blob: b397ec726400c2a52533356363b36bfd275dfb39 [file] [log] [blame]
Oded Gabbay4a488a72014-07-16 21:08:55 +03001/*
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>
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030033#include <linux/kfd_ioctl.h>
Oded Gabbay4a488a72014-07-16 21:08:55 +030034#include <kgd_kfd_interface.h>
35
Evgeny Pinchuk5b5c4e42014-07-16 21:22:32 +030036#define KFD_SYSFS_FILE_MODE 0444
37
Andrew Lewyckyf3a39812015-05-10 12:15:46 +030038#define KFD_MMAP_DOORBELL_MASK 0x8000000000000
39#define KFD_MMAP_EVENTS_MASK 0x4000000000000
40
Ben Gozed6e6a32014-07-17 00:45:35 +030041/*
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
Evgeny Pinchuk5b5c4e42014-07-16 21:22:32 +030051/* 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
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030058#define KFD_MAX_NUM_OF_PROCESSES 512
Oded Gabbayb8cbab02015-01-18 13:18:01 +020059#define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030060
61/*
Oded Gabbayb8cbab02015-01-18 13:18:01 +020062 * Kernel module parameter to specify maximum number of supported queues per
63 * device
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030064 */
Oded Gabbayb8cbab02015-01-18 13:18:01 +020065extern int max_num_of_queues_per_device;
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030066
Oded Gabbayb8cbab02015-01-18 13:18:01 +020067#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)
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030071
Ben Gozed6e6a32014-07-17 00:45:35 +030072#define KFD_KERNEL_QUEUE_SIZE 2048
73
Ben Goz31c21fe2014-07-17 00:48:28 +030074/* Kernel module parameter to specify the scheduling policy */
75extern int sched_policy;
76
Oded Gabbay81663012014-12-24 13:30:52 +020077/*
78 * Kernel module parameter to specify whether to send sigterm to HSA process on
79 * unhandled exception
80 */
81extern int send_sigterm;
82
Ben Goz31c21fe2014-07-17 00:48:28 +030083/**
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
Ben Gozed6e6a32014-07-17 00:45:35 +0300110enum cache_policy {
111 cache_policy_coherent,
112 cache_policy_noncoherent
113};
114
Ben Goz0da75582015-01-01 17:10:01 +0200115enum asic_family_type {
116 CHIP_KAVERI = 0,
117 CHIP_CARRIZO
118};
119
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300120struct 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
Oded Gabbay4a488a72014-07-16 21:08:55 +0300127struct kfd_device_info {
Ben Goz0da75582015-01-01 17:10:01 +0200128 unsigned int asic_family;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300129 const struct kfd_event_interrupt_class *event_interrupt_class;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300130 unsigned int max_pasid_bits;
Yair Shachar992839a2015-05-20 13:43:04 +0300131 unsigned int max_no_of_hqd;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300132 size_t ih_ring_entry_size;
Alexey Skidanovf7c826a2014-10-13 16:35:12 +0300133 uint8_t num_of_watch_points;
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300134 uint16_t mqd_size_aligned;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300135};
136
Oded Gabbay36b5c082014-10-26 09:53:10 +0200137struct 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
Oded Gabbay4a488a72014-07-16 21:08:55 +0300144struct 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
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300152 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
Oded Gabbay4a488a72014-07-16 21:08:55 +0300167 struct kgd2kfd_shared_resources shared_resources;
168
Xihan Zhangcea405b2015-03-17 19:32:53 +0800169 const struct kfd2kgd_calls *kfd2kgd;
170 struct mutex doorbell_mutex;
Joe Perchesf761d8b2015-05-19 18:37:51 -0700171 DECLARE_BITMAP(doorbell_available_index,
172 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
Xihan Zhangcea405b2015-03-17 19:32:53 +0800173
Oded Gabbay36b5c082014-10-26 09:53:10 +0200174 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
Andrew Lewycky2249d552014-07-17 01:37:30 +0300182 /* 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
Ben Gozed6e6a32014-07-17 00:45:35 +0300190 /* QCM Device instance */
191 struct device_queue_manager *dqm;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300192
Ben Gozed6e6a32014-07-17 00:45:35 +0300193 bool init_complete;
Andrew Lewycky2249d552014-07-17 01:37:30 +0300194 /*
195 * Interrupts of interest to KFD are copied
196 * from the HW ring into a SW ring.
197 */
198 bool interrupts_active;
Yair Shacharfbeb6612015-05-20 13:48:26 +0300199
200 /* Debug manager */
201 struct kfd_dbgmgr *dbgmgr;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300202};
203
204/* KGD2KFD callbacks */
205void kgd2kfd_exit(void);
Xihan Zhangcea405b2015-03-17 19:32:53 +0800206struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
207 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g);
Oded Gabbay4a488a72014-07-16 21:08:55 +0300208bool kgd2kfd_device_init(struct kfd_dev *kfd,
Xihan Zhangcea405b2015-03-17 19:32:53 +0800209 const struct kgd2kfd_shared_resources *gpu_resources);
Oded Gabbay4a488a72014-07-16 21:08:55 +0300210void kgd2kfd_device_exit(struct kfd_dev *kfd);
211
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300212enum kfd_mempool {
213 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
214 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
215 KFD_MEMPOOL_FRAMEBUFFER = 3,
216};
217
Oded Gabbay4a488a72014-07-16 21:08:55 +0300218/* Character device interface */
219int kfd_chardev_init(void);
220void kfd_chardev_exit(void);
221struct device *kfd_chardev(void);
222
Ben Goz241f24f2014-07-17 00:55:28 +0300223/**
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,
Yair Shachar992839a2015-05-20 13:43:04 +0300238 KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES,
Ben Goz241f24f2014-07-17 00:55:28 +0300239 KFD_PREEMPT_TYPE_FILTER_BY_PASID
240};
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300241
Ben Gozed8aab42014-07-17 00:18:51 +0300242/**
243 * enum kfd_queue_type
244 *
245 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
246 *
247 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
248 *
249 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
250 *
251 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
252 */
253enum kfd_queue_type {
254 KFD_QUEUE_TYPE_COMPUTE,
255 KFD_QUEUE_TYPE_SDMA,
256 KFD_QUEUE_TYPE_HIQ,
257 KFD_QUEUE_TYPE_DIQ
258};
259
Ben Goz6e99df52014-07-17 00:36:17 +0300260enum kfd_queue_format {
261 KFD_QUEUE_FORMAT_PM4,
262 KFD_QUEUE_FORMAT_AQL
263};
264
Ben Gozed8aab42014-07-17 00:18:51 +0300265/**
266 * struct queue_properties
267 *
268 * @type: The queue type.
269 *
270 * @queue_id: Queue identifier.
271 *
272 * @queue_address: Queue ring buffer address.
273 *
274 * @queue_size: Queue ring buffer size.
275 *
276 * @priority: Defines the queue priority relative to other queues in the
277 * process.
278 * This is just an indication and HW scheduling may override the priority as
279 * necessary while keeping the relative prioritization.
280 * the priority granularity is from 0 to f which f is the highest priority.
281 * currently all queues are initialized with the highest priority.
282 *
283 * @queue_percent: This field is partially implemented and currently a zero in
284 * this field defines that the queue is non active.
285 *
286 * @read_ptr: User space address which points to the number of dwords the
287 * cp read from the ring buffer. This field updates automatically by the H/W.
288 *
289 * @write_ptr: Defines the number of dwords written to the ring buffer.
290 *
291 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
Kent Russell8eabaf52017-08-15 23:00:04 -0400292 * the queue ring buffer. This field should be similar to write_ptr and the
293 * user should update this field after he updated the write_ptr.
Ben Gozed8aab42014-07-17 00:18:51 +0300294 *
295 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
296 *
Kent Russell8eabaf52017-08-15 23:00:04 -0400297 * @is_interop: Defines if this is a interop queue. Interop queue means that
298 * the queue can access both graphics and compute resources.
Ben Gozed8aab42014-07-17 00:18:51 +0300299 *
300 * @is_active: Defines if the queue is active or not.
301 *
302 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
303 * of the queue.
304 *
305 * This structure represents the queue properties for each queue no matter if
306 * it's user mode or kernel mode queue.
307 *
308 */
309struct queue_properties {
310 enum kfd_queue_type type;
Ben Goz6e99df52014-07-17 00:36:17 +0300311 enum kfd_queue_format format;
Ben Gozed8aab42014-07-17 00:18:51 +0300312 unsigned int queue_id;
313 uint64_t queue_address;
314 uint64_t queue_size;
315 uint32_t priority;
316 uint32_t queue_percent;
317 uint32_t *read_ptr;
318 uint32_t *write_ptr;
Oded Gabbay5cd78de2014-11-20 16:14:56 +0200319 uint32_t __iomem *doorbell_ptr;
Ben Gozed8aab42014-07-17 00:18:51 +0300320 uint32_t doorbell_off;
321 bool is_interop;
322 bool is_active;
323 /* Not relevant for user mode queues in cp scheduling */
324 unsigned int vmid;
Ben Goz77669eb2015-01-03 22:12:31 +0200325 /* Relevant only for sdma queues*/
326 uint32_t sdma_engine_id;
327 uint32_t sdma_queue_id;
328 uint32_t sdma_vm_addr;
Ben Gozff3d04a2015-01-04 10:37:18 +0200329 /* Relevant only for VI */
330 uint64_t eop_ring_buffer_address;
331 uint32_t eop_ring_buffer_size;
332 uint64_t ctx_save_restore_area_address;
333 uint32_t ctx_save_restore_area_size;
Ben Gozed8aab42014-07-17 00:18:51 +0300334};
335
336/**
337 * struct queue
338 *
339 * @list: Queue linked list.
340 *
341 * @mqd: The queue MQD.
342 *
343 * @mqd_mem_obj: The MQD local gpu memory object.
344 *
345 * @gart_mqd_addr: The MQD gart mc address.
346 *
347 * @properties: The queue properties.
348 *
349 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
Kent Russell8eabaf52017-08-15 23:00:04 -0400350 * that the queue should be execute on.
Ben Gozed8aab42014-07-17 00:18:51 +0300351 *
Kent Russell8eabaf52017-08-15 23:00:04 -0400352 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
353 * id.
Ben Gozed8aab42014-07-17 00:18:51 +0300354 *
355 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
356 *
357 * @process: The kfd process that created this queue.
358 *
359 * @device: The kfd device that created this queue.
360 *
361 * This structure represents user mode compute queues.
362 * It contains all the necessary data to handle such queues.
363 *
364 */
365
366struct queue {
367 struct list_head list;
368 void *mqd;
369 struct kfd_mem_obj *mqd_mem_obj;
370 uint64_t gart_mqd_addr;
371 struct queue_properties properties;
372
373 uint32_t mec;
374 uint32_t pipe;
375 uint32_t queue;
376
Ben Goz77669eb2015-01-03 22:12:31 +0200377 unsigned int sdma_id;
378
Ben Gozed8aab42014-07-17 00:18:51 +0300379 struct kfd_process *process;
380 struct kfd_dev *device;
381};
382
Ben Goz6e99df52014-07-17 00:36:17 +0300383/*
384 * Please read the kfd_mqd_manager.h description.
385 */
386enum KFD_MQD_TYPE {
Ben Goz85d258f2015-01-04 10:36:30 +0200387 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */
388 KFD_MQD_TYPE_HIQ, /* for hiq */
389 KFD_MQD_TYPE_CP, /* for cp queues and diq */
390 KFD_MQD_TYPE_SDMA, /* for sdma queues */
Ben Goz6e99df52014-07-17 00:36:17 +0300391 KFD_MQD_TYPE_MAX
392};
393
Ben Goz241f24f2014-07-17 00:55:28 +0300394struct scheduling_resources {
395 unsigned int vmid_mask;
396 enum kfd_queue_type type;
397 uint64_t queue_mask;
398 uint64_t gws_mask;
399 uint32_t oac_mask;
400 uint32_t gds_heap_base;
401 uint32_t gds_heap_size;
402};
403
404struct process_queue_manager {
405 /* data */
406 struct kfd_process *process;
407 unsigned int num_concurrent_processes;
408 struct list_head queues;
409 unsigned long *queue_slot_bitmap;
410};
411
412struct qcm_process_device {
413 /* The Device Queue Manager that owns this data */
414 struct device_queue_manager *dqm;
415 struct process_queue_manager *pqm;
Ben Goz241f24f2014-07-17 00:55:28 +0300416 /* Queues list */
417 struct list_head queues_list;
418 struct list_head priv_queue_list;
419
420 unsigned int queue_count;
421 unsigned int vmid;
422 bool is_debug;
423 /*
424 * All the memory management data should be here too
425 */
426 uint64_t gds_context_area;
427 uint32_t sh_mem_config;
428 uint32_t sh_mem_bases;
429 uint32_t sh_mem_ape1_base;
430 uint32_t sh_mem_ape1_limit;
431 uint32_t page_table_base;
432 uint32_t gds_size;
433 uint32_t num_gws;
434 uint32_t num_oac;
Moses Reuben6a1c9512017-08-15 23:00:20 -0400435 uint32_t sh_hidden_private_base;
Ben Goz241f24f2014-07-17 00:55:28 +0300436};
437
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300438/* Data that is per-process-per device. */
439struct kfd_process_device {
440 /*
441 * List of all per-device data for a process.
442 * Starts from kfd_process.per_device_data.
443 */
444 struct list_head per_device_list;
445
446 /* The device that owns this data. */
447 struct kfd_dev *dev;
448
449
Ben Goz45102042014-07-17 01:04:10 +0300450 /* per-process-per device QCM data structure */
451 struct qcm_process_device qpd;
452
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300453 /*Apertures*/
454 uint64_t lds_base;
455 uint64_t lds_limit;
456 uint64_t gpuvm_base;
457 uint64_t gpuvm_limit;
458 uint64_t scratch_base;
459 uint64_t scratch_limit;
460
461 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
462 bool bound;
Ben Goza82918f2015-03-25 13:12:20 +0200463
464 /* This flag tells if we should reset all
465 * wavefronts on process termination
466 */
467 bool reset_wavefronts;
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300468};
469
Alexey Skidanov52a5fdc2014-11-19 17:07:00 +0200470#define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
471
Oded Gabbay4a488a72014-07-16 21:08:55 +0300472/* Process data */
473struct kfd_process {
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300474 /*
475 * kfd_process are stored in an mm_struct*->kfd_process*
476 * hash table (kfd_processes in kfd_process.c)
477 */
478 struct hlist_node kfd_processes;
479
480 struct mm_struct *mm;
481
482 struct mutex mutex;
483
484 /*
485 * In any process, the thread that started main() is the lead
486 * thread and outlives the rest.
487 * It is here because amd_iommu_bind_pasid wants a task_struct.
488 */
489 struct task_struct *lead_thread;
490
491 /* We want to receive a notification when the mm_struct is destroyed */
492 struct mmu_notifier mmu_notifier;
493
494 /* Use for delayed freeing of kfd_process structure */
495 struct rcu_head rcu;
496
497 unsigned int pasid;
498
499 /*
500 * List of kfd_process_device structures,
501 * one for each device the process is using.
502 */
503 struct list_head per_device_data;
504
Ben Goz45102042014-07-17 01:04:10 +0300505 struct process_queue_manager pqm;
506
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300507 /* The process's queues. */
508 size_t queue_array_size;
509
510 /* Size is queue_array_size, up to MAX_PROCESS_QUEUES. */
511 struct kfd_queue **queues;
512
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300513 /*Is the user space process 32 bit?*/
514 bool is_32bit_user_mode;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300515
516 /* Event-related data */
517 struct mutex event_mutex;
518 /* All events in process hashed by ID, linked on kfd_event.events. */
519 DECLARE_HASHTABLE(events, 4);
Kent Russell8eabaf52017-08-15 23:00:04 -0400520 /* struct slot_page_header.event_pages */
521 struct list_head signal_event_pages;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300522 u32 next_nonsignal_event_id;
523 size_t signal_event_count;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300524};
525
Oded Gabbay76baee62014-12-29 14:20:05 +0200526/**
527 * Ioctl function type.
528 *
529 * \param filep pointer to file structure.
530 * \param p amdkfd process pointer.
531 * \param data pointer to arg that was copied from user.
532 */
533typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
534 void *data);
535
536struct amdkfd_ioctl_desc {
537 unsigned int cmd;
538 int flags;
539 amdkfd_ioctl_t *func;
540 unsigned int cmd_drv;
541 const char *name;
542};
543
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300544void kfd_process_create_wq(void);
545void kfd_process_destroy_wq(void);
546struct kfd_process *kfd_create_process(const struct task_struct *);
547struct kfd_process *kfd_get_process(const struct task_struct *);
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300548struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300549
Ben Goz64c7f8c2014-07-17 01:27:00 +0300550struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
551 struct kfd_process *p);
Oded Gabbayb17f0682014-07-17 00:06:27 +0300552void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid);
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300553struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
Alexey Skidanov093c7d82014-11-18 14:00:04 +0200554 struct kfd_process *p);
555struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
556 struct kfd_process *p);
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300557
Alexey Skidanov775921e2014-07-17 01:49:36 +0300558/* Process device data iterator */
Kent Russell8eabaf52017-08-15 23:00:04 -0400559struct kfd_process_device *kfd_get_first_process_device_data(
560 struct kfd_process *p);
561struct kfd_process_device *kfd_get_next_process_device_data(
562 struct kfd_process *p,
Alexey Skidanov775921e2014-07-17 01:49:36 +0300563 struct kfd_process_device *pdd);
564bool kfd_has_process_device_data(struct kfd_process *p);
565
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300566/* PASIDs */
567int kfd_pasid_init(void);
568void kfd_pasid_exit(void);
569bool kfd_set_pasid_limit(unsigned int new_limit);
570unsigned int kfd_get_pasid_limit(void);
571unsigned int kfd_pasid_alloc(void);
572void kfd_pasid_free(unsigned int pasid);
573
574/* Doorbells */
Felix Kuehling735df2b2017-08-15 23:00:10 -0400575int kfd_doorbell_init(struct kfd_dev *kfd);
576void kfd_doorbell_fini(struct kfd_dev *kfd);
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300577int 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
Oded Gabbay6e810902014-10-27 14:36:07 +0200587/* 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
Oded Gabbay4a488a72014-07-16 21:08:55 +0300594extern struct device *kfd_device;
595
Evgeny Pinchuk5b5c4e42014-07-16 21:22:32 +0300596/* 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
Oded Gabbay4a488a72014-07-16 21:08:55 +0300605/* Interrupts */
Andrew Lewycky2249d552014-07-17 01:37:30 +0300606int kfd_interrupt_init(struct kfd_dev *dev);
607void kfd_interrupt_exit(struct kfd_dev *dev);
Andrew Lewyckyb3f5e6b2014-07-17 01:37:30 +0300608void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
Andrew Lewycky2249d552014-07-17 01:37:30 +0300609bool 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);
Oded Gabbay4a488a72014-07-16 21:08:55 +0300611
612/* Power Management */
Andrew Lewyckyb3f5e6b2014-07-17 01:37:30 +0300613void kgd2kfd_suspend(struct kfd_dev *kfd);
614int kgd2kfd_resume(struct kfd_dev *kfd);
Oded Gabbay4a488a72014-07-16 21:08:55 +0300615
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300616/* amdkfd Apertures */
617int kfd_init_apertures(struct kfd_process *process);
618
Ben Gozed6e6a32014-07-17 00:45:35 +0300619/* Queue Context Management */
Ben Goz77669eb2015-01-03 22:12:31 +0200620struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd);
Ben Goz241f24f2014-07-17 00:55:28 +0300621
Edward O'Callaghane88a6142016-09-17 15:01:45 +1000622int init_queue(struct queue **q, const struct queue_properties *properties);
Ben Gozed6e6a32014-07-17 00:45:35 +0300623void uninit_queue(struct queue *q);
Ben Goz45102042014-07-17 01:04:10 +0300624void print_queue_properties(struct queue_properties *q);
Ben Gozed6e6a32014-07-17 00:45:35 +0300625void print_queue(struct queue *q);
626
Ben Goz64c7f8c2014-07-17 01:27:00 +0300627struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
628 struct kfd_dev *dev);
Ben Goz4b8f5892015-01-04 11:24:25 +0200629struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
630 struct kfd_dev *dev);
631struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
632 struct kfd_dev *dev);
Ben Goz64c7f8c2014-07-17 01:27:00 +0300633struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
634void device_queue_manager_uninit(struct device_queue_manager *dqm);
Ben Goz241f24f2014-07-17 00:55:28 +0300635struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
636 enum kfd_queue_type type);
637void kernel_queue_uninit(struct kernel_queue *kq);
638
Ben Goz45102042014-07-17 01:04:10 +0300639/* Process Queue Manager */
640struct process_queue_node {
641 struct queue *q;
642 struct kernel_queue *kq;
643 struct list_head process_queue_list;
644};
645
646int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
647void pqm_uninit(struct process_queue_manager *pqm);
648int pqm_create_queue(struct process_queue_manager *pqm,
649 struct kfd_dev *dev,
650 struct file *f,
651 struct queue_properties *properties,
652 unsigned int flags,
653 enum kfd_queue_type type,
654 unsigned int *qid);
655int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
656int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
657 struct queue_properties *p);
Yair Shacharfbeb6612015-05-20 13:48:26 +0300658struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
659 unsigned int qid);
Ben Goz45102042014-07-17 01:04:10 +0300660
Yair Shachar788bf832015-05-20 13:58:12 +0300661int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
662 unsigned int fence_value,
663 unsigned long timeout);
664
Ben Gozed6e6a32014-07-17 00:45:35 +0300665/* Packet Manager */
666
Ben Goz241f24f2014-07-17 00:55:28 +0300667#define KFD_HIQ_TIMEOUT (500)
668
Ben Goz64c7f8c2014-07-17 01:27:00 +0300669#define KFD_FENCE_COMPLETED (100)
670#define KFD_FENCE_INIT (10)
Ben Goz241f24f2014-07-17 00:55:28 +0300671#define KFD_UNMAP_LATENCY (150)
672
Ben Gozed6e6a32014-07-17 00:45:35 +0300673struct packet_manager {
674 struct device_queue_manager *dqm;
675 struct kernel_queue *priv_queue;
676 struct mutex lock;
677 bool allocated;
678 struct kfd_mem_obj *ib_buffer_obj;
679};
680
Ben Goz64c7f8c2014-07-17 01:27:00 +0300681int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
682void pm_uninit(struct packet_manager *pm);
683int pm_send_set_resources(struct packet_manager *pm,
684 struct scheduling_resources *res);
685int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
686int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
687 uint32_t fence_value);
688
689int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
690 enum kfd_preempt_type_filter mode,
691 uint32_t filter_param, bool reset,
692 unsigned int sdma_engine);
693
Ben Goz241f24f2014-07-17 00:55:28 +0300694void pm_release_ib(struct packet_manager *pm);
695
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300696uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
697phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
698 struct kfd_process *process);
699
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300700/* Events */
701extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
Alexey Skidanov930c5ff2014-11-25 10:34:31 +0200702extern const struct kfd_device_global_init_class device_global_init_class_cik;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300703
704enum kfd_event_wait_result {
705 KFD_WAIT_COMPLETE,
706 KFD_WAIT_TIMEOUT,
707 KFD_WAIT_ERROR
708};
709
710void kfd_event_init_process(struct kfd_process *p);
711void kfd_event_free_process(struct kfd_process *p);
712int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
713int kfd_wait_on_events(struct kfd_process *p,
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300714 uint32_t num_events, void __user *data,
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300715 bool all, uint32_t user_timeout_ms,
716 enum kfd_event_wait_result *wait_result);
717void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
718 uint32_t valid_id_bits);
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300719void kfd_signal_iommu_event(struct kfd_dev *dev,
720 unsigned int pasid, unsigned long address,
721 bool is_write_requested, bool is_execute_requested);
Alexey Skidanov930c5ff2014-11-25 10:34:31 +0200722void kfd_signal_hw_exception_event(unsigned int pasid);
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300723int kfd_set_event(struct kfd_process *p, uint32_t event_id);
724int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
725int kfd_event_create(struct file *devkfd, struct kfd_process *p,
726 uint32_t event_type, bool auto_reset, uint32_t node_id,
727 uint32_t *event_id, uint32_t *event_trigger_data,
728 uint64_t *event_page_offset, uint32_t *event_slot_index);
729int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
730
Ben Gozc3447e82015-05-20 18:05:44 +0300731int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
732
Oded Gabbay4a488a72014-07-16 21:08:55 +0300733#endif