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 | ||
06b89b38 OZ |
277 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
278 | q->properties.type)]; | |
279 | q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties); | |
280 | if (!q->mqd_mem_obj) | |
281 | return -ENOMEM; | |
282 | ||
efeaed4d | 283 | dqm_lock(dqm); |
64c7f8cf | 284 | |
b8cbab04 | 285 | if (dqm->total_queue_count >= max_num_of_queues_per_device) { |
79775b62 | 286 | pr_warn("Can't create new usermode queue because %d queues were already created\n", |
b8cbab04 | 287 | dqm->total_queue_count); |
ab7c1648 KR |
288 | retval = -EPERM; |
289 | goto out_unlock; | |
b8cbab04 OG |
290 | } |
291 | ||
64c7f8cf BG |
292 | if (list_empty(&qpd->queues_list)) { |
293 | retval = allocate_vmid(dqm, qpd, q); | |
ab7c1648 KR |
294 | if (retval) |
295 | goto out_unlock; | |
64c7f8cf | 296 | } |
64c7f8cf | 297 | q->properties.vmid = qpd->vmid; |
26103436 | 298 | /* |
bb2d2128 FK |
299 | * Eviction state logic: mark all queues as evicted, even ones |
300 | * not currently active. Restoring inactive queues later only | |
301 | * updates the is_evicted flag but is a no-op otherwise. | |
26103436 | 302 | */ |
bb2d2128 | 303 | q->properties.is_evicted = !!qpd->evicted; |
64c7f8cf | 304 | |
373d7080 FK |
305 | q->properties.tba_addr = qpd->tba_addr; |
306 | q->properties.tma_addr = qpd->tma_addr; | |
307 | ||
d39b7737 OZ |
308 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) { |
309 | retval = allocate_hqd(dqm, q); | |
310 | if (retval) | |
311 | goto deallocate_vmid; | |
312 | pr_debug("Loading mqd to hqd on pipe %d, queue %d\n", | |
313 | q->pipe, q->queue); | |
314 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || | |
315 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
316 | retval = allocate_sdma_queue(dqm, q); | |
317 | if (retval) | |
318 | goto deallocate_vmid; | |
319 | dqm->asic_ops.init_sdma_vm(dqm, q, qpd); | |
320 | } | |
321 | ||
322 | retval = allocate_doorbell(qpd, q); | |
323 | if (retval) | |
324 | goto out_deallocate_hqd; | |
325 | ||
8636e53c OZ |
326 | mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, |
327 | &q->gart_mqd_addr, &q->properties); | |
d39b7737 OZ |
328 | if (q->properties.is_active) { |
329 | ||
330 | if (WARN(q->process->mm != current->mm, | |
331 | "should only run in user thread")) | |
332 | retval = -EFAULT; | |
333 | else | |
334 | retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe, | |
335 | q->queue, &q->properties, current->mm); | |
336 | if (retval) | |
06b89b38 | 337 | goto out_deallocate_doorbell; |
64c7f8cf BG |
338 | } |
339 | ||
340 | list_add(&q->list, &qpd->queues_list); | |
bc920fd4 | 341 | qpd->queue_count++; |
b6819cec JC |
342 | if (q->properties.is_active) |
343 | dqm->queue_count++; | |
64c7f8cf | 344 | |
bcea3081 BG |
345 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) |
346 | dqm->sdma_queue_count++; | |
1b4670f6 OZ |
347 | else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) |
348 | dqm->xgmi_sdma_queue_count++; | |
64c7f8cf | 349 | |
b8cbab04 OG |
350 | /* |
351 | * Unconditionally increment this counter, regardless of the queue's | |
352 | * type or whether the queue is active. | |
353 | */ | |
354 | dqm->total_queue_count++; | |
355 | pr_debug("Total of %d queues are accountable so far\n", | |
356 | dqm->total_queue_count); | |
06b89b38 OZ |
357 | dqm_unlock(dqm); |
358 | return retval; | |
b8cbab04 | 359 | |
d39b7737 OZ |
360 | out_deallocate_doorbell: |
361 | deallocate_doorbell(qpd, q); | |
362 | out_deallocate_hqd: | |
363 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) | |
364 | deallocate_hqd(dqm, q); | |
365 | else if (q->properties.type == KFD_QUEUE_TYPE_SDMA || | |
366 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) | |
367 | deallocate_sdma_queue(dqm, q); | |
368 | deallocate_vmid: | |
369 | if (list_empty(&qpd->queues_list)) | |
370 | deallocate_vmid(dqm, qpd, q); | |
ab7c1648 | 371 | out_unlock: |
efeaed4d | 372 | dqm_unlock(dqm); |
06b89b38 | 373 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); |
ab7c1648 | 374 | return retval; |
64c7f8cf BG |
375 | } |
376 | ||
377 | static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q) | |
378 | { | |
379 | bool set; | |
f0ec5b99 | 380 | int pipe, bit, i; |
64c7f8cf BG |
381 | |
382 | set = false; | |
383 | ||
8eabaf54 KR |
384 | for (pipe = dqm->next_pipe_to_allocate, i = 0; |
385 | i < get_pipes_per_mec(dqm); | |
d0b63bb3 AR |
386 | pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) { |
387 | ||
388 | if (!is_pipe_enabled(dqm, 0, pipe)) | |
389 | continue; | |
390 | ||
64c7f8cf | 391 | if (dqm->allocated_queues[pipe] != 0) { |
4252bf68 HK |
392 | bit = ffs(dqm->allocated_queues[pipe]) - 1; |
393 | dqm->allocated_queues[pipe] &= ~(1 << bit); | |
64c7f8cf BG |
394 | q->pipe = pipe; |
395 | q->queue = bit; | |
396 | set = true; | |
397 | break; | |
398 | } | |
399 | } | |
400 | ||
991ca8ee | 401 | if (!set) |
64c7f8cf BG |
402 | return -EBUSY; |
403 | ||
79775b62 | 404 | pr_debug("hqd slot - pipe %d, queue %d\n", q->pipe, q->queue); |
64c7f8cf | 405 | /* horizontal hqd allocation */ |
d0b63bb3 | 406 | dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm); |
64c7f8cf BG |
407 | |
408 | return 0; | |
409 | } | |
410 | ||
411 | static inline void deallocate_hqd(struct device_queue_manager *dqm, | |
412 | struct queue *q) | |
413 | { | |
4252bf68 | 414 | dqm->allocated_queues[q->pipe] |= (1 << q->queue); |
64c7f8cf BG |
415 | } |
416 | ||
9fd3f1bf FK |
417 | /* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked |
418 | * to avoid asynchronized access | |
419 | */ | |
420 | static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm, | |
64c7f8cf BG |
421 | struct qcm_process_device *qpd, |
422 | struct queue *q) | |
423 | { | |
424 | int retval; | |
8d5f3552 | 425 | struct mqd_manager *mqd_mgr; |
64c7f8cf | 426 | |
fdfa090b OZ |
427 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
428 | q->properties.type)]; | |
64c7f8cf | 429 | |
c2e1b3a4 | 430 | if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) { |
c2e1b3a4 BG |
431 | deallocate_hqd(dqm, q); |
432 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { | |
c2e1b3a4 | 433 | dqm->sdma_queue_count--; |
1b4670f6 OZ |
434 | deallocate_sdma_queue(dqm, q); |
435 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
436 | dqm->xgmi_sdma_queue_count--; | |
437 | deallocate_sdma_queue(dqm, q); | |
7113cd65 | 438 | } else { |
79775b62 | 439 | pr_debug("q->properties.type %d is invalid\n", |
7113cd65 | 440 | q->properties.type); |
9fd3f1bf | 441 | return -EINVAL; |
64c7f8cf | 442 | } |
9fd3f1bf | 443 | dqm->total_queue_count--; |
64c7f8cf | 444 | |
ef568db7 FK |
445 | deallocate_doorbell(qpd, q); |
446 | ||
8d5f3552 | 447 | retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, |
c2e1b3a4 | 448 | KFD_PREEMPT_TYPE_WAVEFRONT_RESET, |
b90e3fbe | 449 | KFD_UNMAP_LATENCY_MS, |
64c7f8cf | 450 | q->pipe, q->queue); |
9fd3f1bf FK |
451 | if (retval == -ETIME) |
452 | qpd->reset_wavefronts = true; | |
64c7f8cf | 453 | |
8636e53c | 454 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); |
64c7f8cf BG |
455 | |
456 | list_del(&q->list); | |
9fd3f1bf FK |
457 | if (list_empty(&qpd->queues_list)) { |
458 | if (qpd->reset_wavefronts) { | |
459 | pr_warn("Resetting wave fronts (nocpsch) on dev %p\n", | |
460 | dqm->dev); | |
461 | /* dbgdev_wave_reset_wavefronts has to be called before | |
462 | * deallocate_vmid(), i.e. when vmid is still in use. | |
463 | */ | |
464 | dbgdev_wave_reset_wavefronts(dqm->dev, | |
465 | qpd->pqm->process); | |
466 | qpd->reset_wavefronts = false; | |
467 | } | |
468 | ||
64c7f8cf | 469 | deallocate_vmid(dqm, qpd, q); |
9fd3f1bf | 470 | } |
bc920fd4 | 471 | qpd->queue_count--; |
b6819cec JC |
472 | if (q->properties.is_active) |
473 | dqm->queue_count--; | |
b8cbab04 | 474 | |
9fd3f1bf FK |
475 | return retval; |
476 | } | |
b8cbab04 | 477 | |
9fd3f1bf FK |
478 | static int destroy_queue_nocpsch(struct device_queue_manager *dqm, |
479 | struct qcm_process_device *qpd, | |
480 | struct queue *q) | |
481 | { | |
482 | int retval; | |
483 | ||
efeaed4d | 484 | dqm_lock(dqm); |
9fd3f1bf | 485 | retval = destroy_queue_nocpsch_locked(dqm, qpd, q); |
efeaed4d | 486 | dqm_unlock(dqm); |
9fd3f1bf | 487 | |
64c7f8cf BG |
488 | return retval; |
489 | } | |
490 | ||
491 | static int update_queue(struct device_queue_manager *dqm, struct queue *q) | |
492 | { | |
8636e53c | 493 | int retval = 0; |
8d5f3552 | 494 | struct mqd_manager *mqd_mgr; |
26103436 | 495 | struct kfd_process_device *pdd; |
b6ffbab8 | 496 | bool prev_active = false; |
64c7f8cf | 497 | |
efeaed4d | 498 | dqm_lock(dqm); |
26103436 FK |
499 | pdd = kfd_get_process_device_data(q->device, q->process); |
500 | if (!pdd) { | |
501 | retval = -ENODEV; | |
502 | goto out_unlock; | |
503 | } | |
fdfa090b OZ |
504 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
505 | q->properties.type)]; | |
64c7f8cf | 506 | |
60a00956 FK |
507 | /* Save previous activity state for counters */ |
508 | prev_active = q->properties.is_active; | |
509 | ||
510 | /* Make sure the queue is unmapped before updating the MQD */ | |
d146c5a7 | 511 | if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { |
60a00956 FK |
512 | retval = unmap_queues_cpsch(dqm, |
513 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); | |
894a8293 | 514 | if (retval) { |
60a00956 FK |
515 | pr_err("unmap queue failed\n"); |
516 | goto out_unlock; | |
517 | } | |
894a8293 | 518 | } else if (prev_active && |
60a00956 | 519 | (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || |
1b4670f6 OZ |
520 | q->properties.type == KFD_QUEUE_TYPE_SDMA || |
521 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { | |
8d5f3552 | 522 | retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, |
60a00956 FK |
523 | KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN, |
524 | KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); | |
525 | if (retval) { | |
526 | pr_err("destroy mqd failed\n"); | |
527 | goto out_unlock; | |
528 | } | |
529 | } | |
530 | ||
8636e53c | 531 | mqd_mgr->update_mqd(mqd_mgr, q->mqd, &q->properties); |
60a00956 | 532 | |
096d1a3e FK |
533 | /* |
534 | * check active state vs. the previous state and modify | |
535 | * counter accordingly. map_queues_cpsch uses the | |
536 | * dqm->queue_count to determine whether a new runlist must be | |
537 | * uploaded. | |
538 | */ | |
539 | if (q->properties.is_active && !prev_active) | |
540 | dqm->queue_count++; | |
541 | else if (!q->properties.is_active && prev_active) | |
542 | dqm->queue_count--; | |
543 | ||
d146c5a7 | 544 | if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) |
60a00956 | 545 | retval = map_queues_cpsch(dqm); |
894a8293 | 546 | else if (q->properties.is_active && |
60a00956 | 547 | (q->properties.type == KFD_QUEUE_TYPE_COMPUTE || |
1b4670f6 OZ |
548 | q->properties.type == KFD_QUEUE_TYPE_SDMA || |
549 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) { | |
1b19aa5a FK |
550 | if (WARN(q->process->mm != current->mm, |
551 | "should only run in user thread")) | |
552 | retval = -EFAULT; | |
553 | else | |
554 | retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, | |
555 | q->pipe, q->queue, | |
556 | &q->properties, current->mm); | |
557 | } | |
b6ffbab8 | 558 | |
ab7c1648 | 559 | out_unlock: |
efeaed4d | 560 | dqm_unlock(dqm); |
64c7f8cf BG |
561 | return retval; |
562 | } | |
563 | ||
26103436 FK |
564 | static int evict_process_queues_nocpsch(struct device_queue_manager *dqm, |
565 | struct qcm_process_device *qpd) | |
566 | { | |
567 | struct queue *q; | |
8d5f3552 | 568 | struct mqd_manager *mqd_mgr; |
26103436 | 569 | struct kfd_process_device *pdd; |
bb2d2128 | 570 | int retval, ret = 0; |
26103436 | 571 | |
efeaed4d | 572 | dqm_lock(dqm); |
26103436 FK |
573 | if (qpd->evicted++ > 0) /* already evicted, do nothing */ |
574 | goto out; | |
575 | ||
576 | pdd = qpd_to_pdd(qpd); | |
577 | pr_info_ratelimited("Evicting PASID %u queues\n", | |
578 | pdd->process->pasid); | |
579 | ||
bb2d2128 FK |
580 | /* Mark all queues as evicted. Deactivate all active queues on |
581 | * the qpd. | |
582 | */ | |
26103436 | 583 | list_for_each_entry(q, &qpd->queues_list, list) { |
bb2d2128 | 584 | q->properties.is_evicted = true; |
26103436 FK |
585 | if (!q->properties.is_active) |
586 | continue; | |
bb2d2128 | 587 | |
fdfa090b OZ |
588 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
589 | q->properties.type)]; | |
26103436 | 590 | q->properties.is_active = false; |
8d5f3552 | 591 | retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd, |
26103436 FK |
592 | KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN, |
593 | KFD_UNMAP_LATENCY_MS, q->pipe, q->queue); | |
bb2d2128 FK |
594 | if (retval && !ret) |
595 | /* Return the first error, but keep going to | |
596 | * maintain a consistent eviction state | |
597 | */ | |
598 | ret = retval; | |
26103436 FK |
599 | dqm->queue_count--; |
600 | } | |
601 | ||
602 | out: | |
efeaed4d | 603 | dqm_unlock(dqm); |
bb2d2128 | 604 | return ret; |
26103436 FK |
605 | } |
606 | ||
607 | static int evict_process_queues_cpsch(struct device_queue_manager *dqm, | |
608 | struct qcm_process_device *qpd) | |
609 | { | |
610 | struct queue *q; | |
611 | struct kfd_process_device *pdd; | |
612 | int retval = 0; | |
613 | ||
efeaed4d | 614 | dqm_lock(dqm); |
26103436 FK |
615 | if (qpd->evicted++ > 0) /* already evicted, do nothing */ |
616 | goto out; | |
617 | ||
618 | pdd = qpd_to_pdd(qpd); | |
619 | pr_info_ratelimited("Evicting PASID %u queues\n", | |
620 | pdd->process->pasid); | |
621 | ||
bb2d2128 FK |
622 | /* Mark all queues as evicted. Deactivate all active queues on |
623 | * the qpd. | |
624 | */ | |
26103436 | 625 | list_for_each_entry(q, &qpd->queues_list, list) { |
bb2d2128 | 626 | q->properties.is_evicted = true; |
26103436 FK |
627 | if (!q->properties.is_active) |
628 | continue; | |
bb2d2128 | 629 | |
26103436 FK |
630 | q->properties.is_active = false; |
631 | dqm->queue_count--; | |
632 | } | |
633 | retval = execute_queues_cpsch(dqm, | |
634 | qpd->is_debug ? | |
635 | KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES : | |
636 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); | |
637 | ||
638 | out: | |
efeaed4d | 639 | dqm_unlock(dqm); |
26103436 FK |
640 | return retval; |
641 | } | |
642 | ||
643 | static int restore_process_queues_nocpsch(struct device_queue_manager *dqm, | |
644 | struct qcm_process_device *qpd) | |
645 | { | |
1b19aa5a | 646 | struct mm_struct *mm = NULL; |
26103436 | 647 | struct queue *q; |
8d5f3552 | 648 | struct mqd_manager *mqd_mgr; |
26103436 | 649 | struct kfd_process_device *pdd; |
e715c6d0 | 650 | uint64_t pd_base; |
bb2d2128 | 651 | int retval, ret = 0; |
26103436 FK |
652 | |
653 | pdd = qpd_to_pdd(qpd); | |
654 | /* Retrieve PD base */ | |
5b87245f | 655 | pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm); |
26103436 | 656 | |
efeaed4d | 657 | dqm_lock(dqm); |
26103436 FK |
658 | if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ |
659 | goto out; | |
660 | if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */ | |
661 | qpd->evicted--; | |
662 | goto out; | |
663 | } | |
664 | ||
665 | pr_info_ratelimited("Restoring PASID %u queues\n", | |
666 | pdd->process->pasid); | |
667 | ||
668 | /* Update PD Base in QPD */ | |
669 | qpd->page_table_base = pd_base; | |
e715c6d0 | 670 | pr_debug("Updated PD address to 0x%llx\n", pd_base); |
26103436 FK |
671 | |
672 | if (!list_empty(&qpd->queues_list)) { | |
673 | dqm->dev->kfd2kgd->set_vm_context_page_table_base( | |
674 | dqm->dev->kgd, | |
675 | qpd->vmid, | |
676 | qpd->page_table_base); | |
677 | kfd_flush_tlb(pdd); | |
678 | } | |
679 | ||
1b19aa5a FK |
680 | /* Take a safe reference to the mm_struct, which may otherwise |
681 | * disappear even while the kfd_process is still referenced. | |
682 | */ | |
683 | mm = get_task_mm(pdd->process->lead_thread); | |
684 | if (!mm) { | |
bb2d2128 | 685 | ret = -EFAULT; |
1b19aa5a FK |
686 | goto out; |
687 | } | |
688 | ||
bb2d2128 FK |
689 | /* Remove the eviction flags. Activate queues that are not |
690 | * inactive for other reasons. | |
691 | */ | |
26103436 | 692 | list_for_each_entry(q, &qpd->queues_list, list) { |
bb2d2128 FK |
693 | q->properties.is_evicted = false; |
694 | if (!QUEUE_IS_ACTIVE(q->properties)) | |
26103436 | 695 | continue; |
bb2d2128 | 696 | |
fdfa090b OZ |
697 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
698 | q->properties.type)]; | |
26103436 | 699 | q->properties.is_active = true; |
8d5f3552 | 700 | retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe, |
1b19aa5a | 701 | q->queue, &q->properties, mm); |
bb2d2128 FK |
702 | if (retval && !ret) |
703 | /* Return the first error, but keep going to | |
704 | * maintain a consistent eviction state | |
705 | */ | |
706 | ret = retval; | |
26103436 FK |
707 | dqm->queue_count++; |
708 | } | |
709 | qpd->evicted = 0; | |
710 | out: | |
1b19aa5a FK |
711 | if (mm) |
712 | mmput(mm); | |
efeaed4d | 713 | dqm_unlock(dqm); |
bb2d2128 | 714 | return ret; |
26103436 FK |
715 | } |
716 | ||
717 | static int restore_process_queues_cpsch(struct device_queue_manager *dqm, | |
718 | struct qcm_process_device *qpd) | |
719 | { | |
720 | struct queue *q; | |
721 | struct kfd_process_device *pdd; | |
e715c6d0 | 722 | uint64_t pd_base; |
26103436 FK |
723 | int retval = 0; |
724 | ||
725 | pdd = qpd_to_pdd(qpd); | |
726 | /* Retrieve PD base */ | |
5b87245f | 727 | pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm); |
26103436 | 728 | |
efeaed4d | 729 | dqm_lock(dqm); |
26103436 FK |
730 | if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */ |
731 | goto out; | |
732 | if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */ | |
733 | qpd->evicted--; | |
734 | goto out; | |
735 | } | |
736 | ||
737 | pr_info_ratelimited("Restoring PASID %u queues\n", | |
738 | pdd->process->pasid); | |
739 | ||
740 | /* Update PD Base in QPD */ | |
741 | qpd->page_table_base = pd_base; | |
e715c6d0 | 742 | pr_debug("Updated PD address to 0x%llx\n", pd_base); |
26103436 FK |
743 | |
744 | /* activate all active queues on the qpd */ | |
745 | list_for_each_entry(q, &qpd->queues_list, list) { | |
26103436 | 746 | q->properties.is_evicted = false; |
bb2d2128 FK |
747 | if (!QUEUE_IS_ACTIVE(q->properties)) |
748 | continue; | |
749 | ||
26103436 FK |
750 | q->properties.is_active = true; |
751 | dqm->queue_count++; | |
752 | } | |
753 | retval = execute_queues_cpsch(dqm, | |
754 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); | |
bb2d2128 | 755 | qpd->evicted = 0; |
26103436 | 756 | out: |
efeaed4d | 757 | dqm_unlock(dqm); |
26103436 FK |
758 | return retval; |
759 | } | |
760 | ||
58dcd5bf | 761 | static int register_process(struct device_queue_manager *dqm, |
64c7f8cf BG |
762 | struct qcm_process_device *qpd) |
763 | { | |
764 | struct device_process_node *n; | |
403575c4 | 765 | struct kfd_process_device *pdd; |
e715c6d0 | 766 | uint64_t pd_base; |
a22fc854 | 767 | int retval; |
64c7f8cf | 768 | |
dbf56ab1 | 769 | n = kzalloc(sizeof(*n), GFP_KERNEL); |
64c7f8cf BG |
770 | if (!n) |
771 | return -ENOMEM; | |
772 | ||
773 | n->qpd = qpd; | |
774 | ||
403575c4 FK |
775 | pdd = qpd_to_pdd(qpd); |
776 | /* Retrieve PD base */ | |
5b87245f | 777 | pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm); |
403575c4 | 778 | |
efeaed4d | 779 | dqm_lock(dqm); |
64c7f8cf BG |
780 | list_add(&n->list, &dqm->queues); |
781 | ||
403575c4 FK |
782 | /* Update PD Base in QPD */ |
783 | qpd->page_table_base = pd_base; | |
e715c6d0 | 784 | pr_debug("Updated PD address to 0x%llx\n", pd_base); |
403575c4 | 785 | |
bfd5e378 | 786 | retval = dqm->asic_ops.update_qpd(dqm, qpd); |
a22fc854 | 787 | |
f756e631 | 788 | dqm->processes_count++; |
64c7f8cf | 789 | |
efeaed4d | 790 | dqm_unlock(dqm); |
64c7f8cf | 791 | |
32cce8bc FK |
792 | /* Outside the DQM lock because under the DQM lock we can't do |
793 | * reclaim or take other locks that others hold while reclaiming. | |
794 | */ | |
795 | kfd_inc_compute_active(dqm->dev); | |
796 | ||
a22fc854 | 797 | return retval; |
64c7f8cf BG |
798 | } |
799 | ||
58dcd5bf | 800 | static int unregister_process(struct device_queue_manager *dqm, |
64c7f8cf BG |
801 | struct qcm_process_device *qpd) |
802 | { | |
803 | int retval; | |
804 | struct device_process_node *cur, *next; | |
805 | ||
1e5ec956 OG |
806 | pr_debug("qpd->queues_list is %s\n", |
807 | list_empty(&qpd->queues_list) ? "empty" : "not empty"); | |
64c7f8cf BG |
808 | |
809 | retval = 0; | |
efeaed4d | 810 | dqm_lock(dqm); |
64c7f8cf BG |
811 | |
812 | list_for_each_entry_safe(cur, next, &dqm->queues, list) { | |
813 | if (qpd == cur->qpd) { | |
814 | list_del(&cur->list); | |
f5d896bb | 815 | kfree(cur); |
f756e631 | 816 | dqm->processes_count--; |
64c7f8cf BG |
817 | goto out; |
818 | } | |
819 | } | |
820 | /* qpd not found in dqm list */ | |
821 | retval = 1; | |
822 | out: | |
efeaed4d | 823 | dqm_unlock(dqm); |
32cce8bc FK |
824 | |
825 | /* Outside the DQM lock because under the DQM lock we can't do | |
826 | * reclaim or take other locks that others hold while reclaiming. | |
827 | */ | |
828 | if (!retval) | |
829 | kfd_dec_compute_active(dqm->dev); | |
830 | ||
64c7f8cf BG |
831 | return retval; |
832 | } | |
833 | ||
834 | static int | |
835 | set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid, | |
836 | unsigned int vmid) | |
837 | { | |
cea405b1 | 838 | return dqm->dev->kfd2kgd->set_pasid_vmid_mapping( |
deb99d7c | 839 | dqm->dev->kgd, pasid, vmid); |
64c7f8cf BG |
840 | } |
841 | ||
2249d558 AL |
842 | static void init_interrupts(struct device_queue_manager *dqm) |
843 | { | |
844 | unsigned int i; | |
845 | ||
d0b63bb3 AR |
846 | for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++) |
847 | if (is_pipe_enabled(dqm, 0, i)) | |
848 | dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd, i); | |
2249d558 AL |
849 | } |
850 | ||
64c7f8cf BG |
851 | static int initialize_nocpsch(struct device_queue_manager *dqm) |
852 | { | |
86194cf8 | 853 | int pipe, queue; |
64c7f8cf | 854 | |
79775b62 | 855 | pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm)); |
64c7f8cf | 856 | |
ab7c1648 KR |
857 | dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm), |
858 | sizeof(unsigned int), GFP_KERNEL); | |
859 | if (!dqm->allocated_queues) | |
860 | return -ENOMEM; | |
861 | ||
efeaed4d | 862 | mutex_init(&dqm->lock_hidden); |
64c7f8cf BG |
863 | INIT_LIST_HEAD(&dqm->queues); |
864 | dqm->queue_count = dqm->next_pipe_to_allocate = 0; | |
bcea3081 | 865 | dqm->sdma_queue_count = 0; |
1b4670f6 | 866 | dqm->xgmi_sdma_queue_count = 0; |
64c7f8cf | 867 | |
86194cf8 FK |
868 | for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) { |
869 | int pipe_offset = pipe * get_queues_per_pipe(dqm); | |
870 | ||
871 | for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) | |
872 | if (test_bit(pipe_offset + queue, | |
873 | dqm->dev->shared_resources.queue_bitmap)) | |
874 | dqm->allocated_queues[pipe] |= 1 << queue; | |
875 | } | |
64c7f8cf | 876 | |
44008d7a | 877 | dqm->vmid_bitmap = (1 << dqm->dev->vm_info.vmid_num_kfd) - 1; |
cb77ee7c | 878 | dqm->sdma_bitmap = (1ULL << get_num_sdma_queues(dqm)) - 1; |
1b4670f6 | 879 | dqm->xgmi_sdma_bitmap = (1ULL << get_num_xgmi_sdma_queues(dqm)) - 1; |
64c7f8cf | 880 | |
64c7f8cf BG |
881 | return 0; |
882 | } | |
883 | ||
58dcd5bf | 884 | static void uninitialize(struct device_queue_manager *dqm) |
64c7f8cf | 885 | { |
6f9d54fd OG |
886 | int i; |
887 | ||
32fa8219 | 888 | WARN_ON(dqm->queue_count > 0 || dqm->processes_count > 0); |
64c7f8cf BG |
889 | |
890 | kfree(dqm->allocated_queues); | |
6f9d54fd | 891 | for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++) |
8d5f3552 | 892 | kfree(dqm->mqd_mgrs[i]); |
efeaed4d | 893 | mutex_destroy(&dqm->lock_hidden); |
a86aa3ca | 894 | kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem); |
64c7f8cf BG |
895 | } |
896 | ||
897 | static int start_nocpsch(struct device_queue_manager *dqm) | |
898 | { | |
2249d558 | 899 | init_interrupts(dqm); |
552764b6 | 900 | return pm_init(&dqm->packets, dqm); |
64c7f8cf BG |
901 | } |
902 | ||
903 | static int stop_nocpsch(struct device_queue_manager *dqm) | |
904 | { | |
552764b6 | 905 | pm_uninit(&dqm->packets); |
64c7f8cf BG |
906 | return 0; |
907 | } | |
908 | ||
bcea3081 | 909 | static int allocate_sdma_queue(struct device_queue_manager *dqm, |
e78579aa | 910 | struct queue *q) |
bcea3081 BG |
911 | { |
912 | int bit; | |
913 | ||
1b4670f6 OZ |
914 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { |
915 | if (dqm->sdma_bitmap == 0) | |
916 | return -ENOMEM; | |
917 | bit = __ffs64(dqm->sdma_bitmap); | |
918 | dqm->sdma_bitmap &= ~(1ULL << bit); | |
919 | q->sdma_id = bit; | |
920 | q->properties.sdma_engine_id = q->sdma_id % | |
921 | get_num_sdma_engines(dqm); | |
922 | q->properties.sdma_queue_id = q->sdma_id / | |
923 | get_num_sdma_engines(dqm); | |
924 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
925 | if (dqm->xgmi_sdma_bitmap == 0) | |
926 | return -ENOMEM; | |
927 | bit = __ffs64(dqm->xgmi_sdma_bitmap); | |
928 | dqm->xgmi_sdma_bitmap &= ~(1ULL << bit); | |
929 | q->sdma_id = bit; | |
930 | /* sdma_engine_id is sdma id including | |
931 | * both PCIe-optimized SDMAs and XGMI- | |
932 | * optimized SDMAs. The calculation below | |
933 | * assumes the first N engines are always | |
934 | * PCIe-optimized ones | |
935 | */ | |
936 | q->properties.sdma_engine_id = get_num_sdma_engines(dqm) + | |
937 | q->sdma_id % get_num_xgmi_sdma_engines(dqm); | |
938 | q->properties.sdma_queue_id = q->sdma_id / | |
939 | get_num_xgmi_sdma_engines(dqm); | |
940 | } | |
e78579aa | 941 | |
e78579aa YZ |
942 | pr_debug("SDMA engine id: %d\n", q->properties.sdma_engine_id); |
943 | pr_debug("SDMA queue id: %d\n", q->properties.sdma_queue_id); | |
bcea3081 BG |
944 | |
945 | return 0; | |
946 | } | |
947 | ||
948 | static void deallocate_sdma_queue(struct device_queue_manager *dqm, | |
1b4670f6 | 949 | struct queue *q) |
bcea3081 | 950 | { |
1b4670f6 OZ |
951 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { |
952 | if (q->sdma_id >= get_num_sdma_queues(dqm)) | |
953 | return; | |
954 | dqm->sdma_bitmap |= (1ULL << q->sdma_id); | |
955 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
956 | if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm)) | |
957 | return; | |
958 | dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id); | |
959 | } | |
bcea3081 BG |
960 | } |
961 | ||
64c7f8cf BG |
962 | /* |
963 | * Device Queue Manager implementation for cp scheduler | |
964 | */ | |
965 | ||
966 | static int set_sched_resources(struct device_queue_manager *dqm) | |
967 | { | |
d0b63bb3 | 968 | int i, mec; |
64c7f8cf | 969 | struct scheduling_resources res; |
64c7f8cf | 970 | |
44008d7a | 971 | res.vmid_mask = dqm->dev->shared_resources.compute_vmid_bitmap; |
d0b63bb3 AR |
972 | |
973 | res.queue_mask = 0; | |
974 | for (i = 0; i < KGD_MAX_QUEUES; ++i) { | |
975 | mec = (i / dqm->dev->shared_resources.num_queue_per_pipe) | |
976 | / dqm->dev->shared_resources.num_pipe_per_mec; | |
977 | ||
978 | if (!test_bit(i, dqm->dev->shared_resources.queue_bitmap)) | |
979 | continue; | |
980 | ||
981 | /* only acquire queues from the first MEC */ | |
982 | if (mec > 0) | |
983 | continue; | |
984 | ||
985 | /* This situation may be hit in the future if a new HW | |
986 | * generation exposes more than 64 queues. If so, the | |
8eabaf54 KR |
987 | * definition of res.queue_mask needs updating |
988 | */ | |
1d11ee89 | 989 | if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) { |
d0b63bb3 AR |
990 | pr_err("Invalid queue enabled by amdgpu: %d\n", i); |
991 | break; | |
992 | } | |
993 | ||
994 | res.queue_mask |= (1ull << i); | |
995 | } | |
64c7f8cf BG |
996 | res.gws_mask = res.oac_mask = res.gds_heap_base = |
997 | res.gds_heap_size = 0; | |
998 | ||
79775b62 KR |
999 | pr_debug("Scheduling resources:\n" |
1000 | "vmid mask: 0x%8X\n" | |
1001 | "queue mask: 0x%8llX\n", | |
64c7f8cf BG |
1002 | res.vmid_mask, res.queue_mask); |
1003 | ||
1004 | return pm_send_set_resources(&dqm->packets, &res); | |
1005 | } | |
1006 | ||
1007 | static int initialize_cpsch(struct device_queue_manager *dqm) | |
1008 | { | |
79775b62 | 1009 | pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm)); |
64c7f8cf | 1010 | |
efeaed4d | 1011 | mutex_init(&dqm->lock_hidden); |
64c7f8cf BG |
1012 | INIT_LIST_HEAD(&dqm->queues); |
1013 | dqm->queue_count = dqm->processes_count = 0; | |
bcea3081 | 1014 | dqm->sdma_queue_count = 0; |
1b4670f6 | 1015 | dqm->xgmi_sdma_queue_count = 0; |
64c7f8cf | 1016 | dqm->active_runlist = false; |
cb77ee7c | 1017 | dqm->sdma_bitmap = (1ULL << get_num_sdma_queues(dqm)) - 1; |
1b4670f6 | 1018 | dqm->xgmi_sdma_bitmap = (1ULL << get_num_xgmi_sdma_queues(dqm)) - 1; |
64c7f8cf | 1019 | |
73ea648d SL |
1020 | INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception); |
1021 | ||
bfd5e378 | 1022 | return 0; |
64c7f8cf BG |
1023 | } |
1024 | ||
1025 | static int start_cpsch(struct device_queue_manager *dqm) | |
1026 | { | |
64c7f8cf BG |
1027 | int retval; |
1028 | ||
64c7f8cf BG |
1029 | retval = 0; |
1030 | ||
1031 | retval = pm_init(&dqm->packets, dqm); | |
4eacc26b | 1032 | if (retval) |
64c7f8cf BG |
1033 | goto fail_packet_manager_init; |
1034 | ||
1035 | retval = set_sched_resources(dqm); | |
4eacc26b | 1036 | if (retval) |
64c7f8cf BG |
1037 | goto fail_set_sched_resources; |
1038 | ||
79775b62 | 1039 | pr_debug("Allocating fence memory\n"); |
64c7f8cf BG |
1040 | |
1041 | /* allocate fence memory on the gart */ | |
a86aa3ca OG |
1042 | retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr), |
1043 | &dqm->fence_mem); | |
64c7f8cf | 1044 | |
4eacc26b | 1045 | if (retval) |
64c7f8cf BG |
1046 | goto fail_allocate_vidmem; |
1047 | ||
1048 | dqm->fence_addr = dqm->fence_mem->cpu_ptr; | |
1049 | dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr; | |
2249d558 AL |
1050 | |
1051 | init_interrupts(dqm); | |
1052 | ||
efeaed4d | 1053 | dqm_lock(dqm); |
73ea648d SL |
1054 | /* clear hang status when driver try to start the hw scheduler */ |
1055 | dqm->is_hws_hang = false; | |
c4744e24 | 1056 | execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); |
efeaed4d | 1057 | dqm_unlock(dqm); |
64c7f8cf BG |
1058 | |
1059 | return 0; | |
1060 | fail_allocate_vidmem: | |
1061 | fail_set_sched_resources: | |
1062 | pm_uninit(&dqm->packets); | |
1063 | fail_packet_manager_init: | |
1064 | return retval; | |
1065 | } | |
1066 | ||
1067 | static int stop_cpsch(struct device_queue_manager *dqm) | |
1068 | { | |
efeaed4d | 1069 | dqm_lock(dqm); |
4465f466 | 1070 | unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); |
efeaed4d | 1071 | dqm_unlock(dqm); |
64c7f8cf | 1072 | |
a86aa3ca | 1073 | kfd_gtt_sa_free(dqm->dev, dqm->fence_mem); |
64c7f8cf BG |
1074 | pm_uninit(&dqm->packets); |
1075 | ||
1076 | return 0; | |
1077 | } | |
1078 | ||
1079 | static int create_kernel_queue_cpsch(struct device_queue_manager *dqm, | |
1080 | struct kernel_queue *kq, | |
1081 | struct qcm_process_device *qpd) | |
1082 | { | |
efeaed4d | 1083 | dqm_lock(dqm); |
b8cbab04 | 1084 | if (dqm->total_queue_count >= max_num_of_queues_per_device) { |
79775b62 | 1085 | pr_warn("Can't create new kernel queue because %d queues were already created\n", |
b8cbab04 | 1086 | dqm->total_queue_count); |
efeaed4d | 1087 | dqm_unlock(dqm); |
b8cbab04 OG |
1088 | return -EPERM; |
1089 | } | |
1090 | ||
1091 | /* | |
1092 | * Unconditionally increment this counter, regardless of the queue's | |
1093 | * type or whether the queue is active. | |
1094 | */ | |
1095 | dqm->total_queue_count++; | |
1096 | pr_debug("Total of %d queues are accountable so far\n", | |
1097 | dqm->total_queue_count); | |
1098 | ||
64c7f8cf BG |
1099 | list_add(&kq->list, &qpd->priv_queue_list); |
1100 | dqm->queue_count++; | |
1101 | qpd->is_debug = true; | |
c4744e24 | 1102 | execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); |
efeaed4d | 1103 | dqm_unlock(dqm); |
64c7f8cf BG |
1104 | |
1105 | return 0; | |
1106 | } | |
1107 | ||
1108 | static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm, | |
1109 | struct kernel_queue *kq, | |
1110 | struct qcm_process_device *qpd) | |
1111 | { | |
efeaed4d | 1112 | dqm_lock(dqm); |
64c7f8cf BG |
1113 | list_del(&kq->list); |
1114 | dqm->queue_count--; | |
1115 | qpd->is_debug = false; | |
c4744e24 | 1116 | execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); |
b8cbab04 OG |
1117 | /* |
1118 | * Unconditionally decrement this counter, regardless of the queue's | |
1119 | * type. | |
1120 | */ | |
8b58f261 | 1121 | dqm->total_queue_count--; |
b8cbab04 OG |
1122 | pr_debug("Total of %d queues are accountable so far\n", |
1123 | dqm->total_queue_count); | |
efeaed4d | 1124 | dqm_unlock(dqm); |
64c7f8cf BG |
1125 | } |
1126 | ||
1127 | static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, | |
b46cb7d7 | 1128 | struct qcm_process_device *qpd) |
64c7f8cf BG |
1129 | { |
1130 | int retval; | |
8d5f3552 | 1131 | struct mqd_manager *mqd_mgr; |
64c7f8cf | 1132 | |
b8cbab04 | 1133 | if (dqm->total_queue_count >= max_num_of_queues_per_device) { |
79775b62 | 1134 | pr_warn("Can't create new usermode queue because %d queues were already created\n", |
b8cbab04 | 1135 | dqm->total_queue_count); |
70d488fb OZ |
1136 | retval = -EPERM; |
1137 | goto out; | |
b8cbab04 OG |
1138 | } |
1139 | ||
1b4670f6 OZ |
1140 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
1141 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
e78579aa | 1142 | retval = allocate_sdma_queue(dqm, q); |
894a8293 | 1143 | if (retval) |
70d488fb | 1144 | goto out; |
e139cd2a | 1145 | } |
ef568db7 FK |
1146 | |
1147 | retval = allocate_doorbell(qpd, q); | |
1148 | if (retval) | |
1149 | goto out_deallocate_sdma_queue; | |
1150 | ||
70d488fb OZ |
1151 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
1152 | q->properties.type)]; | |
26103436 | 1153 | /* |
bb2d2128 FK |
1154 | * Eviction state logic: mark all queues as evicted, even ones |
1155 | * not currently active. Restoring inactive queues later only | |
1156 | * updates the is_evicted flag but is a no-op otherwise. | |
26103436 | 1157 | */ |
bb2d2128 | 1158 | q->properties.is_evicted = !!qpd->evicted; |
eec0b4cf OZ |
1159 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
1160 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) | |
1161 | dqm->asic_ops.init_sdma_vm(dqm, q, qpd); | |
373d7080 FK |
1162 | q->properties.tba_addr = qpd->tba_addr; |
1163 | q->properties.tma_addr = qpd->tma_addr; | |
70d488fb OZ |
1164 | q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties); |
1165 | if (!q->mqd_mem_obj) { | |
1166 | retval = -ENOMEM; | |
1167 | goto out_deallocate_doorbell; | |
1168 | } | |
8636e53c OZ |
1169 | mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, |
1170 | &q->gart_mqd_addr, &q->properties); | |
70d488fb | 1171 | dqm_lock(dqm); |
89cd9d23 | 1172 | |
64c7f8cf | 1173 | list_add(&q->list, &qpd->queues_list); |
bc920fd4 | 1174 | qpd->queue_count++; |
64c7f8cf BG |
1175 | if (q->properties.is_active) { |
1176 | dqm->queue_count++; | |
c4744e24 YZ |
1177 | retval = execute_queues_cpsch(dqm, |
1178 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); | |
64c7f8cf BG |
1179 | } |
1180 | ||
bcea3081 | 1181 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) |
8eabaf54 | 1182 | dqm->sdma_queue_count++; |
1b4670f6 OZ |
1183 | else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) |
1184 | dqm->xgmi_sdma_queue_count++; | |
b8cbab04 OG |
1185 | /* |
1186 | * Unconditionally increment this counter, regardless of the queue's | |
1187 | * type or whether the queue is active. | |
1188 | */ | |
1189 | dqm->total_queue_count++; | |
1190 | ||
1191 | pr_debug("Total of %d queues are accountable so far\n", | |
1192 | dqm->total_queue_count); | |
1193 | ||
efeaed4d | 1194 | dqm_unlock(dqm); |
72a01d23 FK |
1195 | return retval; |
1196 | ||
70d488fb OZ |
1197 | out_deallocate_doorbell: |
1198 | deallocate_doorbell(qpd, q); | |
72a01d23 | 1199 | out_deallocate_sdma_queue: |
1b4670f6 OZ |
1200 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA || |
1201 | q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) | |
1202 | deallocate_sdma_queue(dqm, q); | |
70d488fb | 1203 | out: |
64c7f8cf BG |
1204 | return retval; |
1205 | } | |
1206 | ||
788bf83d | 1207 | int amdkfd_fence_wait_timeout(unsigned int *fence_addr, |
d80d19bd | 1208 | unsigned int fence_value, |
8c72c3d7 | 1209 | unsigned int timeout_ms) |
64c7f8cf | 1210 | { |
8c72c3d7 | 1211 | unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies; |
64c7f8cf BG |
1212 | |
1213 | while (*fence_addr != fence_value) { | |
8c72c3d7 | 1214 | if (time_after(jiffies, end_jiffies)) { |
79775b62 | 1215 | pr_err("qcm fence wait loop timeout expired\n"); |
0e9a860c YZ |
1216 | /* In HWS case, this is used to halt the driver thread |
1217 | * in order not to mess up CP states before doing | |
1218 | * scandumps for FW debugging. | |
1219 | */ | |
1220 | while (halt_if_hws_hang) | |
1221 | schedule(); | |
1222 | ||
64c7f8cf BG |
1223 | return -ETIME; |
1224 | } | |
99331a51 | 1225 | schedule(); |
64c7f8cf BG |
1226 | } |
1227 | ||
1228 | return 0; | |
1229 | } | |
1230 | ||
065e4bdf | 1231 | static int unmap_sdma_queues(struct device_queue_manager *dqm) |
bcea3081 | 1232 | { |
065e4bdf OZ |
1233 | int i, retval = 0; |
1234 | ||
1b4670f6 OZ |
1235 | for (i = 0; i < dqm->dev->device_info->num_sdma_engines + |
1236 | dqm->dev->device_info->num_xgmi_sdma_engines; i++) { | |
065e4bdf OZ |
1237 | retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_SDMA, |
1238 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, false, i); | |
1239 | if (retval) | |
1240 | return retval; | |
1241 | } | |
1242 | return retval; | |
bcea3081 BG |
1243 | } |
1244 | ||
60a00956 FK |
1245 | /* dqm->lock mutex has to be locked before calling this function */ |
1246 | static int map_queues_cpsch(struct device_queue_manager *dqm) | |
1247 | { | |
1248 | int retval; | |
1249 | ||
1250 | if (dqm->queue_count <= 0 || dqm->processes_count <= 0) | |
1251 | return 0; | |
1252 | ||
1253 | if (dqm->active_runlist) | |
1254 | return 0; | |
1255 | ||
1256 | retval = pm_send_runlist(&dqm->packets, &dqm->queues); | |
1257 | if (retval) { | |
1258 | pr_err("failed to execute runlist\n"); | |
1259 | return retval; | |
1260 | } | |
1261 | dqm->active_runlist = true; | |
1262 | ||
1263 | return retval; | |
1264 | } | |
1265 | ||
ac30c783 | 1266 | /* dqm->lock mutex has to be locked before calling this function */ |
7da2bcf8 | 1267 | static int unmap_queues_cpsch(struct device_queue_manager *dqm, |
4465f466 YZ |
1268 | enum kfd_unmap_queues_filter filter, |
1269 | uint32_t filter_param) | |
64c7f8cf | 1270 | { |
9fd3f1bf | 1271 | int retval = 0; |
64c7f8cf | 1272 | |
73ea648d SL |
1273 | if (dqm->is_hws_hang) |
1274 | return -EIO; | |
991ca8ee | 1275 | if (!dqm->active_runlist) |
ac30c783 | 1276 | return retval; |
bcea3081 | 1277 | |
1b4670f6 OZ |
1278 | pr_debug("Before destroying queues, sdma queue count is : %u, xgmi sdma queue count is : %u\n", |
1279 | dqm->sdma_queue_count, dqm->xgmi_sdma_queue_count); | |
bcea3081 | 1280 | |
1b4670f6 | 1281 | if (dqm->sdma_queue_count > 0 || dqm->xgmi_sdma_queue_count) |
065e4bdf | 1282 | unmap_sdma_queues(dqm); |
bcea3081 | 1283 | |
64c7f8cf | 1284 | retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE, |
4465f466 | 1285 | filter, filter_param, false, 0); |
4eacc26b | 1286 | if (retval) |
ac30c783 | 1287 | return retval; |
64c7f8cf BG |
1288 | |
1289 | *dqm->fence_addr = KFD_FENCE_INIT; | |
1290 | pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr, | |
1291 | KFD_FENCE_COMPLETED); | |
1292 | /* should be timed out */ | |
c3447e81 | 1293 | retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED, |
64c7f8cf | 1294 | QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS); |
9fd3f1bf | 1295 | if (retval) |
ac30c783 | 1296 | return retval; |
9fd3f1bf | 1297 | |
64c7f8cf BG |
1298 | pm_release_ib(&dqm->packets); |
1299 | dqm->active_runlist = false; | |
1300 | ||
64c7f8cf BG |
1301 | return retval; |
1302 | } | |
1303 | ||
ac30c783 | 1304 | /* dqm->lock mutex has to be locked before calling this function */ |
c4744e24 YZ |
1305 | static int execute_queues_cpsch(struct device_queue_manager *dqm, |
1306 | enum kfd_unmap_queues_filter filter, | |
1307 | uint32_t filter_param) | |
64c7f8cf BG |
1308 | { |
1309 | int retval; | |
1310 | ||
73ea648d SL |
1311 | if (dqm->is_hws_hang) |
1312 | return -EIO; | |
c4744e24 | 1313 | retval = unmap_queues_cpsch(dqm, filter, filter_param); |
4eacc26b | 1314 | if (retval) { |
c4744e24 | 1315 | pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n"); |
73ea648d SL |
1316 | dqm->is_hws_hang = true; |
1317 | schedule_work(&dqm->hw_exception_work); | |
ac30c783 | 1318 | return retval; |
64c7f8cf BG |
1319 | } |
1320 | ||
60a00956 | 1321 | return map_queues_cpsch(dqm); |
64c7f8cf BG |
1322 | } |
1323 | ||
1324 | static int destroy_queue_cpsch(struct device_queue_manager *dqm, | |
1325 | struct qcm_process_device *qpd, | |
1326 | struct queue *q) | |
1327 | { | |
1328 | int retval; | |
8d5f3552 | 1329 | struct mqd_manager *mqd_mgr; |
992839ad | 1330 | |
64c7f8cf BG |
1331 | retval = 0; |
1332 | ||
1333 | /* remove queue from list to prevent rescheduling after preemption */ | |
efeaed4d | 1334 | dqm_lock(dqm); |
992839ad YS |
1335 | |
1336 | if (qpd->is_debug) { | |
1337 | /* | |
1338 | * error, currently we do not allow to destroy a queue | |
1339 | * of a currently debugged process | |
1340 | */ | |
1341 | retval = -EBUSY; | |
1342 | goto failed_try_destroy_debugged_queue; | |
1343 | ||
1344 | } | |
1345 | ||
fdfa090b OZ |
1346 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
1347 | q->properties.type)]; | |
64c7f8cf | 1348 | |
ef568db7 FK |
1349 | deallocate_doorbell(qpd, q); |
1350 | ||
e139cd2a | 1351 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { |
bcea3081 | 1352 | dqm->sdma_queue_count--; |
1b4670f6 OZ |
1353 | deallocate_sdma_queue(dqm, q); |
1354 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
1355 | dqm->xgmi_sdma_queue_count--; | |
1356 | deallocate_sdma_queue(dqm, q); | |
e139cd2a | 1357 | } |
bcea3081 | 1358 | |
64c7f8cf | 1359 | list_del(&q->list); |
bc920fd4 | 1360 | qpd->queue_count--; |
40a526dc | 1361 | if (q->properties.is_active) { |
b6819cec | 1362 | dqm->queue_count--; |
40a526dc | 1363 | retval = execute_queues_cpsch(dqm, |
9fd3f1bf | 1364 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); |
40a526dc YZ |
1365 | if (retval == -ETIME) |
1366 | qpd->reset_wavefronts = true; | |
1367 | } | |
64c7f8cf | 1368 | |
b8cbab04 OG |
1369 | /* |
1370 | * Unconditionally decrement this counter, regardless of the queue's | |
1371 | * type | |
1372 | */ | |
1373 | dqm->total_queue_count--; | |
1374 | pr_debug("Total of %d queues are accountable so far\n", | |
1375 | dqm->total_queue_count); | |
64c7f8cf | 1376 | |
efeaed4d | 1377 | dqm_unlock(dqm); |
64c7f8cf | 1378 | |
8636e53c OZ |
1379 | /* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */ |
1380 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); | |
89cd9d23 | 1381 | |
9e827224 | 1382 | return retval; |
64c7f8cf | 1383 | |
992839ad YS |
1384 | failed_try_destroy_debugged_queue: |
1385 | ||
efeaed4d | 1386 | dqm_unlock(dqm); |
64c7f8cf BG |
1387 | return retval; |
1388 | } | |
1389 | ||
1390 | /* | |
1391 | * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to | |
1392 | * stay in user mode. | |
1393 | */ | |
1394 | #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL | |
1395 | /* APE1 limit is inclusive and 64K aligned. */ | |
1396 | #define APE1_LIMIT_ALIGNMENT 0xFFFF | |
1397 | ||
1398 | static bool set_cache_memory_policy(struct device_queue_manager *dqm, | |
1399 | struct qcm_process_device *qpd, | |
1400 | enum cache_policy default_policy, | |
1401 | enum cache_policy alternate_policy, | |
1402 | void __user *alternate_aperture_base, | |
1403 | uint64_t alternate_aperture_size) | |
1404 | { | |
bed4f110 FK |
1405 | bool retval = true; |
1406 | ||
1407 | if (!dqm->asic_ops.set_cache_memory_policy) | |
1408 | return retval; | |
64c7f8cf | 1409 | |
efeaed4d | 1410 | dqm_lock(dqm); |
64c7f8cf BG |
1411 | |
1412 | if (alternate_aperture_size == 0) { | |
1413 | /* base > limit disables APE1 */ | |
1414 | qpd->sh_mem_ape1_base = 1; | |
1415 | qpd->sh_mem_ape1_limit = 0; | |
1416 | } else { | |
1417 | /* | |
1418 | * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]}, | |
1419 | * SH_MEM_APE1_BASE[31:0], 0x0000 } | |
1420 | * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]}, | |
1421 | * SH_MEM_APE1_LIMIT[31:0], 0xFFFF } | |
1422 | * Verify that the base and size parameters can be | |
1423 | * represented in this format and convert them. | |
1424 | * Additionally restrict APE1 to user-mode addresses. | |
1425 | */ | |
1426 | ||
1427 | uint64_t base = (uintptr_t)alternate_aperture_base; | |
1428 | uint64_t limit = base + alternate_aperture_size - 1; | |
1429 | ||
ab7c1648 KR |
1430 | if (limit <= base || (base & APE1_FIXED_BITS_MASK) != 0 || |
1431 | (limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT) { | |
1432 | retval = false; | |
64c7f8cf | 1433 | goto out; |
ab7c1648 | 1434 | } |
64c7f8cf BG |
1435 | |
1436 | qpd->sh_mem_ape1_base = base >> 16; | |
1437 | qpd->sh_mem_ape1_limit = limit >> 16; | |
1438 | } | |
1439 | ||
bfd5e378 | 1440 | retval = dqm->asic_ops.set_cache_memory_policy( |
a22fc854 BG |
1441 | dqm, |
1442 | qpd, | |
1443 | default_policy, | |
1444 | alternate_policy, | |
1445 | alternate_aperture_base, | |
1446 | alternate_aperture_size); | |
64c7f8cf | 1447 | |
d146c5a7 | 1448 | if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0)) |
64c7f8cf BG |
1449 | program_sh_mem_settings(dqm, qpd); |
1450 | ||
79775b62 | 1451 | pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n", |
64c7f8cf BG |
1452 | qpd->sh_mem_config, qpd->sh_mem_ape1_base, |
1453 | qpd->sh_mem_ape1_limit); | |
1454 | ||
64c7f8cf | 1455 | out: |
efeaed4d | 1456 | dqm_unlock(dqm); |
ab7c1648 | 1457 | return retval; |
64c7f8cf BG |
1458 | } |
1459 | ||
d7b9bd22 FK |
1460 | static int set_trap_handler(struct device_queue_manager *dqm, |
1461 | struct qcm_process_device *qpd, | |
1462 | uint64_t tba_addr, | |
1463 | uint64_t tma_addr) | |
1464 | { | |
1465 | uint64_t *tma; | |
1466 | ||
1467 | if (dqm->dev->cwsr_enabled) { | |
1468 | /* Jump from CWSR trap handler to user trap */ | |
1469 | tma = (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET); | |
1470 | tma[0] = tba_addr; | |
1471 | tma[1] = tma_addr; | |
1472 | } else { | |
1473 | qpd->tba_addr = tba_addr; | |
1474 | qpd->tma_addr = tma_addr; | |
1475 | } | |
1476 | ||
1477 | return 0; | |
1478 | } | |
1479 | ||
9fd3f1bf FK |
1480 | static int process_termination_nocpsch(struct device_queue_manager *dqm, |
1481 | struct qcm_process_device *qpd) | |
1482 | { | |
1483 | struct queue *q, *next; | |
1484 | struct device_process_node *cur, *next_dpn; | |
1485 | int retval = 0; | |
32cce8bc | 1486 | bool found = false; |
9fd3f1bf | 1487 | |
efeaed4d | 1488 | dqm_lock(dqm); |
9fd3f1bf FK |
1489 | |
1490 | /* Clear all user mode queues */ | |
1491 | list_for_each_entry_safe(q, next, &qpd->queues_list, list) { | |
1492 | int ret; | |
1493 | ||
1494 | ret = destroy_queue_nocpsch_locked(dqm, qpd, q); | |
1495 | if (ret) | |
1496 | retval = ret; | |
1497 | } | |
1498 | ||
1499 | /* Unregister process */ | |
1500 | list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) { | |
1501 | if (qpd == cur->qpd) { | |
1502 | list_del(&cur->list); | |
1503 | kfree(cur); | |
1504 | dqm->processes_count--; | |
32cce8bc | 1505 | found = true; |
9fd3f1bf FK |
1506 | break; |
1507 | } | |
1508 | } | |
1509 | ||
efeaed4d | 1510 | dqm_unlock(dqm); |
32cce8bc FK |
1511 | |
1512 | /* Outside the DQM lock because under the DQM lock we can't do | |
1513 | * reclaim or take other locks that others hold while reclaiming. | |
1514 | */ | |
1515 | if (found) | |
1516 | kfd_dec_compute_active(dqm->dev); | |
1517 | ||
9fd3f1bf FK |
1518 | return retval; |
1519 | } | |
1520 | ||
5df099e8 JC |
1521 | static int get_wave_state(struct device_queue_manager *dqm, |
1522 | struct queue *q, | |
1523 | void __user *ctl_stack, | |
1524 | u32 *ctl_stack_used_size, | |
1525 | u32 *save_area_used_size) | |
1526 | { | |
4e6c6fc1 | 1527 | struct mqd_manager *mqd_mgr; |
5df099e8 JC |
1528 | int r; |
1529 | ||
1530 | dqm_lock(dqm); | |
1531 | ||
1532 | if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE || | |
1533 | q->properties.is_active || !q->device->cwsr_enabled) { | |
1534 | r = -EINVAL; | |
1535 | goto dqm_unlock; | |
1536 | } | |
1537 | ||
fdfa090b | 1538 | mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_COMPUTE]; |
5df099e8 | 1539 | |
4e6c6fc1 | 1540 | if (!mqd_mgr->get_wave_state) { |
5df099e8 JC |
1541 | r = -EINVAL; |
1542 | goto dqm_unlock; | |
1543 | } | |
1544 | ||
4e6c6fc1 YZ |
1545 | r = mqd_mgr->get_wave_state(mqd_mgr, q->mqd, ctl_stack, |
1546 | ctl_stack_used_size, save_area_used_size); | |
5df099e8 JC |
1547 | |
1548 | dqm_unlock: | |
1549 | dqm_unlock(dqm); | |
1550 | return r; | |
1551 | } | |
9fd3f1bf FK |
1552 | |
1553 | static int process_termination_cpsch(struct device_queue_manager *dqm, | |
1554 | struct qcm_process_device *qpd) | |
1555 | { | |
1556 | int retval; | |
1557 | struct queue *q, *next; | |
1558 | struct kernel_queue *kq, *kq_next; | |
8d5f3552 | 1559 | struct mqd_manager *mqd_mgr; |
9fd3f1bf FK |
1560 | struct device_process_node *cur, *next_dpn; |
1561 | enum kfd_unmap_queues_filter filter = | |
1562 | KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES; | |
32cce8bc | 1563 | bool found = false; |
9fd3f1bf FK |
1564 | |
1565 | retval = 0; | |
1566 | ||
efeaed4d | 1567 | dqm_lock(dqm); |
9fd3f1bf FK |
1568 | |
1569 | /* Clean all kernel queues */ | |
1570 | list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) { | |
1571 | list_del(&kq->list); | |
1572 | dqm->queue_count--; | |
1573 | qpd->is_debug = false; | |
1574 | dqm->total_queue_count--; | |
1575 | filter = KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES; | |
1576 | } | |
1577 | ||
1578 | /* Clear all user mode queues */ | |
1579 | list_for_each_entry(q, &qpd->queues_list, list) { | |
72a01d23 | 1580 | if (q->properties.type == KFD_QUEUE_TYPE_SDMA) { |
9fd3f1bf | 1581 | dqm->sdma_queue_count--; |
1b4670f6 OZ |
1582 | deallocate_sdma_queue(dqm, q); |
1583 | } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) { | |
1584 | dqm->xgmi_sdma_queue_count--; | |
1585 | deallocate_sdma_queue(dqm, q); | |
72a01d23 | 1586 | } |
9fd3f1bf FK |
1587 | |
1588 | if (q->properties.is_active) | |
1589 | dqm->queue_count--; | |
1590 | ||
1591 | dqm->total_queue_count--; | |
1592 | } | |
1593 | ||
1594 | /* Unregister process */ | |
1595 | list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) { | |
1596 | if (qpd == cur->qpd) { | |
1597 | list_del(&cur->list); | |
1598 | kfree(cur); | |
1599 | dqm->processes_count--; | |
32cce8bc | 1600 | found = true; |
9fd3f1bf FK |
1601 | break; |
1602 | } | |
1603 | } | |
1604 | ||
1605 | retval = execute_queues_cpsch(dqm, filter, 0); | |
73ea648d | 1606 | if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) { |
9fd3f1bf FK |
1607 | pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev); |
1608 | dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process); | |
1609 | qpd->reset_wavefronts = false; | |
1610 | } | |
1611 | ||
89cd9d23 PY |
1612 | dqm_unlock(dqm); |
1613 | ||
32cce8bc FK |
1614 | /* Outside the DQM lock because under the DQM lock we can't do |
1615 | * reclaim or take other locks that others hold while reclaiming. | |
1616 | */ | |
1617 | if (found) | |
1618 | kfd_dec_compute_active(dqm->dev); | |
1619 | ||
89cd9d23 | 1620 | /* Lastly, free mqd resources. |
8636e53c | 1621 | * Do free_mqd() after dqm_unlock to avoid circular locking. |
89cd9d23 | 1622 | */ |
9fd3f1bf | 1623 | list_for_each_entry_safe(q, next, &qpd->queues_list, list) { |
fdfa090b OZ |
1624 | mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type( |
1625 | q->properties.type)]; | |
9fd3f1bf | 1626 | list_del(&q->list); |
bc920fd4 | 1627 | qpd->queue_count--; |
8636e53c | 1628 | mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj); |
9fd3f1bf FK |
1629 | } |
1630 | ||
9fd3f1bf FK |
1631 | return retval; |
1632 | } | |
1633 | ||
fdfa090b OZ |
1634 | static int init_mqd_managers(struct device_queue_manager *dqm) |
1635 | { | |
1636 | int i, j; | |
1637 | struct mqd_manager *mqd_mgr; | |
1638 | ||
1639 | for (i = 0; i < KFD_MQD_TYPE_MAX; i++) { | |
1640 | mqd_mgr = dqm->asic_ops.mqd_manager_init(i, dqm->dev); | |
1641 | if (!mqd_mgr) { | |
1642 | pr_err("mqd manager [%d] initialization failed\n", i); | |
1643 | goto out_free; | |
1644 | } | |
1645 | dqm->mqd_mgrs[i] = mqd_mgr; | |
1646 | } | |
1647 | ||
1648 | return 0; | |
1649 | ||
1650 | out_free: | |
1651 | for (j = 0; j < i; j++) { | |
1652 | kfree(dqm->mqd_mgrs[j]); | |
1653 | dqm->mqd_mgrs[j] = NULL; | |
1654 | } | |
1655 | ||
1656 | return -ENOMEM; | |
1657 | } | |
11614c36 OZ |
1658 | |
1659 | /* Allocate one hiq mqd (HWS) and all SDMA mqd in a continuous trunk*/ | |
1660 | static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm) | |
1661 | { | |
1662 | int retval; | |
1663 | struct kfd_dev *dev = dqm->dev; | |
1664 | struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd; | |
1665 | uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size * | |
1666 | dev->device_info->num_sdma_engines * | |
1667 | dev->device_info->num_sdma_queues_per_engine + | |
1668 | dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size; | |
1669 | ||
1670 | retval = amdgpu_amdkfd_alloc_gtt_mem(dev->kgd, size, | |
1671 | &(mem_obj->gtt_mem), &(mem_obj->gpu_addr), | |
1672 | (void *)&(mem_obj->cpu_ptr), true); | |
1673 | ||
1674 | return retval; | |
1675 | } | |
1676 | ||
64c7f8cf BG |
1677 | struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev) |
1678 | { | |
1679 | struct device_queue_manager *dqm; | |
1680 | ||
79775b62 | 1681 | pr_debug("Loading device queue manager\n"); |
a22fc854 | 1682 | |
dbf56ab1 | 1683 | dqm = kzalloc(sizeof(*dqm), GFP_KERNEL); |
64c7f8cf BG |
1684 | if (!dqm) |
1685 | return NULL; | |
1686 | ||
d146c5a7 FK |
1687 | switch (dev->device_info->asic_family) { |
1688 | /* HWS is not available on Hawaii. */ | |
1689 | case CHIP_HAWAII: | |
1690 | /* HWS depends on CWSR for timely dequeue. CWSR is not | |
1691 | * available on Tonga. | |
1692 | * | |
1693 | * FIXME: This argument also applies to Kaveri. | |
1694 | */ | |
1695 | case CHIP_TONGA: | |
1696 | dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS; | |
1697 | break; | |
1698 | default: | |
1699 | dqm->sched_policy = sched_policy; | |
1700 | break; | |
1701 | } | |
1702 | ||
64c7f8cf | 1703 | dqm->dev = dev; |
d146c5a7 | 1704 | switch (dqm->sched_policy) { |
64c7f8cf BG |
1705 | case KFD_SCHED_POLICY_HWS: |
1706 | case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: | |
1707 | /* initialize dqm for cp scheduling */ | |
45c9a5e4 OG |
1708 | dqm->ops.create_queue = create_queue_cpsch; |
1709 | dqm->ops.initialize = initialize_cpsch; | |
1710 | dqm->ops.start = start_cpsch; | |
1711 | dqm->ops.stop = stop_cpsch; | |
1712 | dqm->ops.destroy_queue = destroy_queue_cpsch; | |
1713 | dqm->ops.update_queue = update_queue; | |
58dcd5bf YZ |
1714 | dqm->ops.register_process = register_process; |
1715 | dqm->ops.unregister_process = unregister_process; | |
1716 | dqm->ops.uninitialize = uninitialize; | |
45c9a5e4 OG |
1717 | dqm->ops.create_kernel_queue = create_kernel_queue_cpsch; |
1718 | dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch; | |
1719 | dqm->ops.set_cache_memory_policy = set_cache_memory_policy; | |
d7b9bd22 | 1720 | dqm->ops.set_trap_handler = set_trap_handler; |
9fd3f1bf | 1721 | dqm->ops.process_termination = process_termination_cpsch; |
26103436 FK |
1722 | dqm->ops.evict_process_queues = evict_process_queues_cpsch; |
1723 | dqm->ops.restore_process_queues = restore_process_queues_cpsch; | |
5df099e8 | 1724 | dqm->ops.get_wave_state = get_wave_state; |
64c7f8cf BG |
1725 | break; |
1726 | case KFD_SCHED_POLICY_NO_HWS: | |
1727 | /* initialize dqm for no cp scheduling */ | |
45c9a5e4 OG |
1728 | dqm->ops.start = start_nocpsch; |
1729 | dqm->ops.stop = stop_nocpsch; | |
1730 | dqm->ops.create_queue = create_queue_nocpsch; | |
1731 | dqm->ops.destroy_queue = destroy_queue_nocpsch; | |
1732 | dqm->ops.update_queue = update_queue; | |
58dcd5bf YZ |
1733 | dqm->ops.register_process = register_process; |
1734 | dqm->ops.unregister_process = unregister_process; | |
45c9a5e4 | 1735 | dqm->ops.initialize = initialize_nocpsch; |
58dcd5bf | 1736 | dqm->ops.uninitialize = uninitialize; |
45c9a5e4 | 1737 | dqm->ops.set_cache_memory_policy = set_cache_memory_policy; |
d7b9bd22 | 1738 | dqm->ops.set_trap_handler = set_trap_handler; |
9fd3f1bf | 1739 | dqm->ops.process_termination = process_termination_nocpsch; |
26103436 FK |
1740 | dqm->ops.evict_process_queues = evict_process_queues_nocpsch; |
1741 | dqm->ops.restore_process_queues = | |
1742 | restore_process_queues_nocpsch; | |
5df099e8 | 1743 | dqm->ops.get_wave_state = get_wave_state; |
64c7f8cf BG |
1744 | break; |
1745 | default: | |
d146c5a7 | 1746 | pr_err("Invalid scheduling policy %d\n", dqm->sched_policy); |
32fa8219 | 1747 | goto out_free; |
64c7f8cf BG |
1748 | } |
1749 | ||
a22fc854 BG |
1750 | switch (dev->device_info->asic_family) { |
1751 | case CHIP_CARRIZO: | |
bfd5e378 | 1752 | device_queue_manager_init_vi(&dqm->asic_ops); |
300dec95 OG |
1753 | break; |
1754 | ||
a22fc854 | 1755 | case CHIP_KAVERI: |
bfd5e378 | 1756 | device_queue_manager_init_cik(&dqm->asic_ops); |
300dec95 | 1757 | break; |
97672cbe FK |
1758 | |
1759 | case CHIP_HAWAII: | |
1760 | device_queue_manager_init_cik_hawaii(&dqm->asic_ops); | |
1761 | break; | |
1762 | ||
1763 | case CHIP_TONGA: | |
1764 | case CHIP_FIJI: | |
1765 | case CHIP_POLARIS10: | |
1766 | case CHIP_POLARIS11: | |
846a44d7 | 1767 | case CHIP_POLARIS12: |
ed81cd6e | 1768 | case CHIP_VEGAM: |
97672cbe FK |
1769 | device_queue_manager_init_vi_tonga(&dqm->asic_ops); |
1770 | break; | |
bed4f110 FK |
1771 | |
1772 | case CHIP_VEGA10: | |
846a44d7 | 1773 | case CHIP_VEGA12: |
22a3a294 | 1774 | case CHIP_VEGA20: |
bed4f110 FK |
1775 | case CHIP_RAVEN: |
1776 | device_queue_manager_init_v9(&dqm->asic_ops); | |
1777 | break; | |
e596b903 YZ |
1778 | default: |
1779 | WARN(1, "Unexpected ASIC family %u", | |
1780 | dev->device_info->asic_family); | |
1781 | goto out_free; | |
a22fc854 BG |
1782 | } |
1783 | ||
fdfa090b OZ |
1784 | if (init_mqd_managers(dqm)) |
1785 | goto out_free; | |
1786 | ||
11614c36 OZ |
1787 | if (allocate_hiq_sdma_mqd(dqm)) { |
1788 | pr_err("Failed to allocate hiq sdma mqd trunk buffer\n"); | |
1789 | goto out_free; | |
1790 | } | |
1791 | ||
32fa8219 FK |
1792 | if (!dqm->ops.initialize(dqm)) |
1793 | return dqm; | |
64c7f8cf | 1794 | |
32fa8219 FK |
1795 | out_free: |
1796 | kfree(dqm); | |
1797 | return NULL; | |
64c7f8cf BG |
1798 | } |
1799 | ||
11614c36 OZ |
1800 | void deallocate_hiq_sdma_mqd(struct kfd_dev *dev, struct kfd_mem_obj *mqd) |
1801 | { | |
1802 | WARN(!mqd, "No hiq sdma mqd trunk to free"); | |
1803 | ||
1804 | amdgpu_amdkfd_free_gtt_mem(dev->kgd, mqd->gtt_mem); | |
1805 | } | |
1806 | ||
64c7f8cf BG |
1807 | void device_queue_manager_uninit(struct device_queue_manager *dqm) |
1808 | { | |
45c9a5e4 | 1809 | dqm->ops.uninitialize(dqm); |
11614c36 | 1810 | deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd); |
64c7f8cf BG |
1811 | kfree(dqm); |
1812 | } | |
851a645e | 1813 | |
2640c3fa | 1814 | int kfd_process_vm_fault(struct device_queue_manager *dqm, |
1815 | unsigned int pasid) | |
1816 | { | |
1817 | struct kfd_process_device *pdd; | |
1818 | struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); | |
1819 | int ret = 0; | |
1820 | ||
1821 | if (!p) | |
1822 | return -EINVAL; | |
1823 | pdd = kfd_get_process_device_data(dqm->dev, p); | |
1824 | if (pdd) | |
1825 | ret = dqm->ops.evict_process_queues(dqm, &pdd->qpd); | |
1826 | kfd_unref_process(p); | |
1827 | ||
1828 | return ret; | |
1829 | } | |
1830 | ||
73ea648d SL |
1831 | static void kfd_process_hw_exception(struct work_struct *work) |
1832 | { | |
1833 | struct device_queue_manager *dqm = container_of(work, | |
1834 | struct device_queue_manager, hw_exception_work); | |
5b87245f | 1835 | amdgpu_amdkfd_gpu_reset(dqm->dev->kgd); |
73ea648d SL |
1836 | } |
1837 | ||
851a645e FK |
1838 | #if defined(CONFIG_DEBUG_FS) |
1839 | ||
1840 | static void seq_reg_dump(struct seq_file *m, | |
1841 | uint32_t (*dump)[2], uint32_t n_regs) | |
1842 | { | |
1843 | uint32_t i, count; | |
1844 | ||
1845 | for (i = 0, count = 0; i < n_regs; i++) { | |
1846 | if (count == 0 || | |
1847 | dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) { | |
1848 | seq_printf(m, "%s %08x: %08x", | |
1849 | i ? "\n" : "", | |
1850 | dump[i][0], dump[i][1]); | |
1851 | count = 7; | |
1852 | } else { | |
1853 | seq_printf(m, " %08x", dump[i][1]); | |
1854 | count--; | |
1855 | } | |
1856 | } | |
1857 | ||
1858 | seq_puts(m, "\n"); | |
1859 | } | |
1860 | ||
1861 | int dqm_debugfs_hqds(struct seq_file *m, void *data) | |
1862 | { | |
1863 | struct device_queue_manager *dqm = data; | |
1864 | uint32_t (*dump)[2], n_regs; | |
1865 | int pipe, queue; | |
1866 | int r = 0; | |
1867 | ||
24f48a42 OZ |
1868 | r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->kgd, |
1869 | KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE, &dump, &n_regs); | |
1870 | if (!r) { | |
1871 | seq_printf(m, " HIQ on MEC %d Pipe %d Queue %d\n", | |
1872 | KFD_CIK_HIQ_PIPE/get_pipes_per_mec(dqm)+1, | |
1873 | KFD_CIK_HIQ_PIPE%get_pipes_per_mec(dqm), | |
1874 | KFD_CIK_HIQ_QUEUE); | |
1875 | seq_reg_dump(m, dump, n_regs); | |
1876 | ||
1877 | kfree(dump); | |
1878 | } | |
1879 | ||
851a645e FK |
1880 | for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) { |
1881 | int pipe_offset = pipe * get_queues_per_pipe(dqm); | |
1882 | ||
1883 | for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) { | |
1884 | if (!test_bit(pipe_offset + queue, | |
1885 | dqm->dev->shared_resources.queue_bitmap)) | |
1886 | continue; | |
1887 | ||
1888 | r = dqm->dev->kfd2kgd->hqd_dump( | |
1889 | dqm->dev->kgd, pipe, queue, &dump, &n_regs); | |
1890 | if (r) | |
1891 | break; | |
1892 | ||
1893 | seq_printf(m, " CP Pipe %d, Queue %d\n", | |
1894 | pipe, queue); | |
1895 | seq_reg_dump(m, dump, n_regs); | |
1896 | ||
1897 | kfree(dump); | |
1898 | } | |
1899 | } | |
1900 | ||
98bb9222 | 1901 | for (pipe = 0; pipe < get_num_sdma_engines(dqm); pipe++) { |
d5094189 SL |
1902 | for (queue = 0; |
1903 | queue < dqm->dev->device_info->num_sdma_queues_per_engine; | |
1904 | queue++) { | |
851a645e FK |
1905 | r = dqm->dev->kfd2kgd->hqd_sdma_dump( |
1906 | dqm->dev->kgd, pipe, queue, &dump, &n_regs); | |
1907 | if (r) | |
1908 | break; | |
1909 | ||
1910 | seq_printf(m, " SDMA Engine %d, RLC %d\n", | |
1911 | pipe, queue); | |
1912 | seq_reg_dump(m, dump, n_regs); | |
1913 | ||
1914 | kfree(dump); | |
1915 | } | |
1916 | } | |
1917 | ||
1918 | return r; | |
1919 | } | |
1920 | ||
a29ec470 SL |
1921 | int dqm_debugfs_execute_queues(struct device_queue_manager *dqm) |
1922 | { | |
1923 | int r = 0; | |
1924 | ||
1925 | dqm_lock(dqm); | |
1926 | dqm->active_runlist = true; | |
1927 | r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); | |
1928 | dqm_unlock(dqm); | |
1929 | ||
1930 | return r; | |
1931 | } | |
1932 | ||
851a645e | 1933 | #endif |