xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_device.c (revision 442bc81bd344dc52c37d8f80b854cc6da062b2d0)
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 "amdgpu_amdkfd.h"
33 #include "kfd_smi_events.h"
34 #include "kfd_svm.h"
35 #include "kfd_migrate.h"
36 #include "amdgpu.h"
37 #include "amdgpu_xcp.h"
38 
39 #define MQD_SIZE_ALIGNED 768
40 
41 /*
42  * kfd_locked is used to lock the kfd driver during suspend or reset
43  * once locked, kfd driver will stop any further GPU execution.
44  * create process (open) will return -EAGAIN.
45  */
46 static int kfd_locked;
47 
48 #ifdef CONFIG_DRM_AMDGPU_CIK
49 extern const struct kfd2kgd_calls gfx_v7_kfd2kgd;
50 #endif
51 extern const struct kfd2kgd_calls gfx_v8_kfd2kgd;
52 extern const struct kfd2kgd_calls gfx_v9_kfd2kgd;
53 extern const struct kfd2kgd_calls arcturus_kfd2kgd;
54 extern const struct kfd2kgd_calls aldebaran_kfd2kgd;
55 extern const struct kfd2kgd_calls gc_9_4_3_kfd2kgd;
56 extern const struct kfd2kgd_calls gfx_v10_kfd2kgd;
57 extern const struct kfd2kgd_calls gfx_v10_3_kfd2kgd;
58 extern const struct kfd2kgd_calls gfx_v11_kfd2kgd;
59 extern const struct kfd2kgd_calls gfx_v12_kfd2kgd;
60 
61 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
62 				unsigned int chunk_size);
63 static void kfd_gtt_sa_fini(struct kfd_dev *kfd);
64 
65 static int kfd_resume(struct kfd_node *kfd);
66 
kfd_device_info_set_sdma_info(struct kfd_dev * kfd)67 static void kfd_device_info_set_sdma_info(struct kfd_dev *kfd)
68 {
69 	uint32_t sdma_version = amdgpu_ip_version(kfd->adev, SDMA0_HWIP, 0);
70 
71 	switch (sdma_version) {
72 	case IP_VERSION(4, 0, 0):/* VEGA10 */
73 	case IP_VERSION(4, 0, 1):/* VEGA12 */
74 	case IP_VERSION(4, 1, 0):/* RAVEN */
75 	case IP_VERSION(4, 1, 1):/* RAVEN */
76 	case IP_VERSION(4, 1, 2):/* RENOIR */
77 	case IP_VERSION(5, 2, 1):/* VANGOGH */
78 	case IP_VERSION(5, 2, 3):/* YELLOW_CARP */
79 	case IP_VERSION(5, 2, 6):/* GC 10.3.6 */
80 	case IP_VERSION(5, 2, 7):/* GC 10.3.7 */
81 		kfd->device_info.num_sdma_queues_per_engine = 2;
82 		break;
83 	case IP_VERSION(4, 2, 0):/* VEGA20 */
84 	case IP_VERSION(4, 2, 2):/* ARCTURUS */
85 	case IP_VERSION(4, 4, 0):/* ALDEBARAN */
86 	case IP_VERSION(4, 4, 2):
87 	case IP_VERSION(4, 4, 5):
88 	case IP_VERSION(4, 4, 4):
89 	case IP_VERSION(5, 0, 0):/* NAVI10 */
90 	case IP_VERSION(5, 0, 1):/* CYAN_SKILLFISH */
91 	case IP_VERSION(5, 0, 2):/* NAVI14 */
92 	case IP_VERSION(5, 0, 5):/* NAVI12 */
93 	case IP_VERSION(5, 2, 0):/* SIENNA_CICHLID */
94 	case IP_VERSION(5, 2, 2):/* NAVY_FLOUNDER */
95 	case IP_VERSION(5, 2, 4):/* DIMGREY_CAVEFISH */
96 	case IP_VERSION(5, 2, 5):/* BEIGE_GOBY */
97 	case IP_VERSION(6, 0, 0):
98 	case IP_VERSION(6, 0, 1):
99 	case IP_VERSION(6, 0, 2):
100 	case IP_VERSION(6, 0, 3):
101 	case IP_VERSION(6, 1, 0):
102 	case IP_VERSION(6, 1, 1):
103 	case IP_VERSION(6, 1, 2):
104 	case IP_VERSION(7, 0, 0):
105 	case IP_VERSION(7, 0, 1):
106 		kfd->device_info.num_sdma_queues_per_engine = 8;
107 		break;
108 	default:
109 		dev_warn(kfd_device,
110 			"Default sdma queue per engine(8) is set due to mismatch of sdma ip block(SDMA_HWIP:0x%x).\n",
111 			sdma_version);
112 		kfd->device_info.num_sdma_queues_per_engine = 8;
113 	}
114 
115 	bitmap_zero(kfd->device_info.reserved_sdma_queues_bitmap, KFD_MAX_SDMA_QUEUES);
116 
117 	switch (sdma_version) {
118 	case IP_VERSION(6, 0, 0):
119 	case IP_VERSION(6, 0, 1):
120 	case IP_VERSION(6, 0, 2):
121 	case IP_VERSION(6, 0, 3):
122 	case IP_VERSION(6, 1, 0):
123 	case IP_VERSION(6, 1, 1):
124 	case IP_VERSION(6, 1, 2):
125 	case IP_VERSION(7, 0, 0):
126 	case IP_VERSION(7, 0, 1):
127 		/* Reserve 1 for paging and 1 for gfx */
128 		kfd->device_info.num_reserved_sdma_queues_per_engine = 2;
129 		/* BIT(0)=engine-0 queue-0; BIT(1)=engine-1 queue-0; BIT(2)=engine-0 queue-1; ... */
130 		bitmap_set(kfd->device_info.reserved_sdma_queues_bitmap, 0,
131 			   kfd->adev->sdma.num_instances *
132 			   kfd->device_info.num_reserved_sdma_queues_per_engine);
133 		break;
134 	default:
135 		break;
136 	}
137 }
138 
kfd_device_info_set_event_interrupt_class(struct kfd_dev * kfd)139 static void kfd_device_info_set_event_interrupt_class(struct kfd_dev *kfd)
140 {
141 	uint32_t gc_version = KFD_GC_VERSION(kfd);
142 
143 	switch (gc_version) {
144 	case IP_VERSION(9, 0, 1): /* VEGA10 */
145 	case IP_VERSION(9, 1, 0): /* RAVEN */
146 	case IP_VERSION(9, 2, 1): /* VEGA12 */
147 	case IP_VERSION(9, 2, 2): /* RAVEN */
148 	case IP_VERSION(9, 3, 0): /* RENOIR */
149 	case IP_VERSION(9, 4, 0): /* VEGA20 */
150 	case IP_VERSION(9, 4, 1): /* ARCTURUS */
151 	case IP_VERSION(9, 4, 2): /* ALDEBARAN */
152 		kfd->device_info.event_interrupt_class = &event_interrupt_class_v9;
153 		break;
154 	case IP_VERSION(9, 4, 3): /* GC 9.4.3 */
155 	case IP_VERSION(9, 4, 4): /* GC 9.4.4 */
156 	case IP_VERSION(9, 5, 0): /* GC 9.5.0 */
157 		kfd->device_info.event_interrupt_class =
158 						&event_interrupt_class_v9_4_3;
159 		break;
160 	case IP_VERSION(10, 3, 1): /* VANGOGH */
161 	case IP_VERSION(10, 3, 3): /* YELLOW_CARP */
162 	case IP_VERSION(10, 3, 6): /* GC 10.3.6 */
163 	case IP_VERSION(10, 3, 7): /* GC 10.3.7 */
164 	case IP_VERSION(10, 1, 3): /* CYAN_SKILLFISH */
165 	case IP_VERSION(10, 1, 4):
166 	case IP_VERSION(10, 1, 10): /* NAVI10 */
167 	case IP_VERSION(10, 1, 2): /* NAVI12 */
168 	case IP_VERSION(10, 1, 1): /* NAVI14 */
169 	case IP_VERSION(10, 3, 0): /* SIENNA_CICHLID */
170 	case IP_VERSION(10, 3, 2): /* NAVY_FLOUNDER */
171 	case IP_VERSION(10, 3, 4): /* DIMGREY_CAVEFISH */
172 	case IP_VERSION(10, 3, 5): /* BEIGE_GOBY */
173 		kfd->device_info.event_interrupt_class = &event_interrupt_class_v10;
174 		break;
175 	case IP_VERSION(11, 0, 0):
176 	case IP_VERSION(11, 0, 1):
177 	case IP_VERSION(11, 0, 2):
178 	case IP_VERSION(11, 0, 3):
179 	case IP_VERSION(11, 0, 4):
180 	case IP_VERSION(11, 5, 0):
181 	case IP_VERSION(11, 5, 1):
182 	case IP_VERSION(11, 5, 2):
183 		kfd->device_info.event_interrupt_class = &event_interrupt_class_v11;
184 		break;
185 	case IP_VERSION(12, 0, 0):
186 	case IP_VERSION(12, 0, 1):
187 		/* GFX12_TODO: Change to v12 version. */
188 		kfd->device_info.event_interrupt_class = &event_interrupt_class_v11;
189 		break;
190 	default:
191 		dev_warn(kfd_device, "v9 event interrupt handler is set due to "
192 			"mismatch of gc ip block(GC_HWIP:0x%x).\n", gc_version);
193 		kfd->device_info.event_interrupt_class = &event_interrupt_class_v9;
194 	}
195 }
196 
kfd_device_info_init(struct kfd_dev * kfd,bool vf,uint32_t gfx_target_version)197 static void kfd_device_info_init(struct kfd_dev *kfd,
198 				 bool vf, uint32_t gfx_target_version)
199 {
200 	uint32_t gc_version = KFD_GC_VERSION(kfd);
201 	uint32_t asic_type = kfd->adev->asic_type;
202 
203 	kfd->device_info.max_pasid_bits = 16;
204 	kfd->device_info.max_no_of_hqd = 24;
205 	kfd->device_info.num_of_watch_points = 4;
206 	kfd->device_info.mqd_size_aligned = MQD_SIZE_ALIGNED;
207 	kfd->device_info.gfx_target_version = gfx_target_version;
208 
209 	if (KFD_IS_SOC15(kfd)) {
210 		kfd->device_info.doorbell_size = 8;
211 		kfd->device_info.ih_ring_entry_size = 8 * sizeof(uint32_t);
212 		kfd->device_info.supports_cwsr = true;
213 
214 		kfd_device_info_set_sdma_info(kfd);
215 
216 		kfd_device_info_set_event_interrupt_class(kfd);
217 
218 		if (gc_version < IP_VERSION(11, 0, 0)) {
219 			/* Navi2x+, Navi1x+ */
220 			if (gc_version == IP_VERSION(10, 3, 6))
221 				kfd->device_info.no_atomic_fw_version = 14;
222 			else if (gc_version == IP_VERSION(10, 3, 7))
223 				kfd->device_info.no_atomic_fw_version = 3;
224 			else if (gc_version >= IP_VERSION(10, 3, 0))
225 				kfd->device_info.no_atomic_fw_version = 92;
226 			else if (gc_version >= IP_VERSION(10, 1, 1))
227 				kfd->device_info.no_atomic_fw_version = 145;
228 
229 			/* Navi1x+ */
230 			if (gc_version >= IP_VERSION(10, 1, 1))
231 				kfd->device_info.needs_pci_atomics = true;
232 		} else if (gc_version < IP_VERSION(12, 0, 0)) {
233 			/*
234 			 * PCIe atomics support acknowledgment in GFX11 RS64 CPFW requires
235 			 * MEC version >= 509. Prior RS64 CPFW versions (and all F32) require
236 			 * PCIe atomics support.
237 			 */
238 			kfd->device_info.needs_pci_atomics = true;
239 			kfd->device_info.no_atomic_fw_version = kfd->adev->gfx.rs64_enable ? 509 : 0;
240 		} else if (gc_version < IP_VERSION(13, 0, 0)) {
241 			kfd->device_info.needs_pci_atomics = true;
242 			kfd->device_info.no_atomic_fw_version = 2090;
243 		} else {
244 			kfd->device_info.needs_pci_atomics = true;
245 		}
246 	} else {
247 		kfd->device_info.doorbell_size = 4;
248 		kfd->device_info.ih_ring_entry_size = 4 * sizeof(uint32_t);
249 		kfd->device_info.event_interrupt_class = &event_interrupt_class_cik;
250 		kfd->device_info.num_sdma_queues_per_engine = 2;
251 
252 		if (asic_type != CHIP_KAVERI &&
253 		    asic_type != CHIP_HAWAII &&
254 		    asic_type != CHIP_TONGA)
255 			kfd->device_info.supports_cwsr = true;
256 
257 		if (asic_type != CHIP_HAWAII && !vf)
258 			kfd->device_info.needs_pci_atomics = true;
259 	}
260 }
261 
kgd2kfd_probe(struct amdgpu_device * adev,bool vf)262 struct kfd_dev *kgd2kfd_probe(struct amdgpu_device *adev, bool vf)
263 {
264 	struct kfd_dev *kfd = NULL;
265 	const struct kfd2kgd_calls *f2g = NULL;
266 	uint32_t gfx_target_version = 0;
267 
268 	switch (adev->asic_type) {
269 #ifdef CONFIG_DRM_AMDGPU_CIK
270 	case CHIP_KAVERI:
271 		gfx_target_version = 70000;
272 		if (!vf)
273 			f2g = &gfx_v7_kfd2kgd;
274 		break;
275 #endif
276 	case CHIP_CARRIZO:
277 		gfx_target_version = 80001;
278 		if (!vf)
279 			f2g = &gfx_v8_kfd2kgd;
280 		break;
281 #ifdef CONFIG_DRM_AMDGPU_CIK
282 	case CHIP_HAWAII:
283 		gfx_target_version = 70001;
284 		if (!amdgpu_exp_hw_support)
285 			pr_info(
286 	"KFD support on Hawaii is experimental. See modparam exp_hw_support\n"
287 				);
288 		else if (!vf)
289 			f2g = &gfx_v7_kfd2kgd;
290 		break;
291 #endif
292 	case CHIP_TONGA:
293 		gfx_target_version = 80002;
294 		if (!vf)
295 			f2g = &gfx_v8_kfd2kgd;
296 		break;
297 	case CHIP_FIJI:
298 	case CHIP_POLARIS10:
299 		gfx_target_version = 80003;
300 		f2g = &gfx_v8_kfd2kgd;
301 		break;
302 	case CHIP_POLARIS11:
303 	case CHIP_POLARIS12:
304 	case CHIP_VEGAM:
305 		gfx_target_version = 80003;
306 		if (!vf)
307 			f2g = &gfx_v8_kfd2kgd;
308 		break;
309 	default:
310 		switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
311 		/* Vega 10 */
312 		case IP_VERSION(9, 0, 1):
313 			gfx_target_version = 90000;
314 			f2g = &gfx_v9_kfd2kgd;
315 			break;
316 		/* Raven */
317 		case IP_VERSION(9, 1, 0):
318 		case IP_VERSION(9, 2, 2):
319 			gfx_target_version = 90002;
320 			if (!vf)
321 				f2g = &gfx_v9_kfd2kgd;
322 			break;
323 		/* Vega12 */
324 		case IP_VERSION(9, 2, 1):
325 			gfx_target_version = 90004;
326 			if (!vf)
327 				f2g = &gfx_v9_kfd2kgd;
328 			break;
329 		/* Renoir */
330 		case IP_VERSION(9, 3, 0):
331 			gfx_target_version = 90012;
332 			if (!vf)
333 				f2g = &gfx_v9_kfd2kgd;
334 			break;
335 		/* Vega20 */
336 		case IP_VERSION(9, 4, 0):
337 			gfx_target_version = 90006;
338 			if (!vf)
339 				f2g = &gfx_v9_kfd2kgd;
340 			break;
341 		/* Arcturus */
342 		case IP_VERSION(9, 4, 1):
343 			gfx_target_version = 90008;
344 			f2g = &arcturus_kfd2kgd;
345 			break;
346 		/* Aldebaran */
347 		case IP_VERSION(9, 4, 2):
348 			gfx_target_version = 90010;
349 			f2g = &aldebaran_kfd2kgd;
350 			break;
351 		case IP_VERSION(9, 4, 3):
352 			gfx_target_version = adev->rev_id >= 1 ? 90402
353 					   : adev->flags & AMD_IS_APU ? 90400
354 					   : 90401;
355 			f2g = &gc_9_4_3_kfd2kgd;
356 			break;
357 		case IP_VERSION(9, 4, 4):
358 			gfx_target_version = 90402;
359 			f2g = &gc_9_4_3_kfd2kgd;
360 			break;
361 		case IP_VERSION(9, 5, 0):
362 			gfx_target_version = 90500;
363 			f2g = &gc_9_4_3_kfd2kgd;
364 			break;
365 		/* Navi10 */
366 		case IP_VERSION(10, 1, 10):
367 			gfx_target_version = 100100;
368 			if (!vf)
369 				f2g = &gfx_v10_kfd2kgd;
370 			break;
371 		/* Navi12 */
372 		case IP_VERSION(10, 1, 2):
373 			gfx_target_version = 100101;
374 			f2g = &gfx_v10_kfd2kgd;
375 			break;
376 		/* Navi14 */
377 		case IP_VERSION(10, 1, 1):
378 			gfx_target_version = 100102;
379 			if (!vf)
380 				f2g = &gfx_v10_kfd2kgd;
381 			break;
382 		/* Cyan Skillfish */
383 		case IP_VERSION(10, 1, 3):
384 		case IP_VERSION(10, 1, 4):
385 			gfx_target_version = 100103;
386 			if (!vf)
387 				f2g = &gfx_v10_kfd2kgd;
388 			break;
389 		/* Sienna Cichlid */
390 		case IP_VERSION(10, 3, 0):
391 			gfx_target_version = 100300;
392 			f2g = &gfx_v10_3_kfd2kgd;
393 			break;
394 		/* Navy Flounder */
395 		case IP_VERSION(10, 3, 2):
396 			gfx_target_version = 100301;
397 			f2g = &gfx_v10_3_kfd2kgd;
398 			break;
399 		/* Van Gogh */
400 		case IP_VERSION(10, 3, 1):
401 			gfx_target_version = 100303;
402 			if (!vf)
403 				f2g = &gfx_v10_3_kfd2kgd;
404 			break;
405 		/* Dimgrey Cavefish */
406 		case IP_VERSION(10, 3, 4):
407 			gfx_target_version = 100302;
408 			f2g = &gfx_v10_3_kfd2kgd;
409 			break;
410 		/* Beige Goby */
411 		case IP_VERSION(10, 3, 5):
412 			gfx_target_version = 100304;
413 			f2g = &gfx_v10_3_kfd2kgd;
414 			break;
415 		/* Yellow Carp */
416 		case IP_VERSION(10, 3, 3):
417 			gfx_target_version = 100305;
418 			if (!vf)
419 				f2g = &gfx_v10_3_kfd2kgd;
420 			break;
421 		case IP_VERSION(10, 3, 6):
422 		case IP_VERSION(10, 3, 7):
423 			gfx_target_version = 100306;
424 			if (!vf)
425 				f2g = &gfx_v10_3_kfd2kgd;
426 			break;
427 		case IP_VERSION(11, 0, 0):
428 			gfx_target_version = 110000;
429 			f2g = &gfx_v11_kfd2kgd;
430 			break;
431 		case IP_VERSION(11, 0, 1):
432 		case IP_VERSION(11, 0, 4):
433 			gfx_target_version = 110003;
434 			f2g = &gfx_v11_kfd2kgd;
435 			break;
436 		case IP_VERSION(11, 0, 2):
437 			gfx_target_version = 110002;
438 			f2g = &gfx_v11_kfd2kgd;
439 			break;
440 		case IP_VERSION(11, 0, 3):
441 			/* Note: Compiler version is 11.0.1 while HW version is 11.0.3 */
442 			gfx_target_version = 110001;
443 			f2g = &gfx_v11_kfd2kgd;
444 			break;
445 		case IP_VERSION(11, 5, 0):
446 			gfx_target_version = 110500;
447 			f2g = &gfx_v11_kfd2kgd;
448 			break;
449 		case IP_VERSION(11, 5, 1):
450 			gfx_target_version = 110501;
451 			f2g = &gfx_v11_kfd2kgd;
452 			break;
453 		case IP_VERSION(11, 5, 2):
454 			gfx_target_version = 110502;
455 			f2g = &gfx_v11_kfd2kgd;
456 			break;
457 		case IP_VERSION(12, 0, 0):
458 			gfx_target_version = 120000;
459 			f2g = &gfx_v12_kfd2kgd;
460 			break;
461 		case IP_VERSION(12, 0, 1):
462 			gfx_target_version = 120001;
463 			f2g = &gfx_v12_kfd2kgd;
464 			break;
465 		default:
466 			break;
467 		}
468 		break;
469 	}
470 
471 	if (!f2g) {
472 		if (amdgpu_ip_version(adev, GC_HWIP, 0))
473 			dev_info(kfd_device,
474 				"GC IP %06x %s not supported in kfd\n",
475 				amdgpu_ip_version(adev, GC_HWIP, 0),
476 				vf ? "VF" : "");
477 		else
478 			dev_info(kfd_device, "%s %s not supported in kfd\n",
479 				amdgpu_asic_name[adev->asic_type], vf ? "VF" : "");
480 		return NULL;
481 	}
482 
483 	kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
484 	if (!kfd)
485 		return NULL;
486 
487 	kfd->adev = adev;
488 	kfd_device_info_init(kfd, vf, gfx_target_version);
489 	kfd->init_complete = false;
490 	kfd->kfd2kgd = f2g;
491 	atomic_set(&kfd->compute_profile, 0);
492 
493 	mutex_init(&kfd->doorbell_mutex);
494 
495 	ida_init(&kfd->doorbell_ida);
496 
497 	return kfd;
498 }
499 
kfd_cwsr_init(struct kfd_dev * kfd)500 static void kfd_cwsr_init(struct kfd_dev *kfd)
501 {
502 	if (cwsr_enable && kfd->device_info.supports_cwsr) {
503 		if (KFD_GC_VERSION(kfd) < IP_VERSION(9, 0, 1)) {
504 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx8_hex)
505 					     > KFD_CWSR_TMA_OFFSET);
506 			kfd->cwsr_isa = cwsr_trap_gfx8_hex;
507 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx8_hex);
508 		} else if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 1)) {
509 			BUILD_BUG_ON(sizeof(cwsr_trap_arcturus_hex)
510 					     > KFD_CWSR_TMA_OFFSET);
511 			kfd->cwsr_isa = cwsr_trap_arcturus_hex;
512 			kfd->cwsr_isa_size = sizeof(cwsr_trap_arcturus_hex);
513 		} else if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 2)) {
514 			BUILD_BUG_ON(sizeof(cwsr_trap_aldebaran_hex)
515 					     > KFD_CWSR_TMA_OFFSET);
516 			kfd->cwsr_isa = cwsr_trap_aldebaran_hex;
517 			kfd->cwsr_isa_size = sizeof(cwsr_trap_aldebaran_hex);
518 		} else if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 3) ||
519 			   KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 4)) {
520 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_4_3_hex)
521 					     > KFD_CWSR_TMA_OFFSET);
522 			kfd->cwsr_isa = cwsr_trap_gfx9_4_3_hex;
523 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_4_3_hex);
524 		} else if (KFD_GC_VERSION(kfd) == IP_VERSION(9, 5, 0)) {
525 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_5_0_hex) > PAGE_SIZE);
526 			kfd->cwsr_isa = cwsr_trap_gfx9_5_0_hex;
527 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_5_0_hex);
528 		} else if (KFD_GC_VERSION(kfd) < IP_VERSION(10, 1, 1)) {
529 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx9_hex)
530 					     > KFD_CWSR_TMA_OFFSET);
531 			kfd->cwsr_isa = cwsr_trap_gfx9_hex;
532 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx9_hex);
533 		} else if (KFD_GC_VERSION(kfd) < IP_VERSION(10, 3, 0)) {
534 			BUILD_BUG_ON(sizeof(cwsr_trap_nv1x_hex)
535 					     > KFD_CWSR_TMA_OFFSET);
536 			kfd->cwsr_isa = cwsr_trap_nv1x_hex;
537 			kfd->cwsr_isa_size = sizeof(cwsr_trap_nv1x_hex);
538 		} else if (KFD_GC_VERSION(kfd) < IP_VERSION(11, 0, 0)) {
539 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx10_hex)
540 					     > KFD_CWSR_TMA_OFFSET);
541 			kfd->cwsr_isa = cwsr_trap_gfx10_hex;
542 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx10_hex);
543 		} else if (KFD_GC_VERSION(kfd) < IP_VERSION(12, 0, 0)) {
544 			/* The gfx11 cwsr trap handler must fit inside a single
545 			   page. */
546 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx11_hex) > PAGE_SIZE);
547 			kfd->cwsr_isa = cwsr_trap_gfx11_hex;
548 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx11_hex);
549 		} else {
550 			BUILD_BUG_ON(sizeof(cwsr_trap_gfx12_hex)
551 					     > KFD_CWSR_TMA_OFFSET);
552 			kfd->cwsr_isa = cwsr_trap_gfx12_hex;
553 			kfd->cwsr_isa_size = sizeof(cwsr_trap_gfx12_hex);
554 		}
555 
556 		kfd->cwsr_enabled = true;
557 	}
558 }
559 
kfd_gws_init(struct kfd_node * node)560 static int kfd_gws_init(struct kfd_node *node)
561 {
562 	int ret = 0;
563 	struct kfd_dev *kfd = node->kfd;
564 	uint32_t mes_rev = node->adev->mes.sched_version & AMDGPU_MES_VERSION_MASK;
565 
566 	if (node->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS)
567 		return 0;
568 
569 	if (hws_gws_support || (KFD_IS_SOC15(node) &&
570 		((KFD_GC_VERSION(node) == IP_VERSION(9, 0, 1)
571 			&& kfd->mec2_fw_version >= 0x81b3) ||
572 		(KFD_GC_VERSION(node) <= IP_VERSION(9, 4, 0)
573 			&& kfd->mec2_fw_version >= 0x1b3)  ||
574 		(KFD_GC_VERSION(node) == IP_VERSION(9, 4, 1)
575 			&& kfd->mec2_fw_version >= 0x30)   ||
576 		(KFD_GC_VERSION(node) == IP_VERSION(9, 4, 2)
577 			&& kfd->mec2_fw_version >= 0x28) ||
578 		(KFD_GC_VERSION(node) == IP_VERSION(9, 4, 3) ||
579 		 KFD_GC_VERSION(node) == IP_VERSION(9, 4, 4)) ||
580 		(KFD_GC_VERSION(node) == IP_VERSION(9, 5, 0)) ||
581 		(KFD_GC_VERSION(node) >= IP_VERSION(10, 3, 0)
582 			&& KFD_GC_VERSION(node) < IP_VERSION(11, 0, 0)
583 			&& kfd->mec2_fw_version >= 0x6b) ||
584 		(KFD_GC_VERSION(node) >= IP_VERSION(11, 0, 0)
585 			&& KFD_GC_VERSION(node) < IP_VERSION(12, 0, 0)
586 			&& mes_rev >= 68))))
587 		ret = amdgpu_amdkfd_alloc_gws(node->adev,
588 				node->adev->gds.gws_size, &node->gws);
589 
590 	return ret;
591 }
592 
kfd_smi_init(struct kfd_node * dev)593 static void kfd_smi_init(struct kfd_node *dev)
594 {
595 	INIT_LIST_HEAD(&dev->smi_clients);
596 	spin_lock_init(&dev->smi_lock);
597 }
598 
kfd_init_node(struct kfd_node * node)599 static int kfd_init_node(struct kfd_node *node)
600 {
601 	int err = -1;
602 
603 	if (kfd_interrupt_init(node)) {
604 		dev_err(kfd_device, "Error initializing interrupts\n");
605 		goto kfd_interrupt_error;
606 	}
607 
608 	node->dqm = device_queue_manager_init(node);
609 	if (!node->dqm) {
610 		dev_err(kfd_device, "Error initializing queue manager\n");
611 		goto device_queue_manager_error;
612 	}
613 
614 	if (kfd_gws_init(node)) {
615 		dev_err(kfd_device, "Could not allocate %d gws\n",
616 			node->adev->gds.gws_size);
617 		goto gws_error;
618 	}
619 
620 	if (kfd_resume(node))
621 		goto kfd_resume_error;
622 
623 	if (kfd_topology_add_device(node)) {
624 		dev_err(kfd_device, "Error adding device to topology\n");
625 		goto kfd_topology_add_device_error;
626 	}
627 
628 	kfd_smi_init(node);
629 
630 	return 0;
631 
632 kfd_topology_add_device_error:
633 kfd_resume_error:
634 gws_error:
635 	device_queue_manager_uninit(node->dqm);
636 device_queue_manager_error:
637 	kfd_interrupt_exit(node);
638 kfd_interrupt_error:
639 	if (node->gws)
640 		amdgpu_amdkfd_free_gws(node->adev, node->gws);
641 
642 	/* Cleanup the node memory here */
643 	kfree(node);
644 	return err;
645 }
646 
kfd_cleanup_nodes(struct kfd_dev * kfd,unsigned int num_nodes)647 static void kfd_cleanup_nodes(struct kfd_dev *kfd, unsigned int num_nodes)
648 {
649 	struct kfd_node *knode;
650 	unsigned int i;
651 
652 	/*
653 	 * flush_work ensures that there are no outstanding
654 	 * work-queue items that will access interrupt_ring. New work items
655 	 * can't be created because we stopped interrupt handling above.
656 	 */
657 	flush_workqueue(kfd->ih_wq);
658 	destroy_workqueue(kfd->ih_wq);
659 
660 	for (i = 0; i < num_nodes; i++) {
661 		knode = kfd->nodes[i];
662 		device_queue_manager_uninit(knode->dqm);
663 		kfd_interrupt_exit(knode);
664 		kfd_topology_remove_device(knode);
665 		if (knode->gws)
666 			amdgpu_amdkfd_free_gws(knode->adev, knode->gws);
667 		kfree(knode);
668 		kfd->nodes[i] = NULL;
669 	}
670 }
671 
kfd_setup_interrupt_bitmap(struct kfd_node * node,unsigned int kfd_node_idx)672 static void kfd_setup_interrupt_bitmap(struct kfd_node *node,
673 				       unsigned int kfd_node_idx)
674 {
675 	struct amdgpu_device *adev = node->adev;
676 	uint32_t xcc_mask = node->xcc_mask;
677 	uint32_t xcc, mapped_xcc;
678 	/*
679 	 * Interrupt bitmap is setup for processing interrupts from
680 	 * different XCDs and AIDs.
681 	 * Interrupt bitmap is defined as follows:
682 	 * 1. Bits 0-15 - correspond to the NodeId field.
683 	 *    Each bit corresponds to NodeId number. For example, if
684 	 *    a KFD node has interrupt bitmap set to 0x7, then this
685 	 *    KFD node will process interrupts with NodeId = 0, 1 and 2
686 	 *    in the IH cookie.
687 	 * 2. Bits 16-31 - unused.
688 	 *
689 	 * Please note that the kfd_node_idx argument passed to this
690 	 * function is not related to NodeId field received in the
691 	 * IH cookie.
692 	 *
693 	 * In CPX mode, a KFD node will process an interrupt if:
694 	 * - the Node Id matches the corresponding bit set in
695 	 *   Bits 0-15.
696 	 * - AND VMID reported in the interrupt lies within the
697 	 *   VMID range of the node.
698 	 */
699 	for_each_inst(xcc, xcc_mask) {
700 		mapped_xcc = GET_INST(GC, xcc);
701 		node->interrupt_bitmap |= (mapped_xcc % 2 ? 5 : 3) << (4 * (mapped_xcc / 2));
702 	}
703 	dev_info(kfd_device, "Node: %d, interrupt_bitmap: %x\n", kfd_node_idx,
704 							node->interrupt_bitmap);
705 }
706 
kgd2kfd_device_init(struct kfd_dev * kfd,const struct kgd2kfd_shared_resources * gpu_resources)707 bool kgd2kfd_device_init(struct kfd_dev *kfd,
708 			 const struct kgd2kfd_shared_resources *gpu_resources)
709 {
710 	unsigned int size, map_process_packet_size, i;
711 	struct kfd_node *node;
712 	uint32_t first_vmid_kfd, last_vmid_kfd, vmid_num_kfd;
713 	unsigned int max_proc_per_quantum;
714 	int partition_mode;
715 	int xcp_idx;
716 
717 	kfd->mec_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev,
718 			KGD_ENGINE_MEC1);
719 	kfd->mec2_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev,
720 			KGD_ENGINE_MEC2);
721 	kfd->sdma_fw_version = amdgpu_amdkfd_get_fw_version(kfd->adev,
722 			KGD_ENGINE_SDMA1);
723 	kfd->shared_resources = *gpu_resources;
724 
725 	kfd->num_nodes = amdgpu_xcp_get_num_xcp(kfd->adev->xcp_mgr);
726 
727 	if (kfd->num_nodes == 0) {
728 		dev_err(kfd_device,
729 			"KFD num nodes cannot be 0, num_xcc_in_node: %d\n",
730 			kfd->adev->gfx.num_xcc_per_xcp);
731 		goto out;
732 	}
733 
734 	/* Allow BIF to recode atomics to PCIe 3.0 AtomicOps.
735 	 * 32 and 64-bit requests are possible and must be
736 	 * supported.
737 	 */
738 	kfd->pci_atomic_requested = amdgpu_amdkfd_have_atomics_support(kfd->adev);
739 	if (!kfd->pci_atomic_requested &&
740 	    kfd->device_info.needs_pci_atomics &&
741 	    (!kfd->device_info.no_atomic_fw_version ||
742 	     kfd->mec_fw_version < kfd->device_info.no_atomic_fw_version)) {
743 		dev_info(kfd_device,
744 			 "skipped device %x:%x, PCI rejects atomics %d<%d\n",
745 			 kfd->adev->pdev->vendor, kfd->adev->pdev->device,
746 			 kfd->mec_fw_version,
747 			 kfd->device_info.no_atomic_fw_version);
748 		return false;
749 	}
750 
751 	first_vmid_kfd = ffs(gpu_resources->compute_vmid_bitmap)-1;
752 	last_vmid_kfd = fls(gpu_resources->compute_vmid_bitmap)-1;
753 	vmid_num_kfd = last_vmid_kfd - first_vmid_kfd + 1;
754 
755 	/* For multi-partition capable GPUs, we need special handling for VMIDs
756 	 * depending on partition mode.
757 	 * In CPX mode, the VMID range needs to be shared between XCDs.
758 	 * Additionally, there are 13 VMIDs (3-15) available for KFD. To
759 	 * divide them equally, we change starting VMID to 4 and not use
760 	 * VMID 3.
761 	 * If the VMID range changes for multi-partition capable GPUs, then
762 	 * this code MUST be revisited.
763 	 */
764 	if (kfd->adev->xcp_mgr) {
765 		partition_mode = amdgpu_xcp_query_partition_mode(kfd->adev->xcp_mgr,
766 								 AMDGPU_XCP_FL_LOCKED);
767 		if (partition_mode == AMDGPU_CPX_PARTITION_MODE &&
768 		    kfd->num_nodes != 1) {
769 			vmid_num_kfd /= 2;
770 			first_vmid_kfd = last_vmid_kfd + 1 - vmid_num_kfd*2;
771 		}
772 	}
773 
774 	/* Verify module parameters regarding mapped process number*/
775 	if (hws_max_conc_proc >= 0)
776 		max_proc_per_quantum = min((u32)hws_max_conc_proc, vmid_num_kfd);
777 	else
778 		max_proc_per_quantum = vmid_num_kfd;
779 
780 	/* calculate max size of mqds needed for queues */
781 	size = max_num_of_queues_per_device *
782 			kfd->device_info.mqd_size_aligned;
783 
784 	/*
785 	 * calculate max size of runlist packet.
786 	 * There can be only 2 packets at once
787 	 */
788 	map_process_packet_size = KFD_GC_VERSION(kfd) == IP_VERSION(9, 4, 2) ?
789 				sizeof(struct pm4_mes_map_process_aldebaran) :
790 				sizeof(struct pm4_mes_map_process);
791 	size += (KFD_MAX_NUM_OF_PROCESSES * map_process_packet_size +
792 		max_num_of_queues_per_device * sizeof(struct pm4_mes_map_queues)
793 		+ sizeof(struct pm4_mes_runlist)) * 2;
794 
795 	/* Add size of HIQ & DIQ */
796 	size += KFD_KERNEL_QUEUE_SIZE * 2;
797 
798 	/* add another 512KB for all other allocations on gart (HPD, fences) */
799 	size += 512 * 1024;
800 
801 	if (amdgpu_amdkfd_alloc_gtt_mem(
802 			kfd->adev, size, &kfd->gtt_mem,
803 			&kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr,
804 			false)) {
805 		dev_err(kfd_device, "Could not allocate %d bytes\n", size);
806 		goto alloc_gtt_mem_failure;
807 	}
808 
809 	dev_info(kfd_device, "Allocated %d bytes on gart\n", size);
810 
811 	/* Initialize GTT sa with 512 byte chunk size */
812 	if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
813 		dev_err(kfd_device, "Error initializing gtt sub-allocator\n");
814 		goto kfd_gtt_sa_init_error;
815 	}
816 
817 	if (kfd_doorbell_init(kfd)) {
818 		dev_err(kfd_device,
819 			"Error initializing doorbell aperture\n");
820 		goto kfd_doorbell_error;
821 	}
822 
823 	if (amdgpu_use_xgmi_p2p)
824 		kfd->hive_id = kfd->adev->gmc.xgmi.hive_id;
825 
826 	/*
827 	 * For multi-partition capable GPUs, the KFD abstracts all partitions
828 	 * within a socket as xGMI connected in the topology so assign a unique
829 	 * hive id per device based on the pci device location if device is in
830 	 * PCIe mode.
831 	 */
832 	if (!kfd->hive_id && kfd->num_nodes > 1)
833 		kfd->hive_id = pci_dev_id(kfd->adev->pdev);
834 
835 	kfd->noretry = kfd->adev->gmc.noretry;
836 
837 	kfd_cwsr_init(kfd);
838 
839 	dev_info(kfd_device, "Total number of KFD nodes to be created: %d\n",
840 				kfd->num_nodes);
841 
842 	/* Allocate the KFD nodes */
843 	for (i = 0, xcp_idx = 0; i < kfd->num_nodes; i++) {
844 		node = kzalloc(sizeof(struct kfd_node), GFP_KERNEL);
845 		if (!node)
846 			goto node_alloc_error;
847 
848 		node->node_id = i;
849 		node->adev = kfd->adev;
850 		node->kfd = kfd;
851 		node->kfd2kgd = kfd->kfd2kgd;
852 		node->vm_info.vmid_num_kfd = vmid_num_kfd;
853 		node->xcp = amdgpu_get_next_xcp(kfd->adev->xcp_mgr, &xcp_idx);
854 		/* TODO : Check if error handling is needed */
855 		if (node->xcp) {
856 			amdgpu_xcp_get_inst_details(node->xcp, AMDGPU_XCP_GFX,
857 						    &node->xcc_mask);
858 			++xcp_idx;
859 		} else {
860 			node->xcc_mask =
861 				(1U << NUM_XCC(kfd->adev->gfx.xcc_mask)) - 1;
862 		}
863 
864 		if (node->xcp) {
865 			dev_info(kfd_device, "KFD node %d partition %d size %lldM\n",
866 				node->node_id, node->xcp->mem_id,
867 				KFD_XCP_MEMORY_SIZE(node->adev, node->node_id) >> 20);
868 		}
869 
870 		if (partition_mode == AMDGPU_CPX_PARTITION_MODE &&
871 		    kfd->num_nodes != 1) {
872 			/* For multi-partition capable GPUs and CPX mode, first
873 			 * XCD gets VMID range 4-9 and second XCD gets VMID
874 			 * range 10-15.
875 			 */
876 
877 			node->vm_info.first_vmid_kfd = (i%2 == 0) ?
878 						first_vmid_kfd :
879 						first_vmid_kfd+vmid_num_kfd;
880 			node->vm_info.last_vmid_kfd = (i%2 == 0) ?
881 						last_vmid_kfd-vmid_num_kfd :
882 						last_vmid_kfd;
883 			node->compute_vmid_bitmap =
884 				((0x1 << (node->vm_info.last_vmid_kfd + 1)) - 1) -
885 				((0x1 << (node->vm_info.first_vmid_kfd)) - 1);
886 		} else {
887 			node->vm_info.first_vmid_kfd = first_vmid_kfd;
888 			node->vm_info.last_vmid_kfd = last_vmid_kfd;
889 			node->compute_vmid_bitmap =
890 				gpu_resources->compute_vmid_bitmap;
891 		}
892 		node->max_proc_per_quantum = max_proc_per_quantum;
893 		atomic_set(&node->sram_ecc_flag, 0);
894 
895 		amdgpu_amdkfd_get_local_mem_info(kfd->adev,
896 					&node->local_mem_info, node->xcp);
897 
898 		if (kfd->adev->xcp_mgr)
899 			kfd_setup_interrupt_bitmap(node, i);
900 
901 		/* Initialize the KFD node */
902 		if (kfd_init_node(node)) {
903 			dev_err(kfd_device, "Error initializing KFD node\n");
904 			goto node_init_error;
905 		}
906 
907 		spin_lock_init(&node->watch_points_lock);
908 
909 		kfd->nodes[i] = node;
910 	}
911 
912 	svm_range_set_max_pages(kfd->adev);
913 
914 	kfd->init_complete = true;
915 	dev_info(kfd_device, "added device %x:%x\n", kfd->adev->pdev->vendor,
916 		 kfd->adev->pdev->device);
917 
918 	pr_debug("Starting kfd with the following scheduling policy %d\n",
919 		node->dqm->sched_policy);
920 
921 	goto out;
922 
923 node_init_error:
924 node_alloc_error:
925 	kfd_cleanup_nodes(kfd, i);
926 	kfd_doorbell_fini(kfd);
927 kfd_doorbell_error:
928 	kfd_gtt_sa_fini(kfd);
929 kfd_gtt_sa_init_error:
930 	amdgpu_amdkfd_free_gtt_mem(kfd->adev, &kfd->gtt_mem);
931 alloc_gtt_mem_failure:
932 	dev_err(kfd_device,
933 		"device %x:%x NOT added due to errors\n",
934 		kfd->adev->pdev->vendor, kfd->adev->pdev->device);
935 out:
936 	return kfd->init_complete;
937 }
938 
kgd2kfd_device_exit(struct kfd_dev * kfd)939 void kgd2kfd_device_exit(struct kfd_dev *kfd)
940 {
941 	if (kfd->init_complete) {
942 		/* Cleanup KFD nodes */
943 		kfd_cleanup_nodes(kfd, kfd->num_nodes);
944 		/* Cleanup common/shared resources */
945 		kfd_doorbell_fini(kfd);
946 		ida_destroy(&kfd->doorbell_ida);
947 		kfd_gtt_sa_fini(kfd);
948 		amdgpu_amdkfd_free_gtt_mem(kfd->adev, &kfd->gtt_mem);
949 	}
950 
951 	kfree(kfd);
952 }
953 
kgd2kfd_pre_reset(struct kfd_dev * kfd,struct amdgpu_reset_context * reset_context)954 int kgd2kfd_pre_reset(struct kfd_dev *kfd,
955 		      struct amdgpu_reset_context *reset_context)
956 {
957 	struct kfd_node *node;
958 	int i;
959 
960 	if (!kfd->init_complete)
961 		return 0;
962 
963 	for (i = 0; i < kfd->num_nodes; i++) {
964 		node = kfd->nodes[i];
965 		kfd_smi_event_update_gpu_reset(node, false, reset_context);
966 	}
967 
968 	kgd2kfd_suspend(kfd, false);
969 
970 	for (i = 0; i < kfd->num_nodes; i++)
971 		kfd_signal_reset_event(kfd->nodes[i]);
972 
973 	return 0;
974 }
975 
976 /*
977  * Fix me. KFD won't be able to resume existing process for now.
978  * We will keep all existing process in a evicted state and
979  * wait the process to be terminated.
980  */
981 
kgd2kfd_post_reset(struct kfd_dev * kfd)982 int kgd2kfd_post_reset(struct kfd_dev *kfd)
983 {
984 	int ret;
985 	struct kfd_node *node;
986 	int i;
987 
988 	if (!kfd->init_complete)
989 		return 0;
990 
991 	for (i = 0; i < kfd->num_nodes; i++) {
992 		ret = kfd_resume(kfd->nodes[i]);
993 		if (ret)
994 			return ret;
995 	}
996 
997 	mutex_lock(&kfd_processes_mutex);
998 	--kfd_locked;
999 	mutex_unlock(&kfd_processes_mutex);
1000 
1001 	for (i = 0; i < kfd->num_nodes; i++) {
1002 		node = kfd->nodes[i];
1003 		atomic_set(&node->sram_ecc_flag, 0);
1004 		kfd_smi_event_update_gpu_reset(node, true, NULL);
1005 	}
1006 
1007 	return 0;
1008 }
1009 
kfd_is_locked(void)1010 bool kfd_is_locked(void)
1011 {
1012 	lockdep_assert_held(&kfd_processes_mutex);
1013 	return  (kfd_locked > 0);
1014 }
1015 
kgd2kfd_suspend(struct kfd_dev * kfd,bool run_pm)1016 void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
1017 {
1018 	struct kfd_node *node;
1019 	int i;
1020 
1021 	if (!kfd->init_complete)
1022 		return;
1023 
1024 	/* for runtime suspend, skip locking kfd */
1025 	if (!run_pm) {
1026 		mutex_lock(&kfd_processes_mutex);
1027 		/* For first KFD device suspend all the KFD processes */
1028 		if (++kfd_locked == 1)
1029 			kfd_suspend_all_processes();
1030 		mutex_unlock(&kfd_processes_mutex);
1031 	}
1032 
1033 	for (i = 0; i < kfd->num_nodes; i++) {
1034 		node = kfd->nodes[i];
1035 		node->dqm->ops.stop(node->dqm);
1036 	}
1037 }
1038 
kgd2kfd_resume(struct kfd_dev * kfd,bool run_pm)1039 int kgd2kfd_resume(struct kfd_dev *kfd, bool run_pm)
1040 {
1041 	int ret, i;
1042 
1043 	if (!kfd->init_complete)
1044 		return 0;
1045 
1046 	for (i = 0; i < kfd->num_nodes; i++) {
1047 		ret = kfd_resume(kfd->nodes[i]);
1048 		if (ret)
1049 			return ret;
1050 	}
1051 
1052 	/* for runtime resume, skip unlocking kfd */
1053 	if (!run_pm) {
1054 		mutex_lock(&kfd_processes_mutex);
1055 		if (--kfd_locked == 0)
1056 			ret = kfd_resume_all_processes();
1057 		WARN_ONCE(kfd_locked < 0, "KFD suspend / resume ref. error");
1058 		mutex_unlock(&kfd_processes_mutex);
1059 	}
1060 
1061 	return ret;
1062 }
1063 
kfd_resume(struct kfd_node * node)1064 static int kfd_resume(struct kfd_node *node)
1065 {
1066 	int err = 0;
1067 
1068 	err = node->dqm->ops.start(node->dqm);
1069 	if (err)
1070 		dev_err(kfd_device,
1071 			"Error starting queue manager for device %x:%x\n",
1072 			node->adev->pdev->vendor, node->adev->pdev->device);
1073 
1074 	return err;
1075 }
1076 
1077 /* This is called directly from KGD at ISR. */
kgd2kfd_interrupt(struct kfd_dev * kfd,const void * ih_ring_entry)1078 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
1079 {
1080 	uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE], i;
1081 	bool is_patched = false;
1082 	unsigned long flags;
1083 	struct kfd_node *node;
1084 
1085 	if (!kfd->init_complete)
1086 		return;
1087 
1088 	if (kfd->device_info.ih_ring_entry_size > sizeof(patched_ihre)) {
1089 		dev_err_once(kfd_device, "Ring entry too small\n");
1090 		return;
1091 	}
1092 
1093 	for (i = 0; i < kfd->num_nodes; i++) {
1094 		node = kfd->nodes[i];
1095 		spin_lock_irqsave(&node->interrupt_lock, flags);
1096 
1097 		if (node->interrupts_active
1098 		    && interrupt_is_wanted(node, ih_ring_entry,
1099 			    	patched_ihre, &is_patched)
1100 		    && enqueue_ih_ring_entry(node,
1101 			    	is_patched ? patched_ihre : ih_ring_entry)) {
1102 			queue_work(node->kfd->ih_wq, &node->interrupt_work);
1103 			spin_unlock_irqrestore(&node->interrupt_lock, flags);
1104 			return;
1105 		}
1106 		spin_unlock_irqrestore(&node->interrupt_lock, flags);
1107 	}
1108 
1109 }
1110 
kgd2kfd_quiesce_mm(struct mm_struct * mm,uint32_t trigger)1111 int kgd2kfd_quiesce_mm(struct mm_struct *mm, uint32_t trigger)
1112 {
1113 	struct kfd_process *p;
1114 	int r;
1115 
1116 	/* Because we are called from arbitrary context (workqueue) as opposed
1117 	 * to process context, kfd_process could attempt to exit while we are
1118 	 * running so the lookup function increments the process ref count.
1119 	 */
1120 	p = kfd_lookup_process_by_mm(mm);
1121 	if (!p)
1122 		return -ESRCH;
1123 
1124 	WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid);
1125 	r = kfd_process_evict_queues(p, trigger);
1126 
1127 	kfd_unref_process(p);
1128 	return r;
1129 }
1130 
kgd2kfd_resume_mm(struct mm_struct * mm)1131 int kgd2kfd_resume_mm(struct mm_struct *mm)
1132 {
1133 	struct kfd_process *p;
1134 	int r;
1135 
1136 	/* Because we are called from arbitrary context (workqueue) as opposed
1137 	 * to process context, kfd_process could attempt to exit while we are
1138 	 * running so the lookup function increments the process ref count.
1139 	 */
1140 	p = kfd_lookup_process_by_mm(mm);
1141 	if (!p)
1142 		return -ESRCH;
1143 
1144 	r = kfd_process_restore_queues(p);
1145 
1146 	kfd_unref_process(p);
1147 	return r;
1148 }
1149 
1150 /** kgd2kfd_schedule_evict_and_restore_process - Schedules work queue that will
1151  *   prepare for safe eviction of KFD BOs that belong to the specified
1152  *   process.
1153  *
1154  * @mm: mm_struct that identifies the specified KFD process
1155  * @fence: eviction fence attached to KFD process BOs
1156  *
1157  */
kgd2kfd_schedule_evict_and_restore_process(struct mm_struct * mm,struct dma_fence * fence)1158 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
1159 					       struct dma_fence *fence)
1160 {
1161 	struct kfd_process *p;
1162 	unsigned long active_time;
1163 	unsigned long delay_jiffies = msecs_to_jiffies(PROCESS_ACTIVE_TIME_MS);
1164 
1165 	if (!fence)
1166 		return -EINVAL;
1167 
1168 	if (dma_fence_is_signaled(fence))
1169 		return 0;
1170 
1171 	p = kfd_lookup_process_by_mm(mm);
1172 	if (!p)
1173 		return -ENODEV;
1174 
1175 	if (fence->seqno == p->last_eviction_seqno)
1176 		goto out;
1177 
1178 	p->last_eviction_seqno = fence->seqno;
1179 
1180 	/* Avoid KFD process starvation. Wait for at least
1181 	 * PROCESS_ACTIVE_TIME_MS before evicting the process again
1182 	 */
1183 	active_time = get_jiffies_64() - p->last_restore_timestamp;
1184 	if (delay_jiffies > active_time)
1185 		delay_jiffies -= active_time;
1186 	else
1187 		delay_jiffies = 0;
1188 
1189 	/* During process initialization eviction_work.dwork is initialized
1190 	 * to kfd_evict_bo_worker
1191 	 */
1192 	WARN(debug_evictions, "Scheduling eviction of pid %d in %ld jiffies",
1193 	     p->lead_thread->pid, delay_jiffies);
1194 	schedule_delayed_work(&p->eviction_work, delay_jiffies);
1195 out:
1196 	kfd_unref_process(p);
1197 	return 0;
1198 }
1199 
kfd_gtt_sa_init(struct kfd_dev * kfd,unsigned int buf_size,unsigned int chunk_size)1200 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
1201 				unsigned int chunk_size)
1202 {
1203 	if (WARN_ON(buf_size < chunk_size))
1204 		return -EINVAL;
1205 	if (WARN_ON(buf_size == 0))
1206 		return -EINVAL;
1207 	if (WARN_ON(chunk_size == 0))
1208 		return -EINVAL;
1209 
1210 	kfd->gtt_sa_chunk_size = chunk_size;
1211 	kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
1212 
1213 	kfd->gtt_sa_bitmap = bitmap_zalloc(kfd->gtt_sa_num_of_chunks,
1214 					   GFP_KERNEL);
1215 	if (!kfd->gtt_sa_bitmap)
1216 		return -ENOMEM;
1217 
1218 	pr_debug("gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
1219 			kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);
1220 
1221 	mutex_init(&kfd->gtt_sa_lock);
1222 
1223 	return 0;
1224 }
1225 
kfd_gtt_sa_fini(struct kfd_dev * kfd)1226 static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
1227 {
1228 	mutex_destroy(&kfd->gtt_sa_lock);
1229 	bitmap_free(kfd->gtt_sa_bitmap);
1230 }
1231 
kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,unsigned int bit_num,unsigned int chunk_size)1232 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
1233 						unsigned int bit_num,
1234 						unsigned int chunk_size)
1235 {
1236 	return start_addr + bit_num * chunk_size;
1237 }
1238 
kfd_gtt_sa_calc_cpu_addr(void * start_addr,unsigned int bit_num,unsigned int chunk_size)1239 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
1240 						unsigned int bit_num,
1241 						unsigned int chunk_size)
1242 {
1243 	return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
1244 }
1245 
kfd_gtt_sa_allocate(struct kfd_node * node,unsigned int size,struct kfd_mem_obj ** mem_obj)1246 int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size,
1247 			struct kfd_mem_obj **mem_obj)
1248 {
1249 	unsigned int found, start_search, cur_size;
1250 	struct kfd_dev *kfd = node->kfd;
1251 
1252 	if (size == 0)
1253 		return -EINVAL;
1254 
1255 	if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
1256 		return -ENOMEM;
1257 
1258 	*mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
1259 	if (!(*mem_obj))
1260 		return -ENOMEM;
1261 
1262 	pr_debug("Allocated mem_obj = %p for size = %d\n", *mem_obj, size);
1263 
1264 	start_search = 0;
1265 
1266 	mutex_lock(&kfd->gtt_sa_lock);
1267 
1268 kfd_gtt_restart_search:
1269 	/* Find the first chunk that is free */
1270 	found = find_next_zero_bit(kfd->gtt_sa_bitmap,
1271 					kfd->gtt_sa_num_of_chunks,
1272 					start_search);
1273 
1274 	pr_debug("Found = %d\n", found);
1275 
1276 	/* If there wasn't any free chunk, bail out */
1277 	if (found == kfd->gtt_sa_num_of_chunks)
1278 		goto kfd_gtt_no_free_chunk;
1279 
1280 	/* Update fields of mem_obj */
1281 	(*mem_obj)->range_start = found;
1282 	(*mem_obj)->range_end = found;
1283 	(*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
1284 					kfd->gtt_start_gpu_addr,
1285 					found,
1286 					kfd->gtt_sa_chunk_size);
1287 	(*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
1288 					kfd->gtt_start_cpu_ptr,
1289 					found,
1290 					kfd->gtt_sa_chunk_size);
1291 
1292 	pr_debug("gpu_addr = %p, cpu_addr = %p\n",
1293 			(uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
1294 
1295 	/* If we need only one chunk, mark it as allocated and get out */
1296 	if (size <= kfd->gtt_sa_chunk_size) {
1297 		pr_debug("Single bit\n");
1298 		__set_bit(found, kfd->gtt_sa_bitmap);
1299 		goto kfd_gtt_out;
1300 	}
1301 
1302 	/* Otherwise, try to see if we have enough contiguous chunks */
1303 	cur_size = size - kfd->gtt_sa_chunk_size;
1304 	do {
1305 		(*mem_obj)->range_end =
1306 			find_next_zero_bit(kfd->gtt_sa_bitmap,
1307 					kfd->gtt_sa_num_of_chunks, ++found);
1308 		/*
1309 		 * If next free chunk is not contiguous than we need to
1310 		 * restart our search from the last free chunk we found (which
1311 		 * wasn't contiguous to the previous ones
1312 		 */
1313 		if ((*mem_obj)->range_end != found) {
1314 			start_search = found;
1315 			goto kfd_gtt_restart_search;
1316 		}
1317 
1318 		/*
1319 		 * If we reached end of buffer, bail out with error
1320 		 */
1321 		if (found == kfd->gtt_sa_num_of_chunks)
1322 			goto kfd_gtt_no_free_chunk;
1323 
1324 		/* Check if we don't need another chunk */
1325 		if (cur_size <= kfd->gtt_sa_chunk_size)
1326 			cur_size = 0;
1327 		else
1328 			cur_size -= kfd->gtt_sa_chunk_size;
1329 
1330 	} while (cur_size > 0);
1331 
1332 	pr_debug("range_start = %d, range_end = %d\n",
1333 		(*mem_obj)->range_start, (*mem_obj)->range_end);
1334 
1335 	/* Mark the chunks as allocated */
1336 	bitmap_set(kfd->gtt_sa_bitmap, (*mem_obj)->range_start,
1337 		   (*mem_obj)->range_end - (*mem_obj)->range_start + 1);
1338 
1339 kfd_gtt_out:
1340 	mutex_unlock(&kfd->gtt_sa_lock);
1341 	return 0;
1342 
1343 kfd_gtt_no_free_chunk:
1344 	pr_debug("Allocation failed with mem_obj = %p\n", *mem_obj);
1345 	mutex_unlock(&kfd->gtt_sa_lock);
1346 	kfree(*mem_obj);
1347 	return -ENOMEM;
1348 }
1349 
kfd_gtt_sa_free(struct kfd_node * node,struct kfd_mem_obj * mem_obj)1350 int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj)
1351 {
1352 	struct kfd_dev *kfd = node->kfd;
1353 
1354 	/* Act like kfree when trying to free a NULL object */
1355 	if (!mem_obj)
1356 		return 0;
1357 
1358 	pr_debug("Free mem_obj = %p, range_start = %d, range_end = %d\n",
1359 			mem_obj, mem_obj->range_start, mem_obj->range_end);
1360 
1361 	mutex_lock(&kfd->gtt_sa_lock);
1362 
1363 	/* Mark the chunks as free */
1364 	bitmap_clear(kfd->gtt_sa_bitmap, mem_obj->range_start,
1365 		     mem_obj->range_end - mem_obj->range_start + 1);
1366 
1367 	mutex_unlock(&kfd->gtt_sa_lock);
1368 
1369 	kfree(mem_obj);
1370 	return 0;
1371 }
1372 
kgd2kfd_set_sram_ecc_flag(struct kfd_dev * kfd)1373 void kgd2kfd_set_sram_ecc_flag(struct kfd_dev *kfd)
1374 {
1375 	/*
1376 	 * TODO: Currently update SRAM ECC flag for first node.
1377 	 * This needs to be updated later when we can
1378 	 * identify SRAM ECC error on other nodes also.
1379 	 */
1380 	if (kfd)
1381 		atomic_inc(&kfd->nodes[0]->sram_ecc_flag);
1382 }
1383 
kfd_inc_compute_active(struct kfd_node * node)1384 void kfd_inc_compute_active(struct kfd_node *node)
1385 {
1386 	if (atomic_inc_return(&node->kfd->compute_profile) == 1)
1387 		amdgpu_amdkfd_set_compute_idle(node->adev, false);
1388 }
1389 
kfd_dec_compute_active(struct kfd_node * node)1390 void kfd_dec_compute_active(struct kfd_node *node)
1391 {
1392 	int count = atomic_dec_return(&node->kfd->compute_profile);
1393 
1394 	if (count == 0)
1395 		amdgpu_amdkfd_set_compute_idle(node->adev, true);
1396 	WARN_ONCE(count < 0, "Compute profile ref. count error");
1397 }
1398 
kfd_compute_active(struct kfd_node * node)1399 static bool kfd_compute_active(struct kfd_node *node)
1400 {
1401 	if (atomic_read(&node->kfd->compute_profile))
1402 		return true;
1403 	return false;
1404 }
1405 
kgd2kfd_smi_event_throttle(struct kfd_dev * kfd,uint64_t throttle_bitmask)1406 void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd, uint64_t throttle_bitmask)
1407 {
1408 	/*
1409 	 * TODO: For now, raise the throttling event only on first node.
1410 	 * This will need to change after we are able to determine
1411 	 * which node raised the throttling event.
1412 	 */
1413 	if (kfd && kfd->init_complete)
1414 		kfd_smi_event_update_thermal_throttling(kfd->nodes[0],
1415 							throttle_bitmask);
1416 }
1417 
1418 /* kfd_get_num_sdma_engines returns the number of PCIe optimized SDMA and
1419  * kfd_get_num_xgmi_sdma_engines returns the number of XGMI SDMA.
1420  * When the device has more than two engines, we reserve two for PCIe to enable
1421  * full-duplex and the rest are used as XGMI.
1422  */
kfd_get_num_sdma_engines(struct kfd_node * node)1423 unsigned int kfd_get_num_sdma_engines(struct kfd_node *node)
1424 {
1425 	/* If XGMI is not supported, all SDMA engines are PCIe */
1426 	if (!node->adev->gmc.xgmi.supported)
1427 		return node->adev->sdma.num_instances/(int)node->kfd->num_nodes;
1428 
1429 	return min(node->adev->sdma.num_instances/(int)node->kfd->num_nodes, 2);
1430 }
1431 
kfd_get_num_xgmi_sdma_engines(struct kfd_node * node)1432 unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *node)
1433 {
1434 	/* After reserved for PCIe, the rest of engines are XGMI */
1435 	return node->adev->sdma.num_instances/(int)node->kfd->num_nodes -
1436 		kfd_get_num_sdma_engines(node);
1437 }
1438 
kgd2kfd_check_and_lock_kfd(void)1439 int kgd2kfd_check_and_lock_kfd(void)
1440 {
1441 	mutex_lock(&kfd_processes_mutex);
1442 	if (!hash_empty(kfd_processes_table) || kfd_is_locked()) {
1443 		mutex_unlock(&kfd_processes_mutex);
1444 		return -EBUSY;
1445 	}
1446 
1447 	++kfd_locked;
1448 	mutex_unlock(&kfd_processes_mutex);
1449 
1450 	return 0;
1451 }
1452 
kgd2kfd_unlock_kfd(void)1453 void kgd2kfd_unlock_kfd(void)
1454 {
1455 	mutex_lock(&kfd_processes_mutex);
1456 	--kfd_locked;
1457 	mutex_unlock(&kfd_processes_mutex);
1458 }
1459 
kgd2kfd_start_sched(struct kfd_dev * kfd,uint32_t node_id)1460 int kgd2kfd_start_sched(struct kfd_dev *kfd, uint32_t node_id)
1461 {
1462 	struct kfd_node *node;
1463 	int ret;
1464 
1465 	if (!kfd->init_complete)
1466 		return 0;
1467 
1468 	if (node_id >= kfd->num_nodes) {
1469 		dev_warn(kfd->adev->dev, "Invalid node ID: %u exceeds %u\n",
1470 			 node_id, kfd->num_nodes - 1);
1471 		return -EINVAL;
1472 	}
1473 	node = kfd->nodes[node_id];
1474 
1475 	ret = node->dqm->ops.unhalt(node->dqm);
1476 	if (ret)
1477 		dev_err(kfd_device, "Error in starting scheduler\n");
1478 
1479 	return ret;
1480 }
1481 
kgd2kfd_stop_sched(struct kfd_dev * kfd,uint32_t node_id)1482 int kgd2kfd_stop_sched(struct kfd_dev *kfd, uint32_t node_id)
1483 {
1484 	struct kfd_node *node;
1485 
1486 	if (!kfd->init_complete)
1487 		return 0;
1488 
1489 	if (node_id >= kfd->num_nodes) {
1490 		dev_warn(kfd->adev->dev, "Invalid node ID: %u exceeds %u\n",
1491 			 node_id, kfd->num_nodes - 1);
1492 		return -EINVAL;
1493 	}
1494 
1495 	node = kfd->nodes[node_id];
1496 	return node->dqm->ops.halt(node->dqm);
1497 }
1498 
kgd2kfd_compute_active(struct kfd_dev * kfd,uint32_t node_id)1499 bool kgd2kfd_compute_active(struct kfd_dev *kfd, uint32_t node_id)
1500 {
1501 	struct kfd_node *node;
1502 
1503 	if (!kfd->init_complete)
1504 		return false;
1505 
1506 	if (node_id >= kfd->num_nodes) {
1507 		dev_warn(kfd->adev->dev, "Invalid node ID: %u exceeds %u\n",
1508 			 node_id, kfd->num_nodes - 1);
1509 		return false;
1510 	}
1511 
1512 	node = kfd->nodes[node_id];
1513 
1514 	return kfd_compute_active(node);
1515 }
1516 
1517 /**
1518  * kgd2kfd_vmfault_fast_path() - KFD vm page fault interrupt handling fast path for gmc v9
1519  * @adev: amdgpu device
1520  * @entry: vm fault interrupt vector
1521  * @retry_fault: if this is retry fault
1522  *
1523  * retry fault -
1524  *    with CAM enabled, adev primary ring
1525  *                           |  gmc_v9_0_process_interrupt()
1526  *                      adev soft_ring
1527  *                           |  gmc_v9_0_process_interrupt() worker failed to recover page fault
1528  *                      KFD node ih_fifo
1529  *                           |  KFD interrupt_wq worker
1530  *                      kfd_signal_vm_fault_event
1531  *
1532  *    without CAM,      adev primary ring1
1533  *                           |  gmc_v9_0_process_interrupt worker failed to recvoer page fault
1534  *                      KFD node ih_fifo
1535  *                           |  KFD interrupt_wq worker
1536  *                      kfd_signal_vm_fault_event
1537  *
1538  * no-retry fault -
1539  *                      adev primary ring
1540  *                           |  gmc_v9_0_process_interrupt()
1541  *                      KFD node ih_fifo
1542  *                           |  KFD interrupt_wq worker
1543  *                      kfd_signal_vm_fault_event
1544  *
1545  * fast path - After kfd_signal_vm_fault_event, gmc_v9_0_process_interrupt drop the page fault
1546  *            of same process, don't copy interrupt to KFD node ih_fifo.
1547  *            With gdb debugger enabled, need convert the retry fault to no-retry fault for
1548  *            debugger, cannot use the fast path.
1549  *
1550  * Return:
1551  *   true - use the fast path to handle this fault
1552  *   false - use normal path to handle it
1553  */
kgd2kfd_vmfault_fast_path(struct amdgpu_device * adev,struct amdgpu_iv_entry * entry,bool retry_fault)1554 bool kgd2kfd_vmfault_fast_path(struct amdgpu_device *adev, struct amdgpu_iv_entry *entry,
1555 			       bool retry_fault)
1556 {
1557 	struct kfd_process *p;
1558 	u32 cam_index;
1559 
1560 	if (entry->ih == &adev->irq.ih_soft || entry->ih == &adev->irq.ih1) {
1561 		p = kfd_lookup_process_by_pasid(entry->pasid);
1562 		if (!p)
1563 			return true;
1564 
1565 		if (p->gpu_page_fault && !p->debug_trap_enabled) {
1566 			if (retry_fault && adev->irq.retry_cam_enabled) {
1567 				cam_index = entry->src_data[2] & 0x3ff;
1568 				WDOORBELL32(adev->irq.retry_cam_doorbell_index, cam_index);
1569 			}
1570 
1571 			kfd_unref_process(p);
1572 			return true;
1573 		}
1574 
1575 		/*
1576 		 * This is the first page fault, set flag and then signal user space
1577 		 */
1578 		p->gpu_page_fault = true;
1579 		kfd_unref_process(p);
1580 	}
1581 	return false;
1582 }
1583 
1584 #if defined(CONFIG_DEBUG_FS)
1585 
1586 /* This function will send a package to HIQ to hang the HWS
1587  * which will trigger a GPU reset and bring the HWS back to normal state
1588  */
kfd_debugfs_hang_hws(struct kfd_node * dev)1589 int kfd_debugfs_hang_hws(struct kfd_node *dev)
1590 {
1591 	if (dev->dqm->sched_policy != KFD_SCHED_POLICY_HWS) {
1592 		pr_err("HWS is not enabled");
1593 		return -EINVAL;
1594 	}
1595 
1596 	return dqm_debugfs_hang_hws(dev->dqm);
1597 }
1598 
1599 #endif
1600