Loading...
Note: File does not exist in v3.1.
1/*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
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:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
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.
21 */
22
23#ifndef KFD_PRIV_H_INCLUDED
24#define KFD_PRIV_H_INCLUDED
25
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>
39
40#include "amd_shared.h"
41
42#define KFD_SYSFS_FILE_MODE 0444
43
44#define KFD_MMAP_DOORBELL_MASK 0x8000000000000ull
45#define KFD_MMAP_EVENTS_MASK 0x4000000000000ull
46#define KFD_MMAP_RESERVED_MEM_MASK 0x2000000000000ull
47
48/*
49 * When working with cp scheduler we should assign the HIQ manually or via
50 * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot
51 * definitions for Kaveri. In Kaveri only the first ME queues participates
52 * in the cp scheduling taking that in mind we set the HIQ slot in the
53 * second ME.
54 */
55#define KFD_CIK_HIQ_PIPE 4
56#define KFD_CIK_HIQ_QUEUE 0
57
58/* GPU ID hash width in bits */
59#define KFD_GPU_ID_HASH_WIDTH 16
60
61/* Macro for allocating structures */
62#define kfd_alloc_struct(ptr_to_struct) \
63 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
64
65#define KFD_MAX_NUM_OF_PROCESSES 512
66#define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
67
68/*
69 * Size of the per-process TBA+TMA buffer: 2 pages
70 *
71 * The first page is the TBA used for the CWSR ISA code. The second
72 * page is used as TMA for daisy changing a user-mode trap handler.
73 */
74#define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
75#define KFD_CWSR_TMA_OFFSET PAGE_SIZE
76
77/*
78 * Kernel module parameter to specify maximum number of supported queues per
79 * device
80 */
81extern int max_num_of_queues_per_device;
82
83#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096
84#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \
85 (KFD_MAX_NUM_OF_PROCESSES * \
86 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
87
88#define KFD_KERNEL_QUEUE_SIZE 2048
89
90/* Kernel module parameter to specify the scheduling policy */
91extern int sched_policy;
92
93/*
94 * Kernel module parameter to specify the maximum process
95 * number per HW scheduler
96 */
97extern int hws_max_conc_proc;
98
99extern int cwsr_enable;
100
101/*
102 * Kernel module parameter to specify whether to send sigterm to HSA process on
103 * unhandled exception
104 */
105extern int send_sigterm;
106
107/*
108 * This kernel module is used to simulate large bar machine on non-large bar
109 * enabled machines.
110 */
111extern int debug_largebar;
112
113/*
114 * Ignore CRAT table during KFD initialization, can be used to work around
115 * broken CRAT tables on some AMD systems
116 */
117extern int ignore_crat;
118
119/**
120 * enum kfd_sched_policy
121 *
122 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp)
123 * scheduling. In this scheduling mode we're using the firmware code to
124 * schedule the user mode queues and kernel queues such as HIQ and DIQ.
125 * the HIQ queue is used as a special queue that dispatches the configuration
126 * to the cp and the user mode queues list that are currently running.
127 * the DIQ queue is a debugging queue that dispatches debugging commands to the
128 * firmware.
129 * in this scheduling mode user mode queues over subscription feature is
130 * enabled.
131 *
132 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over
133 * subscription feature disabled.
134 *
135 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly
136 * set the command processor registers and sets the queues "manually". This
137 * mode is used *ONLY* for debugging proposes.
138 *
139 */
140enum kfd_sched_policy {
141 KFD_SCHED_POLICY_HWS = 0,
142 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION,
143 KFD_SCHED_POLICY_NO_HWS
144};
145
146enum cache_policy {
147 cache_policy_coherent,
148 cache_policy_noncoherent
149};
150
151struct kfd_event_interrupt_class {
152 bool (*interrupt_isr)(struct kfd_dev *dev,
153 const uint32_t *ih_ring_entry);
154 void (*interrupt_wq)(struct kfd_dev *dev,
155 const uint32_t *ih_ring_entry);
156};
157
158struct kfd_device_info {
159 enum amd_asic_type asic_family;
160 const struct kfd_event_interrupt_class *event_interrupt_class;
161 unsigned int max_pasid_bits;
162 unsigned int max_no_of_hqd;
163 size_t ih_ring_entry_size;
164 uint8_t num_of_watch_points;
165 uint16_t mqd_size_aligned;
166 bool supports_cwsr;
167 bool needs_iommu_device;
168 bool needs_pci_atomics;
169};
170
171struct kfd_mem_obj {
172 uint32_t range_start;
173 uint32_t range_end;
174 uint64_t gpu_addr;
175 uint32_t *cpu_ptr;
176};
177
178struct kfd_vmid_info {
179 uint32_t first_vmid_kfd;
180 uint32_t last_vmid_kfd;
181 uint32_t vmid_num_kfd;
182};
183
184struct kfd_dev {
185 struct kgd_dev *kgd;
186
187 const struct kfd_device_info *device_info;
188 struct pci_dev *pdev;
189
190 unsigned int id; /* topology stub index */
191
192 phys_addr_t doorbell_base; /* Start of actual doorbells used by
193 * KFD. It is aligned for mapping
194 * into user mode
195 */
196 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell
197 * to HW doorbell, GFX reserved some
198 * at the start)
199 */
200 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
201 * page used by kernel queue
202 */
203
204 struct kgd2kfd_shared_resources shared_resources;
205 struct kfd_vmid_info vm_info;
206
207 const struct kfd2kgd_calls *kfd2kgd;
208 struct mutex doorbell_mutex;
209 DECLARE_BITMAP(doorbell_available_index,
210 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
211
212 void *gtt_mem;
213 uint64_t gtt_start_gpu_addr;
214 void *gtt_start_cpu_ptr;
215 void *gtt_sa_bitmap;
216 struct mutex gtt_sa_lock;
217 unsigned int gtt_sa_chunk_size;
218 unsigned int gtt_sa_num_of_chunks;
219
220 /* Interrupts */
221 struct kfifo ih_fifo;
222 struct workqueue_struct *ih_wq;
223 struct work_struct interrupt_work;
224 spinlock_t interrupt_lock;
225
226 /* QCM Device instance */
227 struct device_queue_manager *dqm;
228
229 bool init_complete;
230 /*
231 * Interrupts of interest to KFD are copied
232 * from the HW ring into a SW ring.
233 */
234 bool interrupts_active;
235
236 /* Debug manager */
237 struct kfd_dbgmgr *dbgmgr;
238
239 /* Maximum process number mapped to HW scheduler */
240 unsigned int max_proc_per_quantum;
241
242 /* CWSR */
243 bool cwsr_enabled;
244 const void *cwsr_isa;
245 unsigned int cwsr_isa_size;
246};
247
248/* KGD2KFD callbacks */
249void kgd2kfd_exit(void);
250struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
251 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g);
252bool kgd2kfd_device_init(struct kfd_dev *kfd,
253 const struct kgd2kfd_shared_resources *gpu_resources);
254void kgd2kfd_device_exit(struct kfd_dev *kfd);
255
256enum kfd_mempool {
257 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
258 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
259 KFD_MEMPOOL_FRAMEBUFFER = 3,
260};
261
262/* Character device interface */
263int kfd_chardev_init(void);
264void kfd_chardev_exit(void);
265struct device *kfd_chardev(void);
266
267/**
268 * enum kfd_unmap_queues_filter
269 *
270 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
271 *
272 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
273 * running queues list.
274 *
275 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
276 * specific process.
277 *
278 */
279enum kfd_unmap_queues_filter {
280 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
281 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
282 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
283 KFD_UNMAP_QUEUES_FILTER_BY_PASID
284};
285
286/**
287 * enum kfd_queue_type
288 *
289 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
290 *
291 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
292 *
293 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
294 *
295 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
296 */
297enum kfd_queue_type {
298 KFD_QUEUE_TYPE_COMPUTE,
299 KFD_QUEUE_TYPE_SDMA,
300 KFD_QUEUE_TYPE_HIQ,
301 KFD_QUEUE_TYPE_DIQ
302};
303
304enum kfd_queue_format {
305 KFD_QUEUE_FORMAT_PM4,
306 KFD_QUEUE_FORMAT_AQL
307};
308
309/**
310 * struct queue_properties
311 *
312 * @type: The queue type.
313 *
314 * @queue_id: Queue identifier.
315 *
316 * @queue_address: Queue ring buffer address.
317 *
318 * @queue_size: Queue ring buffer size.
319 *
320 * @priority: Defines the queue priority relative to other queues in the
321 * process.
322 * This is just an indication and HW scheduling may override the priority as
323 * necessary while keeping the relative prioritization.
324 * the priority granularity is from 0 to f which f is the highest priority.
325 * currently all queues are initialized with the highest priority.
326 *
327 * @queue_percent: This field is partially implemented and currently a zero in
328 * this field defines that the queue is non active.
329 *
330 * @read_ptr: User space address which points to the number of dwords the
331 * cp read from the ring buffer. This field updates automatically by the H/W.
332 *
333 * @write_ptr: Defines the number of dwords written to the ring buffer.
334 *
335 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
336 * the queue ring buffer. This field should be similar to write_ptr and the
337 * user should update this field after he updated the write_ptr.
338 *
339 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
340 *
341 * @is_interop: Defines if this is a interop queue. Interop queue means that
342 * the queue can access both graphics and compute resources.
343 *
344 * @is_evicted: Defines if the queue is evicted. Only active queues
345 * are evicted, rendering them inactive.
346 *
347 * @is_active: Defines if the queue is active or not. @is_active and
348 * @is_evicted are protected by the DQM lock.
349 *
350 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
351 * of the queue.
352 *
353 * This structure represents the queue properties for each queue no matter if
354 * it's user mode or kernel mode queue.
355 *
356 */
357struct queue_properties {
358 enum kfd_queue_type type;
359 enum kfd_queue_format format;
360 unsigned int queue_id;
361 uint64_t queue_address;
362 uint64_t queue_size;
363 uint32_t priority;
364 uint32_t queue_percent;
365 uint32_t *read_ptr;
366 uint32_t *write_ptr;
367 uint32_t __iomem *doorbell_ptr;
368 uint32_t doorbell_off;
369 bool is_interop;
370 bool is_evicted;
371 bool is_active;
372 /* Not relevant for user mode queues in cp scheduling */
373 unsigned int vmid;
374 /* Relevant only for sdma queues*/
375 uint32_t sdma_engine_id;
376 uint32_t sdma_queue_id;
377 uint32_t sdma_vm_addr;
378 /* Relevant only for VI */
379 uint64_t eop_ring_buffer_address;
380 uint32_t eop_ring_buffer_size;
381 uint64_t ctx_save_restore_area_address;
382 uint32_t ctx_save_restore_area_size;
383 uint32_t ctl_stack_size;
384 uint64_t tba_addr;
385 uint64_t tma_addr;
386};
387
388/**
389 * struct queue
390 *
391 * @list: Queue linked list.
392 *
393 * @mqd: The queue MQD.
394 *
395 * @mqd_mem_obj: The MQD local gpu memory object.
396 *
397 * @gart_mqd_addr: The MQD gart mc address.
398 *
399 * @properties: The queue properties.
400 *
401 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
402 * that the queue should be execute on.
403 *
404 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
405 * id.
406 *
407 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
408 *
409 * @process: The kfd process that created this queue.
410 *
411 * @device: The kfd device that created this queue.
412 *
413 * This structure represents user mode compute queues.
414 * It contains all the necessary data to handle such queues.
415 *
416 */
417
418struct queue {
419 struct list_head list;
420 void *mqd;
421 struct kfd_mem_obj *mqd_mem_obj;
422 uint64_t gart_mqd_addr;
423 struct queue_properties properties;
424
425 uint32_t mec;
426 uint32_t pipe;
427 uint32_t queue;
428
429 unsigned int sdma_id;
430
431 struct kfd_process *process;
432 struct kfd_dev *device;
433};
434
435/*
436 * Please read the kfd_mqd_manager.h description.
437 */
438enum KFD_MQD_TYPE {
439 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */
440 KFD_MQD_TYPE_HIQ, /* for hiq */
441 KFD_MQD_TYPE_CP, /* for cp queues and diq */
442 KFD_MQD_TYPE_SDMA, /* for sdma queues */
443 KFD_MQD_TYPE_MAX
444};
445
446struct scheduling_resources {
447 unsigned int vmid_mask;
448 enum kfd_queue_type type;
449 uint64_t queue_mask;
450 uint64_t gws_mask;
451 uint32_t oac_mask;
452 uint32_t gds_heap_base;
453 uint32_t gds_heap_size;
454};
455
456struct process_queue_manager {
457 /* data */
458 struct kfd_process *process;
459 struct list_head queues;
460 unsigned long *queue_slot_bitmap;
461};
462
463struct qcm_process_device {
464 /* The Device Queue Manager that owns this data */
465 struct device_queue_manager *dqm;
466 struct process_queue_manager *pqm;
467 /* Queues list */
468 struct list_head queues_list;
469 struct list_head priv_queue_list;
470
471 unsigned int queue_count;
472 unsigned int vmid;
473 bool is_debug;
474 unsigned int evicted; /* eviction counter, 0=active */
475
476 /* This flag tells if we should reset all wavefronts on
477 * process termination
478 */
479 bool reset_wavefronts;
480
481 /*
482 * All the memory management data should be here too
483 */
484 uint64_t gds_context_area;
485 uint32_t sh_mem_config;
486 uint32_t sh_mem_bases;
487 uint32_t sh_mem_ape1_base;
488 uint32_t sh_mem_ape1_limit;
489 uint32_t page_table_base;
490 uint32_t gds_size;
491 uint32_t num_gws;
492 uint32_t num_oac;
493 uint32_t sh_hidden_private_base;
494
495 /* CWSR memory */
496 void *cwsr_kaddr;
497 uint64_t cwsr_base;
498 uint64_t tba_addr;
499 uint64_t tma_addr;
500
501 /* IB memory */
502 uint64_t ib_base;
503 void *ib_kaddr;
504};
505
506/* KFD Memory Eviction */
507
508/* Approx. wait time before attempting to restore evicted BOs */
509#define PROCESS_RESTORE_TIME_MS 100
510/* Approx. back off time if restore fails due to lack of memory */
511#define PROCESS_BACK_OFF_TIME_MS 100
512/* Approx. time before evicting the process again */
513#define PROCESS_ACTIVE_TIME_MS 10
514
515int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
516 struct dma_fence *fence);
517
518/* 8 byte handle containing GPU ID in the most significant 4 bytes and
519 * idr_handle in the least significant 4 bytes
520 */
521#define MAKE_HANDLE(gpu_id, idr_handle) \
522 (((uint64_t)(gpu_id) << 32) + idr_handle)
523#define GET_GPU_ID(handle) (handle >> 32)
524#define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
525
526enum kfd_pdd_bound {
527 PDD_UNBOUND = 0,
528 PDD_BOUND,
529 PDD_BOUND_SUSPENDED,
530};
531
532/* Data that is per-process-per device. */
533struct kfd_process_device {
534 /*
535 * List of all per-device data for a process.
536 * Starts from kfd_process.per_device_data.
537 */
538 struct list_head per_device_list;
539
540 /* The device that owns this data. */
541 struct kfd_dev *dev;
542
543 /* The process that owns this kfd_process_device. */
544 struct kfd_process *process;
545
546 /* per-process-per device QCM data structure */
547 struct qcm_process_device qpd;
548
549 /*Apertures*/
550 uint64_t lds_base;
551 uint64_t lds_limit;
552 uint64_t gpuvm_base;
553 uint64_t gpuvm_limit;
554 uint64_t scratch_base;
555 uint64_t scratch_limit;
556
557 /* VM context for GPUVM allocations */
558 struct file *drm_file;
559 void *vm;
560
561 /* GPUVM allocations storage */
562 struct idr alloc_idr;
563
564 /* Flag used to tell the pdd has dequeued from the dqm.
565 * This is used to prevent dev->dqm->ops.process_termination() from
566 * being called twice when it is already called in IOMMU callback
567 * function.
568 */
569 bool already_dequeued;
570
571 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
572 enum kfd_pdd_bound bound;
573};
574
575#define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
576
577/* Process data */
578struct kfd_process {
579 /*
580 * kfd_process are stored in an mm_struct*->kfd_process*
581 * hash table (kfd_processes in kfd_process.c)
582 */
583 struct hlist_node kfd_processes;
584
585 /*
586 * Opaque pointer to mm_struct. We don't hold a reference to
587 * it so it should never be dereferenced from here. This is
588 * only used for looking up processes by their mm.
589 */
590 void *mm;
591
592 struct kref ref;
593 struct work_struct release_work;
594
595 struct mutex mutex;
596
597 /*
598 * In any process, the thread that started main() is the lead
599 * thread and outlives the rest.
600 * It is here because amd_iommu_bind_pasid wants a task_struct.
601 * It can also be used for safely getting a reference to the
602 * mm_struct of the process.
603 */
604 struct task_struct *lead_thread;
605
606 /* We want to receive a notification when the mm_struct is destroyed */
607 struct mmu_notifier mmu_notifier;
608
609 /* Use for delayed freeing of kfd_process structure */
610 struct rcu_head rcu;
611
612 unsigned int pasid;
613 unsigned int doorbell_index;
614
615 /*
616 * List of kfd_process_device structures,
617 * one for each device the process is using.
618 */
619 struct list_head per_device_data;
620
621 struct process_queue_manager pqm;
622
623 /*Is the user space process 32 bit?*/
624 bool is_32bit_user_mode;
625
626 /* Event-related data */
627 struct mutex event_mutex;
628 /* Event ID allocator and lookup */
629 struct idr event_idr;
630 /* Event page */
631 struct kfd_signal_page *signal_page;
632 size_t signal_mapped_size;
633 size_t signal_event_count;
634 bool signal_event_limit_reached;
635
636 /* Information used for memory eviction */
637 void *kgd_process_info;
638 /* Eviction fence that is attached to all the BOs of this process. The
639 * fence will be triggered during eviction and new one will be created
640 * during restore
641 */
642 struct dma_fence *ef;
643
644 /* Work items for evicting and restoring BOs */
645 struct delayed_work eviction_work;
646 struct delayed_work restore_work;
647 /* seqno of the last scheduled eviction */
648 unsigned int last_eviction_seqno;
649 /* Approx. the last timestamp (in jiffies) when the process was
650 * restored after an eviction
651 */
652 unsigned long last_restore_timestamp;
653};
654
655#define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
656extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
657extern struct srcu_struct kfd_processes_srcu;
658
659/**
660 * Ioctl function type.
661 *
662 * \param filep pointer to file structure.
663 * \param p amdkfd process pointer.
664 * \param data pointer to arg that was copied from user.
665 */
666typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
667 void *data);
668
669struct amdkfd_ioctl_desc {
670 unsigned int cmd;
671 int flags;
672 amdkfd_ioctl_t *func;
673 unsigned int cmd_drv;
674 const char *name;
675};
676
677int kfd_process_create_wq(void);
678void kfd_process_destroy_wq(void);
679struct kfd_process *kfd_create_process(struct file *filep);
680struct kfd_process *kfd_get_process(const struct task_struct *);
681struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
682struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
683void kfd_unref_process(struct kfd_process *p);
684void kfd_suspend_all_processes(void);
685int kfd_resume_all_processes(void);
686
687int kfd_process_device_init_vm(struct kfd_process_device *pdd,
688 struct file *drm_file);
689struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
690 struct kfd_process *p);
691struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
692 struct kfd_process *p);
693struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
694 struct kfd_process *p);
695
696int kfd_reserved_mem_mmap(struct kfd_process *process,
697 struct vm_area_struct *vma);
698
699/* KFD process API for creating and translating handles */
700int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
701 void *mem);
702void *kfd_process_device_translate_handle(struct kfd_process_device *p,
703 int handle);
704void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
705 int handle);
706
707/* Process device data iterator */
708struct kfd_process_device *kfd_get_first_process_device_data(
709 struct kfd_process *p);
710struct kfd_process_device *kfd_get_next_process_device_data(
711 struct kfd_process *p,
712 struct kfd_process_device *pdd);
713bool kfd_has_process_device_data(struct kfd_process *p);
714
715/* PASIDs */
716int kfd_pasid_init(void);
717void kfd_pasid_exit(void);
718bool kfd_set_pasid_limit(unsigned int new_limit);
719unsigned int kfd_get_pasid_limit(void);
720unsigned int kfd_pasid_alloc(void);
721void kfd_pasid_free(unsigned int pasid);
722
723/* Doorbells */
724int kfd_doorbell_init(struct kfd_dev *kfd);
725void kfd_doorbell_fini(struct kfd_dev *kfd);
726int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma);
727u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
728 unsigned int *doorbell_off);
729void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
730u32 read_kernel_doorbell(u32 __iomem *db);
731void write_kernel_doorbell(u32 __iomem *db, u32 value);
732unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
733 struct kfd_process *process,
734 unsigned int queue_id);
735phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
736 struct kfd_process *process);
737int kfd_alloc_process_doorbells(struct kfd_process *process);
738void kfd_free_process_doorbells(struct kfd_process *process);
739
740/* GTT Sub-Allocator */
741
742int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
743 struct kfd_mem_obj **mem_obj);
744
745int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
746
747extern struct device *kfd_device;
748
749/* Topology */
750int kfd_topology_init(void);
751void kfd_topology_shutdown(void);
752int kfd_topology_add_device(struct kfd_dev *gpu);
753int kfd_topology_remove_device(struct kfd_dev *gpu);
754struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
755 uint32_t proximity_domain);
756struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
757struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
758int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
759int kfd_numa_node_to_apic_id(int numa_node_id);
760
761/* Interrupts */
762int kfd_interrupt_init(struct kfd_dev *dev);
763void kfd_interrupt_exit(struct kfd_dev *dev);
764void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
765bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry);
766bool interrupt_is_wanted(struct kfd_dev *dev, const uint32_t *ih_ring_entry);
767
768/* Power Management */
769void kgd2kfd_suspend(struct kfd_dev *kfd);
770int kgd2kfd_resume(struct kfd_dev *kfd);
771
772/* amdkfd Apertures */
773int kfd_init_apertures(struct kfd_process *process);
774
775/* Queue Context Management */
776int init_queue(struct queue **q, const struct queue_properties *properties);
777void uninit_queue(struct queue *q);
778void print_queue_properties(struct queue_properties *q);
779void print_queue(struct queue *q);
780
781struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
782 struct kfd_dev *dev);
783struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
784 struct kfd_dev *dev);
785struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
786 struct kfd_dev *dev);
787struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
788 struct kfd_dev *dev);
789struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
790 struct kfd_dev *dev);
791struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
792void device_queue_manager_uninit(struct device_queue_manager *dqm);
793struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
794 enum kfd_queue_type type);
795void kernel_queue_uninit(struct kernel_queue *kq);
796
797/* Process Queue Manager */
798struct process_queue_node {
799 struct queue *q;
800 struct kernel_queue *kq;
801 struct list_head process_queue_list;
802};
803
804void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
805void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
806int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
807void pqm_uninit(struct process_queue_manager *pqm);
808int pqm_create_queue(struct process_queue_manager *pqm,
809 struct kfd_dev *dev,
810 struct file *f,
811 struct queue_properties *properties,
812 unsigned int *qid);
813int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
814int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
815 struct queue_properties *p);
816struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
817 unsigned int qid);
818
819int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
820 unsigned int fence_value,
821 unsigned int timeout_ms);
822
823/* Packet Manager */
824
825#define KFD_FENCE_COMPLETED (100)
826#define KFD_FENCE_INIT (10)
827
828struct packet_manager {
829 struct device_queue_manager *dqm;
830 struct kernel_queue *priv_queue;
831 struct mutex lock;
832 bool allocated;
833 struct kfd_mem_obj *ib_buffer_obj;
834 unsigned int ib_size_bytes;
835};
836
837int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
838void pm_uninit(struct packet_manager *pm);
839int pm_send_set_resources(struct packet_manager *pm,
840 struct scheduling_resources *res);
841int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
842int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
843 uint32_t fence_value);
844
845int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
846 enum kfd_unmap_queues_filter mode,
847 uint32_t filter_param, bool reset,
848 unsigned int sdma_engine);
849
850void pm_release_ib(struct packet_manager *pm);
851
852uint32_t pm_create_release_mem(uint64_t gpu_addr, uint32_t *buffer);
853
854uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
855
856/* Events */
857extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
858extern const struct kfd_device_global_init_class device_global_init_class_cik;
859
860void kfd_event_init_process(struct kfd_process *p);
861void kfd_event_free_process(struct kfd_process *p);
862int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
863int kfd_wait_on_events(struct kfd_process *p,
864 uint32_t num_events, void __user *data,
865 bool all, uint32_t user_timeout_ms,
866 uint32_t *wait_result);
867void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
868 uint32_t valid_id_bits);
869void kfd_signal_iommu_event(struct kfd_dev *dev,
870 unsigned int pasid, unsigned long address,
871 bool is_write_requested, bool is_execute_requested);
872void kfd_signal_hw_exception_event(unsigned int pasid);
873int kfd_set_event(struct kfd_process *p, uint32_t event_id);
874int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
875int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
876 uint64_t size);
877int kfd_event_create(struct file *devkfd, struct kfd_process *p,
878 uint32_t event_type, bool auto_reset, uint32_t node_id,
879 uint32_t *event_id, uint32_t *event_trigger_data,
880 uint64_t *event_page_offset, uint32_t *event_slot_index);
881int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
882
883void kfd_flush_tlb(struct kfd_process_device *pdd);
884
885int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
886
887/* Debugfs */
888#if defined(CONFIG_DEBUG_FS)
889
890void kfd_debugfs_init(void);
891void kfd_debugfs_fini(void);
892int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
893int pqm_debugfs_mqds(struct seq_file *m, void *data);
894int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
895int dqm_debugfs_hqds(struct seq_file *m, void *data);
896int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
897int pm_debugfs_runlist(struct seq_file *m, void *data);
898
899#else
900
901static inline void kfd_debugfs_init(void) {}
902static inline void kfd_debugfs_fini(void) {}
903
904#endif
905
906#endif