1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include <linux/bsearch.h>
25 #include <linux/pci.h>
26 #include <linux/slab.h>
27 #include "kfd_priv.h"
28 #include "kfd_device_queue_manager.h"
29 #include "kfd_pm4_headers_vi.h"
30 #include "kfd_pm4_headers_aldebaran.h"
31 #include "cwsr_trap_handler.h"
32 #include "kfd_iommu.h"
33 #include "amdgpu_amdkfd.h"
34 #include "kfd_smi_events.h"
35 #include "kfd_migrate.h"
36 #include "amdgpu.h"
37 
38 #define MQD_SIZE_ALIGNED 768
39 
40 /*
41  * kfd_locked is used to lock the kfd driver during suspend or reset
42  * once locked, kfd driver will stop any further GPU execution.
43  * create process (open) will return -EAGAIN.
44  */
45 static atomic_t kfd_locked = ATOMIC_INIT(0);
46 
47 #ifdef CONFIG_DRM_AMDGPU_CIK
48 extern const struct kfd2kgd_calls gfx_v7_kfd2kgd;
49 #endif
50 extern const struct kfd2kgd_calls gfx_v8_kfd2kgd;
51 extern const struct kfd2kgd_calls gfx_v9_kfd2kgd;
52 extern const struct kfd2kgd_calls arcturus_kfd2kgd;
53 extern const struct kfd2kgd_calls aldebaran_kfd2kgd;
54 extern const struct kfd2kgd_calls gfx_v10_kfd2kgd;
55 extern const struct kfd2kgd_calls gfx_v10_3_kfd2kgd;
56 
57 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
58 				unsigned int chunk_size);
59 static void kfd_gtt_sa_fini(struct kfd_dev *kfd);
60 
61 static int kfd_resume(struct kfd_dev *kfd);
62 
63 static void kfd_device_info_set_sdma_queue_num(struct kfd_dev *kfd)
64 {
65 	uint32_t sdma_version = kfd->adev->ip_versions[SDMA0_HWIP][0];
66 
67 	switch (sdma_version) {
68 	case IP_VERSION(4, 0, 0):/* VEGA10 */
69 	case IP_VERSION(4, 0, 1):/* VEGA12 */
70 	case IP_VERSION(4, 1, 0):/* RAVEN */
71 	case IP_VERSION(4, 1, 1):/* RAVEN */
72 	case IP_VERSION(4, 1, 2):/* RENOIR */
73 	case IP_VERSION(5, 2, 1):/* VANGOGH */
74 	case IP_VERSION(5, 2, 3):/* YELLOW_CARP */
75 		kfd->device_info.num_sdma_queues_per_engine = 2;
76 		break;
77 	case IP_VERSION(4, 2, 0):/* VEGA20 */
78 	case IP_VERSION(4, 2, 2):/* ARCTURUS */
79 	case IP_VERSION(4, 4, 0):/* ALDEBARAN */
80 	case IP_VERSION(5, 0, 0):/* NAVI10 */
81 	case IP_VERSION(5, 0, 1):/* CYAN_SKILLFISH */
82 	case IP_VERSION(5, 0, 2):/* NAVI14 */
83 	case IP_VERSION(5, 0, 5):/* NAVI12 */
84 	case IP_VERSION(5, 2, 0):/* SIENNA_CICHLID */
85 	case IP_VERSION(5, 2, 2):/* NAVY_FLOUNDER */
86 	case IP_VERSION(5, 2, 4):/* DIMGREY_CAVEFISH */
87 	case IP_VERSION(5, 2, 5):/* BEIGE_GOBY */
88 		kfd->device_info.num_sdma_queues_per_engine = 8;
89 		break;
90 	default:
91 		dev_warn(kfd_device,
92 			"Default sdma queue per engine(8) is set due to mismatch of sdma ip block(SDMA_HWIP:0x%x).\n",
93 			sdma_version);
94 		kfd->device_info.num_sdma_queues_per_engine = 8;
95 	}
96 }
97 
98 static void kfd_device_info_set_event_interrupt_class(struct kfd_dev *kfd)
99 {
100 	uint32_t gc_version = KFD_GC_VERSION(kfd);
101 
102 	switch (gc_version) {
103 	case IP_VERSION(9, 0, 1): /* VEGA10 */
104 	case IP_VERSION(9, 1, 0): /* RAVEN */
105 	case IP_VERSION(9, 2, 1): /* VEGA12 */
106 	case IP_VERSION(9, 2, 2): /* RAVEN */
107 	case IP_VERSION(9, 3, 0): /* RENOIR */
108 	case IP_VERSION(9, 4, 0): /* VEGA20 */
109 	case IP_VERSION(9, 4, 1): /* ARCTURUS */
110 	case IP_VERSION(9, 4, 2): /* ALDEBARAN */
111 	case IP_VERSION(10, 3, 1): /* VANGOGH */
112 	case IP_VERSION(10, 3, 3): /* YELLOW_CARP */
113 	case IP_VERSION(10, 1, 3): /* CYAN_SKILLFISH */
114 	case IP_VERSION(10, 1, 4):
115 	case IP_VERSION(10, 1, 10): /* NAVI10 */
116 	case IP_VERSION(10, 1, 2): /* NAVI12 */
117 	case IP_VERSION(10, 1, 1): /* NAVI14 */
118 	case IP_VERSION(10, 3, 0): /* SIENNA_CICHLID */
119 	case IP_VERSION(10, 3, 2): /* NAVY_FLOUNDER */
120 	case IP_VERSION(10, 3, 4): /* DIMGREY_CAVEFISH */
121 	case IP_VERSION(10, 3, 5): /* BEIGE_GOBY */
122 		kfd->device_info.event_interrupt_class = &event_interrupt_class_v9;
123 		break;
124 	default:
125 		dev_warn(kfd_device, "v9 event interrupt handler is set due to "
126 			"mismatch of gc ip block(GC_HWIP:0x%x).\n", gc_version);
127 		kfd->device_info.event_interrupt_class = &event_interrupt_class_v9;
128 	}
129 }
130 
131 static void kfd_device_info_init(struct kfd_dev *kfd,
132 				 bool vf, uint32_t gfx_target_version)
133 {
134 	uint32_t gc_version = KFD_GC_VERSION(kfd);
135 	uint32_t asic_type = kfd->adev->asic_type;
136 
137 	kfd->device_info.max_pasid_bits = 16;
138 	kfd->device_info.max_no_of_hqd = 24;
139 	kfd->device_info.num_of_watch_points = 4;
140 	kfd->device_info.mqd_size_aligned = MQD_SIZE_ALIGNED;
141 	kfd->device_info.gfx_target_version = gfx_target_version;
142 
143 	if (KFD_IS_SOC15(kfd)) {
144 		kfd->device_info.doorbell_size = 8;
145 		kfd->device_info.ih_ring_entry_size = 8 * sizeof(uint32_t);
146 		kfd->device_info.supports_cwsr = true;
147 
148 		kfd_device_info_set_sdma_queue_num(kfd);
149 
150 		kfd_device_info_set_event_interrupt_class(kfd);
151 
152 		/* Raven */
153 		if (gc_version == IP_VERSION(9, 1, 0) ||
154 		    gc_version == IP_VERSION(9, 2, 2))
155 			kfd->device_info.needs_iommu_device = true;
156 
157 		if (gc_version < IP_VERSION(11, 0, 0)) {
158 			/* Navi2x+, Navi1x+ */
159 			if (gc_version >= IP_VERSION(10, 3, 0))
160 				kfd->device_info.no_atomic_fw_version = 92;
161 			else if (gc_version >= IP_VERSION(10, 1, 1))
162 				kfd->device_info.no_atomic_fw_version = 145;
163 
164 			/* Navi1x+ */
165 			if (gc_version >= IP_VERSION(10, 1, 1))
166 				kfd->device_info.needs_pci_atomics = true;
167 		}
168 	} else {
169 		kfd->device_info.doorbell_size = 4;
170 		kfd->device_info.ih_ring_entry_size = 4 * sizeof(uint32_t);
171 		kfd->device_info.event_interrupt_class = &event_interrupt_class_cik;
172 		kfd->device_info.num_sdma_queues_per_engine = 2;
173 
174 		if (asic_type != CHIP_KAVERI &&
175 		    asic_type != CHIP_HAWAII &&
176 		    asic_type != CHIP_TONGA)
177 			kfd->device_info.supports_cwsr = true;
178 
179 		if (asic_type == CHIP_KAVERI ||
180 		    asic_type == CHIP_CARRIZO)
181 			kfd->device_info.needs_iommu_device = true;
182 
183 		if (asic_type != CHIP_HAWAII && !vf)
184 			kfd->device_info.needs_pci_atomics = true;
185 	}
186 }
187 
188 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf)
189 {
190 	struct kfd_dev *kfd = NULL;
191 	const struct kfd2kgd_calls *f2g = NULL;
192 	struct pci_dev *pdev = adev->pdev;
193 	uint32_t gfx_target_version = 0;
194 
195 	switch (adev->asic_type) {
196 #ifdef KFD_SUPPORT_IOMMU_V2
197 #ifdef CONFIG_DRM_AMDGPU_CIK
198 	case CHIP_KAVERI:
199 		gfx_target_version = 70000;
200 		if (!vf)
201 			f2g = &gfx_v7_kfd2kgd;
202 		break;
203 #endif
204 	case CHIP_CARRIZO:
205 		gfx_target_version = 80001;
206 		if (!vf)
207 			f2g = &gfx_v8_kfd2kgd;
208 		break;
209 #endif
210 #ifdef CONFIG_DRM_AMDGPU_CIK
211 	case CHIP_HAWAII:
212 		gfx_target_version = 70001;
213 		if (!amdgpu_exp_hw_support)
214 			pr_info(
215 	"KFD support on Hawaii is experimental. See modparam exp_hw_support\n"
216 				);
217 		else if (!vf)
218 			f2g = &gfx_v7_kfd2kgd;
219 		break;
220 #endif
221 	case CHIP_TONGA:
222 		gfx_target_version = 80002;
223 		if (!vf)
224 			f2g = &gfx_v8_kfd2kgd;
225 		break;
226 	case CHIP_FIJI:
227 		gfx_target_version = 80003;
228 		f2g = &gfx_v8_kfd2kgd;
229 		break;
230 	case CHIP_POLARIS10:
231 		gfx_target_version = 80003;
232 		f2g = &gfx_v8_kfd2kgd;
233 		break;
234 	case CHIP_POLARIS11:
235 		gfx_target_version = 80003;
236 		if (!vf)
237 			f2g = &gfx_v8_kfd2kgd;
238 		break;
239 	case CHIP_POLARIS12:
240 		gfx_target_version = 80003;
241 		if (!vf)
242 			f2g = &gfx_v8_kfd2kgd;
243 		break;
244 	case CHIP_VEGAM:
245 		gfx_target_version = 80003;
246 		if (!vf)
247 			f2g = &gfx_v8_kfd2kgd;
248 		break;
249 	default:
250 		switch (adev->ip_versions[GC_HWIP][0]) {
251 		/* Vega 10 */
252 		case IP_VERSION(9, 0, 1):
253 			gfx_target_version = 90000;
254 			f2g = &gfx_v9_kfd2kgd;
255 			break;
256 #ifdef KFD_SUPPORT_IOMMU_V2
257 		/* Raven */
258 		case IP_VERSION(9, 1, 0):
259 		case IP_VERSION(9, 2, 2):
260 			gfx_target_version = 90002;
261 			if (!vf)
262 				f2g = &gfx_v9_kfd2kgd;
263 			break;
264 #endif
265 		/* Vega12 */
266 		case IP_VERSION(9, 2, 1):
267 			gfx_target_version = 90004;
268 			if (!vf)
269 				f2g = &gfx_v9_kfd2kgd;
270 			break;
271 		/* Renoir */
272 		case IP_VERSION(9, 3, 0):
273 			gfx_target_version = 90012;
274 			if (!vf)
275 				f2g = &gfx_v9_kfd2kgd;
276 			break;
277 		/* Vega20 */
278 		case IP_VERSION(9, 4, 0):
279 			gfx_target_version = 90006;
280 			if (!vf)
281 				f2g = &gfx_v9_kfd2kgd;
282 			break;
283 		/* Arcturus */
284 		case IP_VERSION(9, 4, 1):
285 			gfx_target_version = 90008;
286 			f2g = &arcturus_kfd2kgd;
287 			break;
288 		/* Aldebaran */
289 		case IP_VERSION(9, 4, 2):
290 			gfx_target_version = 90010;
291 			f2g = &aldebaran_kfd2kgd;
292 			break;
293 		/* Navi10 */
294 		case IP_VERSION(10, 1, 10):
295 			gfx_target_version = 100100;
296 			if (!vf)
297 				f2g = &gfx_v10_kfd2kgd;
298 			break;
299 		/* Navi12 */
300 		case IP_VERSION(10, 1, 2):
301 			gfx_target_version = 100101;
302 			f2g = &gfx_v10_kfd2kgd;
303 			break;
304 		/* Navi14 */
305 		case IP_VERSION(10, 1, 1):
306 			gfx_target_version = 100102;
307 			if (!vf)
308 				f2g = &gfx_v10_kfd2kgd;
309 			break;
310 		/* Cyan Skillfish */
311 		case IP_VERSION(10, 1, 3):
312 		case IP_VERSION(10, 1, 4):
313 			gfx_target_version = 100103;
314 			if (!vf)
315 				f2g = &gfx_v10_kfd2kgd;
316 			break;
317 		/* Sienna Cichlid */
318 		case IP_VERSION(10, 3, 0):
319 			gfx_target_version = 100300;
320 			f2g = &gfx_v10_3_kfd2kgd;
321 			break;
322 		/* Navy Flounder */
323 		case IP_VERSION(10, 3, 2):
324 			gfx_target_version = 100301;
325 			f2g = &gfx_v10_3_kfd2kgd;
326 			break;
327 		/* Van Gogh */
328 		case IP_VERSION(10, 3, 1):
329 			gfx_target_version = 100303;
330 			if (!vf)
331 				f2g = &gfx_v10_3_kfd2kgd;
332 			break;
333 		/* Dimgrey Cavefish */
334 		case IP_VERSION(10, 3, 4):
335 			gfx_target_version = 100302;
336 			f2g = &gfx_v10_3_kfd2kgd;
337 			break;
338 		/* Beige Goby */
339 		case IP_VERSION(10, 3, 5):
340 			gfx_target_version = 100304;
341 			f2g = &gfx_v10_3_kfd2kgd;
342 			break;
343 		/* Yellow Carp */
344 		case IP_VERSION(10, 3, 3):
345 			gfx_target_version = 100305;
346 			if (!vf)
347 				f2g = &gfx_v10_3_kfd2kgd;
348 			break;
349 		default:
350 			break;
351 		}
352 		break;
353 	}
354 
355 	if (!f2g) {
356 		if (adev->ip_versions[GC_HWIP][0])
357 			dev_err(kfd_device, "GC IP %06x %s not supported in kfd\n",
358 				adev->ip_versions[GC_HWIP][0], vf ? "VF" : "");
359 		else
360 			dev_err(kfd_device, "%s %s not supported in kfd\n",
361 				amdgpu_asic_name[adev->asic_type], vf ? "VF" : "");
362 		return NULL;
363 	}
364 
365 	kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
366 	if (!kfd)
367 		return NULL;
368 
369 	kfd->adev = adev;
370 	kfd_device_info_init(kfd, vf, gfx_target_version);
371 	kfd->pdev = pdev;
372 	kfd->init_complete = false;
373 	kfd->kfd2kgd = f2g;
374 	atomic_set(&kfd->compute_profile, 0);
375 
376 	mutex_init(&kfd->doorbell_mutex);
377 	memset(&kfd->doorbell_available_index, 0,
378 		sizeof(kfd->doorbell_available_index));
379 
380 	atomic_set(&kfd->sram_ecc_flag, 0);
381 
382 	ida_init(&kfd->doorbell_ida);
383 
384 	return kfd;
385 }
386 
387 static void kfd_cwsr_init(struct kfd_dev *kfd)
388 {
389 	if (cwsr_enable && kfd->device_info.supports_cwsr) {
390 		if (KFD_GC_VERSION(kfd) < IP_VERSION(9, 0, 1)) {
391 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx8_hex) > PAGE_SIZE);
392 			kfd->cwsr_isa = cwsr_trap_gfx8_hex;
393 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx8_hex);
394 		} else if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 1)) {
395 			BUILD_BUG_ON(sizeof(cwsr_trap_arcturus_hex) > PAGE_SIZE);
396 			kfd->cwsr_isa = cwsr_trap_arcturus_hex;
397 			kfd->cwsr_isa_size = sizeof(cwsr_trap_arcturus_hex);
398 		} else if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 2)) {
399 			BUILD_BUG_ON(sizeof(cwsr_trap_aldebaran_hex) > PAGE_SIZE);
400 			kfd->cwsr_isa = cwsr_trap_aldebaran_hex;
401 			kfd->cwsr_isa_size = sizeof(cwsr_trap_aldebaran_hex);
402 		} else if (KFD_GC_VERSION(kfd) < IP_VERSION(10, 1, 1)) {
403 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_hex) > PAGE_SIZE);
404 			kfd->cwsr_isa = cwsr_trap_gfx9_hex;
405 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_hex);
406 		} else if (KFD_GC_VERSION(kfd) < IP_VERSION(10, 3, 0)) {
407 			BUILD_BUG_ON(sizeof(cwsr_trap_nv1x_hex) > PAGE_SIZE);
408 			kfd->cwsr_isa = cwsr_trap_nv1x_hex;
409 			kfd->cwsr_isa_size = sizeof(cwsr_trap_nv1x_hex);
410 		} else {
411 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx10_hex) > PAGE_SIZE);
412 			kfd->cwsr_isa = cwsr_trap_gfx10_hex;
413 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx10_hex);
414 		}
415 
416 		kfd->cwsr_enabled = true;
417 	}
418 }
419 
420 static int kfd_gws_init(struct kfd_dev *kfd)
421 {
422 	int ret = 0;
423 
424 	if (kfd->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS)
425 		return 0;
426 
427 	if (hws_gws_support || (KFD_IS_SOC15(kfd) &&
428 		((KFD_GC_VERSION(kfd) == IP_VERSION(9, 0, 1)
429 			&& kfd->mec2_fw_version >= 0x81b3) ||
430 		(KFD_GC_VERSION(kfd) <= IP_VERSION(9, 4, 0)
431 			&& kfd->mec2_fw_version >= 0x1b3)  ||
432 		(KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 1)
433 			&& kfd->mec2_fw_version >= 0x30)   ||
434 		(KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 2)
435 			&& kfd->mec2_fw_version >= 0x28))))
436 		ret = amdgpu_amdkfd_alloc_gws(kfd->adev,
437 				kfd->adev->gds.gws_size, &kfd->gws);
438 
439 	return ret;
440 }
441 
442 static void kfd_smi_init(struct kfd_dev *dev)
443 {
444 	INIT_LIST_HEAD(&dev->smi_clients);
445 	spin_lock_init(&dev->smi_lock);
446 }
447 
448 bool kgd2kfd_device_init(struct kfd_dev *kfd,
449 			 struct drm_device *ddev,
450 			 const struct kgd2kfd_shared_resources *gpu_resources)
451 {
452 	unsigned int size, map_process_packet_size;
453 
454 	kfd->ddev = ddev;
455 	kfd->mec_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev,
456 			KGD_ENGINE_MEC1);
457 	kfd->mec2_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev,
458 			KGD_ENGINE_MEC2);
459 	kfd->sdma_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev,
460 			KGD_ENGINE_SDMA1);
461 	kfd->shared_resources = *gpu_resources;
462 
463 	kfd->vm_info.first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1;
464 	kfd->vm_info.last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1;
465 	kfd->vm_info.vmid_num_kfd = kfd->vm_info.last_vmid_kfd
466 			- kfd->vm_info.first_vmid_kfd + 1;
467 
468 	/* Allow BIF to recode atomics to PCIe 3.0 AtomicOps.
469 	 * 32 and 64-bit requests are possible and must be
470 	 * supported.
471 	 */
472 	kfd->pci_atomic_requested = amdgpu_amdkfd_have_atomics_support(kfd->adev);
473 	if (!kfd->pci_atomic_requested &&
474 	    kfd->device_info.needs_pci_atomics &&
475 	    (!kfd->device_info.no_atomic_fw_version ||
476 	     kfd->mec_fw_version < kfd->device_info.no_atomic_fw_version)) {
477 		dev_info(kfd_device,
478 			 "skipped device %x:%x, PCI rejects atomics %d<%d\n",
479 			 kfd->pdev->vendor, kfd->pdev->device,
480 			 kfd->mec_fw_version,
481 			 kfd->device_info.no_atomic_fw_version);
482 		return false;
483 	}
484 
485 	/* Verify module parameters regarding mapped process number*/
486 	if ((hws_max_conc_proc < 0)
487 			|| (hws_max_conc_proc > kfd->vm_info.vmid_num_kfd)) {
488 		dev_err(kfd_device,
489 			"hws_max_conc_proc %d must be between 0 and %d, use %d instead\n",
490 			hws_max_conc_proc, kfd->vm_info.vmid_num_kfd,
491 			kfd->vm_info.vmid_num_kfd);
492 		kfd->max_proc_per_quantum = kfd->vm_info.vmid_num_kfd;
493 	} else
494 		kfd->max_proc_per_quantum = hws_max_conc_proc;
495 
496 	/* calculate max size of mqds needed for queues */
497 	size = max_num_of_queues_per_device *
498 			kfd->device_info.mqd_size_aligned;
499 
500 	/*
501 	 * calculate max size of runlist packet.
502 	 * There can be only 2 packets at once
503 	 */
504 	map_process_packet_size = KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 2) ?
505 				sizeof(struct pm4_mes_map_process_aldebaran) :
506 				sizeof(struct pm4_mes_map_process);
507 	size += (KFD_MAX_NUM_OF_PROCESSES * map_process_packet_size +
508 		max_num_of_queues_per_device * sizeof(struct pm4_mes_map_queues)
509 		+ sizeof(struct pm4_mes_runlist)) * 2;
510 
511 	/* Add size of HIQ & DIQ */
512 	size += KFD_KERNEL_QUEUE_SIZE * 2;
513 
514 	/* add another 512KB for all other allocations on gart (HPD, fences) */
515 	size += 512 * 1024;
516 
517 	if (amdgpu_amdkfd_alloc_gtt_mem(
518 			kfd->adev, size, &kfd->gtt_mem,
519 			&kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr,
520 			false)) {
521 		dev_err(kfd_device, "Could not allocate %d bytes\n", size);
522 		goto alloc_gtt_mem_failure;
523 	}
524 
525 	dev_info(kfd_device, "Allocated %d bytes on gart\n", size);
526 
527 	/* Initialize GTT sa with 512 byte chunk size */
528 	if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
529 		dev_err(kfd_device, "Error initializing gtt sub-allocator\n");
530 		goto kfd_gtt_sa_init_error;
531 	}
532 
533 	if (kfd_doorbell_init(kfd)) {
534 		dev_err(kfd_device,
535 			"Error initializing doorbell aperture\n");
536 		goto kfd_doorbell_error;
537 	}
538 
539 	kfd->hive_id = kfd->adev->gmc.xgmi.hive_id;
540 
541 	kfd->noretry = kfd->adev->gmc.noretry;
542 
543 	if (kfd_interrupt_init(kfd)) {
544 		dev_err(kfd_device, "Error initializing interrupts\n");
545 		goto kfd_interrupt_error;
546 	}
547 
548 	kfd->dqm = device_queue_manager_init(kfd);
549 	if (!kfd->dqm) {
550 		dev_err(kfd_device, "Error initializing queue manager\n");
551 		goto device_queue_manager_error;
552 	}
553 
554 	/* If supported on this device, allocate global GWS that is shared
555 	 * by all KFD processes
556 	 */
557 	if (kfd_gws_init(kfd)) {
558 		dev_err(kfd_device, "Could not allocate %d gws\n",
559 			kfd->adev->gds.gws_size);
560 		goto gws_error;
561 	}
562 
563 	/* If CRAT is broken, won't set iommu enabled */
564 	kfd_double_confirm_iommu_support(kfd);
565 
566 	if (kfd_iommu_device_init(kfd)) {
567 		kfd->use_iommu_v2 = false;
568 		dev_err(kfd_device, "Error initializing iommuv2\n");
569 		goto device_iommu_error;
570 	}
571 
572 	kfd_cwsr_init(kfd);
573 
574 	svm_migrate_init(kfd->adev);
575 
576 	if (kgd2kfd_resume_iommu(kfd))
577 		goto device_iommu_error;
578 
579 	if (kfd_resume(kfd))
580 		goto kfd_resume_error;
581 
582 	if (kfd_topology_add_device(kfd)) {
583 		dev_err(kfd_device, "Error adding device to topology\n");
584 		goto kfd_topology_add_device_error;
585 	}
586 
587 	kfd_smi_init(kfd);
588 
589 	kfd->init_complete = true;
590 	dev_info(kfd_device, "added device %x:%x\n", kfd->pdev->vendor,
591 		 kfd->pdev->device);
592 
593 	pr_debug("Starting kfd with the following scheduling policy %d\n",
594 		kfd->dqm->sched_policy);
595 
596 	goto out;
597 
598 kfd_topology_add_device_error:
599 kfd_resume_error:
600 device_iommu_error:
601 gws_error:
602 	device_queue_manager_uninit(kfd->dqm);
603 device_queue_manager_error:
604 	kfd_interrupt_exit(kfd);
605 kfd_interrupt_error:
606 	kfd_doorbell_fini(kfd);
607 kfd_doorbell_error:
608 	kfd_gtt_sa_fini(kfd);
609 kfd_gtt_sa_init_error:
610 	amdgpu_amdkfd_free_gtt_mem(kfd->adev, kfd->gtt_mem);
611 alloc_gtt_mem_failure:
612 	if (kfd->gws)
613 		amdgpu_amdkfd_free_gws(kfd->adev, kfd->gws);
614 	dev_err(kfd_device,
615 		"device %x:%x NOT added due to errors\n",
616 		kfd->pdev->vendor, kfd->pdev->device);
617 out:
618 	return kfd->init_complete;
619 }
620 
621 void kgd2kfd_device_exit(struct kfd_dev *kfd)
622 {
623 	if (kfd->init_complete) {
624 		device_queue_manager_uninit(kfd->dqm);
625 		kfd_interrupt_exit(kfd);
626 		kfd_topology_remove_device(kfd);
627 		kfd_doorbell_fini(kfd);
628 		ida_destroy(&kfd->doorbell_ida);
629 		kfd_gtt_sa_fini(kfd);
630 		amdgpu_amdkfd_free_gtt_mem(kfd->adev, kfd->gtt_mem);
631 		if (kfd->gws)
632 			amdgpu_amdkfd_free_gws(kfd->adev, kfd->gws);
633 	}
634 
635 	kfree(kfd);
636 }
637 
638 int kgd2kfd_pre_reset(struct kfd_dev *kfd)
639 {
640 	if (!kfd->init_complete)
641 		return 0;
642 
643 	kfd_smi_event_update_gpu_reset(kfd, false);
644 
645 	kfd->dqm->ops.pre_reset(kfd->dqm);
646 
647 	kgd2kfd_suspend(kfd, false);
648 
649 	kfd_signal_reset_event(kfd);
650 	return 0;
651 }
652 
653 /*
654  * Fix me. KFD won't be able to resume existing process for now.
655  * We will keep all existing process in a evicted state and
656  * wait the process to be terminated.
657  */
658 
659 int kgd2kfd_post_reset(struct kfd_dev *kfd)
660 {
661 	int ret;
662 
663 	if (!kfd->init_complete)
664 		return 0;
665 
666 	ret = kfd_resume(kfd);
667 	if (ret)
668 		return ret;
669 	atomic_dec(&kfd_locked);
670 
671 	atomic_set(&kfd->sram_ecc_flag, 0);
672 
673 	kfd_smi_event_update_gpu_reset(kfd, true);
674 
675 	return 0;
676 }
677 
678 bool kfd_is_locked(void)
679 {
680 	return  (atomic_read(&kfd_locked) > 0);
681 }
682 
683 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
684 {
685 	if (!kfd->init_complete)
686 		return;
687 
688 	/* for runtime suspend, skip locking kfd */
689 	if (!run_pm) {
690 		/* For first KFD device suspend all the KFD processes */
691 		if (atomic_inc_return(&kfd_locked) == 1)
692 			kfd_suspend_all_processes();
693 	}
694 
695 	kfd->dqm->ops.stop(kfd->dqm);
696 	kfd_iommu_suspend(kfd);
697 }
698 
699 int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
700 {
701 	int ret, count;
702 
703 	if (!kfd->init_complete)
704 		return 0;
705 
706 	ret = kfd_resume(kfd);
707 	if (ret)
708 		return ret;
709 
710 	/* for runtime resume, skip unlocking kfd */
711 	if (!run_pm) {
712 		count = atomic_dec_return(&kfd_locked);
713 		WARN_ONCE(count < 0, "KFD suspend / resume ref. error");
714 		if (count == 0)
715 			ret = kfd_resume_all_processes();
716 	}
717 
718 	return ret;
719 }
720 
721 int kgd2kfd_resume_iommu(struct kfd_dev *kfd)
722 {
723 	int err = 0;
724 
725 	err = kfd_iommu_resume(kfd);
726 	if (err)
727 		dev_err(kfd_device,
728 			"Failed to resume IOMMU for device %x:%x\n",
729 			kfd->pdev->vendor, kfd->pdev->device);
730 	return err;
731 }
732 
733 static int kfd_resume(struct kfd_dev *kfd)
734 {
735 	int err = 0;
736 
737 	err = kfd->dqm->ops.start(kfd->dqm);
738 	if (err)
739 		dev_err(kfd_device,
740 			"Error starting queue manager for device %x:%x\n",
741 			kfd->pdev->vendor, kfd->pdev->device);
742 
743 	return err;
744 }
745 
746 static inline void kfd_queue_work(struct workqueue_struct *wq,
747 				  struct work_struct *work)
748 {
749 	int cpu, new_cpu;
750 
751 	cpu = new_cpu = smp_processor_id();
752 	do {
753 		new_cpu = cpumask_next(new_cpu, cpu_online_mask) % nr_cpu_ids;
754 		if (cpu_to_node(new_cpu) == numa_node_id())
755 			break;
756 	} while (cpu != new_cpu);
757 
758 	queue_work_on(new_cpu, wq, work);
759 }
760 
761 /* This is called directly from KGD at ISR. */
762 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
763 {
764 	uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE];
765 	bool is_patched = false;
766 	unsigned long flags;
767 
768 	if (!kfd->init_complete)
769 		return;
770 
771 	if (kfd->device_info.ih_ring_entry_size > sizeof(patched_ihre)) {
772 		dev_err_once(kfd_device, "Ring entry too small\n");
773 		return;
774 	}
775 
776 	spin_lock_irqsave(&kfd->interrupt_lock, flags);
777 
778 	if (kfd->interrupts_active
779 	    && interrupt_is_wanted(kfd, ih_ring_entry,
780 				   patched_ihre, &is_patched)
781 	    && enqueue_ih_ring_entry(kfd,
782 				     is_patched ? patched_ihre : ih_ring_entry))
783 		kfd_queue_work(kfd->ih_wq, &kfd->interrupt_work);
784 
785 	spin_unlock_irqrestore(&kfd->interrupt_lock, flags);
786 }
787 
788 int kgd2kfd_quiesce_mm(struct mm_struct *mm)
789 {
790 	struct kfd_process *p;
791 	int r;
792 
793 	/* Because we are called from arbitrary context (workqueue) as opposed
794 	 * to process context, kfd_process could attempt to exit while we are
795 	 * running so the lookup function increments the process ref count.
796 	 */
797 	p = kfd_lookup_process_by_mm(mm);
798 	if (!p)
799 		return -ESRCH;
800 
801 	WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid);
802 	r = kfd_process_evict_queues(p);
803 
804 	kfd_unref_process(p);
805 	return r;
806 }
807 
808 int kgd2kfd_resume_mm(struct mm_struct *mm)
809 {
810 	struct kfd_process *p;
811 	int r;
812 
813 	/* Because we are called from arbitrary context (workqueue) as opposed
814 	 * to process context, kfd_process could attempt to exit while we are
815 	 * running so the lookup function increments the process ref count.
816 	 */
817 	p = kfd_lookup_process_by_mm(mm);
818 	if (!p)
819 		return -ESRCH;
820 
821 	r = kfd_process_restore_queues(p);
822 
823 	kfd_unref_process(p);
824 	return r;
825 }
826 
827 /** kgd2kfd_schedule_evict_and_restore_process - Schedules work queue that will
828  *   prepare for safe eviction of KFD BOs that belong to the specified
829  *   process.
830  *
831  * @mm: mm_struct that identifies the specified KFD process
832  * @fence: eviction fence attached to KFD process BOs
833  *
834  */
835 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
836 					       struct dma_fence *fence)
837 {
838 	struct kfd_process *p;
839 	unsigned long active_time;
840 	unsigned long delay_jiffies = msecs_to_jiffies(PROCESS_ACTIVE_TIME_MS);
841 
842 	if (!fence)
843 		return -EINVAL;
844 
845 	if (dma_fence_is_signaled(fence))
846 		return 0;
847 
848 	p = kfd_lookup_process_by_mm(mm);
849 	if (!p)
850 		return -ENODEV;
851 
852 	if (fence->seqno == p->last_eviction_seqno)
853 		goto out;
854 
855 	p->last_eviction_seqno = fence->seqno;
856 
857 	/* Avoid KFD process starvation. Wait for at least
858 	 * PROCESS_ACTIVE_TIME_MS before evicting the process again
859 	 */
860 	active_time = get_jiffies_64() - p->last_restore_timestamp;
861 	if (delay_jiffies > active_time)
862 		delay_jiffies -= active_time;
863 	else
864 		delay_jiffies = 0;
865 
866 	/* During process initialization eviction_work.dwork is initialized
867 	 * to kfd_evict_bo_worker
868 	 */
869 	WARN(debug_evictions, "Scheduling eviction of pid %d in %ld jiffies",
870 	     p->lead_thread->pid, delay_jiffies);
871 	schedule_delayed_work(&p->eviction_work, delay_jiffies);
872 out:
873 	kfd_unref_process(p);
874 	return 0;
875 }
876 
877 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
878 				unsigned int chunk_size)
879 {
880 	unsigned int num_of_longs;
881 
882 	if (WARN_ON(buf_size < chunk_size))
883 		return -EINVAL;
884 	if (WARN_ON(buf_size == 0))
885 		return -EINVAL;
886 	if (WARN_ON(chunk_size == 0))
887 		return -EINVAL;
888 
889 	kfd->gtt_sa_chunk_size = chunk_size;
890 	kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
891 
892 	num_of_longs = (kfd->gtt_sa_num_of_chunks + BITS_PER_LONG - 1) /
893 		BITS_PER_LONG;
894 
895 	kfd->gtt_sa_bitmap = kcalloc(num_of_longs, sizeof(long), GFP_KERNEL);
896 
897 	if (!kfd->gtt_sa_bitmap)
898 		return -ENOMEM;
899 
900 	pr_debug("gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
901 			kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);
902 
903 	mutex_init(&kfd->gtt_sa_lock);
904 
905 	return 0;
906 
907 }
908 
909 static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
910 {
911 	mutex_destroy(&kfd->gtt_sa_lock);
912 	kfree(kfd->gtt_sa_bitmap);
913 }
914 
915 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
916 						unsigned int bit_num,
917 						unsigned int chunk_size)
918 {
919 	return start_addr + bit_num * chunk_size;
920 }
921 
922 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
923 						unsigned int bit_num,
924 						unsigned int chunk_size)
925 {
926 	return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
927 }
928 
929 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
930 			struct kfd_mem_obj **mem_obj)
931 {
932 	unsigned int found, start_search, cur_size;
933 
934 	if (size == 0)
935 		return -EINVAL;
936 
937 	if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
938 		return -ENOMEM;
939 
940 	*mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
941 	if (!(*mem_obj))
942 		return -ENOMEM;
943 
944 	pr_debug("Allocated mem_obj = %p for size = %d\n", *mem_obj, size);
945 
946 	start_search = 0;
947 
948 	mutex_lock(&kfd->gtt_sa_lock);
949 
950 kfd_gtt_restart_search:
951 	/* Find the first chunk that is free */
952 	found = find_next_zero_bit(kfd->gtt_sa_bitmap,
953 					kfd->gtt_sa_num_of_chunks,
954 					start_search);
955 
956 	pr_debug("Found = %d\n", found);
957 
958 	/* If there wasn't any free chunk, bail out */
959 	if (found == kfd->gtt_sa_num_of_chunks)
960 		goto kfd_gtt_no_free_chunk;
961 
962 	/* Update fields of mem_obj */
963 	(*mem_obj)->range_start = found;
964 	(*mem_obj)->range_end = found;
965 	(*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
966 					kfd->gtt_start_gpu_addr,
967 					found,
968 					kfd->gtt_sa_chunk_size);
969 	(*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
970 					kfd->gtt_start_cpu_ptr,
971 					found,
972 					kfd->gtt_sa_chunk_size);
973 
974 	pr_debug("gpu_addr = %p, cpu_addr = %p\n",
975 			(uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
976 
977 	/* If we need only one chunk, mark it as allocated and get out */
978 	if (size <= kfd->gtt_sa_chunk_size) {
979 		pr_debug("Single bit\n");
980 		set_bit(found, kfd->gtt_sa_bitmap);
981 		goto kfd_gtt_out;
982 	}
983 
984 	/* Otherwise, try to see if we have enough contiguous chunks */
985 	cur_size = size - kfd->gtt_sa_chunk_size;
986 	do {
987 		(*mem_obj)->range_end =
988 			find_next_zero_bit(kfd->gtt_sa_bitmap,
989 					kfd->gtt_sa_num_of_chunks, ++found);
990 		/*
991 		 * If next free chunk is not contiguous than we need to
992 		 * restart our search from the last free chunk we found (which
993 		 * wasn't contiguous to the previous ones
994 		 */
995 		if ((*mem_obj)->range_end != found) {
996 			start_search = found;
997 			goto kfd_gtt_restart_search;
998 		}
999 
1000 		/*
1001 		 * If we reached end of buffer, bail out with error
1002 		 */
1003 		if (found == kfd->gtt_sa_num_of_chunks)
1004 			goto kfd_gtt_no_free_chunk;
1005 
1006 		/* Check if we don't need another chunk */
1007 		if (cur_size <= kfd->gtt_sa_chunk_size)
1008 			cur_size = 0;
1009 		else
1010 			cur_size -= kfd->gtt_sa_chunk_size;
1011 
1012 	} while (cur_size > 0);
1013 
1014 	pr_debug("range_start = %d, range_end = %d\n",
1015 		(*mem_obj)->range_start, (*mem_obj)->range_end);
1016 
1017 	/* Mark the chunks as allocated */
1018 	for (found = (*mem_obj)->range_start;
1019 		found <= (*mem_obj)->range_end;
1020 		found++)
1021 		set_bit(found, kfd->gtt_sa_bitmap);
1022 
1023 kfd_gtt_out:
1024 	mutex_unlock(&kfd->gtt_sa_lock);
1025 	return 0;
1026 
1027 kfd_gtt_no_free_chunk:
1028 	pr_debug("Allocation failed with mem_obj = %p\n", *mem_obj);
1029 	mutex_unlock(&kfd->gtt_sa_lock);
1030 	kfree(*mem_obj);
1031 	return -ENOMEM;
1032 }
1033 
1034 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj)
1035 {
1036 	unsigned int bit;
1037 
1038 	/* Act like kfree when trying to free a NULL object */
1039 	if (!mem_obj)
1040 		return 0;
1041 
1042 	pr_debug("Free mem_obj = %p, range_start = %d, range_end = %d\n",
1043 			mem_obj, mem_obj->range_start, mem_obj->range_end);
1044 
1045 	mutex_lock(&kfd->gtt_sa_lock);
1046 
1047 	/* Mark the chunks as free */
1048 	for (bit = mem_obj->range_start;
1049 		bit <= mem_obj->range_end;
1050 		bit++)
1051 		clear_bit(bit, kfd->gtt_sa_bitmap);
1052 
1053 	mutex_unlock(&kfd->gtt_sa_lock);
1054 
1055 	kfree(mem_obj);
1056 	return 0;
1057 }
1058 
1059 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd)
1060 {
1061 	if (kfd)
1062 		atomic_inc(&kfd->sram_ecc_flag);
1063 }
1064 
1065 void kfd_inc_compute_active(struct kfd_dev *kfd)
1066 {
1067 	if (atomic_inc_return(&kfd->compute_profile) == 1)
1068 		amdgpu_amdkfd_set_compute_idle(kfd->adev, false);
1069 }
1070 
1071 void kfd_dec_compute_active(struct kfd_dev *kfd)
1072 {
1073 	int count = atomic_dec_return(&kfd->compute_profile);
1074 
1075 	if (count == 0)
1076 		amdgpu_amdkfd_set_compute_idle(kfd->adev, true);
1077 	WARN_ONCE(count < 0, "Compute profile ref. count error");
1078 }
1079 
1080 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask)
1081 {
1082 	if (kfd && kfd->init_complete)
1083 		kfd_smi_event_update_thermal_throttling(kfd, throttle_bitmask);
1084 }
1085 
1086 /* kfd_get_num_sdma_engines returns the number of PCIe optimized SDMA and
1087  * kfd_get_num_xgmi_sdma_engines returns the number of XGMI SDMA.
1088  * When the device has more than two engines, we reserve two for PCIe to enable
1089  * full-duplex and the rest are used as XGMI.
1090  */
1091 unsigned int kfd_get_num_sdma_engines(struct kfd_dev *kdev)
1092 {
1093 	/* If XGMI is not supported, all SDMA engines are PCIe */
1094 	if (!kdev->adev->gmc.xgmi.supported)
1095 		return kdev->adev->sdma.num_instances;
1096 
1097 	return min(kdev->adev->sdma.num_instances, 2);
1098 }
1099 
1100 unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_dev *kdev)
1101 {
1102 	/* After reserved for PCIe, the rest of engines are XGMI */
1103 	return kdev->adev->sdma.num_instances - kfd_get_num_sdma_engines(kdev);
1104 }
1105 
1106 #if defined(CONFIG_DEBUG_FS)
1107 
1108 /* This function will send a package to HIQ to hang the HWS
1109  * which will trigger a GPU reset and bring the HWS back to normal state
1110  */
1111 int kfd_debugfs_hang_hws(struct kfd_dev *dev)
1112 {
1113 	if (dev->dqm->sched_policy != KFD_SCHED_POLICY_HWS) {
1114 		pr_err("HWS is not enabled");
1115 		return -EINVAL;
1116 	}
1117 
1118 	return dqm_debugfs_hang_hws(dev->dqm);
1119 }
1120 
1121 #endif
1122