xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_priv.h (revision ddc141e5)
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  */
81 extern 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 */
91 extern int sched_policy;
92 
93 /*
94  * Kernel module parameter to specify the maximum process
95  * number per HW scheduler
96  */
97 extern int hws_max_conc_proc;
98 
99 extern int cwsr_enable;
100 
101 /*
102  * Kernel module parameter to specify whether to send sigterm to HSA process on
103  * unhandled exception
104  */
105 extern int send_sigterm;
106 
107 /*
108  * Ignore CRAT table during KFD initialization, can be used to work around
109  * broken CRAT tables on some AMD systems
110  */
111 extern int ignore_crat;
112 
113 /**
114  * enum kfd_sched_policy
115  *
116  * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp)
117  * scheduling. In this scheduling mode we're using the firmware code to
118  * schedule the user mode queues and kernel queues such as HIQ and DIQ.
119  * the HIQ queue is used as a special queue that dispatches the configuration
120  * to the cp and the user mode queues list that are currently running.
121  * the DIQ queue is a debugging queue that dispatches debugging commands to the
122  * firmware.
123  * in this scheduling mode user mode queues over subscription feature is
124  * enabled.
125  *
126  * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over
127  * subscription feature disabled.
128  *
129  * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly
130  * set the command processor registers and sets the queues "manually". This
131  * mode is used *ONLY* for debugging proposes.
132  *
133  */
134 enum kfd_sched_policy {
135 	KFD_SCHED_POLICY_HWS = 0,
136 	KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION,
137 	KFD_SCHED_POLICY_NO_HWS
138 };
139 
140 enum cache_policy {
141 	cache_policy_coherent,
142 	cache_policy_noncoherent
143 };
144 
145 struct kfd_event_interrupt_class {
146 	bool (*interrupt_isr)(struct kfd_dev *dev,
147 				const uint32_t *ih_ring_entry);
148 	void (*interrupt_wq)(struct kfd_dev *dev,
149 				const uint32_t *ih_ring_entry);
150 };
151 
152 struct kfd_device_info {
153 	enum amd_asic_type asic_family;
154 	const struct kfd_event_interrupt_class *event_interrupt_class;
155 	unsigned int max_pasid_bits;
156 	unsigned int max_no_of_hqd;
157 	size_t ih_ring_entry_size;
158 	uint8_t num_of_watch_points;
159 	uint16_t mqd_size_aligned;
160 	bool supports_cwsr;
161 };
162 
163 struct kfd_mem_obj {
164 	uint32_t range_start;
165 	uint32_t range_end;
166 	uint64_t gpu_addr;
167 	uint32_t *cpu_ptr;
168 };
169 
170 struct kfd_vmid_info {
171 	uint32_t first_vmid_kfd;
172 	uint32_t last_vmid_kfd;
173 	uint32_t vmid_num_kfd;
174 };
175 
176 struct kfd_dev {
177 	struct kgd_dev *kgd;
178 
179 	const struct kfd_device_info *device_info;
180 	struct pci_dev *pdev;
181 
182 	unsigned int id;		/* topology stub index */
183 
184 	phys_addr_t doorbell_base;	/* Start of actual doorbells used by
185 					 * KFD. It is aligned for mapping
186 					 * into user mode
187 					 */
188 	size_t doorbell_id_offset;	/* Doorbell offset (from KFD doorbell
189 					 * to HW doorbell, GFX reserved some
190 					 * at the start)
191 					 */
192 	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
193 					   * page used by kernel queue
194 					   */
195 
196 	struct kgd2kfd_shared_resources shared_resources;
197 	struct kfd_vmid_info vm_info;
198 
199 	const struct kfd2kgd_calls *kfd2kgd;
200 	struct mutex doorbell_mutex;
201 	DECLARE_BITMAP(doorbell_available_index,
202 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
203 
204 	void *gtt_mem;
205 	uint64_t gtt_start_gpu_addr;
206 	void *gtt_start_cpu_ptr;
207 	void *gtt_sa_bitmap;
208 	struct mutex gtt_sa_lock;
209 	unsigned int gtt_sa_chunk_size;
210 	unsigned int gtt_sa_num_of_chunks;
211 
212 	/* Interrupts */
213 	struct kfifo ih_fifo;
214 	struct workqueue_struct *ih_wq;
215 	struct work_struct interrupt_work;
216 	spinlock_t interrupt_lock;
217 
218 	/* QCM Device instance */
219 	struct device_queue_manager *dqm;
220 
221 	bool init_complete;
222 	/*
223 	 * Interrupts of interest to KFD are copied
224 	 * from the HW ring into a SW ring.
225 	 */
226 	bool interrupts_active;
227 
228 	/* Debug manager */
229 	struct kfd_dbgmgr           *dbgmgr;
230 
231 	/* Maximum process number mapped to HW scheduler */
232 	unsigned int max_proc_per_quantum;
233 
234 	/* CWSR */
235 	bool cwsr_enabled;
236 	const void *cwsr_isa;
237 	unsigned int cwsr_isa_size;
238 };
239 
240 /* KGD2KFD callbacks */
241 void kgd2kfd_exit(void);
242 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
243 			struct pci_dev *pdev, const struct kfd2kgd_calls *f2g);
244 bool kgd2kfd_device_init(struct kfd_dev *kfd,
245 			const struct kgd2kfd_shared_resources *gpu_resources);
246 void kgd2kfd_device_exit(struct kfd_dev *kfd);
247 
248 enum kfd_mempool {
249 	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
250 	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
251 	KFD_MEMPOOL_FRAMEBUFFER = 3,
252 };
253 
254 /* Character device interface */
255 int kfd_chardev_init(void);
256 void kfd_chardev_exit(void);
257 struct device *kfd_chardev(void);
258 
259 /**
260  * enum kfd_unmap_queues_filter
261  *
262  * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
263  *
264  * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
265  *						running queues list.
266  *
267  * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
268  *						specific process.
269  *
270  */
271 enum kfd_unmap_queues_filter {
272 	KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
273 	KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
274 	KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
275 	KFD_UNMAP_QUEUES_FILTER_BY_PASID
276 };
277 
278 /**
279  * enum kfd_queue_type
280  *
281  * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
282  *
283  * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
284  *
285  * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
286  *
287  * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
288  */
289 enum kfd_queue_type  {
290 	KFD_QUEUE_TYPE_COMPUTE,
291 	KFD_QUEUE_TYPE_SDMA,
292 	KFD_QUEUE_TYPE_HIQ,
293 	KFD_QUEUE_TYPE_DIQ
294 };
295 
296 enum kfd_queue_format {
297 	KFD_QUEUE_FORMAT_PM4,
298 	KFD_QUEUE_FORMAT_AQL
299 };
300 
301 /**
302  * struct queue_properties
303  *
304  * @type: The queue type.
305  *
306  * @queue_id: Queue identifier.
307  *
308  * @queue_address: Queue ring buffer address.
309  *
310  * @queue_size: Queue ring buffer size.
311  *
312  * @priority: Defines the queue priority relative to other queues in the
313  * process.
314  * This is just an indication and HW scheduling may override the priority as
315  * necessary while keeping the relative prioritization.
316  * the priority granularity is from 0 to f which f is the highest priority.
317  * currently all queues are initialized with the highest priority.
318  *
319  * @queue_percent: This field is partially implemented and currently a zero in
320  * this field defines that the queue is non active.
321  *
322  * @read_ptr: User space address which points to the number of dwords the
323  * cp read from the ring buffer. This field updates automatically by the H/W.
324  *
325  * @write_ptr: Defines the number of dwords written to the ring buffer.
326  *
327  * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
328  * the queue ring buffer. This field should be similar to write_ptr and the
329  * user should update this field after he updated the write_ptr.
330  *
331  * @doorbell_off: The doorbell offset in the doorbell pci-bar.
332  *
333  * @is_interop: Defines if this is a interop queue. Interop queue means that
334  * the queue can access both graphics and compute resources.
335  *
336  * @is_active: Defines if the queue is active or not.
337  *
338  * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
339  * of the queue.
340  *
341  * This structure represents the queue properties for each queue no matter if
342  * it's user mode or kernel mode queue.
343  *
344  */
345 struct queue_properties {
346 	enum kfd_queue_type type;
347 	enum kfd_queue_format format;
348 	unsigned int queue_id;
349 	uint64_t queue_address;
350 	uint64_t  queue_size;
351 	uint32_t priority;
352 	uint32_t queue_percent;
353 	uint32_t *read_ptr;
354 	uint32_t *write_ptr;
355 	uint32_t __iomem *doorbell_ptr;
356 	uint32_t doorbell_off;
357 	bool is_interop;
358 	bool is_active;
359 	/* Not relevant for user mode queues in cp scheduling */
360 	unsigned int vmid;
361 	/* Relevant only for sdma queues*/
362 	uint32_t sdma_engine_id;
363 	uint32_t sdma_queue_id;
364 	uint32_t sdma_vm_addr;
365 	/* Relevant only for VI */
366 	uint64_t eop_ring_buffer_address;
367 	uint32_t eop_ring_buffer_size;
368 	uint64_t ctx_save_restore_area_address;
369 	uint32_t ctx_save_restore_area_size;
370 	uint32_t ctl_stack_size;
371 	uint64_t tba_addr;
372 	uint64_t tma_addr;
373 };
374 
375 /**
376  * struct queue
377  *
378  * @list: Queue linked list.
379  *
380  * @mqd: The queue MQD.
381  *
382  * @mqd_mem_obj: The MQD local gpu memory object.
383  *
384  * @gart_mqd_addr: The MQD gart mc address.
385  *
386  * @properties: The queue properties.
387  *
388  * @mec: Used only in no cp scheduling mode and identifies to micro engine id
389  *	 that the queue should be execute on.
390  *
391  * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
392  *	  id.
393  *
394  * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
395  *
396  * @process: The kfd process that created this queue.
397  *
398  * @device: The kfd device that created this queue.
399  *
400  * This structure represents user mode compute queues.
401  * It contains all the necessary data to handle such queues.
402  *
403  */
404 
405 struct queue {
406 	struct list_head list;
407 	void *mqd;
408 	struct kfd_mem_obj *mqd_mem_obj;
409 	uint64_t gart_mqd_addr;
410 	struct queue_properties properties;
411 
412 	uint32_t mec;
413 	uint32_t pipe;
414 	uint32_t queue;
415 
416 	unsigned int sdma_id;
417 
418 	struct kfd_process	*process;
419 	struct kfd_dev		*device;
420 };
421 
422 /*
423  * Please read the kfd_mqd_manager.h description.
424  */
425 enum KFD_MQD_TYPE {
426 	KFD_MQD_TYPE_COMPUTE = 0,	/* for no cp scheduling */
427 	KFD_MQD_TYPE_HIQ,		/* for hiq */
428 	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
429 	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
430 	KFD_MQD_TYPE_MAX
431 };
432 
433 struct scheduling_resources {
434 	unsigned int vmid_mask;
435 	enum kfd_queue_type type;
436 	uint64_t queue_mask;
437 	uint64_t gws_mask;
438 	uint32_t oac_mask;
439 	uint32_t gds_heap_base;
440 	uint32_t gds_heap_size;
441 };
442 
443 struct process_queue_manager {
444 	/* data */
445 	struct kfd_process	*process;
446 	struct list_head	queues;
447 	unsigned long		*queue_slot_bitmap;
448 };
449 
450 struct qcm_process_device {
451 	/* The Device Queue Manager that owns this data */
452 	struct device_queue_manager *dqm;
453 	struct process_queue_manager *pqm;
454 	/* Queues list */
455 	struct list_head queues_list;
456 	struct list_head priv_queue_list;
457 
458 	unsigned int queue_count;
459 	unsigned int vmid;
460 	bool is_debug;
461 
462 	/* This flag tells if we should reset all wavefronts on
463 	 * process termination
464 	 */
465 	bool reset_wavefronts;
466 
467 	/*
468 	 * All the memory management data should be here too
469 	 */
470 	uint64_t gds_context_area;
471 	uint32_t sh_mem_config;
472 	uint32_t sh_mem_bases;
473 	uint32_t sh_mem_ape1_base;
474 	uint32_t sh_mem_ape1_limit;
475 	uint32_t page_table_base;
476 	uint32_t gds_size;
477 	uint32_t num_gws;
478 	uint32_t num_oac;
479 	uint32_t sh_hidden_private_base;
480 
481 	/* CWSR memory */
482 	void *cwsr_kaddr;
483 	uint64_t tba_addr;
484 	uint64_t tma_addr;
485 };
486 
487 
488 enum kfd_pdd_bound {
489 	PDD_UNBOUND = 0,
490 	PDD_BOUND,
491 	PDD_BOUND_SUSPENDED,
492 };
493 
494 /* Data that is per-process-per device. */
495 struct kfd_process_device {
496 	/*
497 	 * List of all per-device data for a process.
498 	 * Starts from kfd_process.per_device_data.
499 	 */
500 	struct list_head per_device_list;
501 
502 	/* The device that owns this data. */
503 	struct kfd_dev *dev;
504 
505 	/* The process that owns this kfd_process_device. */
506 	struct kfd_process *process;
507 
508 	/* per-process-per device QCM data structure */
509 	struct qcm_process_device qpd;
510 
511 	/*Apertures*/
512 	uint64_t lds_base;
513 	uint64_t lds_limit;
514 	uint64_t gpuvm_base;
515 	uint64_t gpuvm_limit;
516 	uint64_t scratch_base;
517 	uint64_t scratch_limit;
518 
519 	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
520 	enum kfd_pdd_bound bound;
521 
522 	/* Flag used to tell the pdd has dequeued from the dqm.
523 	 * This is used to prevent dev->dqm->ops.process_termination() from
524 	 * being called twice when it is already called in IOMMU callback
525 	 * function.
526 	 */
527 	bool already_dequeued;
528 };
529 
530 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
531 
532 /* Process data */
533 struct kfd_process {
534 	/*
535 	 * kfd_process are stored in an mm_struct*->kfd_process*
536 	 * hash table (kfd_processes in kfd_process.c)
537 	 */
538 	struct hlist_node kfd_processes;
539 
540 	/*
541 	 * Opaque pointer to mm_struct. We don't hold a reference to
542 	 * it so it should never be dereferenced from here. This is
543 	 * only used for looking up processes by their mm.
544 	 */
545 	void *mm;
546 
547 	struct kref ref;
548 	struct work_struct release_work;
549 
550 	struct mutex mutex;
551 
552 	/*
553 	 * In any process, the thread that started main() is the lead
554 	 * thread and outlives the rest.
555 	 * It is here because amd_iommu_bind_pasid wants a task_struct.
556 	 * It can also be used for safely getting a reference to the
557 	 * mm_struct of the process.
558 	 */
559 	struct task_struct *lead_thread;
560 
561 	/* We want to receive a notification when the mm_struct is destroyed */
562 	struct mmu_notifier mmu_notifier;
563 
564 	/* Use for delayed freeing of kfd_process structure */
565 	struct rcu_head	rcu;
566 
567 	unsigned int pasid;
568 	unsigned int doorbell_index;
569 
570 	/*
571 	 * List of kfd_process_device structures,
572 	 * one for each device the process is using.
573 	 */
574 	struct list_head per_device_data;
575 
576 	struct process_queue_manager pqm;
577 
578 	/*Is the user space process 32 bit?*/
579 	bool is_32bit_user_mode;
580 
581 	/* Event-related data */
582 	struct mutex event_mutex;
583 	/* Event ID allocator and lookup */
584 	struct idr event_idr;
585 	/* Event page */
586 	struct kfd_signal_page *signal_page;
587 	size_t signal_mapped_size;
588 	size_t signal_event_count;
589 	bool signal_event_limit_reached;
590 };
591 
592 /**
593  * Ioctl function type.
594  *
595  * \param filep pointer to file structure.
596  * \param p amdkfd process pointer.
597  * \param data pointer to arg that was copied from user.
598  */
599 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
600 				void *data);
601 
602 struct amdkfd_ioctl_desc {
603 	unsigned int cmd;
604 	int flags;
605 	amdkfd_ioctl_t *func;
606 	unsigned int cmd_drv;
607 	const char *name;
608 };
609 
610 void kfd_process_create_wq(void);
611 void kfd_process_destroy_wq(void);
612 struct kfd_process *kfd_create_process(struct file *filep);
613 struct kfd_process *kfd_get_process(const struct task_struct *);
614 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
615 void kfd_unref_process(struct kfd_process *p);
616 
617 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
618 						struct kfd_process *p);
619 int kfd_bind_processes_to_device(struct kfd_dev *dev);
620 void kfd_unbind_processes_from_device(struct kfd_dev *dev);
621 void kfd_process_iommu_unbind_callback(struct kfd_dev *dev, unsigned int pasid);
622 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
623 							struct kfd_process *p);
624 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
625 							struct kfd_process *p);
626 
627 int kfd_reserved_mem_mmap(struct kfd_process *process,
628 			  struct vm_area_struct *vma);
629 
630 /* Process device data iterator */
631 struct kfd_process_device *kfd_get_first_process_device_data(
632 							struct kfd_process *p);
633 struct kfd_process_device *kfd_get_next_process_device_data(
634 						struct kfd_process *p,
635 						struct kfd_process_device *pdd);
636 bool kfd_has_process_device_data(struct kfd_process *p);
637 
638 /* PASIDs */
639 int kfd_pasid_init(void);
640 void kfd_pasid_exit(void);
641 bool kfd_set_pasid_limit(unsigned int new_limit);
642 unsigned int kfd_get_pasid_limit(void);
643 unsigned int kfd_pasid_alloc(void);
644 void kfd_pasid_free(unsigned int pasid);
645 
646 /* Doorbells */
647 int kfd_doorbell_init(struct kfd_dev *kfd);
648 void kfd_doorbell_fini(struct kfd_dev *kfd);
649 int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma);
650 u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
651 					unsigned int *doorbell_off);
652 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
653 u32 read_kernel_doorbell(u32 __iomem *db);
654 void write_kernel_doorbell(u32 __iomem *db, u32 value);
655 unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
656 					struct kfd_process *process,
657 					unsigned int queue_id);
658 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
659 					struct kfd_process *process);
660 int kfd_alloc_process_doorbells(struct kfd_process *process);
661 void kfd_free_process_doorbells(struct kfd_process *process);
662 
663 /* GTT Sub-Allocator */
664 
665 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
666 			struct kfd_mem_obj **mem_obj);
667 
668 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
669 
670 extern struct device *kfd_device;
671 
672 /* Topology */
673 int kfd_topology_init(void);
674 void kfd_topology_shutdown(void);
675 int kfd_topology_add_device(struct kfd_dev *gpu);
676 int kfd_topology_remove_device(struct kfd_dev *gpu);
677 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
678 						uint32_t proximity_domain);
679 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
680 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
681 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
682 int kfd_numa_node_to_apic_id(int numa_node_id);
683 
684 /* Interrupts */
685 int kfd_interrupt_init(struct kfd_dev *dev);
686 void kfd_interrupt_exit(struct kfd_dev *dev);
687 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
688 bool enqueue_ih_ring_entry(struct kfd_dev *kfd,	const void *ih_ring_entry);
689 bool interrupt_is_wanted(struct kfd_dev *dev, const uint32_t *ih_ring_entry);
690 
691 /* Power Management */
692 void kgd2kfd_suspend(struct kfd_dev *kfd);
693 int kgd2kfd_resume(struct kfd_dev *kfd);
694 
695 /* amdkfd Apertures */
696 int kfd_init_apertures(struct kfd_process *process);
697 
698 /* Queue Context Management */
699 int init_queue(struct queue **q, const struct queue_properties *properties);
700 void uninit_queue(struct queue *q);
701 void print_queue_properties(struct queue_properties *q);
702 void print_queue(struct queue *q);
703 
704 struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
705 					struct kfd_dev *dev);
706 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
707 		struct kfd_dev *dev);
708 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
709 		struct kfd_dev *dev);
710 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
711 void device_queue_manager_uninit(struct device_queue_manager *dqm);
712 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
713 					enum kfd_queue_type type);
714 void kernel_queue_uninit(struct kernel_queue *kq);
715 
716 /* Process Queue Manager */
717 struct process_queue_node {
718 	struct queue *q;
719 	struct kernel_queue *kq;
720 	struct list_head process_queue_list;
721 };
722 
723 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
724 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
725 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
726 void pqm_uninit(struct process_queue_manager *pqm);
727 int pqm_create_queue(struct process_queue_manager *pqm,
728 			    struct kfd_dev *dev,
729 			    struct file *f,
730 			    struct queue_properties *properties,
731 			    unsigned int *qid);
732 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
733 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
734 			struct queue_properties *p);
735 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
736 						unsigned int qid);
737 
738 int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
739 				unsigned int fence_value,
740 				unsigned int timeout_ms);
741 
742 /* Packet Manager */
743 
744 #define KFD_FENCE_COMPLETED (100)
745 #define KFD_FENCE_INIT   (10)
746 
747 struct packet_manager {
748 	struct device_queue_manager *dqm;
749 	struct kernel_queue *priv_queue;
750 	struct mutex lock;
751 	bool allocated;
752 	struct kfd_mem_obj *ib_buffer_obj;
753 	unsigned int ib_size_bytes;
754 };
755 
756 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
757 void pm_uninit(struct packet_manager *pm);
758 int pm_send_set_resources(struct packet_manager *pm,
759 				struct scheduling_resources *res);
760 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
761 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
762 				uint32_t fence_value);
763 
764 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
765 			enum kfd_unmap_queues_filter mode,
766 			uint32_t filter_param, bool reset,
767 			unsigned int sdma_engine);
768 
769 void pm_release_ib(struct packet_manager *pm);
770 
771 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
772 
773 /* Events */
774 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
775 extern const struct kfd_device_global_init_class device_global_init_class_cik;
776 
777 void kfd_event_init_process(struct kfd_process *p);
778 void kfd_event_free_process(struct kfd_process *p);
779 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
780 int kfd_wait_on_events(struct kfd_process *p,
781 		       uint32_t num_events, void __user *data,
782 		       bool all, uint32_t user_timeout_ms,
783 		       uint32_t *wait_result);
784 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
785 				uint32_t valid_id_bits);
786 void kfd_signal_iommu_event(struct kfd_dev *dev,
787 		unsigned int pasid, unsigned long address,
788 		bool is_write_requested, bool is_execute_requested);
789 void kfd_signal_hw_exception_event(unsigned int pasid);
790 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
791 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
792 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
793 		     uint32_t event_type, bool auto_reset, uint32_t node_id,
794 		     uint32_t *event_id, uint32_t *event_trigger_data,
795 		     uint64_t *event_page_offset, uint32_t *event_slot_index);
796 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
797 
798 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
799 
800 /* Debugfs */
801 #if defined(CONFIG_DEBUG_FS)
802 
803 void kfd_debugfs_init(void);
804 void kfd_debugfs_fini(void);
805 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
806 int pqm_debugfs_mqds(struct seq_file *m, void *data);
807 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
808 int dqm_debugfs_hqds(struct seq_file *m, void *data);
809 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
810 int pm_debugfs_runlist(struct seq_file *m, void *data);
811 
812 #else
813 
814 static inline void kfd_debugfs_init(void) {}
815 static inline void kfd_debugfs_fini(void) {}
816 
817 #endif
818 
819 #endif
820