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/amd-iommu.h>
24 #include <linux/bsearch.h>
25 #include <linux/pci.h>
26 #include <linux/slab.h>
28 #include "kfd_device_queue_manager.h"
29 #include "kfd_pm4_headers.h"
31 #define MQD_SIZE_ALIGNED 768
33 static const struct kfd_device_info kaveri_device_info = {
34 .asic_family = CHIP_KAVERI,
36 /* max num of queues for KV.TODO should be a dynamic value */
38 .ih_ring_entry_size = 4 * sizeof(uint32_t),
39 .event_interrupt_class = &event_interrupt_class_cik,
40 .num_of_watch_points = 4,
41 .mqd_size_aligned = MQD_SIZE_ALIGNED
44 static const struct kfd_device_info carrizo_device_info = {
45 .asic_family = CHIP_CARRIZO,
47 /* max num of queues for CZ.TODO should be a dynamic value */
49 .ih_ring_entry_size = 4 * sizeof(uint32_t),
50 .event_interrupt_class = &event_interrupt_class_cik,
51 .num_of_watch_points = 4,
52 .mqd_size_aligned = MQD_SIZE_ALIGNED
57 const struct kfd_device_info *device_info;
60 /* Please keep this sorted by increasing device id. */
61 static const struct kfd_deviceid supported_devices[] = {
62 { 0x1304, &kaveri_device_info }, /* Kaveri */
63 { 0x1305, &kaveri_device_info }, /* Kaveri */
64 { 0x1306, &kaveri_device_info }, /* Kaveri */
65 { 0x1307, &kaveri_device_info }, /* Kaveri */
66 { 0x1309, &kaveri_device_info }, /* Kaveri */
67 { 0x130A, &kaveri_device_info }, /* Kaveri */
68 { 0x130B, &kaveri_device_info }, /* Kaveri */
69 { 0x130C, &kaveri_device_info }, /* Kaveri */
70 { 0x130D, &kaveri_device_info }, /* Kaveri */
71 { 0x130E, &kaveri_device_info }, /* Kaveri */
72 { 0x130F, &kaveri_device_info }, /* Kaveri */
73 { 0x1310, &kaveri_device_info }, /* Kaveri */
74 { 0x1311, &kaveri_device_info }, /* Kaveri */
75 { 0x1312, &kaveri_device_info }, /* Kaveri */
76 { 0x1313, &kaveri_device_info }, /* Kaveri */
77 { 0x1315, &kaveri_device_info }, /* Kaveri */
78 { 0x1316, &kaveri_device_info }, /* Kaveri */
79 { 0x1317, &kaveri_device_info }, /* Kaveri */
80 { 0x1318, &kaveri_device_info }, /* Kaveri */
81 { 0x131B, &kaveri_device_info }, /* Kaveri */
82 { 0x131C, &kaveri_device_info }, /* Kaveri */
83 { 0x131D, &kaveri_device_info } /* Kaveri */
86 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
87 unsigned int chunk_size);
88 static void kfd_gtt_sa_fini(struct kfd_dev *kfd);
90 static const struct kfd_device_info *lookup_device_info(unsigned short did)
94 for (i = 0; i < ARRAY_SIZE(supported_devices); i++) {
95 if (supported_devices[i].did == did) {
96 BUG_ON(supported_devices[i].device_info == NULL);
97 return supported_devices[i].device_info;
104 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
105 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g)
109 const struct kfd_device_info *device_info =
110 lookup_device_info(pdev->device);
115 kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
120 kfd->device_info = device_info;
122 kfd->init_complete = false;
125 mutex_init(&kfd->doorbell_mutex);
126 memset(&kfd->doorbell_available_index, 0,
127 sizeof(kfd->doorbell_available_index));
132 static bool device_iommu_pasid_init(struct kfd_dev *kfd)
134 const u32 required_iommu_flags = AMD_IOMMU_DEVICE_FLAG_ATS_SUP |
135 AMD_IOMMU_DEVICE_FLAG_PRI_SUP |
136 AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
138 struct amd_iommu_device_info iommu_info;
139 unsigned int pasid_limit;
142 err = amd_iommu_device_info(kfd->pdev, &iommu_info);
145 "error getting iommu info. is the iommu enabled?\n");
149 if ((iommu_info.flags & required_iommu_flags) != required_iommu_flags) {
150 dev_err(kfd_device, "error required iommu flags ats(%i), pri(%i), pasid(%i)\n",
151 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_ATS_SUP) != 0,
152 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PRI_SUP) != 0,
153 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PASID_SUP) != 0);
157 pasid_limit = min_t(unsigned int,
158 (unsigned int)1 << kfd->device_info->max_pasid_bits,
159 iommu_info.max_pasids);
161 * last pasid is used for kernel queues doorbells
162 * in the future the last pasid might be used for a kernel thread.
164 pasid_limit = min_t(unsigned int,
166 kfd->doorbell_process_limit - 1);
168 err = amd_iommu_init_device(kfd->pdev, pasid_limit);
170 dev_err(kfd_device, "error initializing iommu device\n");
174 if (!kfd_set_pasid_limit(pasid_limit)) {
175 dev_err(kfd_device, "error setting pasid limit\n");
176 amd_iommu_free_device(kfd->pdev);
183 static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid)
185 struct kfd_dev *dev = kfd_device_by_pci_dev(pdev);
188 kfd_unbind_process_from_device(dev, pasid);
192 * This function called by IOMMU driver on PPR failure
194 static int iommu_invalid_ppr_cb(struct pci_dev *pdev, int pasid,
195 unsigned long address, u16 flags)
200 "Invalid PPR device %x:%x.%x pasid %d address 0x%lX flags 0x%X",
201 PCI_BUS_NUM(pdev->devfn),
202 PCI_SLOT(pdev->devfn),
203 PCI_FUNC(pdev->devfn),
208 dev = kfd_device_by_pci_dev(pdev);
211 kfd_signal_iommu_event(dev, pasid, address,
212 flags & PPR_FAULT_WRITE, flags & PPR_FAULT_EXEC);
214 return AMD_IOMMU_INV_PRI_RSP_INVALID;
217 bool kgd2kfd_device_init(struct kfd_dev *kfd,
218 const struct kgd2kfd_shared_resources *gpu_resources)
222 kfd->shared_resources = *gpu_resources;
224 /* calculate max size of mqds needed for queues */
225 size = max_num_of_queues_per_device *
226 kfd->device_info->mqd_size_aligned;
229 * calculate max size of runlist packet.
230 * There can be only 2 packets at once
232 size += (KFD_MAX_NUM_OF_PROCESSES * sizeof(struct pm4_map_process) +
233 max_num_of_queues_per_device *
234 sizeof(struct pm4_map_queues) + sizeof(struct pm4_runlist)) * 2;
236 /* Add size of HIQ & DIQ */
237 size += KFD_KERNEL_QUEUE_SIZE * 2;
239 /* add another 512KB for all other allocations on gart (HPD, fences) */
242 if (kfd->kfd2kgd->init_gtt_mem_allocation(
243 kfd->kgd, size, &kfd->gtt_mem,
244 &kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr)){
246 "Could not allocate %d bytes for device (%x:%x)\n",
247 size, kfd->pdev->vendor, kfd->pdev->device);
252 "Allocated %d bytes on gart for device(%x:%x)\n",
253 size, kfd->pdev->vendor, kfd->pdev->device);
255 /* Initialize GTT sa with 512 byte chunk size */
256 if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
258 "Error initializing gtt sub-allocator\n");
259 goto kfd_gtt_sa_init_error;
262 kfd_doorbell_init(kfd);
264 if (kfd_topology_add_device(kfd) != 0) {
266 "Error adding device (%x:%x) to topology\n",
267 kfd->pdev->vendor, kfd->pdev->device);
268 goto kfd_topology_add_device_error;
271 if (kfd_interrupt_init(kfd)) {
273 "Error initializing interrupts for device (%x:%x)\n",
274 kfd->pdev->vendor, kfd->pdev->device);
275 goto kfd_interrupt_error;
278 if (!device_iommu_pasid_init(kfd)) {
280 "Error initializing iommuv2 for device (%x:%x)\n",
281 kfd->pdev->vendor, kfd->pdev->device);
282 goto device_iommu_pasid_error;
284 amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
285 iommu_pasid_shutdown_callback);
286 amd_iommu_set_invalid_ppr_cb(kfd->pdev, iommu_invalid_ppr_cb);
288 kfd->dqm = device_queue_manager_init(kfd);
291 "Error initializing queue manager for device (%x:%x)\n",
292 kfd->pdev->vendor, kfd->pdev->device);
293 goto device_queue_manager_error;
296 if (kfd->dqm->ops.start(kfd->dqm) != 0) {
298 "Error starting queuen manager for device (%x:%x)\n",
299 kfd->pdev->vendor, kfd->pdev->device);
300 goto dqm_start_error;
305 kfd->init_complete = true;
306 dev_info(kfd_device, "added device (%x:%x)\n", kfd->pdev->vendor,
309 pr_debug("kfd: Starting kfd with the following scheduling policy %d\n",
315 device_queue_manager_uninit(kfd->dqm);
316 device_queue_manager_error:
317 amd_iommu_free_device(kfd->pdev);
318 device_iommu_pasid_error:
319 kfd_interrupt_exit(kfd);
321 kfd_topology_remove_device(kfd);
322 kfd_topology_add_device_error:
323 kfd_gtt_sa_fini(kfd);
324 kfd_gtt_sa_init_error:
325 kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
327 "device (%x:%x) NOT added due to errors\n",
328 kfd->pdev->vendor, kfd->pdev->device);
330 return kfd->init_complete;
333 void kgd2kfd_device_exit(struct kfd_dev *kfd)
335 if (kfd->init_complete) {
336 device_queue_manager_uninit(kfd->dqm);
337 amd_iommu_free_device(kfd->pdev);
338 kfd_interrupt_exit(kfd);
339 kfd_topology_remove_device(kfd);
340 kfd_gtt_sa_fini(kfd);
341 kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
347 void kgd2kfd_suspend(struct kfd_dev *kfd)
351 if (kfd->init_complete) {
352 kfd->dqm->ops.stop(kfd->dqm);
353 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);
354 amd_iommu_set_invalid_ppr_cb(kfd->pdev, NULL);
355 amd_iommu_free_device(kfd->pdev);
359 int kgd2kfd_resume(struct kfd_dev *kfd)
361 unsigned int pasid_limit;
366 pasid_limit = kfd_get_pasid_limit();
368 if (kfd->init_complete) {
369 err = amd_iommu_init_device(kfd->pdev, pasid_limit);
372 amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
373 iommu_pasid_shutdown_callback);
374 amd_iommu_set_invalid_ppr_cb(kfd->pdev, iommu_invalid_ppr_cb);
375 kfd->dqm->ops.start(kfd->dqm);
381 /* This is called directly from KGD at ISR. */
382 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
384 if (!kfd->init_complete)
387 spin_lock(&kfd->interrupt_lock);
389 if (kfd->interrupts_active
390 && interrupt_is_wanted(kfd, ih_ring_entry)
391 && enqueue_ih_ring_entry(kfd, ih_ring_entry))
392 schedule_work(&kfd->interrupt_work);
394 spin_unlock(&kfd->interrupt_lock);
397 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
398 unsigned int chunk_size)
400 unsigned int num_of_bits;
403 BUG_ON(!kfd->gtt_mem);
404 BUG_ON(buf_size < chunk_size);
405 BUG_ON(buf_size == 0);
406 BUG_ON(chunk_size == 0);
408 kfd->gtt_sa_chunk_size = chunk_size;
409 kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
411 num_of_bits = kfd->gtt_sa_num_of_chunks / BITS_PER_BYTE;
412 BUG_ON(num_of_bits == 0);
414 kfd->gtt_sa_bitmap = kzalloc(num_of_bits, GFP_KERNEL);
416 if (!kfd->gtt_sa_bitmap)
419 pr_debug("kfd: gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
420 kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);
422 mutex_init(&kfd->gtt_sa_lock);
428 static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
430 mutex_destroy(&kfd->gtt_sa_lock);
431 kfree(kfd->gtt_sa_bitmap);
434 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
435 unsigned int bit_num,
436 unsigned int chunk_size)
438 return start_addr + bit_num * chunk_size;
441 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
442 unsigned int bit_num,
443 unsigned int chunk_size)
445 return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
448 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
449 struct kfd_mem_obj **mem_obj)
451 unsigned int found, start_search, cur_size;
458 if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
461 *mem_obj = kmalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
462 if ((*mem_obj) == NULL)
465 pr_debug("kfd: allocated mem_obj = %p for size = %d\n", *mem_obj, size);
469 mutex_lock(&kfd->gtt_sa_lock);
471 kfd_gtt_restart_search:
472 /* Find the first chunk that is free */
473 found = find_next_zero_bit(kfd->gtt_sa_bitmap,
474 kfd->gtt_sa_num_of_chunks,
477 pr_debug("kfd: found = %d\n", found);
479 /* If there wasn't any free chunk, bail out */
480 if (found == kfd->gtt_sa_num_of_chunks)
481 goto kfd_gtt_no_free_chunk;
483 /* Update fields of mem_obj */
484 (*mem_obj)->range_start = found;
485 (*mem_obj)->range_end = found;
486 (*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
487 kfd->gtt_start_gpu_addr,
489 kfd->gtt_sa_chunk_size);
490 (*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
491 kfd->gtt_start_cpu_ptr,
493 kfd->gtt_sa_chunk_size);
495 pr_debug("kfd: gpu_addr = %p, cpu_addr = %p\n",
496 (uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
498 /* If we need only one chunk, mark it as allocated and get out */
499 if (size <= kfd->gtt_sa_chunk_size) {
500 pr_debug("kfd: single bit\n");
501 set_bit(found, kfd->gtt_sa_bitmap);
505 /* Otherwise, try to see if we have enough contiguous chunks */
506 cur_size = size - kfd->gtt_sa_chunk_size;
508 (*mem_obj)->range_end =
509 find_next_zero_bit(kfd->gtt_sa_bitmap,
510 kfd->gtt_sa_num_of_chunks, ++found);
512 * If next free chunk is not contiguous than we need to
513 * restart our search from the last free chunk we found (which
514 * wasn't contiguous to the previous ones
516 if ((*mem_obj)->range_end != found) {
517 start_search = found;
518 goto kfd_gtt_restart_search;
522 * If we reached end of buffer, bail out with error
524 if (found == kfd->gtt_sa_num_of_chunks)
525 goto kfd_gtt_no_free_chunk;
527 /* Check if we don't need another chunk */
528 if (cur_size <= kfd->gtt_sa_chunk_size)
531 cur_size -= kfd->gtt_sa_chunk_size;
533 } while (cur_size > 0);
535 pr_debug("kfd: range_start = %d, range_end = %d\n",
536 (*mem_obj)->range_start, (*mem_obj)->range_end);
538 /* Mark the chunks as allocated */
539 for (found = (*mem_obj)->range_start;
540 found <= (*mem_obj)->range_end;
542 set_bit(found, kfd->gtt_sa_bitmap);
545 mutex_unlock(&kfd->gtt_sa_lock);
548 kfd_gtt_no_free_chunk:
549 pr_debug("kfd: allocation failed with mem_obj = %p\n", mem_obj);
550 mutex_unlock(&kfd->gtt_sa_lock);
555 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj)
561 /* Act like kfree when trying to free a NULL object */
565 pr_debug("kfd: free mem_obj = %p, range_start = %d, range_end = %d\n",
566 mem_obj, mem_obj->range_start, mem_obj->range_end);
568 mutex_lock(&kfd->gtt_sa_lock);
570 /* Mark the chunks as free */
571 for (bit = mem_obj->range_start;
572 bit <= mem_obj->range_end;
574 clear_bit(bit, kfd->gtt_sa_bitmap);
576 mutex_unlock(&kfd->gtt_sa_lock);