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 */
84 { 0x9870, &carrizo_device_info }, /* Carrizo */
85 { 0x9874, &carrizo_device_info }, /* Carrizo */
86 { 0x9875, &carrizo_device_info }, /* Carrizo */
87 { 0x9876, &carrizo_device_info }, /* Carrizo */
88 { 0x9877, &carrizo_device_info } /* Carrizo */
91 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
92 unsigned int chunk_size);
93 static void kfd_gtt_sa_fini(struct kfd_dev *kfd);
95 static const struct kfd_device_info *lookup_device_info(unsigned short did)
99 for (i = 0; i < ARRAY_SIZE(supported_devices); i++) {
100 if (supported_devices[i].did == did) {
101 BUG_ON(supported_devices[i].device_info == NULL);
102 return supported_devices[i].device_info;
109 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
110 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g)
114 const struct kfd_device_info *device_info =
115 lookup_device_info(pdev->device);
120 kfd = kzalloc(sizeof(*kfd), GFP_KERNEL);
125 kfd->device_info = device_info;
127 kfd->init_complete = false;
130 mutex_init(&kfd->doorbell_mutex);
131 memset(&kfd->doorbell_available_index, 0,
132 sizeof(kfd->doorbell_available_index));
137 static bool device_iommu_pasid_init(struct kfd_dev *kfd)
139 const u32 required_iommu_flags = AMD_IOMMU_DEVICE_FLAG_ATS_SUP |
140 AMD_IOMMU_DEVICE_FLAG_PRI_SUP |
141 AMD_IOMMU_DEVICE_FLAG_PASID_SUP;
143 struct amd_iommu_device_info iommu_info;
144 unsigned int pasid_limit;
147 err = amd_iommu_device_info(kfd->pdev, &iommu_info);
150 "error getting iommu info. is the iommu enabled?\n");
154 if ((iommu_info.flags & required_iommu_flags) != required_iommu_flags) {
155 dev_err(kfd_device, "error required iommu flags ats(%i), pri(%i), pasid(%i)\n",
156 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_ATS_SUP) != 0,
157 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PRI_SUP) != 0,
158 (iommu_info.flags & AMD_IOMMU_DEVICE_FLAG_PASID_SUP) != 0);
162 pasid_limit = min_t(unsigned int,
163 (unsigned int)1 << kfd->device_info->max_pasid_bits,
164 iommu_info.max_pasids);
166 * last pasid is used for kernel queues doorbells
167 * in the future the last pasid might be used for a kernel thread.
169 pasid_limit = min_t(unsigned int,
171 kfd->doorbell_process_limit - 1);
173 err = amd_iommu_init_device(kfd->pdev, pasid_limit);
175 dev_err(kfd_device, "error initializing iommu device\n");
179 if (!kfd_set_pasid_limit(pasid_limit)) {
180 dev_err(kfd_device, "error setting pasid limit\n");
181 amd_iommu_free_device(kfd->pdev);
188 static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid)
190 struct kfd_dev *dev = kfd_device_by_pci_dev(pdev);
193 kfd_unbind_process_from_device(dev, pasid);
197 * This function called by IOMMU driver on PPR failure
199 static int iommu_invalid_ppr_cb(struct pci_dev *pdev, int pasid,
200 unsigned long address, u16 flags)
205 "Invalid PPR device %x:%x.%x pasid %d address 0x%lX flags 0x%X",
206 PCI_BUS_NUM(pdev->devfn),
207 PCI_SLOT(pdev->devfn),
208 PCI_FUNC(pdev->devfn),
213 dev = kfd_device_by_pci_dev(pdev);
216 kfd_signal_iommu_event(dev, pasid, address,
217 flags & PPR_FAULT_WRITE, flags & PPR_FAULT_EXEC);
219 return AMD_IOMMU_INV_PRI_RSP_INVALID;
222 bool kgd2kfd_device_init(struct kfd_dev *kfd,
223 const struct kgd2kfd_shared_resources *gpu_resources)
227 kfd->shared_resources = *gpu_resources;
229 /* calculate max size of mqds needed for queues */
230 size = max_num_of_queues_per_device *
231 kfd->device_info->mqd_size_aligned;
234 * calculate max size of runlist packet.
235 * There can be only 2 packets at once
237 size += (KFD_MAX_NUM_OF_PROCESSES * sizeof(struct pm4_map_process) +
238 max_num_of_queues_per_device *
239 sizeof(struct pm4_map_queues) + sizeof(struct pm4_runlist)) * 2;
241 /* Add size of HIQ & DIQ */
242 size += KFD_KERNEL_QUEUE_SIZE * 2;
244 /* add another 512KB for all other allocations on gart (HPD, fences) */
247 if (kfd->kfd2kgd->init_gtt_mem_allocation(
248 kfd->kgd, size, &kfd->gtt_mem,
249 &kfd->gtt_start_gpu_addr, &kfd->gtt_start_cpu_ptr)){
251 "Could not allocate %d bytes for device (%x:%x)\n",
252 size, kfd->pdev->vendor, kfd->pdev->device);
257 "Allocated %d bytes on gart for device(%x:%x)\n",
258 size, kfd->pdev->vendor, kfd->pdev->device);
260 /* Initialize GTT sa with 512 byte chunk size */
261 if (kfd_gtt_sa_init(kfd, size, 512) != 0) {
263 "Error initializing gtt sub-allocator\n");
264 goto kfd_gtt_sa_init_error;
267 kfd_doorbell_init(kfd);
269 if (kfd_topology_add_device(kfd) != 0) {
271 "Error adding device (%x:%x) to topology\n",
272 kfd->pdev->vendor, kfd->pdev->device);
273 goto kfd_topology_add_device_error;
276 if (kfd_interrupt_init(kfd)) {
278 "Error initializing interrupts for device (%x:%x)\n",
279 kfd->pdev->vendor, kfd->pdev->device);
280 goto kfd_interrupt_error;
283 if (!device_iommu_pasid_init(kfd)) {
285 "Error initializing iommuv2 for device (%x:%x)\n",
286 kfd->pdev->vendor, kfd->pdev->device);
287 goto device_iommu_pasid_error;
289 amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
290 iommu_pasid_shutdown_callback);
291 amd_iommu_set_invalid_ppr_cb(kfd->pdev, iommu_invalid_ppr_cb);
293 kfd->dqm = device_queue_manager_init(kfd);
296 "Error initializing queue manager for device (%x:%x)\n",
297 kfd->pdev->vendor, kfd->pdev->device);
298 goto device_queue_manager_error;
301 if (kfd->dqm->ops.start(kfd->dqm) != 0) {
303 "Error starting queuen manager for device (%x:%x)\n",
304 kfd->pdev->vendor, kfd->pdev->device);
305 goto dqm_start_error;
310 kfd->init_complete = true;
311 dev_info(kfd_device, "added device (%x:%x)\n", kfd->pdev->vendor,
314 pr_debug("kfd: Starting kfd with the following scheduling policy %d\n",
320 device_queue_manager_uninit(kfd->dqm);
321 device_queue_manager_error:
322 amd_iommu_free_device(kfd->pdev);
323 device_iommu_pasid_error:
324 kfd_interrupt_exit(kfd);
326 kfd_topology_remove_device(kfd);
327 kfd_topology_add_device_error:
328 kfd_gtt_sa_fini(kfd);
329 kfd_gtt_sa_init_error:
330 kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
332 "device (%x:%x) NOT added due to errors\n",
333 kfd->pdev->vendor, kfd->pdev->device);
335 return kfd->init_complete;
338 void kgd2kfd_device_exit(struct kfd_dev *kfd)
340 if (kfd->init_complete) {
341 device_queue_manager_uninit(kfd->dqm);
342 amd_iommu_free_device(kfd->pdev);
343 kfd_interrupt_exit(kfd);
344 kfd_topology_remove_device(kfd);
345 kfd_gtt_sa_fini(kfd);
346 kfd->kfd2kgd->free_gtt_mem(kfd->kgd, kfd->gtt_mem);
352 void kgd2kfd_suspend(struct kfd_dev *kfd)
356 if (kfd->init_complete) {
357 kfd->dqm->ops.stop(kfd->dqm);
358 amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);
359 amd_iommu_set_invalid_ppr_cb(kfd->pdev, NULL);
360 amd_iommu_free_device(kfd->pdev);
364 int kgd2kfd_resume(struct kfd_dev *kfd)
366 unsigned int pasid_limit;
371 pasid_limit = kfd_get_pasid_limit();
373 if (kfd->init_complete) {
374 err = amd_iommu_init_device(kfd->pdev, pasid_limit);
377 amd_iommu_set_invalidate_ctx_cb(kfd->pdev,
378 iommu_pasid_shutdown_callback);
379 amd_iommu_set_invalid_ppr_cb(kfd->pdev, iommu_invalid_ppr_cb);
380 kfd->dqm->ops.start(kfd->dqm);
386 /* This is called directly from KGD at ISR. */
387 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry)
389 if (!kfd->init_complete)
392 spin_lock(&kfd->interrupt_lock);
394 if (kfd->interrupts_active
395 && interrupt_is_wanted(kfd, ih_ring_entry)
396 && enqueue_ih_ring_entry(kfd, ih_ring_entry))
397 schedule_work(&kfd->interrupt_work);
399 spin_unlock(&kfd->interrupt_lock);
402 static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
403 unsigned int chunk_size)
405 unsigned int num_of_bits;
408 BUG_ON(!kfd->gtt_mem);
409 BUG_ON(buf_size < chunk_size);
410 BUG_ON(buf_size == 0);
411 BUG_ON(chunk_size == 0);
413 kfd->gtt_sa_chunk_size = chunk_size;
414 kfd->gtt_sa_num_of_chunks = buf_size / chunk_size;
416 num_of_bits = kfd->gtt_sa_num_of_chunks / BITS_PER_BYTE;
417 BUG_ON(num_of_bits == 0);
419 kfd->gtt_sa_bitmap = kzalloc(num_of_bits, GFP_KERNEL);
421 if (!kfd->gtt_sa_bitmap)
424 pr_debug("kfd: gtt_sa_num_of_chunks = %d, gtt_sa_bitmap = %p\n",
425 kfd->gtt_sa_num_of_chunks, kfd->gtt_sa_bitmap);
427 mutex_init(&kfd->gtt_sa_lock);
433 static void kfd_gtt_sa_fini(struct kfd_dev *kfd)
435 mutex_destroy(&kfd->gtt_sa_lock);
436 kfree(kfd->gtt_sa_bitmap);
439 static inline uint64_t kfd_gtt_sa_calc_gpu_addr(uint64_t start_addr,
440 unsigned int bit_num,
441 unsigned int chunk_size)
443 return start_addr + bit_num * chunk_size;
446 static inline uint32_t *kfd_gtt_sa_calc_cpu_addr(void *start_addr,
447 unsigned int bit_num,
448 unsigned int chunk_size)
450 return (uint32_t *) ((uint64_t) start_addr + bit_num * chunk_size);
453 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
454 struct kfd_mem_obj **mem_obj)
456 unsigned int found, start_search, cur_size;
463 if (size > kfd->gtt_sa_num_of_chunks * kfd->gtt_sa_chunk_size)
466 *mem_obj = kmalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
467 if ((*mem_obj) == NULL)
470 pr_debug("kfd: allocated mem_obj = %p for size = %d\n", *mem_obj, size);
474 mutex_lock(&kfd->gtt_sa_lock);
476 kfd_gtt_restart_search:
477 /* Find the first chunk that is free */
478 found = find_next_zero_bit(kfd->gtt_sa_bitmap,
479 kfd->gtt_sa_num_of_chunks,
482 pr_debug("kfd: found = %d\n", found);
484 /* If there wasn't any free chunk, bail out */
485 if (found == kfd->gtt_sa_num_of_chunks)
486 goto kfd_gtt_no_free_chunk;
488 /* Update fields of mem_obj */
489 (*mem_obj)->range_start = found;
490 (*mem_obj)->range_end = found;
491 (*mem_obj)->gpu_addr = kfd_gtt_sa_calc_gpu_addr(
492 kfd->gtt_start_gpu_addr,
494 kfd->gtt_sa_chunk_size);
495 (*mem_obj)->cpu_ptr = kfd_gtt_sa_calc_cpu_addr(
496 kfd->gtt_start_cpu_ptr,
498 kfd->gtt_sa_chunk_size);
500 pr_debug("kfd: gpu_addr = %p, cpu_addr = %p\n",
501 (uint64_t *) (*mem_obj)->gpu_addr, (*mem_obj)->cpu_ptr);
503 /* If we need only one chunk, mark it as allocated and get out */
504 if (size <= kfd->gtt_sa_chunk_size) {
505 pr_debug("kfd: single bit\n");
506 set_bit(found, kfd->gtt_sa_bitmap);
510 /* Otherwise, try to see if we have enough contiguous chunks */
511 cur_size = size - kfd->gtt_sa_chunk_size;
513 (*mem_obj)->range_end =
514 find_next_zero_bit(kfd->gtt_sa_bitmap,
515 kfd->gtt_sa_num_of_chunks, ++found);
517 * If next free chunk is not contiguous than we need to
518 * restart our search from the last free chunk we found (which
519 * wasn't contiguous to the previous ones
521 if ((*mem_obj)->range_end != found) {
522 start_search = found;
523 goto kfd_gtt_restart_search;
527 * If we reached end of buffer, bail out with error
529 if (found == kfd->gtt_sa_num_of_chunks)
530 goto kfd_gtt_no_free_chunk;
532 /* Check if we don't need another chunk */
533 if (cur_size <= kfd->gtt_sa_chunk_size)
536 cur_size -= kfd->gtt_sa_chunk_size;
538 } while (cur_size > 0);
540 pr_debug("kfd: range_start = %d, range_end = %d\n",
541 (*mem_obj)->range_start, (*mem_obj)->range_end);
543 /* Mark the chunks as allocated */
544 for (found = (*mem_obj)->range_start;
545 found <= (*mem_obj)->range_end;
547 set_bit(found, kfd->gtt_sa_bitmap);
550 mutex_unlock(&kfd->gtt_sa_lock);
553 kfd_gtt_no_free_chunk:
554 pr_debug("kfd: allocation failed with mem_obj = %p\n", mem_obj);
555 mutex_unlock(&kfd->gtt_sa_lock);
560 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj)
566 /* Act like kfree when trying to free a NULL object */
570 pr_debug("kfd: free mem_obj = %p, range_start = %d, range_end = %d\n",
571 mem_obj, mem_obj->range_start, mem_obj->range_end);
573 mutex_lock(&kfd->gtt_sa_lock);
575 /* Mark the chunks as free */
576 for (bit = mem_obj->range_start;
577 bit <= mem_obj->range_end;
579 clear_bit(bit, kfd->gtt_sa_bitmap);
581 mutex_unlock(&kfd->gtt_sa_lock);