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 	{ 0x686C, &vega10_vf_device_info },	/* Vega10  vf*/
341 	{ 0x687F, &vega10_device_info },	/* Vega10 */
342 	{ 0x66a0, &vega20_device_info },	/* Vega20 */
343 	{ 0x66a1, &vega20_device_info },	/* Vega20 */
344 	{ 0x66a2, &vega20_device_info },	/* Vega20 */
345 	{ 0x66a3, &vega20_device_info },	/* Vega20 */
346 	{ 0x66a7, &vega20_device_info },	/* Vega20 */
347 	{ 0x66af, &vega20_device_info }		/* Vega20 */
348 };
349 
350 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
351 				unsigned int chunk_size);
352 static void kfd_gtt_sa_fini(struct kfd_dev *kfd);
353 
354 static int kfd_resume(struct kfd_dev *kfd);
355 
356 static const struct kfd_device_info *lookup_device_info(unsigned short did)
357 {
358 	size_t i;
359 
360 	for (i = 0; i < ARRAY_SIZE(supported_devices); i++) {
361 		if (supported_devices[i].did == did) {
362 			WARN_ON(!supported_devices[i].device_info);
363 			return supported_devices[i].device_info;
364 		}
365 	}
366 
367 	dev_warn(kfd_device, "DID %04x is missing in supported_devices\n",
368 		 did);
369 
370 	return NULL;
371 }
372 
373 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
374 	struct pci_dev *pdev, const struct kfd2kgd_calls *f2g)
375 {
376 	struct kfd_dev *kfd;
377 	int ret;
378 	const struct kfd_device_info *device_info =
379 					lookup_device_info(pdev->device);
380 
381 	if (!device_info) {
382 		dev_err(kfd_device, "kgd2kfd_probe failed\n");
383 		return NULL;
384 	}
385 
386 	kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
387 	if (!kfd)
388 		return NULL;
389 
390 	/* Allow BIF to recode atomics to PCIe 3.0 AtomicOps.
391 	 * 32 and 64-bit requests are possible and must be
392 	 * supported.
393 	 */
394 	ret = pci_enable_atomic_ops_to_root(pdev,
395 			PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
396 			PCI_EXP_DEVCAP2_ATOMIC_COMP64);
397 	if (device_info->needs_pci_atomics && ret < 0) {
398 		dev_info(kfd_device,
399 			 "skipped device %x:%x, PCI rejects atomics\n",
400 			 pdev->vendor, pdev->device);
401 		kfree(kfd);
402 		return NULL;
403 	} else if (!ret)
404 		kfd->pci_atomic_requested = true;
405 
406 	kfd->kgd = kgd;
407 	kfd->device_info = device_info;
408 	kfd->pdev = pdev;
409 	kfd->init_complete = false;
410 	kfd->kfd2kgd = f2g;
411 
412 	mutex_init(&kfd->doorbell_mutex);
413 	memset(&kfd->doorbell_available_index, 0,
414 		sizeof(kfd->doorbell_available_index));
415 
416 	return kfd;
417 }
418 
419 static void kfd_cwsr_init(struct kfd_dev *kfd)
420 {
421 	if (cwsr_enable && kfd->device_info->supports_cwsr) {
422 		if (kfd->device_info->asic_family < CHIP_VEGA10) {
423 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx8_hex) > PAGE_SIZE);
424 			kfd->cwsr_isa = cwsr_trap_gfx8_hex;
425 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx8_hex);
426 		} else {
427 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_hex) > PAGE_SIZE);
428 			kfd->cwsr_isa = cwsr_trap_gfx9_hex;
429 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_hex);
430 		}
431 
432 		kfd->cwsr_enabled = true;
433 	}
434 }
435 
436 bool kgd2kfd_device_init(struct kfd_dev *kfd,
437 			 const struct kgd2kfd_shared_resources *gpu_resources)
438 {
439 	unsigned int size;
440 
441 	kfd->mec_fw_version = kfd->kfd2kgd->get_fw_version(kfd->kgd,
442 			KGD_ENGINE_MEC1);
443 	kfd->sdma_fw_version = kfd->kfd2kgd->get_fw_version(kfd->kgd,
444 			KGD_ENGINE_SDMA1);
445 	kfd->shared_resources = *gpu_resources;
446 
447 	kfd->vm_info.first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1;
448 	kfd->vm_info.last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1;
449 	kfd->vm_info.vmid_num_kfd = kfd->vm_info.last_vmid_kfd
450 			- kfd->vm_info.first_vmid_kfd + 1;
451 
452 	/* Verify module parameters regarding mapped process number*/
453 	if ((hws_max_conc_proc < 0)
454 			|| (hws_max_conc_proc > kfd->vm_info.vmid_num_kfd)) {
455 		dev_err(kfd_device,
456 			"hws_max_conc_proc %d must be between 0 and %d, use %d instead\n",
457 			hws_max_conc_proc, kfd->vm_info.vmid_num_kfd,
458 			kfd->vm_info.vmid_num_kfd);
459 		kfd->max_proc_per_quantum = kfd->vm_info.vmid_num_kfd;
460 	} else
461 		kfd->max_proc_per_quantum = hws_max_conc_proc;
462 
463 	/* calculate max size of mqds needed for queues */
464 	size = max_num_of_queues_per_device *
465 			kfd->device_info->mqd_size_aligned;
466 
467 	/*
468 	 * calculate max size of runlist packet.
469 	 * There can be only 2 packets at once
470 	 */
471 	size += (KFD_MAX_NUM_OF_PROCESSES * sizeof(struct pm4_mes_map_process) +
472 		max_num_of_queues_per_device * sizeof(struct pm4_mes_map_queues)
473 		+ sizeof(struct pm4_mes_runlist)) * 2;
474 
475 	/* Add size of HIQ & DIQ */
476 	size += KFD_KERNEL_QUEUE_SIZE * 2;
477 
478 	/* add another 512KB for all other allocations on gart (HPD, fences) */
479 	size += 512 * 1024;
480 
481 	if (kfd->kfd2kgd->init_gtt_mem_allocation(
482 			kfd->kgd, size, &kfd->gtt_mem,
483 			&kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr,
484 			false)) {
485 		dev_err(kfd_device, "Could not allocate %d bytes\n", size);
486 		goto out;
487 	}
488 
489 	dev_info(kfd_device, "Allocated %d bytes on gart\n", size);
490 
491 	/* Initialize GTT sa with 512 byte chunk size */
492 	if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
493 		dev_err(kfd_device, "Error initializing gtt sub-allocator\n");
494 		goto kfd_gtt_sa_init_error;
495 	}
496 
497 	if (kfd_doorbell_init(kfd)) {
498 		dev_err(kfd_device,
499 			"Error initializing doorbell aperture\n");
500 		goto kfd_doorbell_error;
501 	}
502 
503 	if (kfd->kfd2kgd->get_hive_id)
504 		kfd->hive_id = kfd->kfd2kgd->get_hive_id(kfd->kgd);
505 
506 	if (kfd_topology_add_device(kfd)) {
507 		dev_err(kfd_device, "Error adding device to topology\n");
508 		goto kfd_topology_add_device_error;
509 	}
510 
511 	if (kfd_interrupt_init(kfd)) {
512 		dev_err(kfd_device, "Error initializing interrupts\n");
513 		goto kfd_interrupt_error;
514 	}
515 
516 	kfd->dqm = device_queue_manager_init(kfd);
517 	if (!kfd->dqm) {
518 		dev_err(kfd_device, "Error initializing queue manager\n");
519 		goto device_queue_manager_error;
520 	}
521 
522 	if (kfd_iommu_device_init(kfd)) {
523 		dev_err(kfd_device, "Error initializing iommuv2\n");
524 		goto device_iommu_error;
525 	}
526 
527 	kfd_cwsr_init(kfd);
528 
529 	if (kfd_resume(kfd))
530 		goto kfd_resume_error;
531 
532 	kfd->dbgmgr = NULL;
533 
534 	kfd->init_complete = true;
535 	dev_info(kfd_device, "added device %x:%x\n", kfd->pdev->vendor,
536 		 kfd->pdev->device);
537 
538 	pr_debug("Starting kfd with the following scheduling policy %d\n",
539 		kfd->dqm->sched_policy);
540 
541 	goto out;
542 
543 kfd_resume_error:
544 device_iommu_error:
545 	device_queue_manager_uninit(kfd->dqm);
546 device_queue_manager_error:
547 	kfd_interrupt_exit(kfd);
548 kfd_interrupt_error:
549 	kfd_topology_remove_device(kfd);
550 kfd_topology_add_device_error:
551 	kfd_doorbell_fini(kfd);
552 kfd_doorbell_error:
553 	kfd_gtt_sa_fini(kfd);
554 kfd_gtt_sa_init_error:
555 	kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
556 	dev_err(kfd_device,
557 		"device %x:%x NOT added due to errors\n",
558 		kfd->pdev->vendor, kfd->pdev->device);
559 out:
560 	return kfd->init_complete;
561 }
562 
563 void kgd2kfd_device_exit(struct kfd_dev *kfd)
564 {
565 	if (kfd->init_complete) {
566 		kgd2kfd_suspend(kfd);
567 		device_queue_manager_uninit(kfd->dqm);
568 		kfd_interrupt_exit(kfd);
569 		kfd_topology_remove_device(kfd);
570 		kfd_doorbell_fini(kfd);
571 		kfd_gtt_sa_fini(kfd);
572 		kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
573 	}
574 
575 	kfree(kfd);
576 }
577 
578 int kgd2kfd_pre_reset(struct kfd_dev *kfd)
579 {
580 	if (!kfd->init_complete)
581 		return 0;
582 	kgd2kfd_suspend(kfd);
583 
584 	/* hold dqm->lock to prevent further execution*/
585 	dqm_lock(kfd->dqm);
586 
587 	kfd_signal_reset_event(kfd);
588 	return 0;
589 }
590 
591 /*
592  * Fix me. KFD won't be able to resume existing process for now.
593  * We will keep all existing process in a evicted state and
594  * wait the process to be terminated.
595  */
596 
597 int kgd2kfd_post_reset(struct kfd_dev *kfd)
598 {
599 	int ret, count;
600 
601 	if (!kfd->init_complete)
602 		return 0;
603 
604 	dqm_unlock(kfd->dqm);
605 
606 	ret = kfd_resume(kfd);
607 	if (ret)
608 		return ret;
609 	count = atomic_dec_return(&kfd_locked);
610 	WARN_ONCE(count != 0, "KFD reset ref. error");
611 	return 0;
612 }
613 
614 bool kfd_is_locked(void)
615 {
616 	return  (atomic_read(&kfd_locked) > 0);
617 }
618 
619 void kgd2kfd_suspend(struct kfd_dev *kfd)
620 {
621 	if (!kfd->init_complete)
622 		return;
623 
624 	/* For first KFD device suspend all the KFD processes */
625 	if (atomic_inc_return(&kfd_locked) == 1)
626 		kfd_suspend_all_processes();
627 
628 	kfd->dqm->ops.stop(kfd->dqm);
629 
630 	kfd_iommu_suspend(kfd);
631 }
632 
633 int kgd2kfd_resume(struct kfd_dev *kfd)
634 {
635 	int ret, count;
636 
637 	if (!kfd->init_complete)
638 		return 0;
639 
640 	ret = kfd_resume(kfd);
641 	if (ret)
642 		return ret;
643 
644 	count = atomic_dec_return(&kfd_locked);
645 	WARN_ONCE(count < 0, "KFD suspend / resume ref. error");
646 	if (count == 0)
647 		ret = kfd_resume_all_processes();
648 
649 	return ret;
650 }
651 
652 static int kfd_resume(struct kfd_dev *kfd)
653 {
654 	int err = 0;
655 
656 	err = kfd_iommu_resume(kfd);
657 	if (err) {
658 		dev_err(kfd_device,
659 			"Failed to resume IOMMU for device %x:%x\n",
660 			kfd->pdev->vendor, kfd->pdev->device);
661 		return err;
662 	}
663 
664 	err = kfd->dqm->ops.start(kfd->dqm);
665 	if (err) {
666 		dev_err(kfd_device,
667 			"Error starting queue manager for device %x:%x\n",
668 			kfd->pdev->vendor, kfd->pdev->device);
669 		goto dqm_start_error;
670 	}
671 
672 	return err;
673 
674 dqm_start_error:
675 	kfd_iommu_suspend(kfd);
676 	return err;
677 }
678 
679 /* This is called directly from KGD at ISR. */
680 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
681 {
682 	uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE];
683 	bool is_patched = false;
684 
685 	if (!kfd->init_complete)
686 		return;
687 
688 	if (kfd->device_info->ih_ring_entry_size > sizeof(patched_ihre)) {
689 		dev_err_once(kfd_device, "Ring entry too small\n");
690 		return;
691 	}
692 
693 	spin_lock(&kfd->interrupt_lock);
694 
695 	if (kfd->interrupts_active
696 	    && interrupt_is_wanted(kfd, ih_ring_entry,
697 				   patched_ihre, &is_patched)
698 	    && enqueue_ih_ring_entry(kfd,
699 				     is_patched ? patched_ihre : ih_ring_entry))
700 		queue_work(kfd->ih_wq, &kfd->interrupt_work);
701 
702 	spin_unlock(&kfd->interrupt_lock);
703 }
704 
705 int kgd2kfd_quiesce_mm(struct mm_struct *mm)
706 {
707 	struct kfd_process *p;
708 	int r;
709 
710 	/* Because we are called from arbitrary context (workqueue) as opposed
711 	 * to process context, kfd_process could attempt to exit while we are
712 	 * running so the lookup function increments the process ref count.
713 	 */
714 	p = kfd_lookup_process_by_mm(mm);
715 	if (!p)
716 		return -ESRCH;
717 
718 	r = kfd_process_evict_queues(p);
719 
720 	kfd_unref_process(p);
721 	return r;
722 }
723 
724 int kgd2kfd_resume_mm(struct mm_struct *mm)
725 {
726 	struct kfd_process *p;
727 	int r;
728 
729 	/* Because we are called from arbitrary context (workqueue) as opposed
730 	 * to process context, kfd_process could attempt to exit while we are
731 	 * running so the lookup function increments the process ref count.
732 	 */
733 	p = kfd_lookup_process_by_mm(mm);
734 	if (!p)
735 		return -ESRCH;
736 
737 	r = kfd_process_restore_queues(p);
738 
739 	kfd_unref_process(p);
740 	return r;
741 }
742 
743 /** kgd2kfd_schedule_evict_and_restore_process - Schedules work queue that will
744  *   prepare for safe eviction of KFD BOs that belong to the specified
745  *   process.
746  *
747  * @mm: mm_struct that identifies the specified KFD process
748  * @fence: eviction fence attached to KFD process BOs
749  *
750  */
751 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
752 					       struct dma_fence *fence)
753 {
754 	struct kfd_process *p;
755 	unsigned long active_time;
756 	unsigned long delay_jiffies = msecs_to_jiffies(PROCESS_ACTIVE_TIME_MS);
757 
758 	if (!fence)
759 		return -EINVAL;
760 
761 	if (dma_fence_is_signaled(fence))
762 		return 0;
763 
764 	p = kfd_lookup_process_by_mm(mm);
765 	if (!p)
766 		return -ENODEV;
767 
768 	if (fence->seqno == p->last_eviction_seqno)
769 		goto out;
770 
771 	p->last_eviction_seqno = fence->seqno;
772 
773 	/* Avoid KFD process starvation. Wait for at least
774 	 * PROCESS_ACTIVE_TIME_MS before evicting the process again
775 	 */
776 	active_time = get_jiffies_64() - p->last_restore_timestamp;
777 	if (delay_jiffies > active_time)
778 		delay_jiffies -= active_time;
779 	else
780 		delay_jiffies = 0;
781 
782 	/* During process initialization eviction_work.dwork is initialized
783 	 * to kfd_evict_bo_worker
784 	 */
785 	schedule_delayed_work(&p->eviction_work, delay_jiffies);
786 out:
787 	kfd_unref_process(p);
788 	return 0;
789 }
790 
791 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
792 				unsigned int chunk_size)
793 {
794 	unsigned int num_of_longs;
795 
796 	if (WARN_ON(buf_size < chunk_size))
797 		return -EINVAL;
798 	if (WARN_ON(buf_size == 0))
799 		return -EINVAL;
800 	if (WARN_ON(chunk_size == 0))
801 		return -EINVAL;
802 
803 	kfd->gtt_sa_chunk_size = chunk_size;
804 	kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
805 
806 	num_of_longs = (kfd->gtt_sa_num_of_chunks + BITS_PER_LONG - 1) /
807 		BITS_PER_LONG;
808 
809 	kfd->gtt_sa_bitmap = kcalloc(num_of_longs, sizeof(long), GFP_KERNEL);
810 
811 	if (!kfd->gtt_sa_bitmap)
812 		return -ENOMEM;
813 
814 	pr_debug("gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
815 			kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);
816 
817 	mutex_init(&kfd->gtt_sa_lock);
818 
819 	return 0;
820 
821 }
822 
823 static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
824 {
825 	mutex_destroy(&kfd->gtt_sa_lock);
826 	kfree(kfd->gtt_sa_bitmap);
827 }
828 
829 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
830 						unsigned int bit_num,
831 						unsigned int chunk_size)
832 {
833 	return start_addr + bit_num * chunk_size;
834 }
835 
836 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
837 						unsigned int bit_num,
838 						unsigned int chunk_size)
839 {
840 	return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
841 }
842 
843 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
844 			struct kfd_mem_obj **mem_obj)
845 {
846 	unsigned int found, start_search, cur_size;
847 
848 	if (size == 0)
849 		return -EINVAL;
850 
851 	if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
852 		return -ENOMEM;
853 
854 	*mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
855 	if (!(*mem_obj))
856 		return -ENOMEM;
857 
858 	pr_debug("Allocated mem_obj = %p for size = %d\n", *mem_obj, size);
859 
860 	start_search = 0;
861 
862 	mutex_lock(&kfd->gtt_sa_lock);
863 
864 kfd_gtt_restart_search:
865 	/* Find the first chunk that is free */
866 	found = find_next_zero_bit(kfd->gtt_sa_bitmap,
867 					kfd->gtt_sa_num_of_chunks,
868 					start_search);
869 
870 	pr_debug("Found = %d\n", found);
871 
872 	/* If there wasn't any free chunk, bail out */
873 	if (found == kfd->gtt_sa_num_of_chunks)
874 		goto kfd_gtt_no_free_chunk;
875 
876 	/* Update fields of mem_obj */
877 	(*mem_obj)->range_start = found;
878 	(*mem_obj)->range_end = found;
879 	(*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
880 					kfd->gtt_start_gpu_addr,
881 					found,
882 					kfd->gtt_sa_chunk_size);
883 	(*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
884 					kfd->gtt_start_cpu_ptr,
885 					found,
886 					kfd->gtt_sa_chunk_size);
887 
888 	pr_debug("gpu_addr = %p, cpu_addr = %p\n",
889 			(uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
890 
891 	/* If we need only one chunk, mark it as allocated and get out */
892 	if (size <= kfd->gtt_sa_chunk_size) {
893 		pr_debug("Single bit\n");
894 		set_bit(found, kfd->gtt_sa_bitmap);
895 		goto kfd_gtt_out;
896 	}
897 
898 	/* Otherwise, try to see if we have enough contiguous chunks */
899 	cur_size = size - kfd->gtt_sa_chunk_size;
900 	do {
901 		(*mem_obj)->range_end =
902 			find_next_zero_bit(kfd->gtt_sa_bitmap,
903 					kfd->gtt_sa_num_of_chunks, ++found);
904 		/*
905 		 * If next free chunk is not contiguous than we need to
906 		 * restart our search from the last free chunk we found (which
907 		 * wasn't contiguous to the previous ones
908 		 */
909 		if ((*mem_obj)->range_end != found) {
910 			start_search = found;
911 			goto kfd_gtt_restart_search;
912 		}
913 
914 		/*
915 		 * If we reached end of buffer, bail out with error
916 		 */
917 		if (found == kfd->gtt_sa_num_of_chunks)
918 			goto kfd_gtt_no_free_chunk;
919 
920 		/* Check if we don't need another chunk */
921 		if (cur_size <= kfd->gtt_sa_chunk_size)
922 			cur_size = 0;
923 		else
924 			cur_size -= kfd->gtt_sa_chunk_size;
925 
926 	} while (cur_size > 0);
927 
928 	pr_debug("range_start = %d, range_end = %d\n",
929 		(*mem_obj)->range_start, (*mem_obj)->range_end);
930 
931 	/* Mark the chunks as allocated */
932 	for (found = (*mem_obj)->range_start;
933 		found <= (*mem_obj)->range_end;
934 		found++)
935 		set_bit(found, kfd->gtt_sa_bitmap);
936 
937 kfd_gtt_out:
938 	mutex_unlock(&kfd->gtt_sa_lock);
939 	return 0;
940 
941 kfd_gtt_no_free_chunk:
942 	pr_debug("Allocation failed with mem_obj = %p\n", mem_obj);
943 	mutex_unlock(&kfd->gtt_sa_lock);
944 	kfree(mem_obj);
945 	return -ENOMEM;
946 }
947 
948 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj)
949 {
950 	unsigned int bit;
951 
952 	/* Act like kfree when trying to free a NULL object */
953 	if (!mem_obj)
954 		return 0;
955 
956 	pr_debug("Free mem_obj = %p, range_start = %d, range_end = %d\n",
957 			mem_obj, mem_obj->range_start, mem_obj->range_end);
958 
959 	mutex_lock(&kfd->gtt_sa_lock);
960 
961 	/* Mark the chunks as free */
962 	for (bit = mem_obj->range_start;
963 		bit <= mem_obj->range_end;
964 		bit++)
965 		clear_bit(bit, kfd->gtt_sa_bitmap);
966 
967 	mutex_unlock(&kfd->gtt_sa_lock);
968 
969 	kfree(mem_obj);
970 	return 0;
971 }
972 
973 #if defined(CONFIG_DEBUG_FS)
974 
975 /* This function will send a package to HIQ to hang the HWS
976  * which will trigger a GPU reset and bring the HWS back to normal state
977  */
978 int kfd_debugfs_hang_hws(struct kfd_dev *dev)
979 {
980 	int r = 0;
981 
982 	if (dev->dqm->sched_policy != KFD_SCHED_POLICY_HWS) {
983 		pr_err("HWS is not enabled");
984 		return -EINVAL;
985 	}
986 
987 	r = pm_debugfs_hang_hws(&dev->dqm->packets);
988 	if (!r)
989 		r = dqm_debugfs_execute_queues(dev->dqm);
990 
991 	return r;
992 }
993 
994 #endif
995