Linux Audio

Check our new training course

Loading...
v4.6
  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 <kgd_kfd_interface.h>
 
 
 
 
 
 
 35
 36#define KFD_SYSFS_FILE_MODE 0444
 37
 38#define KFD_MMAP_DOORBELL_MASK 0x8000000000000
 39#define KFD_MMAP_EVENTS_MASK 0x4000000000000
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 40
 41/*
 42 * When working with cp scheduler we should assign the HIQ manually or via
 43 * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot
 44 * definitions for Kaveri. In Kaveri only the first ME queues participates
 45 * in the cp scheduling taking that in mind we set the HIQ slot in the
 46 * second ME.
 47 */
 48#define KFD_CIK_HIQ_PIPE 4
 49#define KFD_CIK_HIQ_QUEUE 0
 50
 51/* GPU ID hash width in bits */
 52#define KFD_GPU_ID_HASH_WIDTH 16
 53
 54/* Macro for allocating structures */
 55#define kfd_alloc_struct(ptr_to_struct)	\
 56	((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
 57
 58#define KFD_MAX_NUM_OF_PROCESSES 512
 59#define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
 60
 61/*
 62 * Kernel module parameter to specify maximum number of supported queues per
 63 * device
 
 
 64 */
 65extern int max_num_of_queues_per_device;
 
 66
 67#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096
 68#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE		\
 69	(KFD_MAX_NUM_OF_PROCESSES *			\
 70			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
 71
 72#define KFD_KERNEL_QUEUE_SIZE 2048
 73
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 74/* Kernel module parameter to specify the scheduling policy */
 75extern int sched_policy;
 76
 77/*
 
 
 
 
 
 
 
 
 78 * Kernel module parameter to specify whether to send sigterm to HSA process on
 79 * unhandled exception
 80 */
 81extern int send_sigterm;
 82
 83/**
 84 * enum kfd_sched_policy
 85 *
 86 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp)
 87 * scheduling. In this scheduling mode we're using the firmware code to
 88 * schedule the user mode queues and kernel queues such as HIQ and DIQ.
 89 * the HIQ queue is used as a special queue that dispatches the configuration
 90 * to the cp and the user mode queues list that are currently running.
 91 * the DIQ queue is a debugging queue that dispatches debugging commands to the
 92 * firmware.
 93 * in this scheduling mode user mode queues over subscription feature is
 94 * enabled.
 95 *
 96 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over
 97 * subscription feature disabled.
 98 *
 99 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly
100 * set the command processor registers and sets the queues "manually". This
101 * mode is used *ONLY* for debugging proposes.
102 *
103 */
104enum kfd_sched_policy {
105	KFD_SCHED_POLICY_HWS = 0,
106	KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION,
107	KFD_SCHED_POLICY_NO_HWS
108};
 
 
 
 
 
109
110enum cache_policy {
111	cache_policy_coherent,
112	cache_policy_noncoherent
113};
114
115enum asic_family_type {
116	CHIP_KAVERI = 0,
117	CHIP_CARRIZO
118};
119
120struct kfd_event_interrupt_class {
121	bool (*interrupt_isr)(struct kfd_dev *dev,
122				const uint32_t *ih_ring_entry);
 
123	void (*interrupt_wq)(struct kfd_dev *dev,
124				const uint32_t *ih_ring_entry);
125};
126
127struct kfd_device_info {
128	unsigned int asic_family;
 
129	const struct kfd_event_interrupt_class *event_interrupt_class;
130	unsigned int max_pasid_bits;
131	unsigned int max_no_of_hqd;
 
132	size_t ih_ring_entry_size;
133	uint8_t num_of_watch_points;
134	uint16_t mqd_size_aligned;
 
 
 
 
 
 
 
135};
136
137struct kfd_mem_obj {
138	uint32_t range_start;
139	uint32_t range_end;
140	uint64_t gpu_addr;
141	uint32_t *cpu_ptr;
 
 
 
 
 
 
 
142};
143
144struct kfd_dev {
145	struct kgd_dev *kgd;
146
147	const struct kfd_device_info *device_info;
148	struct pci_dev *pdev;
 
149
150	unsigned int id;		/* topology stub index */
151
152	phys_addr_t doorbell_base;	/* Start of actual doorbells used by
153					 * KFD. It is aligned for mapping
154					 * into user mode
155					 */
156	size_t doorbell_id_offset;	/* Doorbell offset (from KFD doorbell
157					 * to HW doorbell, GFX reserved some
158					 * at the start)
159					 */
160	size_t doorbell_process_limit;	/* Number of processes we have doorbell
161					 * space for.
162					 */
163	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
164					   * page used by kernel queue
165					   */
166
167	struct kgd2kfd_shared_resources shared_resources;
 
168
169	const struct kfd2kgd_calls *kfd2kgd;
170	struct mutex doorbell_mutex;
171	DECLARE_BITMAP(doorbell_available_index,
172			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
173
174	void *gtt_mem;
175	uint64_t gtt_start_gpu_addr;
176	void *gtt_start_cpu_ptr;
177	void *gtt_sa_bitmap;
178	struct mutex gtt_sa_lock;
179	unsigned int gtt_sa_chunk_size;
180	unsigned int gtt_sa_num_of_chunks;
181
182	/* Interrupts */
183	void *interrupt_ring;
184	size_t interrupt_ring_size;
185	atomic_t interrupt_ring_rptr;
186	atomic_t interrupt_ring_wptr;
187	struct work_struct interrupt_work;
188	spinlock_t interrupt_lock;
189
190	/* QCM Device instance */
191	struct device_queue_manager *dqm;
192
193	bool init_complete;
194	/*
195	 * Interrupts of interest to KFD are copied
196	 * from the HW ring into a SW ring.
197	 */
198	bool interrupts_active;
199
200	/* Debug manager */
201	struct kfd_dbgmgr           *dbgmgr;
202};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
203
204/* KGD2KFD callbacks */
205void kgd2kfd_exit(void);
206struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
207			struct pci_dev *pdev, const struct kfd2kgd_calls *f2g);
208bool kgd2kfd_device_init(struct kfd_dev *kfd,
209			const struct kgd2kfd_shared_resources *gpu_resources);
210void kgd2kfd_device_exit(struct kfd_dev *kfd);
 
 
 
 
 
 
 
 
 
 
211
212enum kfd_mempool {
213	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
214	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
215	KFD_MEMPOOL_FRAMEBUFFER = 3,
216};
217
218/* Character device interface */
219int kfd_chardev_init(void);
220void kfd_chardev_exit(void);
221struct device *kfd_chardev(void);
222
223/**
224 * enum kfd_preempt_type_filter
225 *
226 * @KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE: Preempts single queue.
227 *
228 * @KFD_PRERMPT_TYPE_FILTER_ALL_QUEUES: Preempts all queues in the
229 *						running queues list.
230 *
231 * @KFD_PRERMPT_TYPE_FILTER_BY_PASID: Preempts queues that belongs to
232 *						specific process.
233 *
234 */
235enum kfd_preempt_type_filter {
236	KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE,
237	KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES,
238	KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES,
239	KFD_PREEMPT_TYPE_FILTER_BY_PASID
240};
241
242enum kfd_preempt_type {
243	KFD_PREEMPT_TYPE_WAVEFRONT,
244	KFD_PREEMPT_TYPE_WAVEFRONT_RESET
245};
246
247/**
248 * enum kfd_queue_type
249 *
250 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
251 *
252 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
253 *
254 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
255 *
256 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
 
 
257 */
258enum kfd_queue_type  {
259	KFD_QUEUE_TYPE_COMPUTE,
260	KFD_QUEUE_TYPE_SDMA,
261	KFD_QUEUE_TYPE_HIQ,
262	KFD_QUEUE_TYPE_DIQ
 
263};
264
265enum kfd_queue_format {
266	KFD_QUEUE_FORMAT_PM4,
267	KFD_QUEUE_FORMAT_AQL
268};
269
 
 
 
 
 
270/**
271 * struct queue_properties
272 *
273 * @type: The queue type.
274 *
275 * @queue_id: Queue identifier.
276 *
277 * @queue_address: Queue ring buffer address.
278 *
279 * @queue_size: Queue ring buffer size.
280 *
281 * @priority: Defines the queue priority relative to other queues in the
282 * process.
283 * This is just an indication and HW scheduling may override the priority as
284 * necessary while keeping the relative prioritization.
285 * the priority granularity is from 0 to f which f is the highest priority.
286 * currently all queues are initialized with the highest priority.
287 *
288 * @queue_percent: This field is partially implemented and currently a zero in
289 * this field defines that the queue is non active.
290 *
291 * @read_ptr: User space address which points to the number of dwords the
292 * cp read from the ring buffer. This field updates automatically by the H/W.
293 *
294 * @write_ptr: Defines the number of dwords written to the ring buffer.
295 *
296 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
297 * the queue ring buffer. This field should be similar to write_ptr and the user
298 * should update this field after he updated the write_ptr.
299 *
300 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
301 *
302 * @is_interop: Defines if this is a interop queue. Interop queue means that the
303 * queue can access both graphics and compute resources.
 
 
 
 
 
 
304 *
305 * @is_active: Defines if the queue is active or not.
 
 
306 *
307 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
308 * of the queue.
309 *
310 * This structure represents the queue properties for each queue no matter if
311 * it's user mode or kernel mode queue.
312 *
313 */
314struct queue_properties {
315	enum kfd_queue_type type;
316	enum kfd_queue_format format;
317	unsigned int queue_id;
318	uint64_t queue_address;
319	uint64_t  queue_size;
320	uint32_t priority;
321	uint32_t queue_percent;
322	uint32_t *read_ptr;
323	uint32_t *write_ptr;
324	uint32_t __iomem *doorbell_ptr;
325	uint32_t doorbell_off;
326	bool is_interop;
 
327	bool is_active;
 
328	/* Not relevant for user mode queues in cp scheduling */
329	unsigned int vmid;
330	/* Relevant only for sdma queues*/
331	uint32_t sdma_engine_id;
332	uint32_t sdma_queue_id;
333	uint32_t sdma_vm_addr;
334	/* Relevant only for VI */
335	uint64_t eop_ring_buffer_address;
336	uint32_t eop_ring_buffer_size;
337	uint64_t ctx_save_restore_area_address;
338	uint32_t ctx_save_restore_area_size;
 
 
 
 
 
 
339};
340
 
 
 
 
 
341/**
342 * struct queue
343 *
344 * @list: Queue linked list.
345 *
346 * @mqd: The queue MQD.
347 *
348 * @mqd_mem_obj: The MQD local gpu memory object.
349 *
350 * @gart_mqd_addr: The MQD gart mc address.
351 *
352 * @properties: The queue properties.
353 *
354 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
355 * that the queue should be execute on.
356 *
357 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe id.
 
358 *
359 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
360 *
361 * @process: The kfd process that created this queue.
362 *
363 * @device: The kfd device that created this queue.
364 *
 
 
 
365 * This structure represents user mode compute queues.
366 * It contains all the necessary data to handle such queues.
367 *
368 */
369
370struct queue {
371	struct list_head list;
372	void *mqd;
373	struct kfd_mem_obj *mqd_mem_obj;
374	uint64_t gart_mqd_addr;
375	struct queue_properties properties;
376
377	uint32_t mec;
378	uint32_t pipe;
379	uint32_t queue;
380
381	unsigned int sdma_id;
 
382
383	struct kfd_process	*process;
384	struct kfd_dev		*device;
 
 
 
 
385};
386
387/*
388 * Please read the kfd_mqd_manager.h description.
389 */
390enum KFD_MQD_TYPE {
391	KFD_MQD_TYPE_COMPUTE = 0,	/* for no cp scheduling */
392	KFD_MQD_TYPE_HIQ,		/* for hiq */
393	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
394	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
 
395	KFD_MQD_TYPE_MAX
396};
397
 
 
 
 
 
 
398struct scheduling_resources {
399	unsigned int vmid_mask;
400	enum kfd_queue_type type;
401	uint64_t queue_mask;
402	uint64_t gws_mask;
403	uint32_t oac_mask;
404	uint32_t gds_heap_base;
405	uint32_t gds_heap_size;
406};
407
408struct process_queue_manager {
409	/* data */
410	struct kfd_process	*process;
411	unsigned int		num_concurrent_processes;
412	struct list_head	queues;
413	unsigned long		*queue_slot_bitmap;
414};
415
416struct qcm_process_device {
417	/* The Device Queue Manager that owns this data */
418	struct device_queue_manager *dqm;
419	struct process_queue_manager *pqm;
420	/* Queues list */
421	struct list_head queues_list;
422	struct list_head priv_queue_list;
423
424	unsigned int queue_count;
425	unsigned int vmid;
426	bool is_debug;
427	/*
428	 * All the memory management data should be here too
 
 
 
 
 
 
 
 
 
 
429	 */
 
 
 
430	uint64_t gds_context_area;
 
 
431	uint32_t sh_mem_config;
432	uint32_t sh_mem_bases;
433	uint32_t sh_mem_ape1_base;
434	uint32_t sh_mem_ape1_limit;
435	uint32_t page_table_base;
436	uint32_t gds_size;
437	uint32_t num_gws;
438	uint32_t num_oac;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
439};
440
 
 
 
 
 
 
 
 
 
441/* Data that is per-process-per device. */
442struct kfd_process_device {
443	/*
444	 * List of all per-device data for a process.
445	 * Starts from kfd_process.per_device_data.
446	 */
447	struct list_head per_device_list;
448
449	/* The device that owns this data. */
450	struct kfd_dev *dev;
451
 
 
452
453	/* per-process-per device QCM data structure */
454	struct qcm_process_device qpd;
455
456	/*Apertures*/
457	uint64_t lds_base;
458	uint64_t lds_limit;
459	uint64_t gpuvm_base;
460	uint64_t gpuvm_limit;
461	uint64_t scratch_base;
462	uint64_t scratch_limit;
463
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
464	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
465	bool bound;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
466
467	/* This flag tells if we should reset all
468	 * wavefronts on process termination
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
469	 */
470	bool reset_wavefronts;
 
 
 
 
 
 
 
 
 
471};
472
473#define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
474
 
 
 
 
 
 
 
 
 
 
 
 
475/* Process data */
476struct kfd_process {
477	/*
478	 * kfd_process are stored in an mm_struct*->kfd_process*
479	 * hash table (kfd_processes in kfd_process.c)
480	 */
481	struct hlist_node kfd_processes;
482
483	struct mm_struct *mm;
 
 
 
 
 
 
 
 
484
485	struct mutex mutex;
486
487	/*
488	 * In any process, the thread that started main() is the lead
489	 * thread and outlives the rest.
490	 * It is here because amd_iommu_bind_pasid wants a task_struct.
 
 
491	 */
492	struct task_struct *lead_thread;
493
494	/* We want to receive a notification when the mm_struct is destroyed */
495	struct mmu_notifier mmu_notifier;
496
497	/* Use for delayed freeing of kfd_process structure */
498	struct rcu_head	rcu;
499
500	unsigned int pasid;
501
502	/*
503	 * List of kfd_process_device structures,
504	 * one for each device the process is using.
505	 */
506	struct list_head per_device_data;
 
507
508	struct process_queue_manager pqm;
509
510	/* The process's queues. */
511	size_t queue_array_size;
512
513	/* Size is queue_array_size, up to MAX_PROCESS_QUEUES. */
514	struct kfd_queue **queues;
515
516	/*Is the user space process 32 bit?*/
517	bool is_32bit_user_mode;
518
519	/* Event-related data */
520	struct mutex event_mutex;
521	/* All events in process hashed by ID, linked on kfd_event.events. */
522	DECLARE_HASHTABLE(events, 4);
523	struct list_head signal_event_pages;	/* struct slot_page_header.
524								event_pages */
525	u32 next_nonsignal_event_id;
526	size_t signal_event_count;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
527};
528
 
 
 
 
529/**
530 * Ioctl function type.
 
 
 
 
531 *
532 * \param filep pointer to file structure.
533 * \param p amdkfd process pointer.
534 * \param data pointer to arg that was copied from user.
535 */
536typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
537				void *data);
538
539struct amdkfd_ioctl_desc {
540	unsigned int cmd;
541	int flags;
542	amdkfd_ioctl_t *func;
543	unsigned int cmd_drv;
544	const char *name;
545};
 
546
547void kfd_process_create_wq(void);
548void kfd_process_destroy_wq(void);
549struct kfd_process *kfd_create_process(const struct task_struct *);
550struct kfd_process *kfd_get_process(const struct task_struct *);
551struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
 
552
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
553struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
554							struct kfd_process *p);
555void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid);
556struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
557							struct kfd_process *p);
558struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
559							struct kfd_process *p);
560
561/* Process device data iterator */
562struct kfd_process_device *kfd_get_first_process_device_data(struct kfd_process *p);
563struct kfd_process_device *kfd_get_next_process_device_data(struct kfd_process *p,
564						struct kfd_process_device *pdd);
565bool kfd_has_process_device_data(struct kfd_process *p);
 
 
 
 
 
 
 
566
567/* PASIDs */
568int kfd_pasid_init(void);
569void kfd_pasid_exit(void);
570bool kfd_set_pasid_limit(unsigned int new_limit);
571unsigned int kfd_get_pasid_limit(void);
572unsigned int kfd_pasid_alloc(void);
573void kfd_pasid_free(unsigned int pasid);
574
575/* Doorbells */
576void kfd_doorbell_init(struct kfd_dev *kfd);
577int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma);
578u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
 
 
 
579					unsigned int *doorbell_off);
580void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
581u32 read_kernel_doorbell(u32 __iomem *db);
582void write_kernel_doorbell(u32 __iomem *db, u32 value);
583unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
584					struct kfd_process *process,
585					unsigned int queue_id);
586
 
 
 
 
 
587/* GTT Sub-Allocator */
588
589int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
590			struct kfd_mem_obj **mem_obj);
591
592int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
593
594extern struct device *kfd_device;
595
 
 
 
 
 
 
596/* Topology */
597int kfd_topology_init(void);
598void kfd_topology_shutdown(void);
599int kfd_topology_add_device(struct kfd_dev *gpu);
600int kfd_topology_remove_device(struct kfd_dev *gpu);
 
 
 
601struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
602struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
603struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx);
 
 
 
604
605/* Interrupts */
606int kfd_interrupt_init(struct kfd_dev *dev);
607void kfd_interrupt_exit(struct kfd_dev *dev);
608void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
609bool enqueue_ih_ring_entry(struct kfd_dev *kfd,	const void *ih_ring_entry);
610bool interrupt_is_wanted(struct kfd_dev *dev, const uint32_t *ih_ring_entry);
611
612/* Power Management */
613void kgd2kfd_suspend(struct kfd_dev *kfd);
614int kgd2kfd_resume(struct kfd_dev *kfd);
615
616/* amdkfd Apertures */
617int kfd_init_apertures(struct kfd_process *process);
618
619/* Queue Context Management */
620inline uint32_t lower_32(uint64_t x);
621inline uint32_t upper_32(uint64_t x);
622struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd);
623inline uint32_t get_sdma_base_addr(struct cik_sdma_rlc_registers *m);
624
625int init_queue(struct queue **q, struct queue_properties properties);
 
626void uninit_queue(struct queue *q);
627void print_queue_properties(struct queue_properties *q);
628void print_queue(struct queue *q);
629
630struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
631					struct kfd_dev *dev);
632struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
633		struct kfd_dev *dev);
 
 
634struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
635		struct kfd_dev *dev);
 
 
 
 
 
 
636struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
637void device_queue_manager_uninit(struct device_queue_manager *dqm);
638struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
639					enum kfd_queue_type type);
640void kernel_queue_uninit(struct kernel_queue *kq);
 
641
642/* Process Queue Manager */
643struct process_queue_node {
644	struct queue *q;
645	struct kernel_queue *kq;
646	struct list_head process_queue_list;
647};
648
 
 
649int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
650void pqm_uninit(struct process_queue_manager *pqm);
651int pqm_create_queue(struct process_queue_manager *pqm,
652			    struct kfd_dev *dev,
653			    struct file *f,
654			    struct queue_properties *properties,
655			    unsigned int flags,
656			    enum kfd_queue_type type,
657			    unsigned int *qid);
658int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
659int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
660			struct queue_properties *p);
 
 
 
 
661struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
662						unsigned int qid);
663
664int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
665				unsigned int fence_value,
666				unsigned long timeout);
 
 
 
 
 
 
 
667
668/* Packet Manager */
669
670#define KFD_HIQ_TIMEOUT (500)
671
672#define KFD_FENCE_COMPLETED (100)
673#define KFD_FENCE_INIT   (10)
674#define KFD_UNMAP_LATENCY (150)
675
676struct packet_manager {
677	struct device_queue_manager *dqm;
678	struct kernel_queue *priv_queue;
679	struct mutex lock;
680	bool allocated;
681	struct kfd_mem_obj *ib_buffer_obj;
 
 
 
 
682};
683
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
684int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
685void pm_uninit(struct packet_manager *pm);
686int pm_send_set_resources(struct packet_manager *pm,
687				struct scheduling_resources *res);
688int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
689int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
690				uint32_t fence_value);
691
692int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
693			enum kfd_preempt_type_filter mode,
694			uint32_t filter_param, bool reset,
695			unsigned int sdma_engine);
696
697void pm_release_ib(struct packet_manager *pm);
698
 
 
 
699uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
700phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
701					struct kfd_process *process);
702
703/* Events */
704extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
705extern const struct kfd_device_global_init_class device_global_init_class_cik;
706
707enum kfd_event_wait_result {
708	KFD_WAIT_COMPLETE,
709	KFD_WAIT_TIMEOUT,
710	KFD_WAIT_ERROR
711};
712
713void kfd_event_init_process(struct kfd_process *p);
714void kfd_event_free_process(struct kfd_process *p);
715int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
716int kfd_wait_on_events(struct kfd_process *p,
717		       uint32_t num_events, void __user *data,
718		       bool all, uint32_t user_timeout_ms,
719		       enum kfd_event_wait_result *wait_result);
720void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
721				uint32_t valid_id_bits);
722void kfd_signal_iommu_event(struct kfd_dev *dev,
723		unsigned int pasid, unsigned long address,
724		bool is_write_requested, bool is_execute_requested);
725void kfd_signal_hw_exception_event(unsigned int pasid);
726int kfd_set_event(struct kfd_process *p, uint32_t event_id);
727int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
 
 
728int kfd_event_create(struct file *devkfd, struct kfd_process *p,
729		     uint32_t event_type, bool auto_reset, uint32_t node_id,
730		     uint32_t *event_id, uint32_t *event_trigger_data,
731		     uint64_t *event_page_offset, uint32_t *event_slot_index);
732int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
733
 
 
 
 
 
 
 
 
 
734int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
735
736#endif
v5.14.15
   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 <linux/sysfs.h>
  39#include <linux/device_cgroup.h>
  40#include <drm/drm_file.h>
  41#include <drm/drm_drv.h>
  42#include <drm/drm_device.h>
  43#include <drm/drm_ioctl.h>
  44#include <kgd_kfd_interface.h>
  45#include <linux/swap.h>
  46
  47#include "amd_shared.h"
  48#include "amdgpu.h"
  49
  50#define KFD_MAX_RING_ENTRY_SIZE	8
  51
  52#define KFD_SYSFS_FILE_MODE 0444
  53
  54/* GPU ID hash width in bits */
  55#define KFD_GPU_ID_HASH_WIDTH 16
  56
  57/* Use upper bits of mmap offset to store KFD driver specific information.
  58 * BITS[63:62] - Encode MMAP type
  59 * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to
  60 * BITS[45:0]  - MMAP offset value
  61 *
  62 * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these
  63 *  defines are w.r.t to PAGE_SIZE
  64 */
  65#define KFD_MMAP_TYPE_SHIFT	62
  66#define KFD_MMAP_TYPE_MASK	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
  67#define KFD_MMAP_TYPE_DOORBELL	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
  68#define KFD_MMAP_TYPE_EVENTS	(0x2ULL << KFD_MMAP_TYPE_SHIFT)
  69#define KFD_MMAP_TYPE_RESERVED_MEM	(0x1ULL << KFD_MMAP_TYPE_SHIFT)
  70#define KFD_MMAP_TYPE_MMIO	(0x0ULL << KFD_MMAP_TYPE_SHIFT)
  71
  72#define KFD_MMAP_GPU_ID_SHIFT 46
  73#define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \
  74				<< KFD_MMAP_GPU_ID_SHIFT)
  75#define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\
  76				& KFD_MMAP_GPU_ID_MASK)
  77#define KFD_MMAP_GET_GPU_ID(offset)    ((offset & KFD_MMAP_GPU_ID_MASK) \
  78				>> KFD_MMAP_GPU_ID_SHIFT)
  79
  80/*
  81 * When working with cp scheduler we should assign the HIQ manually or via
  82 * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot
  83 * definitions for Kaveri. In Kaveri only the first ME queues participates
  84 * in the cp scheduling taking that in mind we set the HIQ slot in the
  85 * second ME.
  86 */
  87#define KFD_CIK_HIQ_PIPE 4
  88#define KFD_CIK_HIQ_QUEUE 0
  89
 
 
 
  90/* Macro for allocating structures */
  91#define kfd_alloc_struct(ptr_to_struct)	\
  92	((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
  93
  94#define KFD_MAX_NUM_OF_PROCESSES 512
  95#define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
  96
  97/*
  98 * Size of the per-process TBA+TMA buffer: 2 pages
  99 *
 100 * The first page is the TBA used for the CWSR ISA code. The second
 101 * page is used as TMA for user-mode trap handler setup in daisy-chain mode.
 102 */
 103#define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
 104#define KFD_CWSR_TMA_OFFSET PAGE_SIZE
 105
 
 106#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE		\
 107	(KFD_MAX_NUM_OF_PROCESSES *			\
 108			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
 109
 110#define KFD_KERNEL_QUEUE_SIZE 2048
 111
 112#define KFD_UNMAP_LATENCY_MS	(4000)
 113
 114/*
 115 * 512 = 0x200
 116 * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
 117 * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.
 118 * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC
 119 * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in
 120 * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.
 121 */
 122#define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
 123
 124
 125/*
 126 * Kernel module parameter to specify maximum number of supported queues per
 127 * device
 128 */
 129extern int max_num_of_queues_per_device;
 130
 131
 132/* Kernel module parameter to specify the scheduling policy */
 133extern int sched_policy;
 134
 135/*
 136 * Kernel module parameter to specify the maximum process
 137 * number per HW scheduler
 138 */
 139extern int hws_max_conc_proc;
 140
 141extern int cwsr_enable;
 142
 143/*
 144 * Kernel module parameter to specify whether to send sigterm to HSA process on
 145 * unhandled exception
 146 */
 147extern int send_sigterm;
 148
 149/*
 150 * This kernel module is used to simulate large bar machine on non-large bar
 151 * enabled machines.
 152 */
 153extern int debug_largebar;
 154
 155/*
 156 * Ignore CRAT table during KFD initialization, can be used to work around
 157 * broken CRAT tables on some AMD systems
 158 */
 159extern int ignore_crat;
 160
 161/* Set sh_mem_config.retry_disable on GFX v9 */
 162extern int amdgpu_noretry;
 163
 164/* Halt if HWS hang is detected */
 165extern int halt_if_hws_hang;
 166
 167/* Whether MEC FW support GWS barriers */
 168extern bool hws_gws_support;
 169
 170/* Queue preemption timeout in ms */
 171extern int queue_preemption_timeout_ms;
 172
 173/*
 174 * Don't evict process queues on vm fault
 175 */
 176extern int amdgpu_no_queue_eviction_on_vm_fault;
 177
 178/* Enable eviction debug messages */
 179extern bool debug_evictions;
 180
 181enum cache_policy {
 182	cache_policy_coherent,
 183	cache_policy_noncoherent
 184};
 185
 186#define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10)
 
 
 
 187
 188struct kfd_event_interrupt_class {
 189	bool (*interrupt_isr)(struct kfd_dev *dev,
 190			const uint32_t *ih_ring_entry, uint32_t *patched_ihre,
 191			bool *patched_flag);
 192	void (*interrupt_wq)(struct kfd_dev *dev,
 193			const uint32_t *ih_ring_entry);
 194};
 195
 196struct kfd_device_info {
 197	enum amd_asic_type asic_family;
 198	const char *asic_name;
 199	const struct kfd_event_interrupt_class *event_interrupt_class;
 200	unsigned int max_pasid_bits;
 201	unsigned int max_no_of_hqd;
 202	unsigned int doorbell_size;
 203	size_t ih_ring_entry_size;
 204	uint8_t num_of_watch_points;
 205	uint16_t mqd_size_aligned;
 206	bool supports_cwsr;
 207	bool needs_iommu_device;
 208	bool needs_pci_atomics;
 209	uint32_t no_atomic_fw_version;
 210	unsigned int num_sdma_engines;
 211	unsigned int num_xgmi_sdma_engines;
 212	unsigned int num_sdma_queues_per_engine;
 213};
 214
 215struct kfd_mem_obj {
 216	uint32_t range_start;
 217	uint32_t range_end;
 218	uint64_t gpu_addr;
 219	uint32_t *cpu_ptr;
 220	void *gtt_mem;
 221};
 222
 223struct kfd_vmid_info {
 224	uint32_t first_vmid_kfd;
 225	uint32_t last_vmid_kfd;
 226	uint32_t vmid_num_kfd;
 227};
 228
 229struct kfd_dev {
 230	struct kgd_dev *kgd;
 231
 232	const struct kfd_device_info *device_info;
 233	struct pci_dev *pdev;
 234	struct drm_device *ddev;
 235
 236	unsigned int id;		/* topology stub index */
 237
 238	phys_addr_t doorbell_base;	/* Start of actual doorbells used by
 239					 * KFD. It is aligned for mapping
 240					 * into user mode
 241					 */
 242	size_t doorbell_base_dw_offset;	/* Offset from the start of the PCI
 243					 * doorbell BAR to the first KFD
 244					 * doorbell in dwords. GFX reserves
 245					 * the segment before this offset.
 
 
 246					 */
 247	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
 248					   * page used by kernel queue
 249					   */
 250
 251	struct kgd2kfd_shared_resources shared_resources;
 252	struct kfd_vmid_info vm_info;
 253
 254	const struct kfd2kgd_calls *kfd2kgd;
 255	struct mutex doorbell_mutex;
 256	DECLARE_BITMAP(doorbell_available_index,
 257			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
 258
 259	void *gtt_mem;
 260	uint64_t gtt_start_gpu_addr;
 261	void *gtt_start_cpu_ptr;
 262	void *gtt_sa_bitmap;
 263	struct mutex gtt_sa_lock;
 264	unsigned int gtt_sa_chunk_size;
 265	unsigned int gtt_sa_num_of_chunks;
 266
 267	/* Interrupts */
 268	struct kfifo ih_fifo;
 269	struct workqueue_struct *ih_wq;
 
 
 270	struct work_struct interrupt_work;
 271	spinlock_t interrupt_lock;
 272
 273	/* QCM Device instance */
 274	struct device_queue_manager *dqm;
 275
 276	bool init_complete;
 277	/*
 278	 * Interrupts of interest to KFD are copied
 279	 * from the HW ring into a SW ring.
 280	 */
 281	bool interrupts_active;
 282
 283	/* Debug manager */
 284	struct kfd_dbgmgr *dbgmgr;
 285
 286	/* Firmware versions */
 287	uint16_t mec_fw_version;
 288	uint16_t mec2_fw_version;
 289	uint16_t sdma_fw_version;
 290
 291	/* Maximum process number mapped to HW scheduler */
 292	unsigned int max_proc_per_quantum;
 293
 294	/* CWSR */
 295	bool cwsr_enabled;
 296	const void *cwsr_isa;
 297	unsigned int cwsr_isa_size;
 298
 299	/* xGMI */
 300	uint64_t hive_id;
 301
 302	bool pci_atomic_requested;
 303
 304	/* Use IOMMU v2 flag */
 305	bool use_iommu_v2;
 306
 307	/* SRAM ECC flag */
 308	atomic_t sram_ecc_flag;
 309
 310	/* Compute Profile ref. count */
 311	atomic_t compute_profile;
 312
 313	/* Global GWS resource shared between processes */
 314	void *gws;
 315
 316	/* Clients watching SMI events */
 317	struct list_head smi_clients;
 318	spinlock_t smi_lock;
 319
 320	uint32_t reset_seq_num;
 321
 322	struct ida doorbell_ida;
 323	unsigned int max_doorbell_slices;
 324
 325	int noretry;
 326
 327	/* HMM page migration MEMORY_DEVICE_PRIVATE mapping */
 328	struct dev_pagemap pgmap;
 329};
 330
 331enum kfd_mempool {
 332	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
 333	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
 334	KFD_MEMPOOL_FRAMEBUFFER = 3,
 335};
 336
 337/* Character device interface */
 338int kfd_chardev_init(void);
 339void kfd_chardev_exit(void);
 340struct device *kfd_chardev(void);
 341
 342/**
 343 * enum kfd_unmap_queues_filter - Enum for queue filters.
 344 *
 345 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
 346 *
 347 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
 348 *						running queues list.
 349 *
 350 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
 351 *						specific process.
 352 *
 353 */
 354enum kfd_unmap_queues_filter {
 355	KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
 356	KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
 357	KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
 358	KFD_UNMAP_QUEUES_FILTER_BY_PASID
 
 
 
 
 
 359};
 360
 361/**
 362 * enum kfd_queue_type - Enum for various queue types.
 363 *
 364 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
 365 *
 366 * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.
 367 *
 368 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
 369 *
 370 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
 371 *
 372 * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.
 373 */
 374enum kfd_queue_type  {
 375	KFD_QUEUE_TYPE_COMPUTE,
 376	KFD_QUEUE_TYPE_SDMA,
 377	KFD_QUEUE_TYPE_HIQ,
 378	KFD_QUEUE_TYPE_DIQ,
 379	KFD_QUEUE_TYPE_SDMA_XGMI
 380};
 381
 382enum kfd_queue_format {
 383	KFD_QUEUE_FORMAT_PM4,
 384	KFD_QUEUE_FORMAT_AQL
 385};
 386
 387enum KFD_QUEUE_PRIORITY {
 388	KFD_QUEUE_PRIORITY_MINIMUM = 0,
 389	KFD_QUEUE_PRIORITY_MAXIMUM = 15
 390};
 391
 392/**
 393 * struct queue_properties
 394 *
 395 * @type: The queue type.
 396 *
 397 * @queue_id: Queue identifier.
 398 *
 399 * @queue_address: Queue ring buffer address.
 400 *
 401 * @queue_size: Queue ring buffer size.
 402 *
 403 * @priority: Defines the queue priority relative to other queues in the
 404 * process.
 405 * This is just an indication and HW scheduling may override the priority as
 406 * necessary while keeping the relative prioritization.
 407 * the priority granularity is from 0 to f which f is the highest priority.
 408 * currently all queues are initialized with the highest priority.
 409 *
 410 * @queue_percent: This field is partially implemented and currently a zero in
 411 * this field defines that the queue is non active.
 412 *
 413 * @read_ptr: User space address which points to the number of dwords the
 414 * cp read from the ring buffer. This field updates automatically by the H/W.
 415 *
 416 * @write_ptr: Defines the number of dwords written to the ring buffer.
 417 *
 418 * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring
 419 * buffer. This field should be similar to write_ptr and the user should
 420 * update this field after updating the write_ptr.
 421 *
 422 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
 423 *
 424 * @is_interop: Defines if this is a interop queue. Interop queue means that
 425 * the queue can access both graphics and compute resources.
 426 *
 427 * @is_evicted: Defines if the queue is evicted. Only active queues
 428 * are evicted, rendering them inactive.
 429 *
 430 * @is_active: Defines if the queue is active or not. @is_active and
 431 * @is_evicted are protected by the DQM lock.
 432 *
 433 * @is_gws: Defines if the queue has been updated to be GWS-capable or not.
 434 * @is_gws should be protected by the DQM lock, since changing it can yield the
 435 * possibility of updating DQM state on number of GWS queues.
 436 *
 437 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
 438 * of the queue.
 439 *
 440 * This structure represents the queue properties for each queue no matter if
 441 * it's user mode or kernel mode queue.
 442 *
 443 */
 444struct queue_properties {
 445	enum kfd_queue_type type;
 446	enum kfd_queue_format format;
 447	unsigned int queue_id;
 448	uint64_t queue_address;
 449	uint64_t  queue_size;
 450	uint32_t priority;
 451	uint32_t queue_percent;
 452	uint32_t *read_ptr;
 453	uint32_t *write_ptr;
 454	void __iomem *doorbell_ptr;
 455	uint32_t doorbell_off;
 456	bool is_interop;
 457	bool is_evicted;
 458	bool is_active;
 459	bool is_gws;
 460	/* Not relevant for user mode queues in cp scheduling */
 461	unsigned int vmid;
 462	/* Relevant only for sdma queues*/
 463	uint32_t sdma_engine_id;
 464	uint32_t sdma_queue_id;
 465	uint32_t sdma_vm_addr;
 466	/* Relevant only for VI */
 467	uint64_t eop_ring_buffer_address;
 468	uint32_t eop_ring_buffer_size;
 469	uint64_t ctx_save_restore_area_address;
 470	uint32_t ctx_save_restore_area_size;
 471	uint32_t ctl_stack_size;
 472	uint64_t tba_addr;
 473	uint64_t tma_addr;
 474	/* Relevant for CU */
 475	uint32_t cu_mask_count; /* Must be a multiple of 32 */
 476	uint32_t *cu_mask;
 477};
 478
 479#define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 &&	\
 480			    (q).queue_address != 0 &&	\
 481			    (q).queue_percent > 0 &&	\
 482			    !(q).is_evicted)
 483
 484/**
 485 * struct queue
 486 *
 487 * @list: Queue linked list.
 488 *
 489 * @mqd: The queue MQD (memory queue descriptor).
 490 *
 491 * @mqd_mem_obj: The MQD local gpu memory object.
 492 *
 493 * @gart_mqd_addr: The MQD gart mc address.
 494 *
 495 * @properties: The queue properties.
 496 *
 497 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
 498 *	 that the queue should be executed on.
 499 *
 500 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
 501 *	  id.
 502 *
 503 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
 504 *
 505 * @process: The kfd process that created this queue.
 506 *
 507 * @device: The kfd device that created this queue.
 508 *
 509 * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
 510 * otherwise.
 511 *
 512 * This structure represents user mode compute queues.
 513 * It contains all the necessary data to handle such queues.
 514 *
 515 */
 516
 517struct queue {
 518	struct list_head list;
 519	void *mqd;
 520	struct kfd_mem_obj *mqd_mem_obj;
 521	uint64_t gart_mqd_addr;
 522	struct queue_properties properties;
 523
 524	uint32_t mec;
 525	uint32_t pipe;
 526	uint32_t queue;
 527
 528	unsigned int sdma_id;
 529	unsigned int doorbell_id;
 530
 531	struct kfd_process	*process;
 532	struct kfd_dev		*device;
 533	void *gws;
 534
 535	/* procfs */
 536	struct kobject kobj;
 537};
 538
 
 
 
 539enum KFD_MQD_TYPE {
 540	KFD_MQD_TYPE_HIQ = 0,		/* for hiq */
 
 541	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
 542	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
 543	KFD_MQD_TYPE_DIQ,		/* for diq */
 544	KFD_MQD_TYPE_MAX
 545};
 546
 547enum KFD_PIPE_PRIORITY {
 548	KFD_PIPE_PRIORITY_CS_LOW = 0,
 549	KFD_PIPE_PRIORITY_CS_MEDIUM,
 550	KFD_PIPE_PRIORITY_CS_HIGH
 551};
 552
 553struct scheduling_resources {
 554	unsigned int vmid_mask;
 555	enum kfd_queue_type type;
 556	uint64_t queue_mask;
 557	uint64_t gws_mask;
 558	uint32_t oac_mask;
 559	uint32_t gds_heap_base;
 560	uint32_t gds_heap_size;
 561};
 562
 563struct process_queue_manager {
 564	/* data */
 565	struct kfd_process	*process;
 
 566	struct list_head	queues;
 567	unsigned long		*queue_slot_bitmap;
 568};
 569
 570struct qcm_process_device {
 571	/* The Device Queue Manager that owns this data */
 572	struct device_queue_manager *dqm;
 573	struct process_queue_manager *pqm;
 574	/* Queues list */
 575	struct list_head queues_list;
 576	struct list_head priv_queue_list;
 577
 578	unsigned int queue_count;
 579	unsigned int vmid;
 580	bool is_debug;
 581	unsigned int evicted; /* eviction counter, 0=active */
 582
 583	/* This flag tells if we should reset all wavefronts on
 584	 * process termination
 585	 */
 586	bool reset_wavefronts;
 587
 588	/* This flag tells us if this process has a GWS-capable
 589	 * queue that will be mapped into the runlist. It's
 590	 * possible to request a GWS BO, but not have the queue
 591	 * currently mapped, and this changes how the MAP_PROCESS
 592	 * PM4 packet is configured.
 593	 */
 594	bool mapped_gws_queue;
 595
 596	/* All the memory management data should be here too */
 597	uint64_t gds_context_area;
 598	/* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
 599	uint64_t page_table_base;
 600	uint32_t sh_mem_config;
 601	uint32_t sh_mem_bases;
 602	uint32_t sh_mem_ape1_base;
 603	uint32_t sh_mem_ape1_limit;
 
 604	uint32_t gds_size;
 605	uint32_t num_gws;
 606	uint32_t num_oac;
 607	uint32_t sh_hidden_private_base;
 608
 609	/* CWSR memory */
 610	void *cwsr_kaddr;
 611	uint64_t cwsr_base;
 612	uint64_t tba_addr;
 613	uint64_t tma_addr;
 614
 615	/* IB memory */
 616	uint64_t ib_base;
 617	void *ib_kaddr;
 618
 619	/* doorbell resources per process per device */
 620	unsigned long *doorbell_bitmap;
 621};
 622
 623/* KFD Memory Eviction */
 624
 625/* Approx. wait time before attempting to restore evicted BOs */
 626#define PROCESS_RESTORE_TIME_MS 100
 627/* Approx. back off time if restore fails due to lack of memory */
 628#define PROCESS_BACK_OFF_TIME_MS 100
 629/* Approx. time before evicting the process again */
 630#define PROCESS_ACTIVE_TIME_MS 10
 631
 632/* 8 byte handle containing GPU ID in the most significant 4 bytes and
 633 * idr_handle in the least significant 4 bytes
 634 */
 635#define MAKE_HANDLE(gpu_id, idr_handle) \
 636	(((uint64_t)(gpu_id) << 32) + idr_handle)
 637#define GET_GPU_ID(handle) (handle >> 32)
 638#define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
 639
 640enum kfd_pdd_bound {
 641	PDD_UNBOUND = 0,
 642	PDD_BOUND,
 643	PDD_BOUND_SUSPENDED,
 644};
 645
 646#define MAX_SYSFS_FILENAME_LEN 15
 647
 648/*
 649 * SDMA counter runs at 100MHz frequency.
 650 * We display SDMA activity in microsecond granularity in sysfs.
 651 * As a result, the divisor is 100.
 652 */
 653#define SDMA_ACTIVITY_DIVISOR  100
 654
 655/* Data that is per-process-per device. */
 656struct kfd_process_device {
 
 
 
 
 
 
 657	/* The device that owns this data. */
 658	struct kfd_dev *dev;
 659
 660	/* The process that owns this kfd_process_device. */
 661	struct kfd_process *process;
 662
 663	/* per-process-per device QCM data structure */
 664	struct qcm_process_device qpd;
 665
 666	/*Apertures*/
 667	uint64_t lds_base;
 668	uint64_t lds_limit;
 669	uint64_t gpuvm_base;
 670	uint64_t gpuvm_limit;
 671	uint64_t scratch_base;
 672	uint64_t scratch_limit;
 673
 674	/* VM context for GPUVM allocations */
 675	struct file *drm_file;
 676	void *drm_priv;
 677
 678	/* GPUVM allocations storage */
 679	struct idr alloc_idr;
 680
 681	/* Flag used to tell the pdd has dequeued from the dqm.
 682	 * This is used to prevent dev->dqm->ops.process_termination() from
 683	 * being called twice when it is already called in IOMMU callback
 684	 * function.
 685	 */
 686	bool already_dequeued;
 687	bool runtime_inuse;
 688
 689	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
 690	enum kfd_pdd_bound bound;
 691
 692	/* VRAM usage */
 693	uint64_t vram_usage;
 694	struct attribute attr_vram;
 695	char vram_filename[MAX_SYSFS_FILENAME_LEN];
 696
 697	/* SDMA activity tracking */
 698	uint64_t sdma_past_activity_counter;
 699	struct attribute attr_sdma;
 700	char sdma_filename[MAX_SYSFS_FILENAME_LEN];
 701
 702	/* Eviction activity tracking */
 703	uint64_t last_evict_timestamp;
 704	atomic64_t evict_duration_counter;
 705	struct attribute attr_evict;
 706
 707	struct kobject *kobj_stats;
 708	unsigned int doorbell_index;
 709
 710	/*
 711	 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
 712	 * that is associated with device encoded by "this" struct instance. The
 713	 * value reflects CU usage by all of the waves launched by this process
 714	 * on this device. A very important property of occupancy parameter is
 715	 * that its value is a snapshot of current use.
 716	 *
 717	 * Following is to be noted regarding how this parameter is reported:
 718	 *
 719	 *  The number of waves that a CU can launch is limited by couple of
 720	 *  parameters. These are encoded by struct amdgpu_cu_info instance
 721	 *  that is part of every device definition. For GFX9 devices this
 722	 *  translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
 723	 *  do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
 724	 *  when they do use scratch memory. This could change for future
 725	 *  devices and therefore this example should be considered as a guide.
 726	 *
 727	 *  All CU's of a device are available for the process. This may not be true
 728	 *  under certain conditions - e.g. CU masking.
 729	 *
 730	 *  Finally number of CU's that are occupied by a process is affected by both
 731	 *  number of CU's a device has along with number of other competing processes
 732	 */
 733	struct attribute attr_cu_occupancy;
 734
 735	/* sysfs counters for GPU retry fault and page migration tracking */
 736	struct kobject *kobj_counters;
 737	struct attribute attr_faults;
 738	struct attribute attr_page_in;
 739	struct attribute attr_page_out;
 740	uint64_t faults;
 741	uint64_t page_in;
 742	uint64_t page_out;
 743};
 744
 745#define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
 746
 747struct svm_range_list {
 748	struct mutex			lock;
 749	struct rb_root_cached		objects;
 750	struct list_head		list;
 751	struct work_struct		deferred_list_work;
 752	struct list_head		deferred_range_list;
 753	spinlock_t			deferred_list_lock;
 754	atomic_t			evicted_ranges;
 755	struct delayed_work		restore_work;
 756	DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE);
 757};
 758
 759/* Process data */
 760struct kfd_process {
 761	/*
 762	 * kfd_process are stored in an mm_struct*->kfd_process*
 763	 * hash table (kfd_processes in kfd_process.c)
 764	 */
 765	struct hlist_node kfd_processes;
 766
 767	/*
 768	 * Opaque pointer to mm_struct. We don't hold a reference to
 769	 * it so it should never be dereferenced from here. This is
 770	 * only used for looking up processes by their mm.
 771	 */
 772	void *mm;
 773
 774	struct kref ref;
 775	struct work_struct release_work;
 776
 777	struct mutex mutex;
 778
 779	/*
 780	 * In any process, the thread that started main() is the lead
 781	 * thread and outlives the rest.
 782	 * It is here because amd_iommu_bind_pasid wants a task_struct.
 783	 * It can also be used for safely getting a reference to the
 784	 * mm_struct of the process.
 785	 */
 786	struct task_struct *lead_thread;
 787
 788	/* We want to receive a notification when the mm_struct is destroyed */
 789	struct mmu_notifier mmu_notifier;
 790
 791	u32 pasid;
 
 
 
 792
 793	/*
 794	 * Array of kfd_process_device pointers,
 795	 * one for each device the process is using.
 796	 */
 797	struct kfd_process_device *pdds[MAX_GPU_INSTANCE];
 798	uint32_t n_pdds;
 799
 800	struct process_queue_manager pqm;
 801
 
 
 
 
 
 
 802	/*Is the user space process 32 bit?*/
 803	bool is_32bit_user_mode;
 804
 805	/* Event-related data */
 806	struct mutex event_mutex;
 807	/* Event ID allocator and lookup */
 808	struct idr event_idr;
 809	/* Event page */
 810	struct kfd_signal_page *signal_page;
 811	size_t signal_mapped_size;
 812	size_t signal_event_count;
 813	bool signal_event_limit_reached;
 814
 815	/* Information used for memory eviction */
 816	void *kgd_process_info;
 817	/* Eviction fence that is attached to all the BOs of this process. The
 818	 * fence will be triggered during eviction and new one will be created
 819	 * during restore
 820	 */
 821	struct dma_fence *ef;
 822
 823	/* Work items for evicting and restoring BOs */
 824	struct delayed_work eviction_work;
 825	struct delayed_work restore_work;
 826	/* seqno of the last scheduled eviction */
 827	unsigned int last_eviction_seqno;
 828	/* Approx. the last timestamp (in jiffies) when the process was
 829	 * restored after an eviction
 830	 */
 831	unsigned long last_restore_timestamp;
 832
 833	/* Kobj for our procfs */
 834	struct kobject *kobj;
 835	struct kobject *kobj_queues;
 836	struct attribute attr_pasid;
 837
 838	/* shared virtual memory registered by this process */
 839	struct svm_range_list svms;
 840
 841	bool xnack_enabled;
 842};
 843
 844#define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
 845extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
 846extern struct srcu_struct kfd_processes_srcu;
 847
 848/**
 849 * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.
 850 *
 851 * @filep: pointer to file structure.
 852 * @p: amdkfd process pointer.
 853 * @data: pointer to arg that was copied from user.
 854 *
 855 * Return: returns ioctl completion code.
 
 
 856 */
 857typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
 858				void *data);
 859
 860struct amdkfd_ioctl_desc {
 861	unsigned int cmd;
 862	int flags;
 863	amdkfd_ioctl_t *func;
 864	unsigned int cmd_drv;
 865	const char *name;
 866};
 867bool kfd_dev_is_large_bar(struct kfd_dev *dev);
 868
 869int kfd_process_create_wq(void);
 870void kfd_process_destroy_wq(void);
 871struct kfd_process *kfd_create_process(struct file *filep);
 872struct kfd_process *kfd_get_process(const struct task_struct *);
 873struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid);
 874struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
 875
 876int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
 877int kfd_process_gpuid_from_kgd(struct kfd_process *p,
 878			       struct amdgpu_device *adev, uint32_t *gpuid,
 879			       uint32_t *gpuidx);
 880static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,
 881				uint32_t gpuidx, uint32_t *gpuid) {
 882	return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;
 883}
 884static inline struct kfd_process_device *kfd_process_device_from_gpuidx(
 885				struct kfd_process *p, uint32_t gpuidx) {
 886	return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL;
 887}
 888
 889void kfd_unref_process(struct kfd_process *p);
 890int kfd_process_evict_queues(struct kfd_process *p);
 891int kfd_process_restore_queues(struct kfd_process *p);
 892void kfd_suspend_all_processes(void);
 893int kfd_resume_all_processes(void);
 894
 895int kfd_process_device_init_vm(struct kfd_process_device *pdd,
 896			       struct file *drm_file);
 897struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
 898						struct kfd_process *p);
 
 899struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
 900							struct kfd_process *p);
 901struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
 902							struct kfd_process *p);
 903
 904bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);
 905
 906int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
 907			  struct vm_area_struct *vma);
 908
 909/* KFD process API for creating and translating handles */
 910int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
 911					void *mem);
 912void *kfd_process_device_translate_handle(struct kfd_process_device *p,
 913					int handle);
 914void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
 915					int handle);
 916
 917/* PASIDs */
 918int kfd_pasid_init(void);
 919void kfd_pasid_exit(void);
 920bool kfd_set_pasid_limit(unsigned int new_limit);
 921unsigned int kfd_get_pasid_limit(void);
 922u32 kfd_pasid_alloc(void);
 923void kfd_pasid_free(u32 pasid);
 924
 925/* Doorbells */
 926size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
 927int kfd_doorbell_init(struct kfd_dev *kfd);
 928void kfd_doorbell_fini(struct kfd_dev *kfd);
 929int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
 930		      struct vm_area_struct *vma);
 931void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
 932					unsigned int *doorbell_off);
 933void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
 934u32 read_kernel_doorbell(u32 __iomem *db);
 935void write_kernel_doorbell(void __iomem *db, u32 value);
 936void write_kernel_doorbell64(void __iomem *db, u64 value);
 937unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
 938					struct kfd_process_device *pdd,
 939					unsigned int doorbell_id);
 940phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);
 941int kfd_alloc_process_doorbells(struct kfd_dev *kfd,
 942				unsigned int *doorbell_index);
 943void kfd_free_process_doorbells(struct kfd_dev *kfd,
 944				unsigned int doorbell_index);
 945/* GTT Sub-Allocator */
 946
 947int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
 948			struct kfd_mem_obj **mem_obj);
 949
 950int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
 951
 952extern struct device *kfd_device;
 953
 954/* KFD's procfs */
 955void kfd_procfs_init(void);
 956void kfd_procfs_shutdown(void);
 957int kfd_procfs_add_queue(struct queue *q);
 958void kfd_procfs_del_queue(struct queue *q);
 959
 960/* Topology */
 961int kfd_topology_init(void);
 962void kfd_topology_shutdown(void);
 963int kfd_topology_add_device(struct kfd_dev *gpu);
 964int kfd_topology_remove_device(struct kfd_dev *gpu);
 965struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
 966						uint32_t proximity_domain);
 967struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
 968struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
 969struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
 970struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd);
 971int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
 972int kfd_numa_node_to_apic_id(int numa_node_id);
 973void kfd_double_confirm_iommu_support(struct kfd_dev *gpu);
 974
 975/* Interrupts */
 976int kfd_interrupt_init(struct kfd_dev *dev);
 977void kfd_interrupt_exit(struct kfd_dev *dev);
 
 978bool enqueue_ih_ring_entry(struct kfd_dev *kfd,	const void *ih_ring_entry);
 979bool interrupt_is_wanted(struct kfd_dev *dev,
 980				const uint32_t *ih_ring_entry,
 981				uint32_t *patched_ihre, bool *flag);
 
 
 982
 983/* amdkfd Apertures */
 984int kfd_init_apertures(struct kfd_process *process);
 985
 986void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
 987				  uint64_t tba_addr,
 988				  uint64_t tma_addr);
 
 
 989
 990/* Queue Context Management */
 991int init_queue(struct queue **q, const struct queue_properties *properties);
 992void uninit_queue(struct queue *q);
 993void print_queue_properties(struct queue_properties *q);
 994void print_queue(struct queue *q);
 995
 
 
 996struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
 997		struct kfd_dev *dev);
 998struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
 999		struct kfd_dev *dev);
1000struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
1001		struct kfd_dev *dev);
1002struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
1003		struct kfd_dev *dev);
1004struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
1005		struct kfd_dev *dev);
1006struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
1007		struct kfd_dev *dev);
1008struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
1009void device_queue_manager_uninit(struct device_queue_manager *dqm);
1010struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
1011					enum kfd_queue_type type);
1012void kernel_queue_uninit(struct kernel_queue *kq, bool hanging);
1013int kfd_process_vm_fault(struct device_queue_manager *dqm, u32 pasid);
1014
1015/* Process Queue Manager */
1016struct process_queue_node {
1017	struct queue *q;
1018	struct kernel_queue *kq;
1019	struct list_head process_queue_list;
1020};
1021
1022void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
1023void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
1024int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
1025void pqm_uninit(struct process_queue_manager *pqm);
1026int pqm_create_queue(struct process_queue_manager *pqm,
1027			    struct kfd_dev *dev,
1028			    struct file *f,
1029			    struct queue_properties *properties,
1030			    unsigned int *qid,
1031			    uint32_t *p_doorbell_offset_in_process);
 
1032int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
1033int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
1034			struct queue_properties *p);
1035int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
1036			struct queue_properties *p);
1037int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
1038			void *gws);
1039struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
1040						unsigned int qid);
1041struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
1042						unsigned int qid);
1043int pqm_get_wave_state(struct process_queue_manager *pqm,
1044		       unsigned int qid,
1045		       void __user *ctl_stack,
1046		       u32 *ctl_stack_used_size,
1047		       u32 *save_area_used_size);
1048
1049int amdkfd_fence_wait_timeout(uint64_t *fence_addr,
1050			      uint64_t fence_value,
1051			      unsigned int timeout_ms);
1052
1053/* Packet Manager */
1054
 
 
1055#define KFD_FENCE_COMPLETED (100)
1056#define KFD_FENCE_INIT   (10)
 
1057
1058struct packet_manager {
1059	struct device_queue_manager *dqm;
1060	struct kernel_queue *priv_queue;
1061	struct mutex lock;
1062	bool allocated;
1063	struct kfd_mem_obj *ib_buffer_obj;
1064	unsigned int ib_size_bytes;
1065	bool is_over_subscription;
1066
1067	const struct packet_manager_funcs *pmf;
1068};
1069
1070struct packet_manager_funcs {
1071	/* Support ASIC-specific packet formats for PM4 packets */
1072	int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
1073			struct qcm_process_device *qpd);
1074	int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
1075			uint64_t ib, size_t ib_size_in_dwords, bool chain);
1076	int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
1077			struct scheduling_resources *res);
1078	int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
1079			struct queue *q, bool is_static);
1080	int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
1081			enum kfd_queue_type type,
1082			enum kfd_unmap_queues_filter mode,
1083			uint32_t filter_param, bool reset,
1084			unsigned int sdma_engine);
1085	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
1086			uint64_t fence_address,	uint64_t fence_value);
1087	int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
1088
1089	/* Packet sizes */
1090	int map_process_size;
1091	int runlist_size;
1092	int set_resources_size;
1093	int map_queues_size;
1094	int unmap_queues_size;
1095	int query_status_size;
1096	int release_mem_size;
1097};
1098
1099extern const struct packet_manager_funcs kfd_vi_pm_funcs;
1100extern const struct packet_manager_funcs kfd_v9_pm_funcs;
1101extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs;
1102
1103int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
1104void pm_uninit(struct packet_manager *pm, bool hanging);
1105int pm_send_set_resources(struct packet_manager *pm,
1106				struct scheduling_resources *res);
1107int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
1108int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
1109				uint64_t fence_value);
1110
1111int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
1112			enum kfd_unmap_queues_filter mode,
1113			uint32_t filter_param, bool reset,
1114			unsigned int sdma_engine);
1115
1116void pm_release_ib(struct packet_manager *pm);
1117
1118/* Following PM funcs can be shared among VI and AI */
1119unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
1120
1121uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
 
 
1122
1123/* Events */
1124extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1125extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1126
1127extern const struct kfd_device_global_init_class device_global_init_class_cik;
 
 
 
 
1128
1129void kfd_event_init_process(struct kfd_process *p);
1130void kfd_event_free_process(struct kfd_process *p);
1131int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1132int kfd_wait_on_events(struct kfd_process *p,
1133		       uint32_t num_events, void __user *data,
1134		       bool all, uint32_t user_timeout_ms,
1135		       uint32_t *wait_result);
1136void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
1137				uint32_t valid_id_bits);
1138void kfd_signal_iommu_event(struct kfd_dev *dev,
1139			    u32 pasid, unsigned long address,
1140			    bool is_write_requested, bool is_execute_requested);
1141void kfd_signal_hw_exception_event(u32 pasid);
1142int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1143int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1144int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
1145		       uint64_t size);
1146int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1147		     uint32_t event_type, bool auto_reset, uint32_t node_id,
1148		     uint32_t *event_id, uint32_t *event_trigger_data,
1149		     uint64_t *event_page_offset, uint32_t *event_slot_index);
1150int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1151
1152void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid,
1153				struct kfd_vm_fault_info *info);
1154
1155void kfd_signal_reset_event(struct kfd_dev *dev);
1156
1157void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);
1158
1159void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type);
1160
1161int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
1162
1163bool kfd_is_locked(void);
1164
1165/* Compute profile */
1166void kfd_inc_compute_active(struct kfd_dev *dev);
1167void kfd_dec_compute_active(struct kfd_dev *dev);
1168
1169/* Cgroup Support */
1170/* Check with device cgroup if @kfd device is accessible */
1171static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)
1172{
1173#if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)
1174	struct drm_device *ddev = kfd->ddev;
1175
1176	return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,
1177					  ddev->render->index,
1178					  DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1179#else
1180	return 0;
1181#endif
1182}
1183
1184/* Debugfs */
1185#if defined(CONFIG_DEBUG_FS)
1186
1187void kfd_debugfs_init(void);
1188void kfd_debugfs_fini(void);
1189int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1190int pqm_debugfs_mqds(struct seq_file *m, void *data);
1191int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1192int dqm_debugfs_hqds(struct seq_file *m, void *data);
1193int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1194int pm_debugfs_runlist(struct seq_file *m, void *data);
1195
1196int kfd_debugfs_hang_hws(struct kfd_dev *dev);
1197int pm_debugfs_hang_hws(struct packet_manager *pm);
1198int dqm_debugfs_execute_queues(struct device_queue_manager *dqm);
1199
1200#else
1201
1202static inline void kfd_debugfs_init(void) {}
1203static inline void kfd_debugfs_fini(void) {}
1204
1205#endif
1206
1207#endif