1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  */
22 
23 #include <linux/bsearch.h>
24 #include <linux/pci.h>
25 #include <linux/slab.h>
26 #include "kfd_priv.h"
27 #include "kfd_device_queue_manager.h"
28 #include "kfd_pm4_headers_vi.h"
29 #include "cwsr_trap_handler.h"
30 #include "kfd_iommu.h"
31 #include "amdgpu_amdkfd.h"
32 
33 #define MQD_SIZE_ALIGNED 768
34 
35 /*
36  * kfd_locked is used to lock the kfd driver during suspend or reset
37  * once locked, kfd driver will stop any further GPU execution.
38  * create process (open) will return -EAGAIN.
39  */
40 static atomic_t kfd_locked = ATOMIC_INIT(0);
41 
42 #ifdef CONFIG_DRM_AMDGPU_CIK
43 extern const struct kfd2kgd_calls gfx_v7_kfd2kgd;
44 #endif
45 extern const struct kfd2kgd_calls gfx_v8_kfd2kgd;
46 extern const struct kfd2kgd_calls gfx_v9_kfd2kgd;
47 extern const struct kfd2kgd_calls arcturus_kfd2kgd;
48 extern const struct kfd2kgd_calls gfx_v10_kfd2kgd;
49 extern const struct kfd2kgd_calls gfx_v10_3_kfd2kgd;
50 
51 static const struct kfd2kgd_calls *kfd2kgd_funcs[] = {
52 #ifdef KFD_SUPPORT_IOMMU_V2
53 #ifdef CONFIG_DRM_AMDGPU_CIK
54 	[CHIP_KAVERI] = &gfx_v7_kfd2kgd,
55 #endif
56 	[CHIP_CARRIZO] = &gfx_v8_kfd2kgd,
57 	[CHIP_RAVEN] = &gfx_v9_kfd2kgd,
58 #endif
59 #ifdef CONFIG_DRM_AMDGPU_CIK
60 	[CHIP_HAWAII] = &gfx_v7_kfd2kgd,
61 #endif
62 	[CHIP_TONGA] = &gfx_v8_kfd2kgd,
63 	[CHIP_FIJI] = &gfx_v8_kfd2kgd,
64 	[CHIP_POLARIS10] = &gfx_v8_kfd2kgd,
65 	[CHIP_POLARIS11] = &gfx_v8_kfd2kgd,
66 	[CHIP_POLARIS12] = &gfx_v8_kfd2kgd,
67 	[CHIP_VEGAM] = &gfx_v8_kfd2kgd,
68 	[CHIP_VEGA10] = &gfx_v9_kfd2kgd,
69 	[CHIP_VEGA12] = &gfx_v9_kfd2kgd,
70 	[CHIP_VEGA20] = &gfx_v9_kfd2kgd,
71 	[CHIP_RENOIR] = &gfx_v9_kfd2kgd,
72 	[CHIP_ARCTURUS] = &arcturus_kfd2kgd,
73 	[CHIP_NAVI10] = &gfx_v10_kfd2kgd,
74 	[CHIP_NAVI12] = &gfx_v10_kfd2kgd,
75 	[CHIP_NAVI14] = &gfx_v10_kfd2kgd,
76 	[CHIP_SIENNA_CICHLID] = &gfx_v10_3_kfd2kgd,
77 	[CHIP_NAVY_FLOUNDER] = &gfx_v10_3_kfd2kgd,
78 };
79 
80 #ifdef KFD_SUPPORT_IOMMU_V2
81 static const struct kfd_device_info kaveri_device_info = {
82 	.asic_family = CHIP_KAVERI,
83 	.asic_name = "kaveri",
84 	.max_pasid_bits = 16,
85 	/* max num of queues for KV.TODO should be a dynamic value */
86 	.max_no_of_hqd	= 24,
87 	.doorbell_size  = 4,
88 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
89 	.event_interrupt_class = &event_interrupt_class_cik,
90 	.num_of_watch_points = 4,
91 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
92 	.supports_cwsr = false,
93 	.needs_iommu_device = true,
94 	.needs_pci_atomics = false,
95 	.num_sdma_engines = 2,
96 	.num_xgmi_sdma_engines = 0,
97 	.num_sdma_queues_per_engine = 2,
98 };
99 
100 static const struct kfd_device_info carrizo_device_info = {
101 	.asic_family = CHIP_CARRIZO,
102 	.asic_name = "carrizo",
103 	.max_pasid_bits = 16,
104 	/* max num of queues for CZ.TODO should be a dynamic value */
105 	.max_no_of_hqd	= 24,
106 	.doorbell_size  = 4,
107 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
108 	.event_interrupt_class = &event_interrupt_class_cik,
109 	.num_of_watch_points = 4,
110 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
111 	.supports_cwsr = true,
112 	.needs_iommu_device = true,
113 	.needs_pci_atomics = false,
114 	.num_sdma_engines = 2,
115 	.num_xgmi_sdma_engines = 0,
116 	.num_sdma_queues_per_engine = 2,
117 };
118 
119 static const struct kfd_device_info raven_device_info = {
120 	.asic_family = CHIP_RAVEN,
121 	.asic_name = "raven",
122 	.max_pasid_bits = 16,
123 	.max_no_of_hqd  = 24,
124 	.doorbell_size  = 8,
125 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
126 	.event_interrupt_class = &event_interrupt_class_v9,
127 	.num_of_watch_points = 4,
128 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
129 	.supports_cwsr = true,
130 	.needs_iommu_device = true,
131 	.needs_pci_atomics = true,
132 	.num_sdma_engines = 1,
133 	.num_xgmi_sdma_engines = 0,
134 	.num_sdma_queues_per_engine = 2,
135 };
136 #endif
137 
138 static const struct kfd_device_info hawaii_device_info = {
139 	.asic_family = CHIP_HAWAII,
140 	.asic_name = "hawaii",
141 	.max_pasid_bits = 16,
142 	/* max num of queues for KV.TODO should be a dynamic value */
143 	.max_no_of_hqd	= 24,
144 	.doorbell_size  = 4,
145 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
146 	.event_interrupt_class = &event_interrupt_class_cik,
147 	.num_of_watch_points = 4,
148 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
149 	.supports_cwsr = false,
150 	.needs_iommu_device = false,
151 	.needs_pci_atomics = false,
152 	.num_sdma_engines = 2,
153 	.num_xgmi_sdma_engines = 0,
154 	.num_sdma_queues_per_engine = 2,
155 };
156 
157 static const struct kfd_device_info tonga_device_info = {
158 	.asic_family = CHIP_TONGA,
159 	.asic_name = "tonga",
160 	.max_pasid_bits = 16,
161 	.max_no_of_hqd  = 24,
162 	.doorbell_size  = 4,
163 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
164 	.event_interrupt_class = &event_interrupt_class_cik,
165 	.num_of_watch_points = 4,
166 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
167 	.supports_cwsr = false,
168 	.needs_iommu_device = false,
169 	.needs_pci_atomics = true,
170 	.num_sdma_engines = 2,
171 	.num_xgmi_sdma_engines = 0,
172 	.num_sdma_queues_per_engine = 2,
173 };
174 
175 static const struct kfd_device_info fiji_device_info = {
176 	.asic_family = CHIP_FIJI,
177 	.asic_name = "fiji",
178 	.max_pasid_bits = 16,
179 	.max_no_of_hqd  = 24,
180 	.doorbell_size  = 4,
181 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
182 	.event_interrupt_class = &event_interrupt_class_cik,
183 	.num_of_watch_points = 4,
184 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
185 	.supports_cwsr = true,
186 	.needs_iommu_device = false,
187 	.needs_pci_atomics = true,
188 	.num_sdma_engines = 2,
189 	.num_xgmi_sdma_engines = 0,
190 	.num_sdma_queues_per_engine = 2,
191 };
192 
193 static const struct kfd_device_info fiji_vf_device_info = {
194 	.asic_family = CHIP_FIJI,
195 	.asic_name = "fiji",
196 	.max_pasid_bits = 16,
197 	.max_no_of_hqd  = 24,
198 	.doorbell_size  = 4,
199 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
200 	.event_interrupt_class = &event_interrupt_class_cik,
201 	.num_of_watch_points = 4,
202 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
203 	.supports_cwsr = true,
204 	.needs_iommu_device = false,
205 	.needs_pci_atomics = false,
206 	.num_sdma_engines = 2,
207 	.num_xgmi_sdma_engines = 0,
208 	.num_sdma_queues_per_engine = 2,
209 };
210 
211 
212 static const struct kfd_device_info polaris10_device_info = {
213 	.asic_family = CHIP_POLARIS10,
214 	.asic_name = "polaris10",
215 	.max_pasid_bits = 16,
216 	.max_no_of_hqd  = 24,
217 	.doorbell_size  = 4,
218 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
219 	.event_interrupt_class = &event_interrupt_class_cik,
220 	.num_of_watch_points = 4,
221 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
222 	.supports_cwsr = true,
223 	.needs_iommu_device = false,
224 	.needs_pci_atomics = true,
225 	.num_sdma_engines = 2,
226 	.num_xgmi_sdma_engines = 0,
227 	.num_sdma_queues_per_engine = 2,
228 };
229 
230 static const struct kfd_device_info polaris10_vf_device_info = {
231 	.asic_family = CHIP_POLARIS10,
232 	.asic_name = "polaris10",
233 	.max_pasid_bits = 16,
234 	.max_no_of_hqd  = 24,
235 	.doorbell_size  = 4,
236 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
237 	.event_interrupt_class = &event_interrupt_class_cik,
238 	.num_of_watch_points = 4,
239 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
240 	.supports_cwsr = true,
241 	.needs_iommu_device = false,
242 	.needs_pci_atomics = false,
243 	.num_sdma_engines = 2,
244 	.num_xgmi_sdma_engines = 0,
245 	.num_sdma_queues_per_engine = 2,
246 };
247 
248 static const struct kfd_device_info polaris11_device_info = {
249 	.asic_family = CHIP_POLARIS11,
250 	.asic_name = "polaris11",
251 	.max_pasid_bits = 16,
252 	.max_no_of_hqd  = 24,
253 	.doorbell_size  = 4,
254 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
255 	.event_interrupt_class = &event_interrupt_class_cik,
256 	.num_of_watch_points = 4,
257 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
258 	.supports_cwsr = true,
259 	.needs_iommu_device = false,
260 	.needs_pci_atomics = true,
261 	.num_sdma_engines = 2,
262 	.num_xgmi_sdma_engines = 0,
263 	.num_sdma_queues_per_engine = 2,
264 };
265 
266 static const struct kfd_device_info polaris12_device_info = {
267 	.asic_family = CHIP_POLARIS12,
268 	.asic_name = "polaris12",
269 	.max_pasid_bits = 16,
270 	.max_no_of_hqd  = 24,
271 	.doorbell_size  = 4,
272 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
273 	.event_interrupt_class = &event_interrupt_class_cik,
274 	.num_of_watch_points = 4,
275 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
276 	.supports_cwsr = true,
277 	.needs_iommu_device = false,
278 	.needs_pci_atomics = true,
279 	.num_sdma_engines = 2,
280 	.num_xgmi_sdma_engines = 0,
281 	.num_sdma_queues_per_engine = 2,
282 };
283 
284 static const struct kfd_device_info vegam_device_info = {
285 	.asic_family = CHIP_VEGAM,
286 	.asic_name = "vegam",
287 	.max_pasid_bits = 16,
288 	.max_no_of_hqd  = 24,
289 	.doorbell_size  = 4,
290 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
291 	.event_interrupt_class = &event_interrupt_class_cik,
292 	.num_of_watch_points = 4,
293 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
294 	.supports_cwsr = true,
295 	.needs_iommu_device = false,
296 	.needs_pci_atomics = true,
297 	.num_sdma_engines = 2,
298 	.num_xgmi_sdma_engines = 0,
299 	.num_sdma_queues_per_engine = 2,
300 };
301 
302 static const struct kfd_device_info vega10_device_info = {
303 	.asic_family = CHIP_VEGA10,
304 	.asic_name = "vega10",
305 	.max_pasid_bits = 16,
306 	.max_no_of_hqd  = 24,
307 	.doorbell_size  = 8,
308 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
309 	.event_interrupt_class = &event_interrupt_class_v9,
310 	.num_of_watch_points = 4,
311 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
312 	.supports_cwsr = true,
313 	.needs_iommu_device = false,
314 	.needs_pci_atomics = false,
315 	.num_sdma_engines = 2,
316 	.num_xgmi_sdma_engines = 0,
317 	.num_sdma_queues_per_engine = 2,
318 };
319 
320 static const struct kfd_device_info vega10_vf_device_info = {
321 	.asic_family = CHIP_VEGA10,
322 	.asic_name = "vega10",
323 	.max_pasid_bits = 16,
324 	.max_no_of_hqd  = 24,
325 	.doorbell_size  = 8,
326 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
327 	.event_interrupt_class = &event_interrupt_class_v9,
328 	.num_of_watch_points = 4,
329 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
330 	.supports_cwsr = true,
331 	.needs_iommu_device = false,
332 	.needs_pci_atomics = false,
333 	.num_sdma_engines = 2,
334 	.num_xgmi_sdma_engines = 0,
335 	.num_sdma_queues_per_engine = 2,
336 };
337 
338 static const struct kfd_device_info vega12_device_info = {
339 	.asic_family = CHIP_VEGA12,
340 	.asic_name = "vega12",
341 	.max_pasid_bits = 16,
342 	.max_no_of_hqd  = 24,
343 	.doorbell_size  = 8,
344 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
345 	.event_interrupt_class = &event_interrupt_class_v9,
346 	.num_of_watch_points = 4,
347 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
348 	.supports_cwsr = true,
349 	.needs_iommu_device = false,
350 	.needs_pci_atomics = false,
351 	.num_sdma_engines = 2,
352 	.num_xgmi_sdma_engines = 0,
353 	.num_sdma_queues_per_engine = 2,
354 };
355 
356 static const struct kfd_device_info vega20_device_info = {
357 	.asic_family = CHIP_VEGA20,
358 	.asic_name = "vega20",
359 	.max_pasid_bits = 16,
360 	.max_no_of_hqd	= 24,
361 	.doorbell_size	= 8,
362 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
363 	.event_interrupt_class = &event_interrupt_class_v9,
364 	.num_of_watch_points = 4,
365 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
366 	.supports_cwsr = true,
367 	.needs_iommu_device = false,
368 	.needs_pci_atomics = false,
369 	.num_sdma_engines = 2,
370 	.num_xgmi_sdma_engines = 0,
371 	.num_sdma_queues_per_engine = 8,
372 };
373 
374 static const struct kfd_device_info arcturus_device_info = {
375 	.asic_family = CHIP_ARCTURUS,
376 	.asic_name = "arcturus",
377 	.max_pasid_bits = 16,
378 	.max_no_of_hqd	= 24,
379 	.doorbell_size	= 8,
380 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
381 	.event_interrupt_class = &event_interrupt_class_v9,
382 	.num_of_watch_points = 4,
383 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
384 	.supports_cwsr = true,
385 	.needs_iommu_device = false,
386 	.needs_pci_atomics = false,
387 	.num_sdma_engines = 2,
388 	.num_xgmi_sdma_engines = 6,
389 	.num_sdma_queues_per_engine = 8,
390 };
391 
392 static const struct kfd_device_info renoir_device_info = {
393 	.asic_family = CHIP_RENOIR,
394 	.asic_name = "renoir",
395 	.max_pasid_bits = 16,
396 	.max_no_of_hqd  = 24,
397 	.doorbell_size  = 8,
398 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
399 	.event_interrupt_class = &event_interrupt_class_v9,
400 	.num_of_watch_points = 4,
401 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
402 	.supports_cwsr = true,
403 	.needs_iommu_device = false,
404 	.needs_pci_atomics = false,
405 	.num_sdma_engines = 1,
406 	.num_xgmi_sdma_engines = 0,
407 	.num_sdma_queues_per_engine = 2,
408 };
409 
410 static const struct kfd_device_info navi10_device_info = {
411 	.asic_family = CHIP_NAVI10,
412 	.asic_name = "navi10",
413 	.max_pasid_bits = 16,
414 	.max_no_of_hqd  = 24,
415 	.doorbell_size  = 8,
416 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
417 	.event_interrupt_class = &event_interrupt_class_v9,
418 	.num_of_watch_points = 4,
419 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
420 	.needs_iommu_device = false,
421 	.supports_cwsr = true,
422 	.needs_pci_atomics = false,
423 	.num_sdma_engines = 2,
424 	.num_xgmi_sdma_engines = 0,
425 	.num_sdma_queues_per_engine = 8,
426 };
427 
428 static const struct kfd_device_info navi12_device_info = {
429 	.asic_family = CHIP_NAVI12,
430 	.asic_name = "navi12",
431 	.max_pasid_bits = 16,
432 	.max_no_of_hqd  = 24,
433 	.doorbell_size  = 8,
434 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
435 	.event_interrupt_class = &event_interrupt_class_v9,
436 	.num_of_watch_points = 4,
437 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
438 	.needs_iommu_device = false,
439 	.supports_cwsr = true,
440 	.needs_pci_atomics = false,
441 	.num_sdma_engines = 2,
442 	.num_xgmi_sdma_engines = 0,
443 	.num_sdma_queues_per_engine = 8,
444 };
445 
446 static const struct kfd_device_info navi14_device_info = {
447 	.asic_family = CHIP_NAVI14,
448 	.asic_name = "navi14",
449 	.max_pasid_bits = 16,
450 	.max_no_of_hqd  = 24,
451 	.doorbell_size  = 8,
452 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
453 	.event_interrupt_class = &event_interrupt_class_v9,
454 	.num_of_watch_points = 4,
455 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
456 	.needs_iommu_device = false,
457 	.supports_cwsr = true,
458 	.needs_pci_atomics = false,
459 	.num_sdma_engines = 2,
460 	.num_xgmi_sdma_engines = 0,
461 	.num_sdma_queues_per_engine = 8,
462 };
463 
464 static const struct kfd_device_info sienna_cichlid_device_info = {
465 	.asic_family = CHIP_SIENNA_CICHLID,
466 	.asic_name = "sienna_cichlid",
467 	.max_pasid_bits = 16,
468 	.max_no_of_hqd  = 24,
469 	.doorbell_size  = 8,
470 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
471 	.event_interrupt_class = &event_interrupt_class_v9,
472 	.num_of_watch_points = 4,
473 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
474 	.needs_iommu_device = false,
475 	.supports_cwsr = true,
476 	.needs_pci_atomics = false,
477 	.num_sdma_engines = 4,
478 	.num_xgmi_sdma_engines = 0,
479 	.num_sdma_queues_per_engine = 8,
480 };
481 
482 static const struct kfd_device_info navy_flounder_device_info = {
483 	.asic_family = CHIP_NAVY_FLOUNDER,
484 	.asic_name = "navy_flounder",
485 	.max_pasid_bits = 16,
486 	.max_no_of_hqd  = 24,
487 	.doorbell_size  = 8,
488 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
489 	.event_interrupt_class = &event_interrupt_class_v9,
490 	.num_of_watch_points = 4,
491 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
492 	.needs_iommu_device = false,
493 	.supports_cwsr = true,
494 	.needs_pci_atomics = false,
495 	.num_sdma_engines = 2,
496 	.num_xgmi_sdma_engines = 0,
497 	.num_sdma_queues_per_engine = 8,
498 };
499 
500 /* For each entry, [0] is regular and [1] is virtualisation device. */
501 static const struct kfd_device_info *kfd_supported_devices[][2] = {
502 #ifdef KFD_SUPPORT_IOMMU_V2
503 	[CHIP_KAVERI] = {&kaveri_device_info, NULL},
504 	[CHIP_CARRIZO] = {&carrizo_device_info, NULL},
505 	[CHIP_RAVEN] = {&raven_device_info, NULL},
506 #endif
507 	[CHIP_HAWAII] = {&hawaii_device_info, NULL},
508 	[CHIP_TONGA] = {&tonga_device_info, NULL},
509 	[CHIP_FIJI] = {&fiji_device_info, &fiji_vf_device_info},
510 	[CHIP_POLARIS10] = {&polaris10_device_info, &polaris10_vf_device_info},
511 	[CHIP_POLARIS11] = {&polaris11_device_info, NULL},
512 	[CHIP_POLARIS12] = {&polaris12_device_info, NULL},
513 	[CHIP_VEGAM] = {&vegam_device_info, NULL},
514 	[CHIP_VEGA10] = {&vega10_device_info, &vega10_vf_device_info},
515 	[CHIP_VEGA12] = {&vega12_device_info, NULL},
516 	[CHIP_VEGA20] = {&vega20_device_info, NULL},
517 	[CHIP_RENOIR] = {&renoir_device_info, NULL},
518 	[CHIP_ARCTURUS] = {&arcturus_device_info, &arcturus_device_info},
519 	[CHIP_NAVI10] = {&navi10_device_info, NULL},
520 	[CHIP_NAVI12] = {&navi12_device_info, &navi12_device_info},
521 	[CHIP_NAVI14] = {&navi14_device_info, NULL},
522 	[CHIP_SIENNA_CICHLID] = {&sienna_cichlid_device_info, &sienna_cichlid_device_info},
523 	[CHIP_NAVY_FLOUNDER] = {&navy_flounder_device_info, &navy_flounder_device_info},
524 };
525 
526 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
527 				unsigned int chunk_size);
528 static void kfd_gtt_sa_fini(struct kfd_dev *kfd);
529 
530 static int kfd_resume(struct kfd_dev *kfd);
531 
532 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
533 	struct pci_dev *pdev, unsigned int asic_type, bool vf)
534 {
535 	struct kfd_dev *kfd;
536 	const struct kfd_device_info *device_info;
537 	const struct kfd2kgd_calls *f2g;
538 
539 	if (asic_type >= sizeof(kfd_supported_devices) / (sizeof(void *) * 2)
540 		|| asic_type >= sizeof(kfd2kgd_funcs) / sizeof(void *)) {
541 		dev_err(kfd_device, "asic_type %d out of range\n", asic_type);
542 		return NULL; /* asic_type out of range */
543 	}
544 
545 	device_info = kfd_supported_devices[asic_type][vf];
546 	f2g = kfd2kgd_funcs[asic_type];
547 
548 	if (!device_info || !f2g) {
549 		dev_err(kfd_device, "%s %s not supported in kfd\n",
550 			amdgpu_asic_name[asic_type], vf ? "VF" : "");
551 		return NULL;
552 	}
553 
554 	kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
555 	if (!kfd)
556 		return NULL;
557 
558 	/* Allow BIF to recode atomics to PCIe 3.0 AtomicOps.
559 	 * 32 and 64-bit requests are possible and must be
560 	 * supported.
561 	 */
562 	kfd->pci_atomic_requested = amdgpu_amdkfd_have_atomics_support(kgd);
563 	if (device_info->needs_pci_atomics &&
564 	    !kfd->pci_atomic_requested) {
565 		dev_info(kfd_device,
566 			 "skipped device %x:%x, PCI rejects atomics\n",
567 			 pdev->vendor, pdev->device);
568 		kfree(kfd);
569 		return NULL;
570 	}
571 
572 	kfd->kgd = kgd;
573 	kfd->device_info = device_info;
574 	kfd->pdev = pdev;
575 	kfd->init_complete = false;
576 	kfd->kfd2kgd = f2g;
577 	atomic_set(&kfd->compute_profile, 0);
578 
579 	mutex_init(&kfd->doorbell_mutex);
580 	memset(&kfd->doorbell_available_index, 0,
581 		sizeof(kfd->doorbell_available_index));
582 
583 	atomic_set(&kfd->sram_ecc_flag, 0);
584 
585 	return kfd;
586 }
587 
588 static void kfd_cwsr_init(struct kfd_dev *kfd)
589 {
590 	if (cwsr_enable && kfd->device_info->supports_cwsr) {
591 		if (kfd->device_info->asic_family < CHIP_VEGA10) {
592 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx8_hex) > PAGE_SIZE);
593 			kfd->cwsr_isa = cwsr_trap_gfx8_hex;
594 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx8_hex);
595 		} else if (kfd->device_info->asic_family == CHIP_ARCTURUS) {
596 			BUILD_BUG_ON(sizeof(cwsr_trap_arcturus_hex) > PAGE_SIZE);
597 			kfd->cwsr_isa = cwsr_trap_arcturus_hex;
598 			kfd->cwsr_isa_size = sizeof(cwsr_trap_arcturus_hex);
599 		} else if (kfd->device_info->asic_family < CHIP_NAVI10) {
600 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_hex) > PAGE_SIZE);
601 			kfd->cwsr_isa = cwsr_trap_gfx9_hex;
602 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_hex);
603 		} else if (kfd->device_info->asic_family < CHIP_SIENNA_CICHLID) {
604 			BUILD_BUG_ON(sizeof(cwsr_trap_nv1x_hex) > PAGE_SIZE);
605 			kfd->cwsr_isa = cwsr_trap_nv1x_hex;
606 			kfd->cwsr_isa_size = sizeof(cwsr_trap_nv1x_hex);
607 		} else {
608 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx10_hex) > PAGE_SIZE);
609 			kfd->cwsr_isa = cwsr_trap_gfx10_hex;
610 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx10_hex);
611 		}
612 
613 		kfd->cwsr_enabled = true;
614 	}
615 }
616 
617 static int kfd_gws_init(struct kfd_dev *kfd)
618 {
619 	int ret = 0;
620 
621 	if (kfd->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS)
622 		return 0;
623 
624 	if (hws_gws_support
625 		|| (kfd->device_info->asic_family == CHIP_VEGA10
626 			&& kfd->mec2_fw_version >= 0x81b3)
627 		|| (kfd->device_info->asic_family >= CHIP_VEGA12
628 			&& kfd->device_info->asic_family <= CHIP_RAVEN
629 			&& kfd->mec2_fw_version >= 0x1b3)
630 		|| (kfd->device_info->asic_family == CHIP_ARCTURUS
631 			&& kfd->mec2_fw_version >= 0x30))
632 		ret = amdgpu_amdkfd_alloc_gws(kfd->kgd,
633 				amdgpu_amdkfd_get_num_gws(kfd->kgd), &kfd->gws);
634 
635 	return ret;
636 }
637 
638 static void kfd_smi_init(struct kfd_dev *dev) {
639 	INIT_LIST_HEAD(&dev->smi_clients);
640 	spin_lock_init(&dev->smi_lock);
641 }
642 
643 bool kgd2kfd_device_init(struct kfd_dev *kfd,
644 			 struct drm_device *ddev,
645 			 const struct kgd2kfd_shared_resources *gpu_resources)
646 {
647 	unsigned int size;
648 
649 	kfd->ddev = ddev;
650 	kfd->mec_fw_version = amdgpu_amdkfd_get_fw_version(kfd->kgd,
651 			KGD_ENGINE_MEC1);
652 	kfd->mec2_fw_version = amdgpu_amdkfd_get_fw_version(kfd->kgd,
653 			KGD_ENGINE_MEC2);
654 	kfd->sdma_fw_version = amdgpu_amdkfd_get_fw_version(kfd->kgd,
655 			KGD_ENGINE_SDMA1);
656 	kfd->shared_resources = *gpu_resources;
657 
658 	kfd->vm_info.first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1;
659 	kfd->vm_info.last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1;
660 	kfd->vm_info.vmid_num_kfd = kfd->vm_info.last_vmid_kfd
661 			- kfd->vm_info.first_vmid_kfd + 1;
662 
663 	/* Verify module parameters regarding mapped process number*/
664 	if ((hws_max_conc_proc < 0)
665 			|| (hws_max_conc_proc > kfd->vm_info.vmid_num_kfd)) {
666 		dev_err(kfd_device,
667 			"hws_max_conc_proc %d must be between 0 and %d, use %d instead\n",
668 			hws_max_conc_proc, kfd->vm_info.vmid_num_kfd,
669 			kfd->vm_info.vmid_num_kfd);
670 		kfd->max_proc_per_quantum = kfd->vm_info.vmid_num_kfd;
671 	} else
672 		kfd->max_proc_per_quantum = hws_max_conc_proc;
673 
674 	/* calculate max size of mqds needed for queues */
675 	size = max_num_of_queues_per_device *
676 			kfd->device_info->mqd_size_aligned;
677 
678 	/*
679 	 * calculate max size of runlist packet.
680 	 * There can be only 2 packets at once
681 	 */
682 	size += (KFD_MAX_NUM_OF_PROCESSES * sizeof(struct pm4_mes_map_process) +
683 		max_num_of_queues_per_device * sizeof(struct pm4_mes_map_queues)
684 		+ sizeof(struct pm4_mes_runlist)) * 2;
685 
686 	/* Add size of HIQ & DIQ */
687 	size += KFD_KERNEL_QUEUE_SIZE * 2;
688 
689 	/* add another 512KB for all other allocations on gart (HPD, fences) */
690 	size += 512 * 1024;
691 
692 	if (amdgpu_amdkfd_alloc_gtt_mem(
693 			kfd->kgd, size, &kfd->gtt_mem,
694 			&kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr,
695 			false)) {
696 		dev_err(kfd_device, "Could not allocate %d bytes\n", size);
697 		goto alloc_gtt_mem_failure;
698 	}
699 
700 	dev_info(kfd_device, "Allocated %d bytes on gart\n", size);
701 
702 	/* Initialize GTT sa with 512 byte chunk size */
703 	if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
704 		dev_err(kfd_device, "Error initializing gtt sub-allocator\n");
705 		goto kfd_gtt_sa_init_error;
706 	}
707 
708 	if (kfd_doorbell_init(kfd)) {
709 		dev_err(kfd_device,
710 			"Error initializing doorbell aperture\n");
711 		goto kfd_doorbell_error;
712 	}
713 
714 	if (kfd->kfd2kgd->get_hive_id)
715 		kfd->hive_id = kfd->kfd2kgd->get_hive_id(kfd->kgd);
716 
717 	if (kfd->kfd2kgd->get_unique_id)
718 		kfd->unique_id = kfd->kfd2kgd->get_unique_id(kfd->kgd);
719 
720 	if (kfd_interrupt_init(kfd)) {
721 		dev_err(kfd_device, "Error initializing interrupts\n");
722 		goto kfd_interrupt_error;
723 	}
724 
725 	kfd->dqm = device_queue_manager_init(kfd);
726 	if (!kfd->dqm) {
727 		dev_err(kfd_device, "Error initializing queue manager\n");
728 		goto device_queue_manager_error;
729 	}
730 
731 	/* If supported on this device, allocate global GWS that is shared
732 	 * by all KFD processes
733 	 */
734 	if (kfd_gws_init(kfd)) {
735 		dev_err(kfd_device, "Could not allocate %d gws\n",
736 			amdgpu_amdkfd_get_num_gws(kfd->kgd));
737 		goto gws_error;
738 	}
739 
740 	if (kfd_iommu_device_init(kfd)) {
741 		dev_err(kfd_device, "Error initializing iommuv2\n");
742 		goto device_iommu_error;
743 	}
744 
745 	kfd_cwsr_init(kfd);
746 
747 	if (kfd_resume(kfd))
748 		goto kfd_resume_error;
749 
750 	kfd->dbgmgr = NULL;
751 
752 	if (kfd_topology_add_device(kfd)) {
753 		dev_err(kfd_device, "Error adding device to topology\n");
754 		goto kfd_topology_add_device_error;
755 	}
756 
757 	kfd_smi_init(kfd);
758 
759 	kfd->init_complete = true;
760 	dev_info(kfd_device, "added device %x:%x\n", kfd->pdev->vendor,
761 		 kfd->pdev->device);
762 
763 	pr_debug("Starting kfd with the following scheduling policy %d\n",
764 		kfd->dqm->sched_policy);
765 
766 	goto out;
767 
768 kfd_topology_add_device_error:
769 kfd_resume_error:
770 device_iommu_error:
771 gws_error:
772 	device_queue_manager_uninit(kfd->dqm);
773 device_queue_manager_error:
774 	kfd_interrupt_exit(kfd);
775 kfd_interrupt_error:
776 	kfd_doorbell_fini(kfd);
777 kfd_doorbell_error:
778 	kfd_gtt_sa_fini(kfd);
779 kfd_gtt_sa_init_error:
780 	amdgpu_amdkfd_free_gtt_mem(kfd->kgd, kfd->gtt_mem);
781 alloc_gtt_mem_failure:
782 	if (kfd->gws)
783 		amdgpu_amdkfd_free_gws(kfd->kgd, kfd->gws);
784 	dev_err(kfd_device,
785 		"device %x:%x NOT added due to errors\n",
786 		kfd->pdev->vendor, kfd->pdev->device);
787 out:
788 	return kfd->init_complete;
789 }
790 
791 void kgd2kfd_device_exit(struct kfd_dev *kfd)
792 {
793 	if (kfd->init_complete) {
794 		kgd2kfd_suspend(kfd, false);
795 		device_queue_manager_uninit(kfd->dqm);
796 		kfd_interrupt_exit(kfd);
797 		kfd_topology_remove_device(kfd);
798 		kfd_doorbell_fini(kfd);
799 		kfd_gtt_sa_fini(kfd);
800 		amdgpu_amdkfd_free_gtt_mem(kfd->kgd, kfd->gtt_mem);
801 		if (kfd->gws)
802 			amdgpu_amdkfd_free_gws(kfd->kgd, kfd->gws);
803 	}
804 
805 	kfree(kfd);
806 }
807 
808 int kgd2kfd_pre_reset(struct kfd_dev *kfd)
809 {
810 	if (!kfd->init_complete)
811 		return 0;
812 
813 	kfd->dqm->ops.pre_reset(kfd->dqm);
814 
815 	kgd2kfd_suspend(kfd, false);
816 
817 	kfd_signal_reset_event(kfd);
818 	return 0;
819 }
820 
821 /*
822  * Fix me. KFD won't be able to resume existing process for now.
823  * We will keep all existing process in a evicted state and
824  * wait the process to be terminated.
825  */
826 
827 int kgd2kfd_post_reset(struct kfd_dev *kfd)
828 {
829 	int ret;
830 
831 	if (!kfd->init_complete)
832 		return 0;
833 
834 	ret = kfd_resume(kfd);
835 	if (ret)
836 		return ret;
837 	atomic_dec(&kfd_locked);
838 
839 	atomic_set(&kfd->sram_ecc_flag, 0);
840 
841 	return 0;
842 }
843 
844 bool kfd_is_locked(void)
845 {
846 	return  (atomic_read(&kfd_locked) > 0);
847 }
848 
849 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
850 {
851 	if (!kfd->init_complete)
852 		return;
853 
854 	/* for runtime suspend, skip locking kfd */
855 	if (!run_pm) {
856 		/* For first KFD device suspend all the KFD processes */
857 		if (atomic_inc_return(&kfd_locked) == 1)
858 			kfd_suspend_all_processes();
859 	}
860 
861 	kfd->dqm->ops.stop(kfd->dqm);
862 	kfd_iommu_suspend(kfd);
863 }
864 
865 int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
866 {
867 	int ret, count;
868 
869 	if (!kfd->init_complete)
870 		return 0;
871 
872 	ret = kfd_resume(kfd);
873 	if (ret)
874 		return ret;
875 
876 	/* for runtime resume, skip unlocking kfd */
877 	if (!run_pm) {
878 		count = atomic_dec_return(&kfd_locked);
879 		WARN_ONCE(count < 0, "KFD suspend / resume ref. error");
880 		if (count == 0)
881 			ret = kfd_resume_all_processes();
882 	}
883 
884 	return ret;
885 }
886 
887 static int kfd_resume(struct kfd_dev *kfd)
888 {
889 	int err = 0;
890 
891 	err = kfd_iommu_resume(kfd);
892 	if (err) {
893 		dev_err(kfd_device,
894 			"Failed to resume IOMMU for device %x:%x\n",
895 			kfd->pdev->vendor, kfd->pdev->device);
896 		return err;
897 	}
898 
899 	err = kfd->dqm->ops.start(kfd->dqm);
900 	if (err) {
901 		dev_err(kfd_device,
902 			"Error starting queue manager for device %x:%x\n",
903 			kfd->pdev->vendor, kfd->pdev->device);
904 		goto dqm_start_error;
905 	}
906 
907 	return err;
908 
909 dqm_start_error:
910 	kfd_iommu_suspend(kfd);
911 	return err;
912 }
913 
914 static inline void kfd_queue_work(struct workqueue_struct *wq,
915 				  struct work_struct *work)
916 {
917 	int cpu, new_cpu;
918 
919 	cpu = new_cpu = smp_processor_id();
920 	do {
921 		new_cpu = cpumask_next(new_cpu, cpu_online_mask) % nr_cpu_ids;
922 		if (cpu_to_node(new_cpu) == numa_node_id())
923 			break;
924 	} while (cpu != new_cpu);
925 
926 	queue_work_on(new_cpu, wq, work);
927 }
928 
929 /* This is called directly from KGD at ISR. */
930 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
931 {
932 	uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE];
933 	bool is_patched = false;
934 	unsigned long flags;
935 
936 	if (!kfd->init_complete)
937 		return;
938 
939 	if (kfd->device_info->ih_ring_entry_size > sizeof(patched_ihre)) {
940 		dev_err_once(kfd_device, "Ring entry too small\n");
941 		return;
942 	}
943 
944 	spin_lock_irqsave(&kfd->interrupt_lock, flags);
945 
946 	if (kfd->interrupts_active
947 	    && interrupt_is_wanted(kfd, ih_ring_entry,
948 				   patched_ihre, &is_patched)
949 	    && enqueue_ih_ring_entry(kfd,
950 				     is_patched ? patched_ihre : ih_ring_entry))
951 		kfd_queue_work(kfd->ih_wq, &kfd->interrupt_work);
952 
953 	spin_unlock_irqrestore(&kfd->interrupt_lock, flags);
954 }
955 
956 int kgd2kfd_quiesce_mm(struct mm_struct *mm)
957 {
958 	struct kfd_process *p;
959 	int r;
960 
961 	/* Because we are called from arbitrary context (workqueue) as opposed
962 	 * to process context, kfd_process could attempt to exit while we are
963 	 * running so the lookup function increments the process ref count.
964 	 */
965 	p = kfd_lookup_process_by_mm(mm);
966 	if (!p)
967 		return -ESRCH;
968 
969 	WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid);
970 	r = kfd_process_evict_queues(p);
971 
972 	kfd_unref_process(p);
973 	return r;
974 }
975 
976 int kgd2kfd_resume_mm(struct mm_struct *mm)
977 {
978 	struct kfd_process *p;
979 	int r;
980 
981 	/* Because we are called from arbitrary context (workqueue) as opposed
982 	 * to process context, kfd_process could attempt to exit while we are
983 	 * running so the lookup function increments the process ref count.
984 	 */
985 	p = kfd_lookup_process_by_mm(mm);
986 	if (!p)
987 		return -ESRCH;
988 
989 	r = kfd_process_restore_queues(p);
990 
991 	kfd_unref_process(p);
992 	return r;
993 }
994 
995 /** kgd2kfd_schedule_evict_and_restore_process - Schedules work queue that will
996  *   prepare for safe eviction of KFD BOs that belong to the specified
997  *   process.
998  *
999  * @mm: mm_struct that identifies the specified KFD process
1000  * @fence: eviction fence attached to KFD process BOs
1001  *
1002  */
1003 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
1004 					       struct dma_fence *fence)
1005 {
1006 	struct kfd_process *p;
1007 	unsigned long active_time;
1008 	unsigned long delay_jiffies = msecs_to_jiffies(PROCESS_ACTIVE_TIME_MS);
1009 
1010 	if (!fence)
1011 		return -EINVAL;
1012 
1013 	if (dma_fence_is_signaled(fence))
1014 		return 0;
1015 
1016 	p = kfd_lookup_process_by_mm(mm);
1017 	if (!p)
1018 		return -ENODEV;
1019 
1020 	if (fence->seqno == p->last_eviction_seqno)
1021 		goto out;
1022 
1023 	p->last_eviction_seqno = fence->seqno;
1024 
1025 	/* Avoid KFD process starvation. Wait for at least
1026 	 * PROCESS_ACTIVE_TIME_MS before evicting the process again
1027 	 */
1028 	active_time = get_jiffies_64() - p->last_restore_timestamp;
1029 	if (delay_jiffies > active_time)
1030 		delay_jiffies -= active_time;
1031 	else
1032 		delay_jiffies = 0;
1033 
1034 	/* During process initialization eviction_work.dwork is initialized
1035 	 * to kfd_evict_bo_worker
1036 	 */
1037 	WARN(debug_evictions, "Scheduling eviction of pid %d in %ld jiffies",
1038 	     p->lead_thread->pid, delay_jiffies);
1039 	schedule_delayed_work(&p->eviction_work, delay_jiffies);
1040 out:
1041 	kfd_unref_process(p);
1042 	return 0;
1043 }
1044 
1045 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
1046 				unsigned int chunk_size)
1047 {
1048 	unsigned int num_of_longs;
1049 
1050 	if (WARN_ON(buf_size < chunk_size))
1051 		return -EINVAL;
1052 	if (WARN_ON(buf_size == 0))
1053 		return -EINVAL;
1054 	if (WARN_ON(chunk_size == 0))
1055 		return -EINVAL;
1056 
1057 	kfd->gtt_sa_chunk_size = chunk_size;
1058 	kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
1059 
1060 	num_of_longs = (kfd->gtt_sa_num_of_chunks + BITS_PER_LONG - 1) /
1061 		BITS_PER_LONG;
1062 
1063 	kfd->gtt_sa_bitmap = kcalloc(num_of_longs, sizeof(long), GFP_KERNEL);
1064 
1065 	if (!kfd->gtt_sa_bitmap)
1066 		return -ENOMEM;
1067 
1068 	pr_debug("gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
1069 			kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);
1070 
1071 	mutex_init(&kfd->gtt_sa_lock);
1072 
1073 	return 0;
1074 
1075 }
1076 
1077 static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
1078 {
1079 	mutex_destroy(&kfd->gtt_sa_lock);
1080 	kfree(kfd->gtt_sa_bitmap);
1081 }
1082 
1083 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
1084 						unsigned int bit_num,
1085 						unsigned int chunk_size)
1086 {
1087 	return start_addr + bit_num * chunk_size;
1088 }
1089 
1090 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
1091 						unsigned int bit_num,
1092 						unsigned int chunk_size)
1093 {
1094 	return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
1095 }
1096 
1097 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
1098 			struct kfd_mem_obj **mem_obj)
1099 {
1100 	unsigned int found, start_search, cur_size;
1101 
1102 	if (size == 0)
1103 		return -EINVAL;
1104 
1105 	if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
1106 		return -ENOMEM;
1107 
1108 	*mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
1109 	if (!(*mem_obj))
1110 		return -ENOMEM;
1111 
1112 	pr_debug("Allocated mem_obj = %p for size = %d\n", *mem_obj, size);
1113 
1114 	start_search = 0;
1115 
1116 	mutex_lock(&kfd->gtt_sa_lock);
1117 
1118 kfd_gtt_restart_search:
1119 	/* Find the first chunk that is free */
1120 	found = find_next_zero_bit(kfd->gtt_sa_bitmap,
1121 					kfd->gtt_sa_num_of_chunks,
1122 					start_search);
1123 
1124 	pr_debug("Found = %d\n", found);
1125 
1126 	/* If there wasn't any free chunk, bail out */
1127 	if (found == kfd->gtt_sa_num_of_chunks)
1128 		goto kfd_gtt_no_free_chunk;
1129 
1130 	/* Update fields of mem_obj */
1131 	(*mem_obj)->range_start = found;
1132 	(*mem_obj)->range_end = found;
1133 	(*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
1134 					kfd->gtt_start_gpu_addr,
1135 					found,
1136 					kfd->gtt_sa_chunk_size);
1137 	(*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
1138 					kfd->gtt_start_cpu_ptr,
1139 					found,
1140 					kfd->gtt_sa_chunk_size);
1141 
1142 	pr_debug("gpu_addr = %p, cpu_addr = %p\n",
1143 			(uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
1144 
1145 	/* If we need only one chunk, mark it as allocated and get out */
1146 	if (size <= kfd->gtt_sa_chunk_size) {
1147 		pr_debug("Single bit\n");
1148 		set_bit(found, kfd->gtt_sa_bitmap);
1149 		goto kfd_gtt_out;
1150 	}
1151 
1152 	/* Otherwise, try to see if we have enough contiguous chunks */
1153 	cur_size = size - kfd->gtt_sa_chunk_size;
1154 	do {
1155 		(*mem_obj)->range_end =
1156 			find_next_zero_bit(kfd->gtt_sa_bitmap,
1157 					kfd->gtt_sa_num_of_chunks, ++found);
1158 		/*
1159 		 * If next free chunk is not contiguous than we need to
1160 		 * restart our search from the last free chunk we found (which
1161 		 * wasn't contiguous to the previous ones
1162 		 */
1163 		if ((*mem_obj)->range_end != found) {
1164 			start_search = found;
1165 			goto kfd_gtt_restart_search;
1166 		}
1167 
1168 		/*
1169 		 * If we reached end of buffer, bail out with error
1170 		 */
1171 		if (found == kfd->gtt_sa_num_of_chunks)
1172 			goto kfd_gtt_no_free_chunk;
1173 
1174 		/* Check if we don't need another chunk */
1175 		if (cur_size <= kfd->gtt_sa_chunk_size)
1176 			cur_size = 0;
1177 		else
1178 			cur_size -= kfd->gtt_sa_chunk_size;
1179 
1180 	} while (cur_size > 0);
1181 
1182 	pr_debug("range_start = %d, range_end = %d\n",
1183 		(*mem_obj)->range_start, (*mem_obj)->range_end);
1184 
1185 	/* Mark the chunks as allocated */
1186 	for (found = (*mem_obj)->range_start;
1187 		found <= (*mem_obj)->range_end;
1188 		found++)
1189 		set_bit(found, kfd->gtt_sa_bitmap);
1190 
1191 kfd_gtt_out:
1192 	mutex_unlock(&kfd->gtt_sa_lock);
1193 	return 0;
1194 
1195 kfd_gtt_no_free_chunk:
1196 	pr_debug("Allocation failed with mem_obj = %p\n", *mem_obj);
1197 	mutex_unlock(&kfd->gtt_sa_lock);
1198 	kfree(*mem_obj);
1199 	return -ENOMEM;
1200 }
1201 
1202 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj)
1203 {
1204 	unsigned int bit;
1205 
1206 	/* Act like kfree when trying to free a NULL object */
1207 	if (!mem_obj)
1208 		return 0;
1209 
1210 	pr_debug("Free mem_obj = %p, range_start = %d, range_end = %d\n",
1211 			mem_obj, mem_obj->range_start, mem_obj->range_end);
1212 
1213 	mutex_lock(&kfd->gtt_sa_lock);
1214 
1215 	/* Mark the chunks as free */
1216 	for (bit = mem_obj->range_start;
1217 		bit <= mem_obj->range_end;
1218 		bit++)
1219 		clear_bit(bit, kfd->gtt_sa_bitmap);
1220 
1221 	mutex_unlock(&kfd->gtt_sa_lock);
1222 
1223 	kfree(mem_obj);
1224 	return 0;
1225 }
1226 
1227 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd)
1228 {
1229 	if (kfd)
1230 		atomic_inc(&kfd->sram_ecc_flag);
1231 }
1232 
1233 void kfd_inc_compute_active(struct kfd_dev *kfd)
1234 {
1235 	if (atomic_inc_return(&kfd->compute_profile) == 1)
1236 		amdgpu_amdkfd_set_compute_idle(kfd->kgd, false);
1237 }
1238 
1239 void kfd_dec_compute_active(struct kfd_dev *kfd)
1240 {
1241 	int count = atomic_dec_return(&kfd->compute_profile);
1242 
1243 	if (count == 0)
1244 		amdgpu_amdkfd_set_compute_idle(kfd->kgd, true);
1245 	WARN_ONCE(count < 0, "Compute profile ref. count error");
1246 }
1247 
1248 #if defined(CONFIG_DEBUG_FS)
1249 
1250 /* This function will send a package to HIQ to hang the HWS
1251  * which will trigger a GPU reset and bring the HWS back to normal state
1252  */
1253 int kfd_debugfs_hang_hws(struct kfd_dev *dev)
1254 {
1255 	int r = 0;
1256 
1257 	if (dev->dqm->sched_policy != KFD_SCHED_POLICY_HWS) {
1258 		pr_err("HWS is not enabled");
1259 		return -EINVAL;
1260 	}
1261 
1262 	r = pm_debugfs_hang_hws(&dev->dqm->packets);
1263 	if (!r)
1264 		r = dqm_debugfs_execute_queues(dev->dqm);
1265 
1266 	return r;
1267 }
1268 
1269 #endif
1270