xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_device.c (revision dc3e0896003ee9b3bcc34c53965dc4bbc8671c44)
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 
32 #define MQD_SIZE_ALIGNED 768
33 
34 /*
35  * kfd_locked is used to lock the kfd driver during suspend or reset
36  * once locked, kfd driver will stop any further GPU execution.
37  * create process (open) will return -EAGAIN.
38  */
39 static atomic_t kfd_locked = ATOMIC_INIT(0);
40 
41 #ifdef KFD_SUPPORT_IOMMU_V2
42 static const struct kfd_device_info kaveri_device_info = {
43 	.asic_family = CHIP_KAVERI,
44 	.max_pasid_bits = 16,
45 	/* max num of queues for KV.TODO should be a dynamic value */
46 	.max_no_of_hqd	= 24,
47 	.doorbell_size  = 4,
48 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
49 	.event_interrupt_class = &event_interrupt_class_cik,
50 	.num_of_watch_points = 4,
51 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
52 	.supports_cwsr = false,
53 	.needs_iommu_device = true,
54 	.needs_pci_atomics = false,
55 	.num_sdma_engines = 2,
56 	.num_sdma_queues_per_engine = 2,
57 };
58 
59 static const struct kfd_device_info carrizo_device_info = {
60 	.asic_family = CHIP_CARRIZO,
61 	.max_pasid_bits = 16,
62 	/* max num of queues for CZ.TODO should be a dynamic value */
63 	.max_no_of_hqd	= 24,
64 	.doorbell_size  = 4,
65 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
66 	.event_interrupt_class = &event_interrupt_class_cik,
67 	.num_of_watch_points = 4,
68 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
69 	.supports_cwsr = true,
70 	.needs_iommu_device = true,
71 	.needs_pci_atomics = false,
72 	.num_sdma_engines = 2,
73 	.num_sdma_queues_per_engine = 2,
74 };
75 
76 static const struct kfd_device_info raven_device_info = {
77 	.asic_family = CHIP_RAVEN,
78 	.max_pasid_bits = 16,
79 	.max_no_of_hqd  = 24,
80 	.doorbell_size  = 8,
81 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
82 	.event_interrupt_class = &event_interrupt_class_v9,
83 	.num_of_watch_points = 4,
84 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
85 	.supports_cwsr = true,
86 	.needs_iommu_device = true,
87 	.needs_pci_atomics = true,
88 	.num_sdma_engines = 1,
89 	.num_sdma_queues_per_engine = 2,
90 };
91 #endif
92 
93 static const struct kfd_device_info hawaii_device_info = {
94 	.asic_family = CHIP_HAWAII,
95 	.max_pasid_bits = 16,
96 	/* max num of queues for KV.TODO should be a dynamic value */
97 	.max_no_of_hqd	= 24,
98 	.doorbell_size  = 4,
99 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
100 	.event_interrupt_class = &event_interrupt_class_cik,
101 	.num_of_watch_points = 4,
102 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
103 	.supports_cwsr = false,
104 	.needs_iommu_device = false,
105 	.needs_pci_atomics = false,
106 	.num_sdma_engines = 2,
107 	.num_sdma_queues_per_engine = 2,
108 };
109 
110 static const struct kfd_device_info tonga_device_info = {
111 	.asic_family = CHIP_TONGA,
112 	.max_pasid_bits = 16,
113 	.max_no_of_hqd  = 24,
114 	.doorbell_size  = 4,
115 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
116 	.event_interrupt_class = &event_interrupt_class_cik,
117 	.num_of_watch_points = 4,
118 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
119 	.supports_cwsr = false,
120 	.needs_iommu_device = false,
121 	.needs_pci_atomics = true,
122 	.num_sdma_engines = 2,
123 	.num_sdma_queues_per_engine = 2,
124 };
125 
126 static const struct kfd_device_info fiji_device_info = {
127 	.asic_family = CHIP_FIJI,
128 	.max_pasid_bits = 16,
129 	.max_no_of_hqd  = 24,
130 	.doorbell_size  = 4,
131 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
132 	.event_interrupt_class = &event_interrupt_class_cik,
133 	.num_of_watch_points = 4,
134 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
135 	.supports_cwsr = true,
136 	.needs_iommu_device = false,
137 	.needs_pci_atomics = true,
138 	.num_sdma_engines = 2,
139 	.num_sdma_queues_per_engine = 2,
140 };
141 
142 static const struct kfd_device_info fiji_vf_device_info = {
143 	.asic_family = CHIP_FIJI,
144 	.max_pasid_bits = 16,
145 	.max_no_of_hqd  = 24,
146 	.doorbell_size  = 4,
147 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
148 	.event_interrupt_class = &event_interrupt_class_cik,
149 	.num_of_watch_points = 4,
150 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
151 	.supports_cwsr = true,
152 	.needs_iommu_device = false,
153 	.needs_pci_atomics = false,
154 	.num_sdma_engines = 2,
155 	.num_sdma_queues_per_engine = 2,
156 };
157 
158 
159 static const struct kfd_device_info polaris10_device_info = {
160 	.asic_family = CHIP_POLARIS10,
161 	.max_pasid_bits = 16,
162 	.max_no_of_hqd  = 24,
163 	.doorbell_size  = 4,
164 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
165 	.event_interrupt_class = &event_interrupt_class_cik,
166 	.num_of_watch_points = 4,
167 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
168 	.supports_cwsr = true,
169 	.needs_iommu_device = false,
170 	.needs_pci_atomics = true,
171 	.num_sdma_engines = 2,
172 	.num_sdma_queues_per_engine = 2,
173 };
174 
175 static const struct kfd_device_info polaris10_vf_device_info = {
176 	.asic_family = CHIP_POLARIS10,
177 	.max_pasid_bits = 16,
178 	.max_no_of_hqd  = 24,
179 	.doorbell_size  = 4,
180 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
181 	.event_interrupt_class = &event_interrupt_class_cik,
182 	.num_of_watch_points = 4,
183 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
184 	.supports_cwsr = true,
185 	.needs_iommu_device = false,
186 	.needs_pci_atomics = false,
187 	.num_sdma_engines = 2,
188 	.num_sdma_queues_per_engine = 2,
189 };
190 
191 static const struct kfd_device_info polaris11_device_info = {
192 	.asic_family = CHIP_POLARIS11,
193 	.max_pasid_bits = 16,
194 	.max_no_of_hqd  = 24,
195 	.doorbell_size  = 4,
196 	.ih_ring_entry_size = 4 * sizeof(uint32_t),
197 	.event_interrupt_class = &event_interrupt_class_cik,
198 	.num_of_watch_points = 4,
199 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
200 	.supports_cwsr = true,
201 	.needs_iommu_device = false,
202 	.needs_pci_atomics = true,
203 	.num_sdma_engines = 2,
204 	.num_sdma_queues_per_engine = 2,
205 };
206 
207 static const struct kfd_device_info vega10_device_info = {
208 	.asic_family = CHIP_VEGA10,
209 	.max_pasid_bits = 16,
210 	.max_no_of_hqd  = 24,
211 	.doorbell_size  = 8,
212 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
213 	.event_interrupt_class = &event_interrupt_class_v9,
214 	.num_of_watch_points = 4,
215 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
216 	.supports_cwsr = true,
217 	.needs_iommu_device = false,
218 	.needs_pci_atomics = false,
219 	.num_sdma_engines = 2,
220 	.num_sdma_queues_per_engine = 2,
221 };
222 
223 static const struct kfd_device_info vega10_vf_device_info = {
224 	.asic_family = CHIP_VEGA10,
225 	.max_pasid_bits = 16,
226 	.max_no_of_hqd  = 24,
227 	.doorbell_size  = 8,
228 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
229 	.event_interrupt_class = &event_interrupt_class_v9,
230 	.num_of_watch_points = 4,
231 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
232 	.supports_cwsr = true,
233 	.needs_iommu_device = false,
234 	.needs_pci_atomics = false,
235 	.num_sdma_engines = 2,
236 	.num_sdma_queues_per_engine = 2,
237 };
238 
239 static const struct kfd_device_info vega20_device_info = {
240 	.asic_family = CHIP_VEGA20,
241 	.max_pasid_bits = 16,
242 	.max_no_of_hqd	= 24,
243 	.doorbell_size	= 8,
244 	.ih_ring_entry_size = 8 * sizeof(uint32_t),
245 	.event_interrupt_class = &event_interrupt_class_v9,
246 	.num_of_watch_points = 4,
247 	.mqd_size_aligned = MQD_SIZE_ALIGNED,
248 	.supports_cwsr = true,
249 	.needs_iommu_device = false,
250 	.needs_pci_atomics = false,
251 	.num_sdma_engines = 2,
252 	.num_sdma_queues_per_engine = 8,
253 };
254 
255 struct kfd_deviceid {
256 	unsigned short did;
257 	const struct kfd_device_info *device_info;
258 };
259 
260 static const struct kfd_deviceid supported_devices[] = {
261 #ifdef KFD_SUPPORT_IOMMU_V2
262 	{ 0x1304, &kaveri_device_info },	/* Kaveri */
263 	{ 0x1305, &kaveri_device_info },	/* Kaveri */
264 	{ 0x1306, &kaveri_device_info },	/* Kaveri */
265 	{ 0x1307, &kaveri_device_info },	/* Kaveri */
266 	{ 0x1309, &kaveri_device_info },	/* Kaveri */
267 	{ 0x130A, &kaveri_device_info },	/* Kaveri */
268 	{ 0x130B, &kaveri_device_info },	/* Kaveri */
269 	{ 0x130C, &kaveri_device_info },	/* Kaveri */
270 	{ 0x130D, &kaveri_device_info },	/* Kaveri */
271 	{ 0x130E, &kaveri_device_info },	/* Kaveri */
272 	{ 0x130F, &kaveri_device_info },	/* Kaveri */
273 	{ 0x1310, &kaveri_device_info },	/* Kaveri */
274 	{ 0x1311, &kaveri_device_info },	/* Kaveri */
275 	{ 0x1312, &kaveri_device_info },	/* Kaveri */
276 	{ 0x1313, &kaveri_device_info },	/* Kaveri */
277 	{ 0x1315, &kaveri_device_info },	/* Kaveri */
278 	{ 0x1316, &kaveri_device_info },	/* Kaveri */
279 	{ 0x1317, &kaveri_device_info },	/* Kaveri */
280 	{ 0x1318, &kaveri_device_info },	/* Kaveri */
281 	{ 0x131B, &kaveri_device_info },	/* Kaveri */
282 	{ 0x131C, &kaveri_device_info },	/* Kaveri */
283 	{ 0x131D, &kaveri_device_info },	/* Kaveri */
284 	{ 0x9870, &carrizo_device_info },	/* Carrizo */
285 	{ 0x9874, &carrizo_device_info },	/* Carrizo */
286 	{ 0x9875, &carrizo_device_info },	/* Carrizo */
287 	{ 0x9876, &carrizo_device_info },	/* Carrizo */
288 	{ 0x9877, &carrizo_device_info },	/* Carrizo */
289 	{ 0x15DD, &raven_device_info },		/* Raven */
290 #endif
291 	{ 0x67A0, &hawaii_device_info },	/* Hawaii */
292 	{ 0x67A1, &hawaii_device_info },	/* Hawaii */
293 	{ 0x67A2, &hawaii_device_info },	/* Hawaii */
294 	{ 0x67A8, &hawaii_device_info },	/* Hawaii */
295 	{ 0x67A9, &hawaii_device_info },	/* Hawaii */
296 	{ 0x67AA, &hawaii_device_info },	/* Hawaii */
297 	{ 0x67B0, &hawaii_device_info },	/* Hawaii */
298 	{ 0x67B1, &hawaii_device_info },	/* Hawaii */
299 	{ 0x67B8, &hawaii_device_info },	/* Hawaii */
300 	{ 0x67B9, &hawaii_device_info },	/* Hawaii */
301 	{ 0x67BA, &hawaii_device_info },	/* Hawaii */
302 	{ 0x67BE, &hawaii_device_info },	/* Hawaii */
303 	{ 0x6920, &tonga_device_info },		/* Tonga */
304 	{ 0x6921, &tonga_device_info },		/* Tonga */
305 	{ 0x6928, &tonga_device_info },		/* Tonga */
306 	{ 0x6929, &tonga_device_info },		/* Tonga */
307 	{ 0x692B, &tonga_device_info },		/* Tonga */
308 	{ 0x6938, &tonga_device_info },		/* Tonga */
309 	{ 0x6939, &tonga_device_info },		/* Tonga */
310 	{ 0x7300, &fiji_device_info },		/* Fiji */
311 	{ 0x730F, &fiji_vf_device_info },	/* Fiji vf*/
312 	{ 0x67C0, &polaris10_device_info },	/* Polaris10 */
313 	{ 0x67C1, &polaris10_device_info },	/* Polaris10 */
314 	{ 0x67C2, &polaris10_device_info },	/* Polaris10 */
315 	{ 0x67C4, &polaris10_device_info },	/* Polaris10 */
316 	{ 0x67C7, &polaris10_device_info },	/* Polaris10 */
317 	{ 0x67C8, &polaris10_device_info },	/* Polaris10 */
318 	{ 0x67C9, &polaris10_device_info },	/* Polaris10 */
319 	{ 0x67CA, &polaris10_device_info },	/* Polaris10 */
320 	{ 0x67CC, &polaris10_device_info },	/* Polaris10 */
321 	{ 0x67CF, &polaris10_device_info },	/* Polaris10 */
322 	{ 0x67D0, &polaris10_vf_device_info },	/* Polaris10 vf*/
323 	{ 0x67DF, &polaris10_device_info },	/* Polaris10 */
324 	{ 0x67E0, &polaris11_device_info },	/* Polaris11 */
325 	{ 0x67E1, &polaris11_device_info },	/* Polaris11 */
326 	{ 0x67E3, &polaris11_device_info },	/* Polaris11 */
327 	{ 0x67E7, &polaris11_device_info },	/* Polaris11 */
328 	{ 0x67E8, &polaris11_device_info },	/* Polaris11 */
329 	{ 0x67E9, &polaris11_device_info },	/* Polaris11 */
330 	{ 0x67EB, &polaris11_device_info },	/* Polaris11 */
331 	{ 0x67EF, &polaris11_device_info },	/* Polaris11 */
332 	{ 0x67FF, &polaris11_device_info },	/* Polaris11 */
333 	{ 0x6860, &vega10_device_info },	/* Vega10 */
334 	{ 0x6861, &vega10_device_info },	/* Vega10 */
335 	{ 0x6862, &vega10_device_info },	/* Vega10 */
336 	{ 0x6863, &vega10_device_info },	/* Vega10 */
337 	{ 0x6864, &vega10_device_info },	/* Vega10 */
338 	{ 0x6867, &vega10_device_info },	/* Vega10 */
339 	{ 0x6868, &vega10_device_info },	/* Vega10 */
340 	{ 0x6869, &vega10_device_info },	/* Vega10 */
341 	{ 0x686A, &vega10_device_info },	/* Vega10 */
342 	{ 0x686B, &vega10_device_info },	/* Vega10 */
343 	{ 0x686C, &vega10_vf_device_info },	/* Vega10  vf*/
344 	{ 0x686D, &vega10_device_info },	/* Vega10 */
345 	{ 0x686E, &vega10_device_info },	/* Vega10 */
346 	{ 0x686F, &vega10_device_info },	/* Vega10 */
347 	{ 0x687F, &vega10_device_info },	/* Vega10 */
348 	{ 0x66a0, &vega20_device_info },	/* Vega20 */
349 	{ 0x66a1, &vega20_device_info },	/* Vega20 */
350 	{ 0x66a2, &vega20_device_info },	/* Vega20 */
351 	{ 0x66a3, &vega20_device_info },	/* Vega20 */
352 	{ 0x66a4, &vega20_device_info },	/* Vega20 */
353 	{ 0x66a7, &vega20_device_info },	/* Vega20 */
354 	{ 0x66af, &vega20_device_info }		/* Vega20 */
355 };
356 
357 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
358 				unsigned int chunk_size);
359 static void kfd_gtt_sa_fini(struct kfd_dev *kfd);
360 
361 static int kfd_resume(struct kfd_dev *kfd);
362 
363 static const struct kfd_device_info *lookup_device_info(unsigned short did)
364 {
365 	size_t i;
366 
367 	for (i = 0; i < ARRAY_SIZE(supported_devices); i++) {
368 		if (supported_devices[i].did == did) {
369 			WARN_ON(!supported_devices[i].device_info);
370 			return supported_devices[i].device_info;
371 		}
372 	}
373 
374 	dev_warn(kfd_device, "DID %04x is missing in supported_devices\n",
375 		 did);
376 
377 	return NULL;
378 }
379 
380 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
381 	struct pci_dev *pdev, const struct kfd2kgd_calls *f2g)
382 {
383 	struct kfd_dev *kfd;
384 	int ret;
385 	const struct kfd_device_info *device_info =
386 					lookup_device_info(pdev->device);
387 
388 	if (!device_info) {
389 		dev_err(kfd_device, "kgd2kfd_probe failed\n");
390 		return NULL;
391 	}
392 
393 	kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
394 	if (!kfd)
395 		return NULL;
396 
397 	/* Allow BIF to recode atomics to PCIe 3.0 AtomicOps.
398 	 * 32 and 64-bit requests are possible and must be
399 	 * supported.
400 	 */
401 	ret = pci_enable_atomic_ops_to_root(pdev,
402 			PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
403 			PCI_EXP_DEVCAP2_ATOMIC_COMP64);
404 	if (device_info->needs_pci_atomics && ret < 0) {
405 		dev_info(kfd_device,
406 			 "skipped device %x:%x, PCI rejects atomics\n",
407 			 pdev->vendor, pdev->device);
408 		kfree(kfd);
409 		return NULL;
410 	} else if (!ret)
411 		kfd->pci_atomic_requested = true;
412 
413 	kfd->kgd = kgd;
414 	kfd->device_info = device_info;
415 	kfd->pdev = pdev;
416 	kfd->init_complete = false;
417 	kfd->kfd2kgd = f2g;
418 
419 	mutex_init(&kfd->doorbell_mutex);
420 	memset(&kfd->doorbell_available_index, 0,
421 		sizeof(kfd->doorbell_available_index));
422 
423 	return kfd;
424 }
425 
426 static void kfd_cwsr_init(struct kfd_dev *kfd)
427 {
428 	if (cwsr_enable && kfd->device_info->supports_cwsr) {
429 		if (kfd->device_info->asic_family < CHIP_VEGA10) {
430 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx8_hex) > PAGE_SIZE);
431 			kfd->cwsr_isa = cwsr_trap_gfx8_hex;
432 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx8_hex);
433 		} else {
434 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_hex) > PAGE_SIZE);
435 			kfd->cwsr_isa = cwsr_trap_gfx9_hex;
436 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_hex);
437 		}
438 
439 		kfd->cwsr_enabled = true;
440 	}
441 }
442 
443 bool kgd2kfd_device_init(struct kfd_dev *kfd,
444 			 const struct kgd2kfd_shared_resources *gpu_resources)
445 {
446 	unsigned int size;
447 
448 	kfd->mec_fw_version = kfd->kfd2kgd->get_fw_version(kfd->kgd,
449 			KGD_ENGINE_MEC1);
450 	kfd->sdma_fw_version = kfd->kfd2kgd->get_fw_version(kfd->kgd,
451 			KGD_ENGINE_SDMA1);
452 	kfd->shared_resources = *gpu_resources;
453 
454 	kfd->vm_info.first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1;
455 	kfd->vm_info.last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1;
456 	kfd->vm_info.vmid_num_kfd = kfd->vm_info.last_vmid_kfd
457 			- kfd->vm_info.first_vmid_kfd + 1;
458 
459 	/* Verify module parameters regarding mapped process number*/
460 	if ((hws_max_conc_proc < 0)
461 			|| (hws_max_conc_proc > kfd->vm_info.vmid_num_kfd)) {
462 		dev_err(kfd_device,
463 			"hws_max_conc_proc %d must be between 0 and %d, use %d instead\n",
464 			hws_max_conc_proc, kfd->vm_info.vmid_num_kfd,
465 			kfd->vm_info.vmid_num_kfd);
466 		kfd->max_proc_per_quantum = kfd->vm_info.vmid_num_kfd;
467 	} else
468 		kfd->max_proc_per_quantum = hws_max_conc_proc;
469 
470 	/* calculate max size of mqds needed for queues */
471 	size = max_num_of_queues_per_device *
472 			kfd->device_info->mqd_size_aligned;
473 
474 	/*
475 	 * calculate max size of runlist packet.
476 	 * There can be only 2 packets at once
477 	 */
478 	size += (KFD_MAX_NUM_OF_PROCESSES * sizeof(struct pm4_mes_map_process) +
479 		max_num_of_queues_per_device * sizeof(struct pm4_mes_map_queues)
480 		+ sizeof(struct pm4_mes_runlist)) * 2;
481 
482 	/* Add size of HIQ & DIQ */
483 	size += KFD_KERNEL_QUEUE_SIZE * 2;
484 
485 	/* add another 512KB for all other allocations on gart (HPD, fences) */
486 	size += 512 * 1024;
487 
488 	if (kfd->kfd2kgd->init_gtt_mem_allocation(
489 			kfd->kgd, size, &kfd->gtt_mem,
490 			&kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr,
491 			false)) {
492 		dev_err(kfd_device, "Could not allocate %d bytes\n", size);
493 		goto out;
494 	}
495 
496 	dev_info(kfd_device, "Allocated %d bytes on gart\n", size);
497 
498 	/* Initialize GTT sa with 512 byte chunk size */
499 	if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
500 		dev_err(kfd_device, "Error initializing gtt sub-allocator\n");
501 		goto kfd_gtt_sa_init_error;
502 	}
503 
504 	if (kfd_doorbell_init(kfd)) {
505 		dev_err(kfd_device,
506 			"Error initializing doorbell aperture\n");
507 		goto kfd_doorbell_error;
508 	}
509 
510 	if (kfd->kfd2kgd->get_hive_id)
511 		kfd->hive_id = kfd->kfd2kgd->get_hive_id(kfd->kgd);
512 
513 	if (kfd_topology_add_device(kfd)) {
514 		dev_err(kfd_device, "Error adding device to topology\n");
515 		goto kfd_topology_add_device_error;
516 	}
517 
518 	if (kfd_interrupt_init(kfd)) {
519 		dev_err(kfd_device, "Error initializing interrupts\n");
520 		goto kfd_interrupt_error;
521 	}
522 
523 	kfd->dqm = device_queue_manager_init(kfd);
524 	if (!kfd->dqm) {
525 		dev_err(kfd_device, "Error initializing queue manager\n");
526 		goto device_queue_manager_error;
527 	}
528 
529 	if (kfd_iommu_device_init(kfd)) {
530 		dev_err(kfd_device, "Error initializing iommuv2\n");
531 		goto device_iommu_error;
532 	}
533 
534 	kfd_cwsr_init(kfd);
535 
536 	if (kfd_resume(kfd))
537 		goto kfd_resume_error;
538 
539 	kfd->dbgmgr = NULL;
540 
541 	kfd->init_complete = true;
542 	dev_info(kfd_device, "added device %x:%x\n", kfd->pdev->vendor,
543 		 kfd->pdev->device);
544 
545 	pr_debug("Starting kfd with the following scheduling policy %d\n",
546 		kfd->dqm->sched_policy);
547 
548 	goto out;
549 
550 kfd_resume_error:
551 device_iommu_error:
552 	device_queue_manager_uninit(kfd->dqm);
553 device_queue_manager_error:
554 	kfd_interrupt_exit(kfd);
555 kfd_interrupt_error:
556 	kfd_topology_remove_device(kfd);
557 kfd_topology_add_device_error:
558 	kfd_doorbell_fini(kfd);
559 kfd_doorbell_error:
560 	kfd_gtt_sa_fini(kfd);
561 kfd_gtt_sa_init_error:
562 	kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
563 	dev_err(kfd_device,
564 		"device %x:%x NOT added due to errors\n",
565 		kfd->pdev->vendor, kfd->pdev->device);
566 out:
567 	return kfd->init_complete;
568 }
569 
570 void kgd2kfd_device_exit(struct kfd_dev *kfd)
571 {
572 	if (kfd->init_complete) {
573 		kgd2kfd_suspend(kfd);
574 		device_queue_manager_uninit(kfd->dqm);
575 		kfd_interrupt_exit(kfd);
576 		kfd_topology_remove_device(kfd);
577 		kfd_doorbell_fini(kfd);
578 		kfd_gtt_sa_fini(kfd);
579 		kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
580 	}
581 
582 	kfree(kfd);
583 }
584 
585 int kgd2kfd_pre_reset(struct kfd_dev *kfd)
586 {
587 	if (!kfd->init_complete)
588 		return 0;
589 	kgd2kfd_suspend(kfd);
590 
591 	/* hold dqm->lock to prevent further execution*/
592 	dqm_lock(kfd->dqm);
593 
594 	kfd_signal_reset_event(kfd);
595 	return 0;
596 }
597 
598 /*
599  * Fix me. KFD won't be able to resume existing process for now.
600  * We will keep all existing process in a evicted state and
601  * wait the process to be terminated.
602  */
603 
604 int kgd2kfd_post_reset(struct kfd_dev *kfd)
605 {
606 	int ret, count;
607 
608 	if (!kfd->init_complete)
609 		return 0;
610 
611 	dqm_unlock(kfd->dqm);
612 
613 	ret = kfd_resume(kfd);
614 	if (ret)
615 		return ret;
616 	count = atomic_dec_return(&kfd_locked);
617 	WARN_ONCE(count != 0, "KFD reset ref. error");
618 	return 0;
619 }
620 
621 bool kfd_is_locked(void)
622 {
623 	return  (atomic_read(&kfd_locked) > 0);
624 }
625 
626 void kgd2kfd_suspend(struct kfd_dev *kfd)
627 {
628 	if (!kfd->init_complete)
629 		return;
630 
631 	/* For first KFD device suspend all the KFD processes */
632 	if (atomic_inc_return(&kfd_locked) == 1)
633 		kfd_suspend_all_processes();
634 
635 	kfd->dqm->ops.stop(kfd->dqm);
636 
637 	kfd_iommu_suspend(kfd);
638 }
639 
640 int kgd2kfd_resume(struct kfd_dev *kfd)
641 {
642 	int ret, count;
643 
644 	if (!kfd->init_complete)
645 		return 0;
646 
647 	ret = kfd_resume(kfd);
648 	if (ret)
649 		return ret;
650 
651 	count = atomic_dec_return(&kfd_locked);
652 	WARN_ONCE(count < 0, "KFD suspend / resume ref. error");
653 	if (count == 0)
654 		ret = kfd_resume_all_processes();
655 
656 	return ret;
657 }
658 
659 static int kfd_resume(struct kfd_dev *kfd)
660 {
661 	int err = 0;
662 
663 	err = kfd_iommu_resume(kfd);
664 	if (err) {
665 		dev_err(kfd_device,
666 			"Failed to resume IOMMU for device %x:%x\n",
667 			kfd->pdev->vendor, kfd->pdev->device);
668 		return err;
669 	}
670 
671 	err = kfd->dqm->ops.start(kfd->dqm);
672 	if (err) {
673 		dev_err(kfd_device,
674 			"Error starting queue manager for device %x:%x\n",
675 			kfd->pdev->vendor, kfd->pdev->device);
676 		goto dqm_start_error;
677 	}
678 
679 	return err;
680 
681 dqm_start_error:
682 	kfd_iommu_suspend(kfd);
683 	return err;
684 }
685 
686 /* This is called directly from KGD at ISR. */
687 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
688 {
689 	uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE];
690 	bool is_patched = false;
691 
692 	if (!kfd->init_complete)
693 		return;
694 
695 	if (kfd->device_info->ih_ring_entry_size > sizeof(patched_ihre)) {
696 		dev_err_once(kfd_device, "Ring entry too small\n");
697 		return;
698 	}
699 
700 	spin_lock(&kfd->interrupt_lock);
701 
702 	if (kfd->interrupts_active
703 	    && interrupt_is_wanted(kfd, ih_ring_entry,
704 				   patched_ihre, &is_patched)
705 	    && enqueue_ih_ring_entry(kfd,
706 				     is_patched ? patched_ihre : ih_ring_entry))
707 		queue_work(kfd->ih_wq, &kfd->interrupt_work);
708 
709 	spin_unlock(&kfd->interrupt_lock);
710 }
711 
712 int kgd2kfd_quiesce_mm(struct mm_struct *mm)
713 {
714 	struct kfd_process *p;
715 	int r;
716 
717 	/* Because we are called from arbitrary context (workqueue) as opposed
718 	 * to process context, kfd_process could attempt to exit while we are
719 	 * running so the lookup function increments the process ref count.
720 	 */
721 	p = kfd_lookup_process_by_mm(mm);
722 	if (!p)
723 		return -ESRCH;
724 
725 	r = kfd_process_evict_queues(p);
726 
727 	kfd_unref_process(p);
728 	return r;
729 }
730 
731 int kgd2kfd_resume_mm(struct mm_struct *mm)
732 {
733 	struct kfd_process *p;
734 	int r;
735 
736 	/* Because we are called from arbitrary context (workqueue) as opposed
737 	 * to process context, kfd_process could attempt to exit while we are
738 	 * running so the lookup function increments the process ref count.
739 	 */
740 	p = kfd_lookup_process_by_mm(mm);
741 	if (!p)
742 		return -ESRCH;
743 
744 	r = kfd_process_restore_queues(p);
745 
746 	kfd_unref_process(p);
747 	return r;
748 }
749 
750 /** kgd2kfd_schedule_evict_and_restore_process - Schedules work queue that will
751  *   prepare for safe eviction of KFD BOs that belong to the specified
752  *   process.
753  *
754  * @mm: mm_struct that identifies the specified KFD process
755  * @fence: eviction fence attached to KFD process BOs
756  *
757  */
758 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
759 					       struct dma_fence *fence)
760 {
761 	struct kfd_process *p;
762 	unsigned long active_time;
763 	unsigned long delay_jiffies = msecs_to_jiffies(PROCESS_ACTIVE_TIME_MS);
764 
765 	if (!fence)
766 		return -EINVAL;
767 
768 	if (dma_fence_is_signaled(fence))
769 		return 0;
770 
771 	p = kfd_lookup_process_by_mm(mm);
772 	if (!p)
773 		return -ENODEV;
774 
775 	if (fence->seqno == p->last_eviction_seqno)
776 		goto out;
777 
778 	p->last_eviction_seqno = fence->seqno;
779 
780 	/* Avoid KFD process starvation. Wait for at least
781 	 * PROCESS_ACTIVE_TIME_MS before evicting the process again
782 	 */
783 	active_time = get_jiffies_64() - p->last_restore_timestamp;
784 	if (delay_jiffies > active_time)
785 		delay_jiffies -= active_time;
786 	else
787 		delay_jiffies = 0;
788 
789 	/* During process initialization eviction_work.dwork is initialized
790 	 * to kfd_evict_bo_worker
791 	 */
792 	schedule_delayed_work(&p->eviction_work, delay_jiffies);
793 out:
794 	kfd_unref_process(p);
795 	return 0;
796 }
797 
798 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
799 				unsigned int chunk_size)
800 {
801 	unsigned int num_of_longs;
802 
803 	if (WARN_ON(buf_size < chunk_size))
804 		return -EINVAL;
805 	if (WARN_ON(buf_size == 0))
806 		return -EINVAL;
807 	if (WARN_ON(chunk_size == 0))
808 		return -EINVAL;
809 
810 	kfd->gtt_sa_chunk_size = chunk_size;
811 	kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
812 
813 	num_of_longs = (kfd->gtt_sa_num_of_chunks + BITS_PER_LONG - 1) /
814 		BITS_PER_LONG;
815 
816 	kfd->gtt_sa_bitmap = kcalloc(num_of_longs, sizeof(long), GFP_KERNEL);
817 
818 	if (!kfd->gtt_sa_bitmap)
819 		return -ENOMEM;
820 
821 	pr_debug("gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
822 			kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);
823 
824 	mutex_init(&kfd->gtt_sa_lock);
825 
826 	return 0;
827 
828 }
829 
830 static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
831 {
832 	mutex_destroy(&kfd->gtt_sa_lock);
833 	kfree(kfd->gtt_sa_bitmap);
834 }
835 
836 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
837 						unsigned int bit_num,
838 						unsigned int chunk_size)
839 {
840 	return start_addr + bit_num * chunk_size;
841 }
842 
843 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
844 						unsigned int bit_num,
845 						unsigned int chunk_size)
846 {
847 	return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
848 }
849 
850 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
851 			struct kfd_mem_obj **mem_obj)
852 {
853 	unsigned int found, start_search, cur_size;
854 
855 	if (size == 0)
856 		return -EINVAL;
857 
858 	if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
859 		return -ENOMEM;
860 
861 	*mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
862 	if (!(*mem_obj))
863 		return -ENOMEM;
864 
865 	pr_debug("Allocated mem_obj = %p for size = %d\n", *mem_obj, size);
866 
867 	start_search = 0;
868 
869 	mutex_lock(&kfd->gtt_sa_lock);
870 
871 kfd_gtt_restart_search:
872 	/* Find the first chunk that is free */
873 	found = find_next_zero_bit(kfd->gtt_sa_bitmap,
874 					kfd->gtt_sa_num_of_chunks,
875 					start_search);
876 
877 	pr_debug("Found = %d\n", found);
878 
879 	/* If there wasn't any free chunk, bail out */
880 	if (found == kfd->gtt_sa_num_of_chunks)
881 		goto kfd_gtt_no_free_chunk;
882 
883 	/* Update fields of mem_obj */
884 	(*mem_obj)->range_start = found;
885 	(*mem_obj)->range_end = found;
886 	(*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
887 					kfd->gtt_start_gpu_addr,
888 					found,
889 					kfd->gtt_sa_chunk_size);
890 	(*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
891 					kfd->gtt_start_cpu_ptr,
892 					found,
893 					kfd->gtt_sa_chunk_size);
894 
895 	pr_debug("gpu_addr = %p, cpu_addr = %p\n",
896 			(uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
897 
898 	/* If we need only one chunk, mark it as allocated and get out */
899 	if (size <= kfd->gtt_sa_chunk_size) {
900 		pr_debug("Single bit\n");
901 		set_bit(found, kfd->gtt_sa_bitmap);
902 		goto kfd_gtt_out;
903 	}
904 
905 	/* Otherwise, try to see if we have enough contiguous chunks */
906 	cur_size = size - kfd->gtt_sa_chunk_size;
907 	do {
908 		(*mem_obj)->range_end =
909 			find_next_zero_bit(kfd->gtt_sa_bitmap,
910 					kfd->gtt_sa_num_of_chunks, ++found);
911 		/*
912 		 * If next free chunk is not contiguous than we need to
913 		 * restart our search from the last free chunk we found (which
914 		 * wasn't contiguous to the previous ones
915 		 */
916 		if ((*mem_obj)->range_end != found) {
917 			start_search = found;
918 			goto kfd_gtt_restart_search;
919 		}
920 
921 		/*
922 		 * If we reached end of buffer, bail out with error
923 		 */
924 		if (found == kfd->gtt_sa_num_of_chunks)
925 			goto kfd_gtt_no_free_chunk;
926 
927 		/* Check if we don't need another chunk */
928 		if (cur_size <= kfd->gtt_sa_chunk_size)
929 			cur_size = 0;
930 		else
931 			cur_size -= kfd->gtt_sa_chunk_size;
932 
933 	} while (cur_size > 0);
934 
935 	pr_debug("range_start = %d, range_end = %d\n",
936 		(*mem_obj)->range_start, (*mem_obj)->range_end);
937 
938 	/* Mark the chunks as allocated */
939 	for (found = (*mem_obj)->range_start;
940 		found <= (*mem_obj)->range_end;
941 		found++)
942 		set_bit(found, kfd->gtt_sa_bitmap);
943 
944 kfd_gtt_out:
945 	mutex_unlock(&kfd->gtt_sa_lock);
946 	return 0;
947 
948 kfd_gtt_no_free_chunk:
949 	pr_debug("Allocation failed with mem_obj = %p\n", mem_obj);
950 	mutex_unlock(&kfd->gtt_sa_lock);
951 	kfree(mem_obj);
952 	return -ENOMEM;
953 }
954 
955 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj)
956 {
957 	unsigned int bit;
958 
959 	/* Act like kfree when trying to free a NULL object */
960 	if (!mem_obj)
961 		return 0;
962 
963 	pr_debug("Free mem_obj = %p, range_start = %d, range_end = %d\n",
964 			mem_obj, mem_obj->range_start, mem_obj->range_end);
965 
966 	mutex_lock(&kfd->gtt_sa_lock);
967 
968 	/* Mark the chunks as free */
969 	for (bit = mem_obj->range_start;
970 		bit <= mem_obj->range_end;
971 		bit++)
972 		clear_bit(bit, kfd->gtt_sa_bitmap);
973 
974 	mutex_unlock(&kfd->gtt_sa_lock);
975 
976 	kfree(mem_obj);
977 	return 0;
978 }
979 
980 #if defined(CONFIG_DEBUG_FS)
981 
982 /* This function will send a package to HIQ to hang the HWS
983  * which will trigger a GPU reset and bring the HWS back to normal state
984  */
985 int kfd_debugfs_hang_hws(struct kfd_dev *dev)
986 {
987 	int r = 0;
988 
989 	if (dev->dqm->sched_policy != KFD_SCHED_POLICY_HWS) {
990 		pr_err("HWS is not enabled");
991 		return -EINVAL;
992 	}
993 
994 	r = pm_debugfs_hang_hws(&dev->dqm->packets);
995 	if (!r)
996 		r = dqm_debugfs_execute_queues(dev->dqm);
997 
998 	return r;
999 }
1000 
1001 #endif
1002