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 #ifndef KFD_PRIV_H_INCLUDED
24 #define KFD_PRIV_H_INCLUDED
26 #include <linux/hashtable.h>
27 #include <linux/mmu_notifier.h>
28 #include <linux/mutex.h>
29 #include <linux/types.h>
30 #include <linux/atomic.h>
31 #include <linux/workqueue.h>
32 #include <linux/spinlock.h>
33 #include <linux/kfd_ioctl.h>
34 #include <linux/idr.h>
35 #include <linux/kfifo.h>
36 #include <linux/seq_file.h>
37 #include <linux/kref.h>
38 #include <kgd_kfd_interface.h>
40 #include "amd_shared.h"
42 #define KFD_MAX_RING_ENTRY_SIZE 8
44 #define KFD_SYSFS_FILE_MODE 0444
46 /* GPU ID hash width in bits */
47 #define KFD_GPU_ID_HASH_WIDTH 16
49 /* Use upper bits of mmap offset to store KFD driver specific information.
50 * BITS[63:62] - Encode MMAP type
51 * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to
52 * BITS[45:0] - MMAP offset value
54 * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these
55 * defines are w.r.t to PAGE_SIZE
57 #define KFD_MMAP_TYPE_SHIFT (62 - PAGE_SHIFT)
58 #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT)
59 #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT)
60 #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT)
61 #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT)
63 #define KFD_MMAP_GPU_ID_SHIFT (46 - PAGE_SHIFT)
64 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \
65 << KFD_MMAP_GPU_ID_SHIFT)
66 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\
67 & KFD_MMAP_GPU_ID_MASK)
68 #define KFD_MMAP_GPU_ID_GET(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \
69 >> KFD_MMAP_GPU_ID_SHIFT)
71 #define KFD_MMAP_OFFSET_VALUE_MASK (0x3FFFFFFFFFFFULL >> PAGE_SHIFT)
72 #define KFD_MMAP_OFFSET_VALUE_GET(offset) (offset & KFD_MMAP_OFFSET_VALUE_MASK)
75 * When working with cp scheduler we should assign the HIQ manually or via
76 * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot
77 * definitions for Kaveri. In Kaveri only the first ME queues participates
78 * in the cp scheduling taking that in mind we set the HIQ slot in the
81 #define KFD_CIK_HIQ_PIPE 4
82 #define KFD_CIK_HIQ_QUEUE 0
84 /* Macro for allocating structures */
85 #define kfd_alloc_struct(ptr_to_struct) \
86 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
88 #define KFD_MAX_NUM_OF_PROCESSES 512
89 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
92 * Size of the per-process TBA+TMA buffer: 2 pages
94 * The first page is the TBA used for the CWSR ISA code. The second
95 * page is used as TMA for daisy changing a user-mode trap handler.
97 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
98 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE
101 * Kernel module parameter to specify maximum number of supported queues per
104 extern int max_num_of_queues_per_device;
106 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \
107 (KFD_MAX_NUM_OF_PROCESSES * \
108 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
110 #define KFD_KERNEL_QUEUE_SIZE 2048
112 /* Kernel module parameter to specify the scheduling policy */
113 extern int sched_policy;
116 * Kernel module parameter to specify the maximum process
117 * number per HW scheduler
119 extern int hws_max_conc_proc;
121 extern int cwsr_enable;
124 * Kernel module parameter to specify whether to send sigterm to HSA process on
125 * unhandled exception
127 extern int send_sigterm;
130 * This kernel module is used to simulate large bar machine on non-large bar
133 extern int debug_largebar;
136 * Ignore CRAT table during KFD initialization, can be used to work around
137 * broken CRAT tables on some AMD systems
139 extern int ignore_crat;
142 * Set sh_mem_config.retry_disable on Vega10
147 * Halt if HWS hang is detected
149 extern int halt_if_hws_hang;
152 cache_policy_coherent,
153 cache_policy_noncoherent
156 #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10)
158 struct kfd_event_interrupt_class {
159 bool (*interrupt_isr)(struct kfd_dev *dev,
160 const uint32_t *ih_ring_entry, uint32_t *patched_ihre,
162 void (*interrupt_wq)(struct kfd_dev *dev,
163 const uint32_t *ih_ring_entry);
166 struct kfd_device_info {
167 enum amd_asic_type asic_family;
168 const struct kfd_event_interrupt_class *event_interrupt_class;
169 unsigned int max_pasid_bits;
170 unsigned int max_no_of_hqd;
171 unsigned int doorbell_size;
172 size_t ih_ring_entry_size;
173 uint8_t num_of_watch_points;
174 uint16_t mqd_size_aligned;
176 bool needs_iommu_device;
177 bool needs_pci_atomics;
178 unsigned int num_sdma_engines;
182 uint32_t range_start;
189 struct kfd_vmid_info {
190 uint32_t first_vmid_kfd;
191 uint32_t last_vmid_kfd;
192 uint32_t vmid_num_kfd;
198 const struct kfd_device_info *device_info;
199 struct pci_dev *pdev;
201 unsigned int id; /* topology stub index */
203 phys_addr_t doorbell_base; /* Start of actual doorbells used by
204 * KFD. It is aligned for mapping
207 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell
208 * to HW doorbell, GFX reserved some
211 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
212 * page used by kernel queue
215 struct kgd2kfd_shared_resources shared_resources;
216 struct kfd_vmid_info vm_info;
218 const struct kfd2kgd_calls *kfd2kgd;
219 struct mutex doorbell_mutex;
220 DECLARE_BITMAP(doorbell_available_index,
221 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
224 uint64_t gtt_start_gpu_addr;
225 void *gtt_start_cpu_ptr;
227 struct mutex gtt_sa_lock;
228 unsigned int gtt_sa_chunk_size;
229 unsigned int gtt_sa_num_of_chunks;
232 struct kfifo ih_fifo;
233 struct workqueue_struct *ih_wq;
234 struct work_struct interrupt_work;
235 spinlock_t interrupt_lock;
237 /* QCM Device instance */
238 struct device_queue_manager *dqm;
242 * Interrupts of interest to KFD are copied
243 * from the HW ring into a SW ring.
245 bool interrupts_active;
248 struct kfd_dbgmgr *dbgmgr;
250 /* Maximum process number mapped to HW scheduler */
251 unsigned int max_proc_per_quantum;
255 const void *cwsr_isa;
256 unsigned int cwsr_isa_size;
261 bool pci_atomic_requested;
264 /* KGD2KFD callbacks */
265 void kgd2kfd_exit(void);
266 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
267 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g);
268 bool kgd2kfd_device_init(struct kfd_dev *kfd,
269 const struct kgd2kfd_shared_resources *gpu_resources);
270 void kgd2kfd_device_exit(struct kfd_dev *kfd);
273 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
274 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
275 KFD_MEMPOOL_FRAMEBUFFER = 3,
278 /* Character device interface */
279 int kfd_chardev_init(void);
280 void kfd_chardev_exit(void);
281 struct device *kfd_chardev(void);
284 * enum kfd_unmap_queues_filter
286 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
288 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
289 * running queues list.
291 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
295 enum kfd_unmap_queues_filter {
296 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
297 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
298 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
299 KFD_UNMAP_QUEUES_FILTER_BY_PASID
303 * enum kfd_queue_type
305 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
307 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
309 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
311 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
313 enum kfd_queue_type {
314 KFD_QUEUE_TYPE_COMPUTE,
320 enum kfd_queue_format {
321 KFD_QUEUE_FORMAT_PM4,
326 * struct queue_properties
328 * @type: The queue type.
330 * @queue_id: Queue identifier.
332 * @queue_address: Queue ring buffer address.
334 * @queue_size: Queue ring buffer size.
336 * @priority: Defines the queue priority relative to other queues in the
338 * This is just an indication and HW scheduling may override the priority as
339 * necessary while keeping the relative prioritization.
340 * the priority granularity is from 0 to f which f is the highest priority.
341 * currently all queues are initialized with the highest priority.
343 * @queue_percent: This field is partially implemented and currently a zero in
344 * this field defines that the queue is non active.
346 * @read_ptr: User space address which points to the number of dwords the
347 * cp read from the ring buffer. This field updates automatically by the H/W.
349 * @write_ptr: Defines the number of dwords written to the ring buffer.
351 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
352 * the queue ring buffer. This field should be similar to write_ptr and the
353 * user should update this field after he updated the write_ptr.
355 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
357 * @is_interop: Defines if this is a interop queue. Interop queue means that
358 * the queue can access both graphics and compute resources.
360 * @is_evicted: Defines if the queue is evicted. Only active queues
361 * are evicted, rendering them inactive.
363 * @is_active: Defines if the queue is active or not. @is_active and
364 * @is_evicted are protected by the DQM lock.
366 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
369 * This structure represents the queue properties for each queue no matter if
370 * it's user mode or kernel mode queue.
373 struct queue_properties {
374 enum kfd_queue_type type;
375 enum kfd_queue_format format;
376 unsigned int queue_id;
377 uint64_t queue_address;
380 uint32_t queue_percent;
383 void __iomem *doorbell_ptr;
384 uint32_t doorbell_off;
388 /* Not relevant for user mode queues in cp scheduling */
390 /* Relevant only for sdma queues*/
391 uint32_t sdma_engine_id;
392 uint32_t sdma_queue_id;
393 uint32_t sdma_vm_addr;
394 /* Relevant only for VI */
395 uint64_t eop_ring_buffer_address;
396 uint32_t eop_ring_buffer_size;
397 uint64_t ctx_save_restore_area_address;
398 uint32_t ctx_save_restore_area_size;
399 uint32_t ctl_stack_size;
402 /* Relevant for CU */
403 uint32_t cu_mask_count; /* Must be a multiple of 32 */
410 * @list: Queue linked list.
412 * @mqd: The queue MQD.
414 * @mqd_mem_obj: The MQD local gpu memory object.
416 * @gart_mqd_addr: The MQD gart mc address.
418 * @properties: The queue properties.
420 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
421 * that the queue should be execute on.
423 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
426 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
428 * @process: The kfd process that created this queue.
430 * @device: The kfd device that created this queue.
432 * This structure represents user mode compute queues.
433 * It contains all the necessary data to handle such queues.
438 struct list_head list;
440 struct kfd_mem_obj *mqd_mem_obj;
441 uint64_t gart_mqd_addr;
442 struct queue_properties properties;
448 unsigned int sdma_id;
449 unsigned int doorbell_id;
451 struct kfd_process *process;
452 struct kfd_dev *device;
456 * Please read the kfd_mqd_manager.h description.
459 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */
460 KFD_MQD_TYPE_HIQ, /* for hiq */
461 KFD_MQD_TYPE_CP, /* for cp queues and diq */
462 KFD_MQD_TYPE_SDMA, /* for sdma queues */
466 struct scheduling_resources {
467 unsigned int vmid_mask;
468 enum kfd_queue_type type;
472 uint32_t gds_heap_base;
473 uint32_t gds_heap_size;
476 struct process_queue_manager {
478 struct kfd_process *process;
479 struct list_head queues;
480 unsigned long *queue_slot_bitmap;
483 struct qcm_process_device {
484 /* The Device Queue Manager that owns this data */
485 struct device_queue_manager *dqm;
486 struct process_queue_manager *pqm;
488 struct list_head queues_list;
489 struct list_head priv_queue_list;
491 unsigned int queue_count;
494 unsigned int evicted; /* eviction counter, 0=active */
496 /* This flag tells if we should reset all wavefronts on
497 * process termination
499 bool reset_wavefronts;
502 * All the memory management data should be here too
504 uint64_t gds_context_area;
505 uint32_t sh_mem_config;
506 uint32_t sh_mem_bases;
507 uint32_t sh_mem_ape1_base;
508 uint32_t sh_mem_ape1_limit;
509 uint32_t page_table_base;
513 uint32_t sh_hidden_private_base;
525 /* doorbell resources per process per device */
526 unsigned long *doorbell_bitmap;
529 /* KFD Memory Eviction */
531 /* Approx. wait time before attempting to restore evicted BOs */
532 #define PROCESS_RESTORE_TIME_MS 100
533 /* Approx. back off time if restore fails due to lack of memory */
534 #define PROCESS_BACK_OFF_TIME_MS 100
535 /* Approx. time before evicting the process again */
536 #define PROCESS_ACTIVE_TIME_MS 10
538 int kgd2kfd_quiesce_mm(struct mm_struct *mm);
539 int kgd2kfd_resume_mm(struct mm_struct *mm);
540 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
541 struct dma_fence *fence);
543 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
544 * idr_handle in the least significant 4 bytes
546 #define MAKE_HANDLE(gpu_id, idr_handle) \
547 (((uint64_t)(gpu_id) << 32) + idr_handle)
548 #define GET_GPU_ID(handle) (handle >> 32)
549 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
557 /* Data that is per-process-per device. */
558 struct kfd_process_device {
560 * List of all per-device data for a process.
561 * Starts from kfd_process.per_device_data.
563 struct list_head per_device_list;
565 /* The device that owns this data. */
568 /* The process that owns this kfd_process_device. */
569 struct kfd_process *process;
571 /* per-process-per device QCM data structure */
572 struct qcm_process_device qpd;
578 uint64_t gpuvm_limit;
579 uint64_t scratch_base;
580 uint64_t scratch_limit;
582 /* VM context for GPUVM allocations */
583 struct file *drm_file;
586 /* GPUVM allocations storage */
587 struct idr alloc_idr;
589 /* Flag used to tell the pdd has dequeued from the dqm.
590 * This is used to prevent dev->dqm->ops.process_termination() from
591 * being called twice when it is already called in IOMMU callback
594 bool already_dequeued;
596 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
597 enum kfd_pdd_bound bound;
600 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
605 * kfd_process are stored in an mm_struct*->kfd_process*
606 * hash table (kfd_processes in kfd_process.c)
608 struct hlist_node kfd_processes;
611 * Opaque pointer to mm_struct. We don't hold a reference to
612 * it so it should never be dereferenced from here. This is
613 * only used for looking up processes by their mm.
618 struct work_struct release_work;
623 * In any process, the thread that started main() is the lead
624 * thread and outlives the rest.
625 * It is here because amd_iommu_bind_pasid wants a task_struct.
626 * It can also be used for safely getting a reference to the
627 * mm_struct of the process.
629 struct task_struct *lead_thread;
631 /* We want to receive a notification when the mm_struct is destroyed */
632 struct mmu_notifier mmu_notifier;
634 /* Use for delayed freeing of kfd_process structure */
638 unsigned int doorbell_index;
641 * List of kfd_process_device structures,
642 * one for each device the process is using.
644 struct list_head per_device_data;
646 struct process_queue_manager pqm;
648 /*Is the user space process 32 bit?*/
649 bool is_32bit_user_mode;
651 /* Event-related data */
652 struct mutex event_mutex;
653 /* Event ID allocator and lookup */
654 struct idr event_idr;
656 struct kfd_signal_page *signal_page;
657 size_t signal_mapped_size;
658 size_t signal_event_count;
659 bool signal_event_limit_reached;
661 /* Information used for memory eviction */
662 void *kgd_process_info;
663 /* Eviction fence that is attached to all the BOs of this process. The
664 * fence will be triggered during eviction and new one will be created
667 struct dma_fence *ef;
669 /* Work items for evicting and restoring BOs */
670 struct delayed_work eviction_work;
671 struct delayed_work restore_work;
672 /* seqno of the last scheduled eviction */
673 unsigned int last_eviction_seqno;
674 /* Approx. the last timestamp (in jiffies) when the process was
675 * restored after an eviction
677 unsigned long last_restore_timestamp;
680 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
681 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
682 extern struct srcu_struct kfd_processes_srcu;
685 * Ioctl function type.
687 * \param filep pointer to file structure.
688 * \param p amdkfd process pointer.
689 * \param data pointer to arg that was copied from user.
691 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
694 struct amdkfd_ioctl_desc {
697 amdkfd_ioctl_t *func;
698 unsigned int cmd_drv;
701 bool kfd_dev_is_large_bar(struct kfd_dev *dev);
703 int kfd_process_create_wq(void);
704 void kfd_process_destroy_wq(void);
705 struct kfd_process *kfd_create_process(struct file *filep);
706 struct kfd_process *kfd_get_process(const struct task_struct *);
707 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
708 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
709 void kfd_unref_process(struct kfd_process *p);
710 int kfd_process_evict_queues(struct kfd_process *p);
711 int kfd_process_restore_queues(struct kfd_process *p);
712 void kfd_suspend_all_processes(void);
713 int kfd_resume_all_processes(void);
715 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
716 struct file *drm_file);
717 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
718 struct kfd_process *p);
719 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
720 struct kfd_process *p);
721 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
722 struct kfd_process *p);
724 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
725 struct vm_area_struct *vma);
727 /* KFD process API for creating and translating handles */
728 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
730 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
732 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
735 /* Process device data iterator */
736 struct kfd_process_device *kfd_get_first_process_device_data(
737 struct kfd_process *p);
738 struct kfd_process_device *kfd_get_next_process_device_data(
739 struct kfd_process *p,
740 struct kfd_process_device *pdd);
741 bool kfd_has_process_device_data(struct kfd_process *p);
744 int kfd_pasid_init(void);
745 void kfd_pasid_exit(void);
746 bool kfd_set_pasid_limit(unsigned int new_limit);
747 unsigned int kfd_get_pasid_limit(void);
748 unsigned int kfd_pasid_alloc(void);
749 void kfd_pasid_free(unsigned int pasid);
752 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
753 int kfd_doorbell_init(struct kfd_dev *kfd);
754 void kfd_doorbell_fini(struct kfd_dev *kfd);
755 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
756 struct vm_area_struct *vma);
757 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
758 unsigned int *doorbell_off);
759 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
760 u32 read_kernel_doorbell(u32 __iomem *db);
761 void write_kernel_doorbell(void __iomem *db, u32 value);
762 void write_kernel_doorbell64(void __iomem *db, u64 value);
763 unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
764 struct kfd_process *process,
765 unsigned int doorbell_id);
766 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
767 struct kfd_process *process);
768 int kfd_alloc_process_doorbells(struct kfd_process *process);
769 void kfd_free_process_doorbells(struct kfd_process *process);
771 /* GTT Sub-Allocator */
773 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
774 struct kfd_mem_obj **mem_obj);
776 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
778 extern struct device *kfd_device;
781 int kfd_topology_init(void);
782 void kfd_topology_shutdown(void);
783 int kfd_topology_add_device(struct kfd_dev *gpu);
784 int kfd_topology_remove_device(struct kfd_dev *gpu);
785 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
786 uint32_t proximity_domain);
787 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
788 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
789 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
790 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
791 int kfd_numa_node_to_apic_id(int numa_node_id);
794 int kfd_interrupt_init(struct kfd_dev *dev);
795 void kfd_interrupt_exit(struct kfd_dev *dev);
796 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
797 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry);
798 bool interrupt_is_wanted(struct kfd_dev *dev,
799 const uint32_t *ih_ring_entry,
800 uint32_t *patched_ihre, bool *flag);
802 /* Power Management */
803 void kgd2kfd_suspend(struct kfd_dev *kfd);
804 int kgd2kfd_resume(struct kfd_dev *kfd);
807 int kgd2kfd_pre_reset(struct kfd_dev *kfd);
808 int kgd2kfd_post_reset(struct kfd_dev *kfd);
810 /* amdkfd Apertures */
811 int kfd_init_apertures(struct kfd_process *process);
813 /* Queue Context Management */
814 int init_queue(struct queue **q, const struct queue_properties *properties);
815 void uninit_queue(struct queue *q);
816 void print_queue_properties(struct queue_properties *q);
817 void print_queue(struct queue *q);
819 struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
820 struct kfd_dev *dev);
821 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
822 struct kfd_dev *dev);
823 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
824 struct kfd_dev *dev);
825 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
826 struct kfd_dev *dev);
827 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
828 struct kfd_dev *dev);
829 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
830 struct kfd_dev *dev);
831 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
832 void device_queue_manager_uninit(struct device_queue_manager *dqm);
833 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
834 enum kfd_queue_type type);
835 void kernel_queue_uninit(struct kernel_queue *kq);
836 int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid);
838 /* Process Queue Manager */
839 struct process_queue_node {
841 struct kernel_queue *kq;
842 struct list_head process_queue_list;
845 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
846 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
847 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
848 void pqm_uninit(struct process_queue_manager *pqm);
849 int pqm_create_queue(struct process_queue_manager *pqm,
852 struct queue_properties *properties,
854 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
855 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
856 struct queue_properties *p);
857 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
858 struct queue_properties *p);
859 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
862 int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
863 unsigned int fence_value,
864 unsigned int timeout_ms);
868 #define KFD_FENCE_COMPLETED (100)
869 #define KFD_FENCE_INIT (10)
871 struct packet_manager {
872 struct device_queue_manager *dqm;
873 struct kernel_queue *priv_queue;
876 struct kfd_mem_obj *ib_buffer_obj;
877 unsigned int ib_size_bytes;
879 const struct packet_manager_funcs *pmf;
882 struct packet_manager_funcs {
883 /* Support ASIC-specific packet formats for PM4 packets */
884 int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
885 struct qcm_process_device *qpd);
886 int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
887 uint64_t ib, size_t ib_size_in_dwords, bool chain);
888 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
889 struct scheduling_resources *res);
890 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
891 struct queue *q, bool is_static);
892 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
893 enum kfd_queue_type type,
894 enum kfd_unmap_queues_filter mode,
895 uint32_t filter_param, bool reset,
896 unsigned int sdma_engine);
897 int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
898 uint64_t fence_address, uint32_t fence_value);
899 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
902 int map_process_size;
904 int set_resources_size;
906 int unmap_queues_size;
907 int query_status_size;
908 int release_mem_size;
911 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
912 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
914 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
915 void pm_uninit(struct packet_manager *pm);
916 int pm_send_set_resources(struct packet_manager *pm,
917 struct scheduling_resources *res);
918 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
919 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
920 uint32_t fence_value);
922 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
923 enum kfd_unmap_queues_filter mode,
924 uint32_t filter_param, bool reset,
925 unsigned int sdma_engine);
927 void pm_release_ib(struct packet_manager *pm);
929 /* Following PM funcs can be shared among VI and AI */
930 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
931 int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer,
932 struct scheduling_resources *res);
934 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
937 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
938 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
940 extern const struct kfd_device_global_init_class device_global_init_class_cik;
942 void kfd_event_init_process(struct kfd_process *p);
943 void kfd_event_free_process(struct kfd_process *p);
944 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
945 int kfd_wait_on_events(struct kfd_process *p,
946 uint32_t num_events, void __user *data,
947 bool all, uint32_t user_timeout_ms,
948 uint32_t *wait_result);
949 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
950 uint32_t valid_id_bits);
951 void kfd_signal_iommu_event(struct kfd_dev *dev,
952 unsigned int pasid, unsigned long address,
953 bool is_write_requested, bool is_execute_requested);
954 void kfd_signal_hw_exception_event(unsigned int pasid);
955 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
956 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
957 int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
959 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
960 uint32_t event_type, bool auto_reset, uint32_t node_id,
961 uint32_t *event_id, uint32_t *event_trigger_data,
962 uint64_t *event_page_offset, uint32_t *event_slot_index);
963 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
965 void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid,
966 struct kfd_vm_fault_info *info);
968 void kfd_signal_reset_event(struct kfd_dev *dev);
970 void kfd_flush_tlb(struct kfd_process_device *pdd);
972 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
974 bool kfd_is_locked(void);
977 #if defined(CONFIG_DEBUG_FS)
979 void kfd_debugfs_init(void);
980 void kfd_debugfs_fini(void);
981 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
982 int pqm_debugfs_mqds(struct seq_file *m, void *data);
983 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
984 int dqm_debugfs_hqds(struct seq_file *m, void *data);
985 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
986 int pm_debugfs_runlist(struct seq_file *m, void *data);
988 int kfd_debugfs_hang_hws(struct kfd_dev *dev);
989 int pm_debugfs_hang_hws(struct packet_manager *pm);
990 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm);
994 static inline void kfd_debugfs_init(void) {}
995 static inline void kfd_debugfs_fini(void) {}