amdkfd: use schedule() in sync_with_hw
[linux-2.6-block.git] / drivers / gpu / drm / amd / amdkfd / kfd_priv.h
CommitLineData
4a488a7a
OG
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>
19f6d2a6 33#include <linux/kfd_ioctl.h>
4a488a7a
OG
34#include <kgd_kfd_interface.h>
35
5b5c4e40
EP
36#define KFD_SYSFS_FILE_MODE 0444
37
ed6e6a34
BG
38/*
39 * When working with cp scheduler we should assign the HIQ manually or via
40 * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot
41 * definitions for Kaveri. In Kaveri only the first ME queues participates
42 * in the cp scheduling taking that in mind we set the HIQ slot in the
43 * second ME.
44 */
45#define KFD_CIK_HIQ_PIPE 4
46#define KFD_CIK_HIQ_QUEUE 0
47
5b5c4e40
EP
48/* GPU ID hash width in bits */
49#define KFD_GPU_ID_HASH_WIDTH 16
50
51/* Macro for allocating structures */
52#define kfd_alloc_struct(ptr_to_struct) \
53 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
54
19f6d2a6
OG
55/* Kernel module parameter to specify maximum number of supported processes */
56extern int max_num_of_processes;
57
58#define KFD_MAX_NUM_OF_PROCESSES_DEFAULT 32
59#define KFD_MAX_NUM_OF_PROCESSES 512
60
61/*
62 * Kernel module parameter to specify maximum number of supported queues
63 * per process
64 */
65extern int max_num_of_queues_per_process;
66
67#define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS_DEFAULT 128
68#define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
69
ed6e6a34
BG
70#define KFD_KERNEL_QUEUE_SIZE 2048
71
31c21fec
BG
72/* Kernel module parameter to specify the scheduling policy */
73extern int sched_policy;
74
75/**
76 * enum kfd_sched_policy
77 *
78 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp)
79 * scheduling. In this scheduling mode we're using the firmware code to
80 * schedule the user mode queues and kernel queues such as HIQ and DIQ.
81 * the HIQ queue is used as a special queue that dispatches the configuration
82 * to the cp and the user mode queues list that are currently running.
83 * the DIQ queue is a debugging queue that dispatches debugging commands to the
84 * firmware.
85 * in this scheduling mode user mode queues over subscription feature is
86 * enabled.
87 *
88 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over
89 * subscription feature disabled.
90 *
91 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly
92 * set the command processor registers and sets the queues "manually". This
93 * mode is used *ONLY* for debugging proposes.
94 *
95 */
96enum kfd_sched_policy {
97 KFD_SCHED_POLICY_HWS = 0,
98 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION,
99 KFD_SCHED_POLICY_NO_HWS
100};
101
ed6e6a34
BG
102enum cache_policy {
103 cache_policy_coherent,
104 cache_policy_noncoherent
105};
106
4a488a7a
OG
107struct kfd_device_info {
108 unsigned int max_pasid_bits;
109 size_t ih_ring_entry_size;
19f6d2a6 110 uint16_t mqd_size_aligned;
4a488a7a
OG
111};
112
113struct kfd_dev {
114 struct kgd_dev *kgd;
115
116 const struct kfd_device_info *device_info;
117 struct pci_dev *pdev;
118
119 unsigned int id; /* topology stub index */
120
19f6d2a6
OG
121 phys_addr_t doorbell_base; /* Start of actual doorbells used by
122 * KFD. It is aligned for mapping
123 * into user mode
124 */
125 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell
126 * to HW doorbell, GFX reserved some
127 * at the start)
128 */
129 size_t doorbell_process_limit; /* Number of processes we have doorbell
130 * space for.
131 */
132 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
133 * page used by kernel queue
134 */
135
4a488a7a
OG
136 struct kgd2kfd_shared_resources shared_resources;
137
b3f5e6b4
AL
138 void *interrupt_ring;
139 size_t interrupt_ring_size;
140 atomic_t interrupt_ring_rptr;
141 atomic_t interrupt_ring_wptr;
142 struct work_struct interrupt_work;
143 spinlock_t interrupt_lock;
144
ed6e6a34
BG
145 /* QCM Device instance */
146 struct device_queue_manager *dqm;
4a488a7a 147
ed6e6a34 148 bool init_complete;
b3f5e6b4
AL
149 /*
150 * Interrupts of interest to KFD are copied
151 * from the HW ring into a SW ring.
152 */
153 bool interrupts_active;
4a488a7a
OG
154};
155
156/* KGD2KFD callbacks */
157void kgd2kfd_exit(void);
158struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev);
159bool kgd2kfd_device_init(struct kfd_dev *kfd,
160 const struct kgd2kfd_shared_resources *gpu_resources);
161void kgd2kfd_device_exit(struct kfd_dev *kfd);
162
163extern const struct kfd2kgd_calls *kfd2kgd;
164
19f6d2a6
OG
165struct kfd_mem_obj {
166 void *bo;
167 uint64_t gpu_addr;
168 uint32_t *cpu_ptr;
169};
170
171enum kfd_mempool {
172 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
173 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
174 KFD_MEMPOOL_FRAMEBUFFER = 3,
175};
176
4a488a7a
OG
177/* Character device interface */
178int kfd_chardev_init(void);
179void kfd_chardev_exit(void);
180struct device *kfd_chardev(void);
181
241f24f8
BG
182/**
183 * enum kfd_preempt_type_filter
184 *
185 * @KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE: Preempts single queue.
186 *
187 * @KFD_PRERMPT_TYPE_FILTER_ALL_QUEUES: Preempts all queues in the
188 * running queues list.
189 *
190 * @KFD_PRERMPT_TYPE_FILTER_BY_PASID: Preempts queues that belongs to
191 * specific process.
192 *
193 */
194enum kfd_preempt_type_filter {
195 KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE,
196 KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES,
197 KFD_PREEMPT_TYPE_FILTER_BY_PASID
198};
19f6d2a6 199
6e99df57
BG
200enum kfd_preempt_type {
201 KFD_PREEMPT_TYPE_WAVEFRONT,
202 KFD_PREEMPT_TYPE_WAVEFRONT_RESET
203};
204
ed8aab45
BG
205/**
206 * enum kfd_queue_type
207 *
208 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
209 *
210 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
211 *
212 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
213 *
214 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
215 */
216enum kfd_queue_type {
217 KFD_QUEUE_TYPE_COMPUTE,
218 KFD_QUEUE_TYPE_SDMA,
219 KFD_QUEUE_TYPE_HIQ,
220 KFD_QUEUE_TYPE_DIQ
221};
222
6e99df57
BG
223enum kfd_queue_format {
224 KFD_QUEUE_FORMAT_PM4,
225 KFD_QUEUE_FORMAT_AQL
226};
227
ed8aab45
BG
228/**
229 * struct queue_properties
230 *
231 * @type: The queue type.
232 *
233 * @queue_id: Queue identifier.
234 *
235 * @queue_address: Queue ring buffer address.
236 *
237 * @queue_size: Queue ring buffer size.
238 *
239 * @priority: Defines the queue priority relative to other queues in the
240 * process.
241 * This is just an indication and HW scheduling may override the priority as
242 * necessary while keeping the relative prioritization.
243 * the priority granularity is from 0 to f which f is the highest priority.
244 * currently all queues are initialized with the highest priority.
245 *
246 * @queue_percent: This field is partially implemented and currently a zero in
247 * this field defines that the queue is non active.
248 *
249 * @read_ptr: User space address which points to the number of dwords the
250 * cp read from the ring buffer. This field updates automatically by the H/W.
251 *
252 * @write_ptr: Defines the number of dwords written to the ring buffer.
253 *
254 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
255 * the queue ring buffer. This field should be similar to write_ptr and the user
256 * should update this field after he updated the write_ptr.
257 *
258 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
259 *
260 * @is_interop: Defines if this is a interop queue. Interop queue means that the
261 * queue can access both graphics and compute resources.
262 *
263 * @is_active: Defines if the queue is active or not.
264 *
265 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
266 * of the queue.
267 *
268 * This structure represents the queue properties for each queue no matter if
269 * it's user mode or kernel mode queue.
270 *
271 */
272struct queue_properties {
273 enum kfd_queue_type type;
6e99df57 274 enum kfd_queue_format format;
ed8aab45
BG
275 unsigned int queue_id;
276 uint64_t queue_address;
277 uint64_t queue_size;
278 uint32_t priority;
279 uint32_t queue_percent;
280 uint32_t *read_ptr;
281 uint32_t *write_ptr;
5cd78de5 282 uint32_t __iomem *doorbell_ptr;
ed8aab45
BG
283 uint32_t doorbell_off;
284 bool is_interop;
285 bool is_active;
286 /* Not relevant for user mode queues in cp scheduling */
287 unsigned int vmid;
288};
289
290/**
291 * struct queue
292 *
293 * @list: Queue linked list.
294 *
295 * @mqd: The queue MQD.
296 *
297 * @mqd_mem_obj: The MQD local gpu memory object.
298 *
299 * @gart_mqd_addr: The MQD gart mc address.
300 *
301 * @properties: The queue properties.
302 *
303 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
304 * that the queue should be execute on.
305 *
306 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe id.
307 *
308 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
309 *
310 * @process: The kfd process that created this queue.
311 *
312 * @device: The kfd device that created this queue.
313 *
314 * This structure represents user mode compute queues.
315 * It contains all the necessary data to handle such queues.
316 *
317 */
318
319struct queue {
320 struct list_head list;
321 void *mqd;
322 struct kfd_mem_obj *mqd_mem_obj;
323 uint64_t gart_mqd_addr;
324 struct queue_properties properties;
325
326 uint32_t mec;
327 uint32_t pipe;
328 uint32_t queue;
329
330 struct kfd_process *process;
331 struct kfd_dev *device;
332};
333
6e99df57
BG
334/*
335 * Please read the kfd_mqd_manager.h description.
336 */
337enum KFD_MQD_TYPE {
338 KFD_MQD_TYPE_CIK_COMPUTE = 0, /* for no cp scheduling */
339 KFD_MQD_TYPE_CIK_HIQ, /* for hiq */
340 KFD_MQD_TYPE_CIK_CP, /* for cp queues and diq */
341 KFD_MQD_TYPE_CIK_SDMA, /* for sdma queues */
342 KFD_MQD_TYPE_MAX
343};
344
241f24f8
BG
345struct scheduling_resources {
346 unsigned int vmid_mask;
347 enum kfd_queue_type type;
348 uint64_t queue_mask;
349 uint64_t gws_mask;
350 uint32_t oac_mask;
351 uint32_t gds_heap_base;
352 uint32_t gds_heap_size;
353};
354
355struct process_queue_manager {
356 /* data */
357 struct kfd_process *process;
358 unsigned int num_concurrent_processes;
359 struct list_head queues;
360 unsigned long *queue_slot_bitmap;
361};
362
363struct qcm_process_device {
364 /* The Device Queue Manager that owns this data */
365 struct device_queue_manager *dqm;
366 struct process_queue_manager *pqm;
367 /* Device Queue Manager lock */
368 struct mutex *lock;
369 /* Queues list */
370 struct list_head queues_list;
371 struct list_head priv_queue_list;
372
373 unsigned int queue_count;
374 unsigned int vmid;
375 bool is_debug;
376 /*
377 * All the memory management data should be here too
378 */
379 uint64_t gds_context_area;
380 uint32_t sh_mem_config;
381 uint32_t sh_mem_bases;
382 uint32_t sh_mem_ape1_base;
383 uint32_t sh_mem_ape1_limit;
384 uint32_t page_table_base;
385 uint32_t gds_size;
386 uint32_t num_gws;
387 uint32_t num_oac;
388};
389
19f6d2a6
OG
390/* Data that is per-process-per device. */
391struct kfd_process_device {
392 /*
393 * List of all per-device data for a process.
394 * Starts from kfd_process.per_device_data.
395 */
396 struct list_head per_device_list;
397
398 /* The device that owns this data. */
399 struct kfd_dev *dev;
400
401
45102048
BG
402 /* per-process-per device QCM data structure */
403 struct qcm_process_device qpd;
404
19f6d2a6
OG
405 /*Apertures*/
406 uint64_t lds_base;
407 uint64_t lds_limit;
408 uint64_t gpuvm_base;
409 uint64_t gpuvm_limit;
410 uint64_t scratch_base;
411 uint64_t scratch_limit;
412
413 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
414 bool bound;
415};
416
4a488a7a
OG
417/* Process data */
418struct kfd_process {
19f6d2a6
OG
419 /*
420 * kfd_process are stored in an mm_struct*->kfd_process*
421 * hash table (kfd_processes in kfd_process.c)
422 */
423 struct hlist_node kfd_processes;
424
425 struct mm_struct *mm;
426
427 struct mutex mutex;
428
429 /*
430 * In any process, the thread that started main() is the lead
431 * thread and outlives the rest.
432 * It is here because amd_iommu_bind_pasid wants a task_struct.
433 */
434 struct task_struct *lead_thread;
435
436 /* We want to receive a notification when the mm_struct is destroyed */
437 struct mmu_notifier mmu_notifier;
438
439 /* Use for delayed freeing of kfd_process structure */
440 struct rcu_head rcu;
441
442 unsigned int pasid;
443
444 /*
445 * List of kfd_process_device structures,
446 * one for each device the process is using.
447 */
448 struct list_head per_device_data;
449
45102048
BG
450 struct process_queue_manager pqm;
451
19f6d2a6
OG
452 /* The process's queues. */
453 size_t queue_array_size;
454
455 /* Size is queue_array_size, up to MAX_PROCESS_QUEUES. */
456 struct kfd_queue **queues;
457
458 unsigned long allocated_queue_bitmap[DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS, BITS_PER_LONG)];
459
460 /*Is the user space process 32 bit?*/
461 bool is_32bit_user_mode;
4a488a7a
OG
462};
463
19f6d2a6
OG
464void kfd_process_create_wq(void);
465void kfd_process_destroy_wq(void);
466struct kfd_process *kfd_create_process(const struct task_struct *);
467struct kfd_process *kfd_get_process(const struct task_struct *);
468
64c7f8cf
BG
469struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
470 struct kfd_process *p);
b17f068a 471void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid);
19f6d2a6
OG
472struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
473 struct kfd_process *p,
474 int create_pdd);
475
775921ed
AS
476/* Process device data iterator */
477struct kfd_process_device *kfd_get_first_process_device_data(struct kfd_process *p);
478struct kfd_process_device *kfd_get_next_process_device_data(struct kfd_process *p,
479 struct kfd_process_device *pdd);
480bool kfd_has_process_device_data(struct kfd_process *p);
481
19f6d2a6
OG
482/* PASIDs */
483int kfd_pasid_init(void);
484void kfd_pasid_exit(void);
485bool kfd_set_pasid_limit(unsigned int new_limit);
486unsigned int kfd_get_pasid_limit(void);
487unsigned int kfd_pasid_alloc(void);
488void kfd_pasid_free(unsigned int pasid);
489
490/* Doorbells */
491void kfd_doorbell_init(struct kfd_dev *kfd);
492int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma);
493u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
494 unsigned int *doorbell_off);
495void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
496u32 read_kernel_doorbell(u32 __iomem *db);
497void write_kernel_doorbell(u32 __iomem *db, u32 value);
498unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
499 struct kfd_process *process,
500 unsigned int queue_id);
501
4a488a7a
OG
502extern struct device *kfd_device;
503
5b5c4e40
EP
504/* Topology */
505int kfd_topology_init(void);
506void kfd_topology_shutdown(void);
507int kfd_topology_add_device(struct kfd_dev *gpu);
508int kfd_topology_remove_device(struct kfd_dev *gpu);
509struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
510struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
511struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx);
512
4a488a7a 513/* Interrupts */
b3f5e6b4
AL
514int kfd_interrupt_init(struct kfd_dev *dev);
515void kfd_interrupt_exit(struct kfd_dev *dev);
516void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
517bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry);
4a488a7a
OG
518
519/* Power Management */
b3f5e6b4
AL
520void kgd2kfd_suspend(struct kfd_dev *kfd);
521int kgd2kfd_resume(struct kfd_dev *kfd);
4a488a7a 522
19f6d2a6
OG
523/* amdkfd Apertures */
524int kfd_init_apertures(struct kfd_process *process);
525
ed6e6a34 526/* Queue Context Management */
241f24f8
BG
527inline uint32_t lower_32(uint64_t x);
528inline uint32_t upper_32(uint64_t x);
529
ed6e6a34
BG
530int init_queue(struct queue **q, struct queue_properties properties);
531void uninit_queue(struct queue *q);
45102048 532void print_queue_properties(struct queue_properties *q);
ed6e6a34
BG
533void print_queue(struct queue *q);
534
64c7f8cf
BG
535struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
536 struct kfd_dev *dev);
537struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
538void device_queue_manager_uninit(struct device_queue_manager *dqm);
241f24f8
BG
539struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
540 enum kfd_queue_type type);
541void kernel_queue_uninit(struct kernel_queue *kq);
542
45102048
BG
543/* Process Queue Manager */
544struct process_queue_node {
545 struct queue *q;
546 struct kernel_queue *kq;
547 struct list_head process_queue_list;
548};
549
550int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
551void pqm_uninit(struct process_queue_manager *pqm);
552int pqm_create_queue(struct process_queue_manager *pqm,
553 struct kfd_dev *dev,
554 struct file *f,
555 struct queue_properties *properties,
556 unsigned int flags,
557 enum kfd_queue_type type,
558 unsigned int *qid);
559int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
560int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
561 struct queue_properties *p);
562
ed6e6a34
BG
563/* Packet Manager */
564
241f24f8
BG
565#define KFD_HIQ_TIMEOUT (500)
566
64c7f8cf
BG
567#define KFD_FENCE_COMPLETED (100)
568#define KFD_FENCE_INIT (10)
241f24f8
BG
569#define KFD_UNMAP_LATENCY (150)
570
ed6e6a34
BG
571struct packet_manager {
572 struct device_queue_manager *dqm;
573 struct kernel_queue *priv_queue;
574 struct mutex lock;
575 bool allocated;
576 struct kfd_mem_obj *ib_buffer_obj;
577};
578
64c7f8cf
BG
579int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
580void pm_uninit(struct packet_manager *pm);
581int pm_send_set_resources(struct packet_manager *pm,
582 struct scheduling_resources *res);
583int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
584int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
585 uint32_t fence_value);
586
587int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
588 enum kfd_preempt_type_filter mode,
589 uint32_t filter_param, bool reset,
590 unsigned int sdma_engine);
591
241f24f8
BG
592void pm_release_ib(struct packet_manager *pm);
593
19f6d2a6
OG
594uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
595phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
596 struct kfd_process *process);
597
4a488a7a 598#endif