2 * Copyright 2014 Advanced Micro Devices, Inc.
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:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
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.
23 #include <linux/bsearch.h>
24 #include <linux/pci.h>
25 #include <linux/slab.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"
32 #define MQD_SIZE_ALIGNED 768
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.
39 static atomic_t kfd_locked = ATOMIC_INIT(0);
41 #ifdef KFD_SUPPORT_IOMMU_V2
42 static const struct kfd_device_info kaveri_device_info = {
43 .asic_family = CHIP_KAVERI,
45 /* max num of queues for KV.TODO should be a dynamic value */
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,
59 static const struct kfd_device_info carrizo_device_info = {
60 .asic_family = CHIP_CARRIZO,
62 /* max num of queues for CZ.TODO should be a dynamic value */
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,
76 static const struct kfd_device_info raven_device_info = {
77 .asic_family = CHIP_RAVEN,
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,
93 static const struct kfd_device_info hawaii_device_info = {
94 .asic_family = CHIP_HAWAII,
96 /* max num of queues for KV.TODO should be a dynamic value */
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,
110 static const struct kfd_device_info tonga_device_info = {
111 .asic_family = CHIP_TONGA,
112 .max_pasid_bits = 16,
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,
126 static const struct kfd_device_info fiji_device_info = {
127 .asic_family = CHIP_FIJI,
128 .max_pasid_bits = 16,
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,
142 static const struct kfd_device_info fiji_vf_device_info = {
143 .asic_family = CHIP_FIJI,
144 .max_pasid_bits = 16,
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,
159 static const struct kfd_device_info polaris10_device_info = {
160 .asic_family = CHIP_POLARIS10,
161 .max_pasid_bits = 16,
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,
175 static const struct kfd_device_info polaris10_vf_device_info = {
176 .asic_family = CHIP_POLARIS10,
177 .max_pasid_bits = 16,
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,
191 static const struct kfd_device_info polaris11_device_info = {
192 .asic_family = CHIP_POLARIS11,
193 .max_pasid_bits = 16,
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,
207 static const struct kfd_device_info vega10_device_info = {
208 .asic_family = CHIP_VEGA10,
209 .max_pasid_bits = 16,
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,
223 static const struct kfd_device_info vega10_vf_device_info = {
224 .asic_family = CHIP_VEGA10,
225 .max_pasid_bits = 16,
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,
239 static const struct kfd_device_info vega20_device_info = {
240 .asic_family = CHIP_VEGA20,
241 .max_pasid_bits = 16,
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,
255 struct kfd_deviceid {
257 const struct kfd_device_info *device_info;
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 */
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 */
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);
354 static int kfd_resume(struct kfd_dev *kfd);
356 static const struct kfd_device_info *lookup_device_info(unsigned short did)
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;
367 dev_warn(kfd_device, "DID %04x is missing in supported_devices\n",
373 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
374 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g)
378 const struct kfd_device_info *device_info =
379 lookup_device_info(pdev->device);
382 dev_err(kfd_device, "kgd2kfd_probe failed\n");
386 kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
390 /* Allow BIF to recode atomics to PCIe 3.0 AtomicOps.
391 * 32 and 64-bit requests are possible and must be
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) {
399 "skipped device %x:%x, PCI rejects atomics\n",
400 pdev->vendor, pdev->device);
404 kfd->pci_atomic_requested = true;
407 kfd->device_info = device_info;
409 kfd->init_complete = false;
412 mutex_init(&kfd->doorbell_mutex);
413 memset(&kfd->doorbell_available_index, 0,
414 sizeof(kfd->doorbell_available_index));
419 static void kfd_cwsr_init(struct kfd_dev *kfd)
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);
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);
432 kfd->cwsr_enabled = true;
436 bool kgd2kfd_device_init(struct kfd_dev *kfd,
437 const struct kgd2kfd_shared_resources *gpu_resources)
441 kfd->mec_fw_version = kfd->kfd2kgd->get_fw_version(kfd->kgd,
443 kfd->sdma_fw_version = kfd->kfd2kgd->get_fw_version(kfd->kgd,
445 kfd->shared_resources = *gpu_resources;
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;
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)) {
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;
461 kfd->max_proc_per_quantum = hws_max_conc_proc;
463 /* calculate max size of mqds needed for queues */
464 size = max_num_of_queues_per_device *
465 kfd->device_info->mqd_size_aligned;
468 * calculate max size of runlist packet.
469 * There can be only 2 packets at once
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;
475 /* Add size of HIQ & DIQ */
476 size += KFD_KERNEL_QUEUE_SIZE * 2;
478 /* add another 512KB for all other allocations on gart (HPD, fences) */
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,
485 dev_err(kfd_device, "Could not allocate %d bytes\n", size);
489 dev_info(kfd_device, "Allocated %d bytes on gart\n", size);
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;
497 if (kfd_doorbell_init(kfd)) {
499 "Error initializing doorbell aperture\n");
500 goto kfd_doorbell_error;
503 if (kfd->kfd2kgd->get_hive_id)
504 kfd->hive_id = kfd->kfd2kgd->get_hive_id(kfd->kgd);
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;
511 if (kfd_interrupt_init(kfd)) {
512 dev_err(kfd_device, "Error initializing interrupts\n");
513 goto kfd_interrupt_error;
516 kfd->dqm = device_queue_manager_init(kfd);
518 dev_err(kfd_device, "Error initializing queue manager\n");
519 goto device_queue_manager_error;
522 if (kfd_iommu_device_init(kfd)) {
523 dev_err(kfd_device, "Error initializing iommuv2\n");
524 goto device_iommu_error;
530 goto kfd_resume_error;
534 kfd->init_complete = true;
535 dev_info(kfd_device, "added device %x:%x\n", kfd->pdev->vendor,
538 pr_debug("Starting kfd with the following scheduling policy %d\n",
539 kfd->dqm->sched_policy);
545 device_queue_manager_uninit(kfd->dqm);
546 device_queue_manager_error:
547 kfd_interrupt_exit(kfd);
549 kfd_topology_remove_device(kfd);
550 kfd_topology_add_device_error:
551 kfd_doorbell_fini(kfd);
553 kfd_gtt_sa_fini(kfd);
554 kfd_gtt_sa_init_error:
555 kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
557 "device %x:%x NOT added due to errors\n",
558 kfd->pdev->vendor, kfd->pdev->device);
560 return kfd->init_complete;
563 void kgd2kfd_device_exit(struct kfd_dev *kfd)
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);
578 int kgd2kfd_pre_reset(struct kfd_dev *kfd)
580 if (!kfd->init_complete)
582 kgd2kfd_suspend(kfd);
584 /* hold dqm->lock to prevent further execution*/
587 kfd_signal_reset_event(kfd);
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.
597 int kgd2kfd_post_reset(struct kfd_dev *kfd)
601 if (!kfd->init_complete)
604 dqm_unlock(kfd->dqm);
606 ret = kfd_resume(kfd);
609 count = atomic_dec_return(&kfd_locked);
610 WARN_ONCE(count != 0, "KFD reset ref. error");
614 bool kfd_is_locked(void)
616 return (atomic_read(&kfd_locked) > 0);
619 void kgd2kfd_suspend(struct kfd_dev *kfd)
621 if (!kfd->init_complete)
624 /* For first KFD device suspend all the KFD processes */
625 if (atomic_inc_return(&kfd_locked) == 1)
626 kfd_suspend_all_processes();
628 kfd->dqm->ops.stop(kfd->dqm);
630 kfd_iommu_suspend(kfd);
633 int kgd2kfd_resume(struct kfd_dev *kfd)
637 if (!kfd->init_complete)
640 ret = kfd_resume(kfd);
644 count = atomic_dec_return(&kfd_locked);
645 WARN_ONCE(count < 0, "KFD suspend / resume ref. error");
647 ret = kfd_resume_all_processes();
652 static int kfd_resume(struct kfd_dev *kfd)
656 err = kfd_iommu_resume(kfd);
659 "Failed to resume IOMMU for device %x:%x\n",
660 kfd->pdev->vendor, kfd->pdev->device);
664 err = kfd->dqm->ops.start(kfd->dqm);
667 "Error starting queue manager for device %x:%x\n",
668 kfd->pdev->vendor, kfd->pdev->device);
669 goto dqm_start_error;
675 kfd_iommu_suspend(kfd);
679 /* This is called directly from KGD at ISR. */
680 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
682 uint32_t patched_ihre[KFD_MAX_RING_ENTRY_SIZE];
683 bool is_patched = false;
685 if (!kfd->init_complete)
688 if (kfd->device_info->ih_ring_entry_size > sizeof(patched_ihre)) {
689 dev_err_once(kfd_device, "Ring entry too small\n");
693 spin_lock(&kfd->interrupt_lock);
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);
702 spin_unlock(&kfd->interrupt_lock);
705 int kgd2kfd_quiesce_mm(struct mm_struct *mm)
707 struct kfd_process *p;
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.
714 p = kfd_lookup_process_by_mm(mm);
718 r = kfd_process_evict_queues(p);
720 kfd_unref_process(p);
724 int kgd2kfd_resume_mm(struct mm_struct *mm)
726 struct kfd_process *p;
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.
733 p = kfd_lookup_process_by_mm(mm);
737 r = kfd_process_restore_queues(p);
739 kfd_unref_process(p);
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
747 * @mm: mm_struct that identifies the specified KFD process
748 * @fence: eviction fence attached to KFD process BOs
751 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
752 struct dma_fence *fence)
754 struct kfd_process *p;
755 unsigned long active_time;
756 unsigned long delay_jiffies = msecs_to_jiffies(PROCESS_ACTIVE_TIME_MS);
761 if (dma_fence_is_signaled(fence))
764 p = kfd_lookup_process_by_mm(mm);
768 if (fence->seqno == p->last_eviction_seqno)
771 p->last_eviction_seqno = fence->seqno;
773 /* Avoid KFD process starvation. Wait for at least
774 * PROCESS_ACTIVE_TIME_MS before evicting the process again
776 active_time = get_jiffies_64() - p->last_restore_timestamp;
777 if (delay_jiffies > active_time)
778 delay_jiffies -= active_time;
782 /* During process initialization eviction_work.dwork is initialized
783 * to kfd_evict_bo_worker
785 schedule_delayed_work(&p->eviction_work, delay_jiffies);
787 kfd_unref_process(p);
791 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
792 unsigned int chunk_size)
794 unsigned int num_of_longs;
796 if (WARN_ON(buf_size < chunk_size))
798 if (WARN_ON(buf_size == 0))
800 if (WARN_ON(chunk_size == 0))
803 kfd->gtt_sa_chunk_size = chunk_size;
804 kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
806 num_of_longs = (kfd->gtt_sa_num_of_chunks + BITS_PER_LONG - 1) /
809 kfd->gtt_sa_bitmap = kcalloc(num_of_longs, sizeof(long), GFP_KERNEL);
811 if (!kfd->gtt_sa_bitmap)
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);
817 mutex_init(&kfd->gtt_sa_lock);
823 static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
825 mutex_destroy(&kfd->gtt_sa_lock);
826 kfree(kfd->gtt_sa_bitmap);
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)
833 return start_addr + bit_num * chunk_size;
836 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
837 unsigned int bit_num,
838 unsigned int chunk_size)
840 return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
843 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
844 struct kfd_mem_obj **mem_obj)
846 unsigned int found, start_search, cur_size;
851 if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
854 *mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
858 pr_debug("Allocated mem_obj = %p for size = %d\n", *mem_obj, size);
862 mutex_lock(&kfd->gtt_sa_lock);
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,
870 pr_debug("Found = %d\n", found);
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;
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,
882 kfd->gtt_sa_chunk_size);
883 (*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
884 kfd->gtt_start_cpu_ptr,
886 kfd->gtt_sa_chunk_size);
888 pr_debug("gpu_addr = %p, cpu_addr = %p\n",
889 (uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
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);
898 /* Otherwise, try to see if we have enough contiguous chunks */
899 cur_size = size - kfd->gtt_sa_chunk_size;
901 (*mem_obj)->range_end =
902 find_next_zero_bit(kfd->gtt_sa_bitmap,
903 kfd->gtt_sa_num_of_chunks, ++found);
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
909 if ((*mem_obj)->range_end != found) {
910 start_search = found;
911 goto kfd_gtt_restart_search;
915 * If we reached end of buffer, bail out with error
917 if (found == kfd->gtt_sa_num_of_chunks)
918 goto kfd_gtt_no_free_chunk;
920 /* Check if we don't need another chunk */
921 if (cur_size <= kfd->gtt_sa_chunk_size)
924 cur_size -= kfd->gtt_sa_chunk_size;
926 } while (cur_size > 0);
928 pr_debug("range_start = %d, range_end = %d\n",
929 (*mem_obj)->range_start, (*mem_obj)->range_end);
931 /* Mark the chunks as allocated */
932 for (found = (*mem_obj)->range_start;
933 found <= (*mem_obj)->range_end;
935 set_bit(found, kfd->gtt_sa_bitmap);
938 mutex_unlock(&kfd->gtt_sa_lock);
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);
948 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj)
952 /* Act like kfree when trying to free a NULL object */
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);
959 mutex_lock(&kfd->gtt_sa_lock);
961 /* Mark the chunks as free */
962 for (bit = mem_obj->range_start;
963 bit <= mem_obj->range_end;
965 clear_bit(bit, kfd->gtt_sa_bitmap);
967 mutex_unlock(&kfd->gtt_sa_lock);
973 #if defined(CONFIG_DEBUG_FS)
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
978 int kfd_debugfs_hang_hws(struct kfd_dev *dev)
982 if (dev->dqm->sched_policy != KFD_SCHED_POLICY_HWS) {
983 pr_err("HWS is not enabled");
987 r = pm_debugfs_hang_hws(&dev->dqm->packets);
989 r = dqm_debugfs_execute_queues(dev->dqm);