1648 KR |
862 | dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm), |
863 | sizeof(unsigned int), GFP_KERNEL); | |
864 | if (!dqm->allocated_queues) | |
865 | return -ENOMEM; | |
866 | ||
efeaed4d | 867 | mutex_init(&dqm->lock_hidden); |
64c7f8cf BG |
868 | INIT_LIST_HEAD(&dqm->queues); |
869 | dqm->queue_count = dqm->next_pipe_to_allocate = 0; | |
bcea3081 | 870 | dqm->sdma_queue_count = 0; |
1b4670f6 | 871 | dqm->xgmi_sdma_queue_count = 0; |
64c7f8cf | 872 | |
86194cf8 FK |
873 | for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) { |
874 | int pipe_offset = pipe * get_queues_per_pipe(dqm); | |
875 | ||
876 | for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) | |
877 | if (test_bit(pipe_offset + queue, | |
878 | dqm->dev->shared_resources.queue_bitmap)) | |
879 | dqm->allocated_queues[pipe] |= 1 << queue; | |
880 | } | |
64c7f8cf | 881 | |
44008d7a | 882 | dqm->vmid_bitmap = (1 << dqm->dev->vm_info.vmid_num_kfd) - 1; |
cb77ee7c | 883 | dqm->sdma_bitmap = (1ULL << get_num_sdma_queues(dqm)) - 1; |
1b4670f6 | 884 | dqm->xgmi_sdma_bitmap = (1ULL << get_num_xgmi_sdma_queues(dqm)) - 1; |
64c7f8cf | 885 | |
64c7f8cf BG |
886 | return 0; |
887 | } | |
888 | ||
58dcd5bf | 889 | static void uninitialize(struct device_queue_manager *dqm) |
64c7f8cf | 890 | { |
6f9d54fd OG |
891 | int i; |
892 | ||
32fa8219 | 893 | WARN_ON(dqm->queue_count > 0 || dqm->processes_count > 0); |
64c7f8cf BG |
894 | |
895 | kfree(dqm->allocated_queues); | |
6f9d54fd | 896 | for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++) |
8d5f3552 | 897 | kfree(dqm->mqd_mgrs[i]); |
efeaed4d | 898 | mutex_destroy(&dqm->lock_hidden); |
a86aa3ca | 899 | kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem); |
64c7f8cf BG |
900 | } |
901 | ||
902 | static int start_nocpsch(struct device_queue_manager *dqm) | |
903 | { | |
2249d558 | 904 | init_interrupts(dqm); |
552764b6 | 905 | return pm_init(&dqm->packets, dqm); |
64c7f8cf BG |
906 | } |
907 | ||
908 | static int stop_nocpsch(struct device_queue_manager *dqm) | |
909 | { | |
552764b6 | 910 | pm_uninit(&dqm->packets); |
64c7f8cf BG |
911 | return 0; |
912 | } | |
913 | ||
bcea3081 | 914 | static int allocate_sdma_queue(struct device_queue_manager *dqm, |
e78579aa | 915 | struct queue *q) |
bcea3081 BG |
916 | { |
917 | int bit; | |
918 | ||
1b4670f6 OZ |
919 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { |
920 | if (dqm->sdma_bitmap == 0) | |
921 | return -ENOMEM; | |
922 | bit = __ffs64(dqm->sdma_bitmap); | |
923 | dqm->sdma_bitmap &= ~(1ULL << bit); | |
924 | q->sdma_id = bit; | |
925 | q->properties.sdma_engine_id = q->sdma_id % | |
926 | get_num_sdma_engines(dqm); | |
927 | q->properties.sdma_queue_id = q->sdma_id / | |
928 | get_num_sdma_engines(dqm); | |
929 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
930 | if (dqm->xgmi_sdma_bitmap == 0) | |
931 | return -ENOMEM; | |
932 | bit = __ffs64(dqm->xgmi_sdma_bitmap); | |
933 | dqm->xgmi_sdma_bitmap &= ~(1ULL << bit); | |
934 | q->sdma_id = bit; | |
935 | /* sdma_engine_id is sdma id including | |
936 | * both PCIe-optimized SDMAs and XGMI- | |
937 | * optimized SDMAs. The calculation below | |
938 | * assumes the first N engines are always | |
939 | * PCIe-optimized ones | |
940 | */ | |
941 | q->properties.sdma_engine_id = get_num_sdma_engines(dqm) + | |
942 | q->sdma_id % get_num_xgmi_sdma_engines(dqm); | |
943 | q->properties.sdma_queue_id = q->sdma_id / | |
944 | get_num_xgmi_sdma_engines(dqm); | |
945 | } | |
e78579aa | 946 | |
e78579aa YZ |
947 | pr_debug("SDMA engine id: %d\n", q->properties.sdma_engine_id); |
948 | pr_debug("SDMA queue id: %d\n", q->properties.sdma_queue_id); | |
bcea3081 BG |
949 | |
950 | return 0; | |
951 | } | |
952 | ||
953 | static void deallocate_sdma_queue(struct device_queue_manager *dqm, | |
1b4670f6 | 954 | struct queue *q) |
bcea3081 | 955 | { |
1b4670f6 OZ |
956 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { |
957 | if (q->sdma_id >= get_num_sdma_queues(dqm)) | |
958 | return; | |
959 | dqm->sdma_bitmap |= (1ULL << q->sdma_id); | |
960 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
961 | if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm)) | |
962 | return; | |
963 | dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id); | |
964 | } | |
bcea3081 BG |
965 | } |
966 | ||
64c7f8cf BG |
967 | /* |
968 | * Device Queue Manager implementation for cp scheduler | |
969 | */ | |
970 | ||
971 | static int set_sched_resources(struct device_queue_manager *dqm) | |
972 | { | |
d0b63bb3 | 973 | int i, mec; |
64c7f8cf | 974 | struct scheduling_resources res; |
64c7f8cf | 975 | |
44008d7a | 976 | res.vmid_mask = dqm->dev->shared_resources.compute_vmid_bitmap; |
d0b63bb3 AR |
977 | |
978 | res.queue_mask = 0; | |
979 | for (i = 0; i < KGD_MAX_QUEUES; ++i) { | |
980 | mec = (i / dqm->dev->shared_resources.num_queue_per_pipe) | |
981 | / dqm->dev->shared_resources.num_pipe_per_mec; | |
982 | ||
983 | if (!test_bit(i, dqm->dev->shared_resources.queue_bitmap)) | |
984 | continue; | |
985 | ||
986 | /* only acquire queues from the first MEC */ | |
987 | if (mec > 0) | |
988 | continue; | |
989 | ||
990 | /* This situation may be hit in the future if a new HW | |
991 | * generation exposes more than 64 queues. If so, the | |
8eabaf54 KR |
992 | * definition of res.queue_mask needs updating |
993 | */ | |
1d11ee89 | 994 | if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) { |
d0b63bb3 AR |
995 | pr_err("Invalid queue enabled by amdgpu: %d\n", i); |
996 | break; | |
997 | } | |
998 | ||
999 | res.queue_mask |= (1ull << i); | |
1000 | } | |
d9848e14 OZ |
1001 | res.gws_mask = ~0ull; |
1002 | res.oac_mask = res.gds_heap_base = res.gds_heap_size = 0; | |
64c7f8cf | 1003 | |
79775b62 KR |
1004 | pr_debug("Scheduling resources:\n" |
1005 | "vmid mask: 0x%8X\n" | |
1006 | "queue mask: 0x%8llX\n", | |
64c7f8cf BG |
1007 | res.vmid_mask, res.queue_mask); |
1008 | ||
1009 | return pm_send_set_resources(&dqm->packets, &res); | |
1010 | } | |
1011 | ||
1012 | static int initialize_cpsch(struct device_queue_manager *dqm) | |
1013 | { | |
79775b62 | 1014 | pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm)); |
64c7f8cf | 1015 | |
efeaed4d | 1016 | mutex_init(&dqm->lock_hidden); |
64c7f8cf BG |
1017 | INIT_LIST_HEAD(&dqm->queues); |
1018 | dqm->queue_count = dqm->processes_count = 0; | |
bcea3081 | 1019 | dqm->sdma_queue_count = 0; |
1b4670f6 | 1020 | dqm->xgmi_sdma_queue_count = 0; |
64c7f8cf | 1021 | dqm->active_runlist = false; |
cb77ee7c | 1022 | dqm->sdma_bitmap = (1ULL << get_num_sdma_queues(dqm)) - 1; |
1b4670f6 | 1023 | dqm->xgmi_sdma_bitmap = (1ULL << get_num_xgmi_sdma_queues(dqm)) - 1; |
64c7f8cf | 1024 | |
73ea648d SL |
1025 | INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception); |
1026 | ||
bfd5e378 | 1027 | return 0; |
64c7f8cf BG |
1028 | } |
1029 | ||
1030 | static int start_cpsch(struct device_queue_manager *dqm) | |
1031 | { | |
64c7f8cf BG |
1032 | int retval; |
1033 | ||
64c7f8cf BG |
1034 | retval = 0; |
1035 | ||
1036 | retval = pm_init(&dqm->packets, dqm); | |
4eacc26b | 1037 | if (retval) |
64c7f8cf BG |
1038 | goto fail_packet_manager_init; |
1039 | ||
1040 | retval = set_sched_resources(dqm); | |
4eacc26b | 1041 | if (retval) |
64c7f8cf BG |
1042 | goto fail_set_sched_resources; |
1043 | ||
79775b62 | 1044 | pr_debug("Allocating fence memory\n"); |
64c7f8cf BG |
1045 | |
1046 | /* allocate fence memory on the gart */ | |
a86aa3ca OG |
1047 | retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr), |
1048 | &dqm->fence_mem); | |
64c7f8cf | 1049 | |
4eacc26b | 1050 | if (retval) |
64c7f8cf BG |
1051 | goto fail_allocate_vidmem; |
1052 | ||
1053 | dqm->fence_addr = dqm->fence_mem->cpu_ptr; | |
1054 | dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr; | |
2249d558 AL |
1055 | |
1056 | init_interrupts(dqm); | |
1057 | ||
efeaed4d | 1058 | dqm_lock(dqm); |
73ea648d SL |
1059 | /* clear hang status when driver try to start the hw scheduler */ |
1060 | dqm->is_hws_hang = false; | |
c4744e24 | 1061 | execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); |
efeaed4d | 1062 | dqm_unlock(dqm); |
64c7f8cf BG |
1063 | |
1064 | return 0; | |
1065 | fail_allocate_vidmem: | |
1066 | fail_set_sched_resources: | |
1067 | pm_uninit(&dqm->packets); | |
1068 | fail_packet_manager_init: | |
1069 | return retval; | |
1070 | } | |
1071 | ||
1072 | static int stop_cpsch(struct device_queue_manager *dqm) | |
1073 | { | |
efeaed4d | 1074 | dqm_lock(dqm); |
4465f466 | 1075 | unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); |
efeaed4d | 1076 | dqm_unlock(dqm); |
64c7f8cf | 1077 | |
a86aa3ca | 1078 | kfd_gtt_sa_free(dqm->dev, dqm->fence_mem); |
64c7f8cf BG |
1079 | pm_uninit(&dqm->packets); |
1080 | ||
1081 | return 0; | |
1082 | } | |
1083 | ||
1084 | static int create_kernel_queue_cpsch(struct device_queue_manager *dqm, | |
1085 | struct kernel_queue *kq, | |
1086 | struct qcm_process_device *qpd) | |
1087 | { | |
efeaed4d | 1088 | dqm_lock(dqm); |
b8cbab04 | 1089 | if (dqm->total_queue_count >= max_num_of_queues_per_device) { |
79775b62 | 1090 | pr_warn("Can't create new kernel queue because %d queues were already created\n", |
b8cbab04 | 1091 | dqm->total_queue_count); |
efeaed4d | 1092 | dqm_unlock(dqm); |
b8cbab04 OG |
1093 | return -EPERM; |
1094 | } | |
1095 | ||
1096 | /* | |
1097 | * Unconditionally increment this counter, regardless of the queue's | |
1098 | * type or whether the queue is active. | |
1099 | */ | |
1100 | dqm->total_queue_count++; | |
1101 | pr_debug("Total of %d queues are accountable so far\n", | |
1102 | dqm->total_queue_count); | |
1103 | ||
64c7f8cf BG |
1104 | list_add(&kq->list, &qpd->priv_queue_list); |
1105 | dqm->queue_count++; | |
1106 | qpd->is_debug = true; | |
c4744e24 | 1107 | execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); |
efeaed4d | 1108 | dqm_unlock(dqm); |
64c7f8cf BG |
1109 | |
1110 | return 0; | |
1111 | } | |
1112 | ||
1113 | static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm, | |
1114 | struct kernel_queue *kq, | |
1115 | struct qcm_process_device *qpd) | |
1116 | { | |
efeaed4d | 1117 | dqm_lock(dqm); |
64c7f8cf BG |
1118 | list_del(&kq->list); |
1119 | dqm->queue_count--; | |
1120 | qpd->is_debug = false; | |
c4744e24 | 1121 | execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); |
b8cbab04 OG |
1122 | /* |
1123 | * Unconditionally decrement this counter, regardless of the queue's | |
1124 | * type. | |
1125 | */ | |
8b58f261 | 1126 | dqm->total_queue_count--; |
b8cbab04 OG |
1127 | pr_debug("Total of %d queues are accountable so far\n", |
1128 | dqm->total_queue_count); | |
efeaed4d | 1129 | dqm_unlock(dqm); |
64c7f8cf BG |
1130 | } |
1131 | ||
1132 | static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, | |
b46cb7d7 | 1133 | struct qcm_process_device *qpd) |
64c7f8cf BG |
1134 | { |
1135 | int retval; | |
8d5f3552 | 1136 | struct mqd_manager *mqd_mgr; |
64c7f8cf | 1137 | |
b8cbab04 | 1138 | if (dqm->total_queue_count >= max_num_of_queues_per_device) { |
79775b62 | 1139 | pr_warn("Can't create new usermode queue because %d queues were already created\n", |
b8cbab04 | 1140 | dqm->total_queue_count); |
70d488fb OZ |
1141 | retval = -EPERM; |
1142 | goto out; | |
b8cbab04 OG |
1143 | } |
1144 | ||
1b4670f6 OZ |
1145 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
1146 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
38bb4226 | 1147 | dqm_lock(dqm); |
e78579aa | 1148 | retval = allocate_sdma_queue(dqm, q); |
38bb4226 | 1149 | dqm_unlock(dqm); |
894a8293 | 1150 | if (retval) |
70d488fb | 1151 | goto out; |
e139cd2a | 1152 | } |
ef568db7 FK |
1153 | |
1154 | retval = allocate_doorbell(qpd, q); | |
1155 | if (retval) | |
1156 | goto out_deallocate_sdma_queue; | |
1157 | ||
70d488fb OZ |
1158 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
1159 | q->properties.type)]; | |
70df8273 | 1160 | |
eec0b4cf OZ |
1161 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
1162 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) | |
1163 | dqm->asic_ops.init_sdma_vm(dqm, q, qpd); | |
373d7080 FK |
1164 | q->properties.tba_addr = qpd->tba_addr; |
1165 | q->properties.tma_addr = qpd->tma_addr; | |
70d488fb OZ |
1166 | q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties); |
1167 | if (!q->mqd_mem_obj) { | |
1168 | retval = -ENOMEM; | |
1169 | goto out_deallocate_doorbell; | |
1170 | } | |
70df8273 EH |
1171 | |
1172 | dqm_lock(dqm); | |
1173 | /* | |
1174 | * Eviction state logic: mark all queues as evicted, even ones | |
1175 | * not currently active. Restoring inactive queues later only | |
1176 | * updates the is_evicted flag but is a no-op otherwise. | |
1177 | */ | |
1178 | q->properties.is_evicted = !!qpd->evicted; | |
8636e53c OZ |
1179 | mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, |
1180 | &q->gart_mqd_addr, &q->properties); | |
89cd9d23 | 1181 | |
64c7f8cf | 1182 | list_add(&q->list, &qpd->queues_list); |
bc920fd4 | 1183 | qpd->queue_count++; |
64c7f8cf BG |
1184 | if (q->properties.is_active) { |
1185 | dqm->queue_count++; | |
c4744e24 YZ |
1186 | retval = execute_queues_cpsch(dqm, |
1187 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); | |
64c7f8cf BG |
1188 | } |
1189 | ||
bcea3081 | 1190 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) |
8eabaf54 | 1191 | dqm->sdma_queue_count++; |
1b4670f6 OZ |
1192 | else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) |
1193 | dqm->xgmi_sdma_queue_count++; | |
b8cbab04 OG |
1194 | /* |
1195 | * Unconditionally increment this counter, regardless of the queue's | |
1196 | * type or whether the queue is active. | |
1197 | */ | |
1198 | dqm->total_queue_count++; | |
1199 | ||
1200 | pr_debug("Total of %d queues are accountable so far\n", | |
1201 | dqm->total_queue_count); | |
1202 | ||
efeaed4d | 1203 | dqm_unlock(dqm); |
72a01d23 FK |
1204 | return retval; |
1205 | ||
70d488fb OZ |
1206 | out_deallocate_doorbell: |
1207 | deallocate_doorbell(qpd, q); | |
72a01d23 | 1208 | out_deallocate_sdma_queue: |
1b4670f6 | 1209 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
38bb4226 OZ |
1210 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { |
1211 | dqm_lock(dqm); | |
1b4670f6 | 1212 | deallocate_sdma_queue(dqm, q); |
38bb4226 OZ |
1213 | dqm_unlock(dqm); |
1214 | } | |
70d488fb | 1215 | out: |
64c7f8cf BG |
1216 | return retval; |
1217 | } | |
1218 | ||
788bf83d | 1219 | int amdkfd_fence_wait_timeout(unsigned int *fence_addr, |
d80d19bd | 1220 | unsigned int fence_value, |
8c72c3d7 | 1221 | unsigned int timeout_ms) |
64c7f8cf | 1222 | { |
8c72c3d7 | 1223 | unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies; |
64c7f8cf BG |
1224 | |
1225 | while (*fence_addr != fence_value) { | |
8c72c3d7 | 1226 | if (time_after(jiffies, end_jiffies)) { |
79775b62 | 1227 | pr_err("qcm fence wait loop timeout expired\n"); |
0e9a860c YZ |
1228 | /* In HWS case, this is used to halt the driver thread |
1229 | * in order not to mess up CP states before doing | |
1230 | * scandumps for FW debugging. | |
1231 | */ | |
1232 | while (halt_if_hws_hang) | |
1233 | schedule(); | |
1234 | ||
64c7f8cf BG |
1235 | return -ETIME; |
1236 | } | |
99331a51 | 1237 | schedule(); |
64c7f8cf BG |
1238 | } |
1239 | ||
1240 | return 0; | |
1241 | } | |
1242 | ||
065e4bdf | 1243 | static int unmap_sdma_queues(struct device_queue_manager *dqm) |
bcea3081 | 1244 | { |
065e4bdf OZ |
1245 | int i, retval = 0; |
1246 | ||
1b4670f6 OZ |
1247 | for (i = 0; i < dqm->dev->device_info->num_sdma_engines + |
1248 | dqm->dev->device_info->num_xgmi_sdma_engines; i++) { | |
065e4bdf OZ |
1249 | retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_SDMA, |
1250 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, false, i); | |
1251 | if (retval) | |
1252 | return retval; | |
1253 | } | |
1254 | return retval; | |
bcea3081 BG |
1255 | } |
1256 | ||
60a00956 FK |
1257 | /* dqm->lock mutex has to be locked before calling this function */ |
1258 | static int map_queues_cpsch(struct device_queue_manager *dqm) | |
1259 | { | |
1260 | int retval; | |
1261 | ||
1262 | if (dqm->queue_count <= 0 || dqm->processes_count <= 0) | |
1263 | return 0; | |
1264 | ||
1265 | if (dqm->active_runlist) | |
1266 | return 0; | |
1267 | ||
1268 | retval = pm_send_runlist(&dqm->packets, &dqm->queues); | |
14328aa5 | 1269 | pr_debug("%s sent runlist\n", __func__); |
60a00956 FK |
1270 | if (retval) { |
1271 | pr_err("failed to execute runlist\n"); | |
1272 | return retval; | |
1273 | } | |
1274 | dqm->active_runlist = true; | |
1275 | ||
1276 | return retval; | |
1277 | } | |
1278 | ||
ac30c783 | 1279 | /* dqm->lock mutex has to be locked before calling this function */ |
7da2bcf8 | 1280 | static int unmap_queues_cpsch(struct device_queue_manager *dqm, |
4465f466 YZ |
1281 | enum kfd_unmap_queues_filter filter, |
1282 | uint32_t filter_param) | |
64c7f8cf | 1283 | { |
9fd3f1bf | 1284 | int retval = 0; |
64c7f8cf | 1285 | |
73ea648d SL |
1286 | if (dqm->is_hws_hang) |
1287 | return -EIO; | |
991ca8ee | 1288 | if (!dqm->active_runlist) |
ac30c783 | 1289 | return retval; |
bcea3081 | 1290 | |
1b4670f6 OZ |
1291 | pr_debug("Before destroying queues, sdma queue count is : %u, xgmi sdma queue count is : %u\n", |
1292 | dqm->sdma_queue_count, dqm->xgmi_sdma_queue_count); | |
bcea3081 | 1293 | |
1b4670f6 | 1294 | if (dqm->sdma_queue_count > 0 || dqm->xgmi_sdma_queue_count) |
065e4bdf | 1295 | unmap_sdma_queues(dqm); |
bcea3081 | 1296 | |
64c7f8cf | 1297 | retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE, |
4465f466 | 1298 | filter, filter_param, false, 0); |
4eacc26b | 1299 | if (retval) |
ac30c783 | 1300 | return retval; |
64c7f8cf BG |
1301 | |
1302 | *dqm->fence_addr = KFD_FENCE_INIT; | |
1303 | pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr, | |
1304 | KFD_FENCE_COMPLETED); | |
1305 | /* should be timed out */ | |
c3447e81 | 1306 | retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED, |
14328aa5 | 1307 | queue_preemption_timeout_ms); |
9fd3f1bf | 1308 | if (retval) |
ac30c783 | 1309 | return retval; |
9fd3f1bf | 1310 | |
64c7f8cf BG |
1311 | pm_release_ib(&dqm->packets); |
1312 | dqm->active_runlist = false; | |
1313 | ||
64c7f8cf BG |
1314 | return retval; |
1315 | } | |
1316 | ||
ac30c783 | 1317 | /* dqm->lock mutex has to be locked before calling this function */ |
c4744e24 YZ |
1318 | static int execute_queues_cpsch(struct device_queue_manager *dqm, |
1319 | enum kfd_unmap_queues_filter filter, | |
1320 | uint32_t filter_param) | |
64c7f8cf BG |
1321 | { |
1322 | int retval; | |
1323 | ||
73ea648d SL |
1324 | if (dqm->is_hws_hang) |
1325 | return -EIO; | |
c4744e24 | 1326 | retval = unmap_queues_cpsch(dqm, filter, filter_param); |
4eacc26b | 1327 | if (retval) { |
c4744e24 | 1328 | pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n"); |
73ea648d SL |
1329 | dqm->is_hws_hang = true; |
1330 | schedule_work(&dqm->hw_exception_work); | |
ac30c783 | 1331 | return retval; |
64c7f8cf BG |
1332 | } |
1333 | ||
60a00956 | 1334 | return map_queues_cpsch(dqm); |
64c7f8cf BG |
1335 | } |
1336 | ||
1337 | static int destroy_queue_cpsch(struct device_queue_manager *dqm, | |
1338 | struct qcm_process_device *qpd, | |
1339 | struct queue *q) | |
1340 | { | |
1341 | int retval; | |
8d5f3552 | 1342 | struct mqd_manager *mqd_mgr; |
992839ad | 1343 | |
64c7f8cf BG |
1344 | retval = 0; |
1345 | ||
1346 | /* remove queue from list to prevent rescheduling after preemption */ | |
efeaed4d | 1347 | dqm_lock(dqm); |
992839ad YS |
1348 | |
1349 | if (qpd->is_debug) { | |
1350 | /* | |
1351 | * error, currently we do not allow to destroy a queue | |
1352 | * of a currently debugged process | |
1353 | */ | |
1354 | retval = -EBUSY; | |
1355 | goto failed_try_destroy_debugged_queue; | |
1356 | ||
1357 | } | |
1358 | ||
fdfa090b OZ |
1359 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
1360 | q->properties.type)]; | |
64c7f8cf | 1361 | |
ef568db7 FK |
1362 | deallocate_doorbell(qpd, q); |
1363 | ||
e139cd2a | 1364 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { |
bcea3081 | 1365 | dqm->sdma_queue_count--; |
1b4670f6 OZ |
1366 | deallocate_sdma_queue(dqm, q); |
1367 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
1368 | dqm->xgmi_sdma_queue_count--; | |
1369 | deallocate_sdma_queue(dqm, q); | |
e139cd2a | 1370 | } |
bcea3081 | 1371 | |
64c7f8cf | 1372 | list_del(&q->list); |
bc920fd4 | 1373 | qpd->queue_count--; |
40a526dc | 1374 | if (q->properties.is_active) { |
b6819cec | 1375 | dqm->queue_count--; |
40a526dc | 1376 | retval = execute_queues_cpsch(dqm, |
9fd3f1bf | 1377 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); |
40a526dc YZ |
1378 | if (retval == -ETIME) |
1379 | qpd->reset_wavefronts = true; | |
1380 | } | |
64c7f8cf | 1381 | |
b8cbab04 OG |
1382 | /* |
1383 | * Unconditionally decrement this counter, regardless of the queue's | |
1384 | * type | |
1385 | */ | |
1386 | dqm->total_queue_count--; | |
1387 | pr_debug("Total of %d queues are accountable so far\n", | |
1388 | dqm->total_queue_count); | |
64c7f8cf | 1389 | |
efeaed4d | 1390 | dqm_unlock(dqm); |
64c7f8cf | 1391 | |
8636e53c OZ |
1392 | /* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */ |
1393 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); | |
89cd9d23 | 1394 | |
9e827224 | 1395 | return retval; |
64c7f8cf | 1396 | |
992839ad YS |
1397 | failed_try_destroy_debugged_queue: |
1398 | ||
efeaed4d | 1399 | dqm_unlock(dqm); |
64c7f8cf BG |
1400 | return retval; |
1401 | } | |
1402 | ||
1403 | /* | |
1404 | * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to | |
1405 | * stay in user mode. | |
1406 | */ | |
1407 | #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL | |
1408 | /* APE1 limit is inclusive and 64K aligned. */ | |
1409 | #define APE1_LIMIT_ALIGNMENT 0xFFFF | |
1410 | ||
1411 | static bool set_cache_memory_policy(struct device_queue_manager *dqm, | |
1412 | struct qcm_process_device *qpd, | |
1413 | enum cache_policy default_policy, | |
1414 | enum cache_policy alternate_policy, | |
1415 | void __user *alternate_aperture_base, | |
1416 | uint64_t alternate_aperture_size) | |
1417 | { | |
bed4f110 FK |
1418 | bool retval = true; |
1419 | ||
1420 | if (!dqm->asic_ops.set_cache_memory_policy) | |
1421 | return retval; | |
64c7f8cf | 1422 | |
efeaed4d | 1423 | dqm_lock(dqm); |
64c7f8cf BG |
1424 | |
1425 | if (alternate_aperture_size == 0) { | |
1426 | /* base > limit disables APE1 */ | |
1427 | qpd->sh_mem_ape1_base = 1; | |
1428 | qpd->sh_mem_ape1_limit = 0; | |
1429 | } else { | |
1430 | /* | |
1431 | * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]}, | |
1432 | * SH_MEM_APE1_BASE[31:0], 0x0000 } | |
1433 | * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]}, | |
1434 | * SH_MEM_APE1_LIMIT[31:0], 0xFFFF } | |
1435 | * Verify that the base and size parameters can be | |
1436 | * represented in this format and convert them. | |
1437 | * Additionally restrict APE1 to user-mode addresses. | |
1438 | */ | |
1439 | ||
1440 | uint64_t base = (uintptr_t)alternate_aperture_base; | |
1441 | uint64_t limit = base + alternate_aperture_size - 1; | |
1442 | ||
ab7c1648 KR |
1443 | if (limit <= base || (base & APE1_FIXED_BITS_MASK) != 0 || |
1444 | (limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT) { | |
1445 | retval = false; | |
64c7f8cf | 1446 | goto out; |
ab7c1648 | 1447 | } |
64c7f8cf BG |
1448 | |
1449 | qpd->sh_mem_ape1_base = base >> 16; | |
1450 | qpd->sh_mem_ape1_limit = limit >> 16; | |
1451 | } | |
1452 | ||
bfd5e378 | 1453 | retval = dqm->asic_ops.set_cache_memory_policy( |
a22fc854 BG |
1454 | dqm, |
1455 | qpd, | |
1456 | default_policy, | |
1457 | alternate_policy, | |
1458 | alternate_aperture_base, | |
1459 | alternate_aperture_size); | |
64c7f8cf | 1460 | |
d146c5a7 | 1461 | if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0)) |
64c7f8cf BG |
1462 | program_sh_mem_settings(dqm, qpd); |
1463 | ||
79775b62 | 1464 | pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n", |
64c7f8cf BG |
1465 | qpd->sh_mem_config, qpd->sh_mem_ape1_base, |
1466 | qpd->sh_mem_ape1_limit); | |
1467 | ||
64c7f8cf | 1468 | out: |
efeaed4d | 1469 | dqm_unlock(dqm); |
ab7c1648 | 1470 | return retval; |
64c7f8cf BG |
1471 | } |
1472 | ||
d7b9bd22 FK |
1473 | static int set_trap_handler(struct device_queue_manager *dqm, |
1474 | struct qcm_process_device *qpd, | |
1475 | uint64_t tba_addr, | |
1476 | uint64_t tma_addr) | |
1477 | { | |
1478 | uint64_t *tma; | |
1479 | ||
1480 | if (dqm->dev->cwsr_enabled) { | |
1481 | /* Jump from CWSR trap handler to user trap */ | |
1482 | tma = (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET); | |
1483 | tma[0] = tba_addr; | |
1484 | tma[1] = tma_addr; | |
1485 | } else { | |
1486 | qpd->tba_addr = tba_addr; | |
1487 | qpd->tma_addr = tma_addr; | |
1488 | } | |
1489 | ||
1490 | return 0; | |
1491 | } | |
1492 | ||
9fd3f1bf FK |
1493 | static int process_termination_nocpsch(struct device_queue_manager *dqm, |
1494 | struct qcm_process_device *qpd) | |
1495 | { | |
1496 | struct queue *q, *next; | |
1497 | struct device_process_node *cur, *next_dpn; | |
1498 | int retval = 0; | |
32cce8bc | 1499 | bool found = false; |
9fd3f1bf | 1500 | |
efeaed4d | 1501 | dqm_lock(dqm); |
9fd3f1bf FK |
1502 | |
1503 | /* Clear all user mode queues */ | |
1504 | list_for_each_entry_safe(q, next, &qpd->queues_list, list) { | |
1505 | int ret; | |
1506 | ||
1507 | ret = destroy_queue_nocpsch_locked(dqm, qpd, q); | |
1508 | if (ret) | |
1509 | retval = ret; | |
1510 | } | |
1511 | ||
1512 | /* Unregister process */ | |
1513 | list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) { | |
1514 | if (qpd == cur->qpd) { | |
1515 | list_del(&cur->list); | |
1516 | kfree(cur); | |
1517 | dqm->processes_count--; | |
32cce8bc | 1518 | found = true; |
9fd3f1bf FK |
1519 | break; |
1520 | } | |
1521 | } | |
1522 | ||
efeaed4d | 1523 | dqm_unlock(dqm); |
32cce8bc FK |
1524 | |
1525 | /* Outside the DQM lock because under the DQM lock we can't do | |
1526 | * reclaim or take other locks that others hold while reclaiming. | |
1527 | */ | |
1528 | if (found) | |
1529 | kfd_dec_compute_active(dqm->dev); | |
1530 | ||
9fd3f1bf FK |
1531 | return retval; |
1532 | } | |
1533 | ||
5df099e8 JC |
1534 | static int get_wave_state(struct device_queue_manager *dqm, |
1535 | struct queue *q, | |
1536 | void __user *ctl_stack, | |
1537 | u32 *ctl_stack_used_size, | |
1538 | u32 *save_area_used_size) | |
1539 | { | |
4e6c6fc1 | 1540 | struct mqd_manager *mqd_mgr; |
5df099e8 JC |
1541 | int r; |
1542 | ||
1543 | dqm_lock(dqm); | |
1544 | ||
1545 | if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE || | |
1546 | q->properties.is_active || !q->device->cwsr_enabled) { | |
1547 | r = -EINVAL; | |
1548 | goto dqm_unlock; | |
1549 | } | |
1550 | ||
fdfa090b | 1551 | mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_COMPUTE]; |
5df099e8 | 1552 | |
4e6c6fc1 | 1553 | if (!mqd_mgr->get_wave_state) { |
5df099e8 JC |
1554 | r = -EINVAL; |
1555 | goto dqm_unlock; | |
1556 | } | |
1557 | ||
4e6c6fc1 YZ |
1558 | r = mqd_mgr->get_wave_state(mqd_mgr, q->mqd, ctl_stack, |
1559 | ctl_stack_used_size, save_area_used_size); | |
5df099e8 JC |
1560 | |
1561 | dqm_unlock: | |
1562 | dqm_unlock(dqm); | |
1563 | return r; | |
1564 | } | |
9fd3f1bf FK |
1565 | |
1566 | static int process_termination_cpsch(struct device_queue_manager *dqm, | |
1567 | struct qcm_process_device *qpd) | |
1568 | { | |
1569 | int retval; | |
1570 | struct queue *q, *next; | |
1571 | struct kernel_queue *kq, *kq_next; | |
8d5f3552 | 1572 | struct mqd_manager *mqd_mgr; |
9fd3f1bf FK |
1573 | struct device_process_node *cur, *next_dpn; |
1574 | enum kfd_unmap_queues_filter filter = | |
1575 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES; | |
32cce8bc | 1576 | bool found = false; |
9fd3f1bf FK |
1577 | |
1578 | retval = 0; | |
1579 | ||
efeaed4d | 1580 | dqm_lock(dqm); |
9fd3f1bf FK |
1581 | |
1582 | /* Clean all kernel queues */ | |
1583 | list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) { | |
1584 | list_del(&kq->list); | |
1585 | dqm->queue_count--; | |
1586 | qpd->is_debug = false; | |
1587 | dqm->total_queue_count--; | |
1588 | filter = KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES; | |
1589 | } | |
1590 | ||
1591 | /* Clear all user mode queues */ | |
1592 | list_for_each_entry(q, &qpd->queues_list, list) { | |
72a01d23 | 1593 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { |
9fd3f1bf | 1594 | dqm->sdma_queue_count--; |
1b4670f6 OZ |
1595 | deallocate_sdma_queue(dqm, q); |
1596 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
1597 | dqm->xgmi_sdma_queue_count--; | |
1598 | deallocate_sdma_queue(dqm, q); | |
72a01d23 | 1599 | } |
9fd3f1bf FK |
1600 | |
1601 | if (q->properties.is_active) | |
1602 | dqm->queue_count--; | |
1603 | ||
1604 | dqm->total_queue_count--; | |
1605 | } | |
1606 | ||
1607 | /* Unregister process */ | |
1608 | list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) { | |
1609 | if (qpd == cur->qpd) { | |
1610 | list_del(&cur->list); | |
1611 | kfree(cur); | |
1612 | dqm->processes_count--; | |
32cce8bc | 1613 | found = true; |
9fd3f1bf FK |
1614 | break; |
1615 | } | |
1616 | } | |
1617 | ||
1618 | retval = execute_queues_cpsch(dqm, filter, 0); | |
73ea648d | 1619 | if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) { |
9fd3f1bf FK |
1620 | pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev); |
1621 | dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process); | |
1622 | qpd->reset_wavefronts = false; | |
1623 | } | |
1624 | ||
89cd9d23 PY |
1625 | dqm_unlock(dqm); |
1626 | ||
32cce8bc FK |
1627 | /* Outside the DQM lock because under the DQM lock we can't do |
1628 | * reclaim or take other locks that others hold while reclaiming. | |
1629 | */ | |
1630 | if (found) | |
1631 | kfd_dec_compute_active(dqm->dev); | |
1632 | ||
89cd9d23 | 1633 | /* Lastly, free mqd resources. |
8636e53c | 1634 | * Do free_mqd() after dqm_unlock to avoid circular locking. |
89cd9d23 | 1635 | */ |
9fd3f1bf | 1636 | list_for_each_entry_safe(q, next, &qpd->queues_list, list) { |
fdfa090b OZ |
1637 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
1638 | q->properties.type)]; | |
9fd3f1bf | 1639 | list_del(&q->list); |
bc920fd4 | 1640 | qpd->queue_count--; |
8636e53c | 1641 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); |
9fd3f1bf FK |
1642 | } |
1643 | ||
9fd3f1bf FK |
1644 | return retval; |
1645 | } | |
1646 | ||
fdfa090b OZ |
1647 | static int init_mqd_managers(struct device_queue_manager *dqm) |
1648 | { | |
1649 | int i, j; | |
1650 | struct mqd_manager *mqd_mgr; | |
1651 | ||
1652 | for (i = 0; i < KFD_MQD_TYPE_MAX; i++) { | |
1653 | mqd_mgr = dqm->asic_ops.mqd_manager_init(i, dqm->dev); | |
1654 | if (!mqd_mgr) { | |
1655 | pr_err("mqd manager [%d] initialization failed\n", i); | |
1656 | goto out_free; | |
1657 | } | |
1658 | dqm->mqd_mgrs[i] = mqd_mgr; | |
1659 | } | |
1660 | ||
1661 | return 0; | |
1662 | ||
1663 | out_free: | |
1664 | for (j = 0; j < i; j++) { | |
1665 | kfree(dqm->mqd_mgrs[j]); | |
1666 | dqm->mqd_mgrs[j] = NULL; | |
1667 | } | |
1668 | ||
1669 | return -ENOMEM; | |
1670 | } | |
11614c36 OZ |
1671 | |
1672 | /* Allocate one hiq mqd (HWS) and all SDMA mqd in a continuous trunk*/ | |
1673 | static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm) | |
1674 | { | |
1675 | int retval; | |
1676 | struct kfd_dev *dev = dqm->dev; | |
1677 | struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd; | |
1678 | uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size * | |
1679 | dev->device_info->num_sdma_engines * | |
1680 | dev->device_info->num_sdma_queues_per_engine + | |
1681 | dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size; | |
1682 | ||
1683 | retval = amdgpu_amdkfd_alloc_gtt_mem(dev->kgd, size, | |
1684 | &(mem_obj->gtt_mem), &(mem_obj->gpu_addr), | |
1685 | (void *)&(mem_obj->cpu_ptr), true); | |
1686 | ||
1687 | return retval; | |
1688 | } | |
1689 | ||
64c7f8cf BG |
1690 | struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev) |
1691 | { | |
1692 | struct device_queue_manager *dqm; | |
1693 | ||
79775b62 | 1694 | pr_debug("Loading device queue manager\n"); |
a22fc854 | 1695 | |
dbf56ab1 | 1696 | dqm = kzalloc(sizeof(*dqm), GFP_KERNEL); |
64c7f8cf BG |
1697 | if (!dqm) |
1698 | return NULL; | |
1699 | ||
d146c5a7 FK |
1700 | switch (dev->device_info->asic_family) { |
1701 | /* HWS is not available on Hawaii. */ | |
1702 | case CHIP_HAWAII: | |
1703 | /* HWS depends on CWSR for timely dequeue. CWSR is not | |
1704 | * available on Tonga. | |
1705 | * | |
1706 | * FIXME: This argument also applies to Kaveri. | |
1707 | */ | |
1708 | case CHIP_TONGA: | |
1709 | dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS; | |
1710 | break; | |
1711 | default: | |
1712 | dqm->sched_policy = sched_policy; | |
1713 | break; | |
1714 | } | |
1715 | ||
64c7f8cf | 1716 | dqm->dev = dev; |
d146c5a7 | 1717 | switch (dqm->sched_policy) { |
64c7f8cf BG |
1718 | case KFD_SCHED_POLICY_HWS: |
1719 | case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: | |
1720 | /* initialize dqm for cp scheduling */ | |
45c9a5e4 OG |
1721 | dqm->ops.create_queue = create_queue_cpsch; |
1722 | dqm->ops.initialize = initialize_cpsch; | |
1723 | dqm->ops.start = start_cpsch; | |
1724 | dqm->ops.stop = stop_cpsch; | |
1725 | dqm->ops.destroy_queue = destroy_queue_cpsch; | |
1726 | dqm->ops.update_queue = update_queue; | |
58dcd5bf YZ |
1727 | dqm->ops.register_process = register_process; |
1728 | dqm->ops.unregister_process = unregister_process; | |
1729 | dqm->ops.uninitialize = uninitialize; | |
45c9a5e4 OG |
1730 | dqm->ops.create_kernel_queue = create_kernel_queue_cpsch; |
1731 | dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch; | |
1732 | dqm->ops.set_cache_memory_policy = set_cache_memory_policy; | |
d7b9bd22 | 1733 | dqm->ops.set_trap_handler = set_trap_handler; |
9fd3f1bf | 1734 | dqm->ops.process_termination = process_termination_cpsch; |
26103436 FK |
1735 | dqm->ops.evict_process_queues = evict_process_queues_cpsch; |
1736 | dqm->ops.restore_process_queues = restore_process_queues_cpsch; | |
5df099e8 | 1737 | dqm->ops.get_wave_state = get_wave_state; |
64c7f8cf BG |
1738 | break; |
1739 | case KFD_SCHED_POLICY_NO_HWS: | |
1740 | /* initialize dqm for no cp scheduling */ | |
45c9a5e4 OG |
1741 | dqm->ops.start = start_nocpsch; |
1742 | dqm->ops.stop = stop_nocpsch; | |
1743 | dqm->ops.create_queue = create_queue_nocpsch; | |
1744 | dqm->ops.destroy_queue = destroy_queue_nocpsch; | |
1745 | dqm->ops.update_queue = update_queue; | |
58dcd5bf YZ |
1746 | dqm->ops.register_process = register_process; |
1747 | dqm->ops.unregister_process = unregister_process; | |
45c9a5e4 | 1748 | dqm->ops.initialize = initialize_nocpsch; |
58dcd5bf | 1749 | dqm->ops.uninitialize = uninitialize; |
45c9a5e4 | 1750 | dqm->ops.set_cache_memory_policy = set_cache_memory_policy; |
d7b9bd22 | 1751 | dqm->ops.set_trap_handler = set_trap_handler; |
9fd3f1bf | 1752 | dqm->ops.process_termination = process_termination_nocpsch; |
26103436 FK |
1753 | dqm->ops.evict_process_queues = evict_process_queues_nocpsch; |
1754 | dqm->ops.restore_process_queues = | |
1755 | restore_process_queues_nocpsch; | |
5df099e8 | 1756 | dqm->ops.get_wave_state = get_wave_state; |
64c7f8cf BG |
1757 | break; |
1758 | default: | |
d146c5a7 | 1759 | pr_err("Invalid scheduling policy %d\n", dqm->sched_policy); |
32fa8219 | 1760 | goto out_free; |
64c7f8cf BG |
1761 | } |
1762 | ||
a22fc854 BG |
1763 | switch (dev->device_info->asic_family) { |
1764 | case CHIP_CARRIZO: | |
bfd5e378 | 1765 | device_queue_manager_init_vi(&dqm->asic_ops); |
300dec95 OG |
1766 | break; |
1767 | ||
a22fc854 | 1768 | case CHIP_KAVERI: |
bfd5e378 | 1769 | device_queue_manager_init_cik(&dqm->asic_ops); |
300dec95 | 1770 | break; |
97672cbe FK |
1771 | |
1772 | case CHIP_HAWAII: | |
1773 | device_queue_manager_init_cik_hawaii(&dqm->asic_ops); | |
1774 | break; | |
1775 | ||
1776 | case CHIP_TONGA: | |
1777 | case CHIP_FIJI: | |
1778 | case CHIP_POLARIS10: | |
1779 | case CHIP_POLARIS11: | |
846a44d7 | 1780 | case CHIP_POLARIS12: |
ed81cd6e | 1781 | case CHIP_VEGAM: |
97672cbe FK |
1782 | device_queue_manager_init_vi_tonga(&dqm->asic_ops); |
1783 | break; | |
bed4f110 FK |
1784 | |
1785 | case CHIP_VEGA10: | |
846a44d7 | 1786 | case CHIP_VEGA12: |
22a3a294 | 1787 | case CHIP_VEGA20: |
bed4f110 | 1788 | case CHIP_RAVEN: |
49adcf8a | 1789 | case CHIP_ARCTURUS: |
bed4f110 FK |
1790 | device_queue_manager_init_v9(&dqm->asic_ops); |
1791 | break; | |
14328aa5 PC |
1792 | case CHIP_NAVI10: |
1793 | device_queue_manager_init_v10_navi10(&dqm->asic_ops); | |
1794 | break; | |
e596b903 YZ |
1795 | default: |
1796 | WARN(1, "Unexpected ASIC family %u", | |
1797 | dev->device_info->asic_family); | |
1798 | goto out_free; | |
a22fc854 BG |
1799 | } |
1800 | ||
fdfa090b OZ |
1801 | if (init_mqd_managers(dqm)) |
1802 | goto out_free; | |
1803 | ||
11614c36 OZ |
1804 | if (allocate_hiq_sdma_mqd(dqm)) { |
1805 | pr_err("Failed to allocate hiq sdma mqd trunk buffer\n"); | |
1806 | goto out_free; | |
1807 | } | |
1808 | ||
32fa8219 FK |
1809 | if (!dqm->ops.initialize(dqm)) |
1810 | return dqm; | |
64c7f8cf | 1811 | |
32fa8219 FK |
1812 | out_free: |
1813 | kfree(dqm); | |
1814 | return NULL; | |
64c7f8cf BG |
1815 | } |
1816 | ||
11614c36 OZ |
1817 | void deallocate_hiq_sdma_mqd(struct kfd_dev *dev, struct kfd_mem_obj *mqd) |
1818 | { | |
1819 | WARN(!mqd, "No hiq sdma mqd trunk to free"); | |
1820 | ||
1821 | amdgpu_amdkfd_free_gtt_mem(dev->kgd, mqd->gtt_mem); | |
1822 | } | |
1823 | ||
64c7f8cf BG |
1824 | void device_queue_manager_uninit(struct device_queue_manager *dqm) |
1825 | { | |
45c9a5e4 | 1826 | dqm->ops.uninitialize(dqm); |
11614c36 | 1827 | deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd); |
64c7f8cf BG |
1828 | kfree(dqm); |
1829 | } | |
851a645e | 1830 | |
2640c3fa | 1831 | int kfd_process_vm_fault(struct device_queue_manager *dqm, |
1832 | unsigned int pasid) | |
1833 | { | |
1834 | struct kfd_process_device *pdd; | |
1835 | struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); | |
1836 | int ret = 0; | |
1837 | ||
1838 | if (!p) | |
1839 | return -EINVAL; | |
1840 | pdd = kfd_get_process_device_data(dqm->dev, p); | |
1841 | if (pdd) | |
1842 | ret = dqm->ops.evict_process_queues(dqm, &pdd->qpd); | |
1843 | kfd_unref_process(p); | |
1844 | ||
1845 | return ret; | |
1846 | } | |
1847 | ||
73ea648d SL |
1848 | static void kfd_process_hw_exception(struct work_struct *work) |
1849 | { | |
1850 | struct device_queue_manager *dqm = container_of(work, | |
1851 | struct device_queue_manager, hw_exception_work); | |
5b87245f | 1852 | amdgpu_amdkfd_gpu_reset(dqm->dev->kgd); |
73ea648d SL |
1853 | } |
1854 | ||
851a645e FK |
1855 | #if defined(CONFIG_DEBUG_FS) |
1856 | ||
1857 | static void seq_reg_dump(struct seq_file *m, | |
1858 | uint32_t (*dump)[2], uint32_t n_regs) | |
1859 | { | |
1860 | uint32_t i, count; | |
1861 | ||
1862 | for (i = 0, count = 0; i < n_regs; i++) { | |
1863 | if (count == 0 || | |
1864 | dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) { | |
1865 | seq_printf(m, "%s %08x: %08x", | |
1866 | i ? "\n" : "", | |
1867 | dump[i][0], dump[i][1]); | |
1868 | count = 7; | |
1869 | } else { | |
1870 | seq_printf(m, " %08x", dump[i][1]); | |
1871 | count--; | |
1872 | } | |
1873 | } | |
1874 | ||
1875 | seq_puts(m, "\n"); | |
1876 | } | |
1877 | ||
1878 | int dqm_debugfs_hqds(struct seq_file *m, void *data) | |
1879 | { | |
1880 | struct device_queue_manager *dqm = data; | |
1881 | uint32_t (*dump)[2], n_regs; | |
1882 | int pipe, queue; | |
1883 | int r = 0; | |
1884 | ||
24f48a42 | 1885 | r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->kgd, |
14328aa5 PC |
1886 | KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE, |
1887 | &dump, &n_regs); | |
24f48a42 OZ |
1888 | if (!r) { |
1889 | seq_printf(m, " HIQ on MEC %d Pipe %d Queue %d\n", | |
14328aa5 PC |
1890 | KFD_CIK_HIQ_PIPE/get_pipes_per_mec(dqm)+1, |
1891 | KFD_CIK_HIQ_PIPE%get_pipes_per_mec(dqm), | |
1892 | KFD_CIK_HIQ_QUEUE); | |
24f48a42 OZ |
1893 | seq_reg_dump(m, dump, n_regs); |
1894 | ||
1895 | kfree(dump); | |
1896 | } | |
1897 | ||
851a645e FK |
1898 | for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) { |
1899 | int pipe_offset = pipe * get_queues_per_pipe(dqm); | |
1900 | ||
1901 | for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) { | |
1902 | if (!test_bit(pipe_offset + queue, | |
1903 | dqm->dev->shared_resources.queue_bitmap)) | |
1904 | continue; | |
1905 | ||
1906 | r = dqm->dev->kfd2kgd->hqd_dump( | |
1907 | dqm->dev->kgd, pipe, queue, &dump, &n_regs); | |
1908 | if (r) | |
1909 | break; | |
1910 | ||
1911 | seq_printf(m, " CP Pipe %d, Queue %d\n", | |
1912 | pipe, queue); | |
1913 | seq_reg_dump(m, dump, n_regs); | |
1914 | ||
1915 | kfree(dump); | |
1916 | } | |
1917 | } | |
1918 | ||
98bb9222 | 1919 | for (pipe = 0; pipe < get_num_sdma_engines(dqm); pipe++) { |
d5094189 SL |
1920 | for (queue = 0; |
1921 | queue < dqm->dev->device_info->num_sdma_queues_per_engine; | |
1922 | queue++) { | |
851a645e FK |
1923 | r = dqm->dev->kfd2kgd->hqd_sdma_dump( |
1924 | dqm->dev->kgd, pipe, queue, &dump, &n_regs); | |
1925 | if (r) | |
1926 | break; | |
1927 | ||
1928 | seq_printf(m, " SDMA Engine %d, RLC %d\n", | |
1929 | pipe, queue); | |
1930 | seq_reg_dump(m, dump, n_regs); | |
1931 | ||
1932 | kfree(dump); | |
1933 | } | |
1934 | } | |
1935 | ||
1936 | return r; | |
1937 | } | |
1938 | ||
a29ec470 SL |
1939 | int dqm_debugfs_execute_queues(struct device_queue_manager *dqm) |
1940 | { | |
1941 | int r = 0; | |
1942 | ||
1943 | dqm_lock(dqm); | |
1944 | dqm->active_runlist = true; | |
1945 | r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); | |
1946 | dqm_unlock(dqm); | |
1947 | ||
1948 | return r; | |
1949 | } | |
1950 | ||
851a645e | 1951 | #endif |