blob: 2e13f5d46d2561d444e87714cb040ab24e742f84 [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
Yong Zhaoe596b902017-09-20 18:10:19 -040036#include "amd_shared.h"
37
Evgeny Pinchuk5b5c4e42014-07-16 21:22:32 +030038#define KFD_SYSFS_FILE_MODE 0444
39
Andrew Lewyckyf3a39812015-05-10 12:15:46 +030040#define KFD_MMAP_DOORBELL_MASK 0x8000000000000
41#define KFD_MMAP_EVENTS_MASK 0x4000000000000
42
Ben Gozed6e6a32014-07-17 00:45:35 +030043/*
44 * When working with cp scheduler we should assign the HIQ manually or via
45 * the radeon driver to a fixed hqd slot, here are the fixed HIQ hqd slot
46 * definitions for Kaveri. In Kaveri only the first ME queues participates
47 * in the cp scheduling taking that in mind we set the HIQ slot in the
48 * second ME.
49 */
50#define KFD_CIK_HIQ_PIPE 4
51#define KFD_CIK_HIQ_QUEUE 0
52
Evgeny Pinchuk5b5c4e42014-07-16 21:22:32 +030053/* GPU ID hash width in bits */
54#define KFD_GPU_ID_HASH_WIDTH 16
55
56/* Macro for allocating structures */
57#define kfd_alloc_struct(ptr_to_struct) \
58 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
59
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030060#define KFD_MAX_NUM_OF_PROCESSES 512
Oded Gabbayb8cbab02015-01-18 13:18:01 +020061#define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030062
63/*
Oded Gabbayb8cbab02015-01-18 13:18:01 +020064 * Kernel module parameter to specify maximum number of supported queues per
65 * device
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030066 */
Oded Gabbayb8cbab02015-01-18 13:18:01 +020067extern int max_num_of_queues_per_device;
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030068
Oded Gabbayb8cbab02015-01-18 13:18:01 +020069#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096
70#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \
71 (KFD_MAX_NUM_OF_PROCESSES * \
72 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
Oded Gabbay19f6d2a2014-07-16 23:25:31 +030073
Ben Gozed6e6a32014-07-17 00:45:35 +030074#define KFD_KERNEL_QUEUE_SIZE 2048
75
Ben Goz31c21fe2014-07-17 00:48:28 +030076/* Kernel module parameter to specify the scheduling policy */
77extern int sched_policy;
78
Oded Gabbay81663012014-12-24 13:30:52 +020079/*
80 * Kernel module parameter to specify whether to send sigterm to HSA process on
81 * unhandled exception
82 */
83extern int send_sigterm;
84
Ben Goz31c21fe2014-07-17 00:48:28 +030085/**
86 * enum kfd_sched_policy
87 *
88 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp)
89 * scheduling. In this scheduling mode we're using the firmware code to
90 * schedule the user mode queues and kernel queues such as HIQ and DIQ.
91 * the HIQ queue is used as a special queue that dispatches the configuration
92 * to the cp and the user mode queues list that are currently running.
93 * the DIQ queue is a debugging queue that dispatches debugging commands to the
94 * firmware.
95 * in this scheduling mode user mode queues over subscription feature is
96 * enabled.
97 *
98 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over
99 * subscription feature disabled.
100 *
101 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly
102 * set the command processor registers and sets the queues "manually". This
103 * mode is used *ONLY* for debugging proposes.
104 *
105 */
106enum kfd_sched_policy {
107 KFD_SCHED_POLICY_HWS = 0,
108 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION,
109 KFD_SCHED_POLICY_NO_HWS
110};
111
Ben Gozed6e6a32014-07-17 00:45:35 +0300112enum cache_policy {
113 cache_policy_coherent,
114 cache_policy_noncoherent
115};
116
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300117struct kfd_event_interrupt_class {
118 bool (*interrupt_isr)(struct kfd_dev *dev,
119 const uint32_t *ih_ring_entry);
120 void (*interrupt_wq)(struct kfd_dev *dev,
121 const uint32_t *ih_ring_entry);
122};
123
Oded Gabbay4a488a72014-07-16 21:08:55 +0300124struct kfd_device_info {
Yong Zhaoe596b902017-09-20 18:10:19 -0400125 enum amd_asic_type asic_family;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300126 const struct kfd_event_interrupt_class *event_interrupt_class;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300127 unsigned int max_pasid_bits;
Yair Shachar992839a2015-05-20 13:43:04 +0300128 unsigned int max_no_of_hqd;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300129 size_t ih_ring_entry_size;
Alexey Skidanovf7c826a2014-10-13 16:35:12 +0300130 uint8_t num_of_watch_points;
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300131 uint16_t mqd_size_aligned;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300132};
133
Oded Gabbay36b5c082014-10-26 09:53:10 +0200134struct kfd_mem_obj {
135 uint32_t range_start;
136 uint32_t range_end;
137 uint64_t gpu_addr;
138 uint32_t *cpu_ptr;
139};
140
Yong Zhao44008d72017-09-20 18:10:18 -0400141struct kfd_vmid_info {
142 uint32_t first_vmid_kfd;
143 uint32_t last_vmid_kfd;
144 uint32_t vmid_num_kfd;
145};
146
Oded Gabbay4a488a72014-07-16 21:08:55 +0300147struct kfd_dev {
148 struct kgd_dev *kgd;
149
150 const struct kfd_device_info *device_info;
151 struct pci_dev *pdev;
152
153 unsigned int id; /* topology stub index */
154
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300155 phys_addr_t doorbell_base; /* Start of actual doorbells used by
156 * KFD. It is aligned for mapping
157 * into user mode
158 */
159 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell
160 * to HW doorbell, GFX reserved some
161 * at the start)
162 */
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300163 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;
Yong Zhao44008d72017-09-20 18:10:18 -0400168 struct kfd_vmid_info vm_info;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300169
Xihan Zhangcea405b2015-03-17 19:32:53 +0800170 const struct kfd2kgd_calls *kfd2kgd;
171 struct mutex doorbell_mutex;
Joe Perchesf761d8b2015-05-19 18:37:51 -0700172 DECLARE_BITMAP(doorbell_available_index,
173 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
Xihan Zhangcea405b2015-03-17 19:32:53 +0800174
Oded Gabbay36b5c082014-10-26 09:53:10 +0200175 void *gtt_mem;
176 uint64_t gtt_start_gpu_addr;
177 void *gtt_start_cpu_ptr;
178 void *gtt_sa_bitmap;
179 struct mutex gtt_sa_lock;
180 unsigned int gtt_sa_chunk_size;
181 unsigned int gtt_sa_num_of_chunks;
182
Andrew Lewycky2249d552014-07-17 01:37:30 +0300183 /* Interrupts */
184 void *interrupt_ring;
185 size_t interrupt_ring_size;
186 atomic_t interrupt_ring_rptr;
187 atomic_t interrupt_ring_wptr;
188 struct work_struct interrupt_work;
189 spinlock_t interrupt_lock;
190
Ben Gozed6e6a32014-07-17 00:45:35 +0300191 /* QCM Device instance */
192 struct device_queue_manager *dqm;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300193
Ben Gozed6e6a32014-07-17 00:45:35 +0300194 bool init_complete;
Andrew Lewycky2249d552014-07-17 01:37:30 +0300195 /*
196 * Interrupts of interest to KFD are copied
197 * from the HW ring into a SW ring.
198 */
199 bool interrupts_active;
Yair Shacharfbeb6612015-05-20 13:48:26 +0300200
201 /* Debug manager */
202 struct kfd_dbgmgr *dbgmgr;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300203};
204
205/* KGD2KFD callbacks */
206void kgd2kfd_exit(void);
Xihan Zhangcea405b2015-03-17 19:32:53 +0800207struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
208 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g);
Oded Gabbay4a488a72014-07-16 21:08:55 +0300209bool kgd2kfd_device_init(struct kfd_dev *kfd,
Xihan Zhangcea405b2015-03-17 19:32:53 +0800210 const struct kgd2kfd_shared_resources *gpu_resources);
Oded Gabbay4a488a72014-07-16 21:08:55 +0300211void kgd2kfd_device_exit(struct kfd_dev *kfd);
212
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300213enum kfd_mempool {
214 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
215 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
216 KFD_MEMPOOL_FRAMEBUFFER = 3,
217};
218
Oded Gabbay4a488a72014-07-16 21:08:55 +0300219/* Character device interface */
220int kfd_chardev_init(void);
221void kfd_chardev_exit(void);
222struct device *kfd_chardev(void);
223
Ben Goz241f24f2014-07-17 00:55:28 +0300224/**
Yong Zhao7da2bcf2017-09-27 00:09:48 -0400225 * enum kfd_unmap_queues_filter
Ben Goz241f24f2014-07-17 00:55:28 +0300226 *
Yong Zhao7da2bcf2017-09-27 00:09:48 -0400227 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
Ben Goz241f24f2014-07-17 00:55:28 +0300228 *
Yong Zhao7da2bcf2017-09-27 00:09:48 -0400229 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
Ben Goz241f24f2014-07-17 00:55:28 +0300230 * running queues list.
231 *
Yong Zhao7da2bcf2017-09-27 00:09:48 -0400232 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
Ben Goz241f24f2014-07-17 00:55:28 +0300233 * specific process.
234 *
235 */
Yong Zhao7da2bcf2017-09-27 00:09:48 -0400236enum kfd_unmap_queues_filter {
237 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
238 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
239 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
240 KFD_UNMAP_QUEUES_FILTER_BY_PASID
Ben Goz241f24f2014-07-17 00:55:28 +0300241};
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300242
Ben Gozed8aab42014-07-17 00:18:51 +0300243/**
244 * enum kfd_queue_type
245 *
246 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
247 *
248 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
249 *
250 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
251 *
252 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
253 */
254enum kfd_queue_type {
255 KFD_QUEUE_TYPE_COMPUTE,
256 KFD_QUEUE_TYPE_SDMA,
257 KFD_QUEUE_TYPE_HIQ,
258 KFD_QUEUE_TYPE_DIQ
259};
260
Ben Goz6e99df52014-07-17 00:36:17 +0300261enum kfd_queue_format {
262 KFD_QUEUE_FORMAT_PM4,
263 KFD_QUEUE_FORMAT_AQL
264};
265
Ben Gozed8aab42014-07-17 00:18:51 +0300266/**
267 * struct queue_properties
268 *
269 * @type: The queue type.
270 *
271 * @queue_id: Queue identifier.
272 *
273 * @queue_address: Queue ring buffer address.
274 *
275 * @queue_size: Queue ring buffer size.
276 *
277 * @priority: Defines the queue priority relative to other queues in the
278 * process.
279 * This is just an indication and HW scheduling may override the priority as
280 * necessary while keeping the relative prioritization.
281 * the priority granularity is from 0 to f which f is the highest priority.
282 * currently all queues are initialized with the highest priority.
283 *
284 * @queue_percent: This field is partially implemented and currently a zero in
285 * this field defines that the queue is non active.
286 *
287 * @read_ptr: User space address which points to the number of dwords the
288 * cp read from the ring buffer. This field updates automatically by the H/W.
289 *
290 * @write_ptr: Defines the number of dwords written to the ring buffer.
291 *
292 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
Kent Russell8eabaf52017-08-15 23:00:04 -0400293 * the queue ring buffer. This field should be similar to write_ptr and the
294 * user should update this field after he updated the write_ptr.
Ben Gozed8aab42014-07-17 00:18:51 +0300295 *
296 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
297 *
Kent Russell8eabaf52017-08-15 23:00:04 -0400298 * @is_interop: Defines if this is a interop queue. Interop queue means that
299 * the queue can access both graphics and compute resources.
Ben Gozed8aab42014-07-17 00:18:51 +0300300 *
301 * @is_active: Defines if the queue is active or not.
302 *
303 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
304 * of the queue.
305 *
306 * This structure represents the queue properties for each queue no matter if
307 * it's user mode or kernel mode queue.
308 *
309 */
310struct queue_properties {
311 enum kfd_queue_type type;
Ben Goz6e99df52014-07-17 00:36:17 +0300312 enum kfd_queue_format format;
Ben Gozed8aab42014-07-17 00:18:51 +0300313 unsigned int queue_id;
314 uint64_t queue_address;
315 uint64_t queue_size;
316 uint32_t priority;
317 uint32_t queue_percent;
318 uint32_t *read_ptr;
319 uint32_t *write_ptr;
Oded Gabbay5cd78de2014-11-20 16:14:56 +0200320 uint32_t __iomem *doorbell_ptr;
Ben Gozed8aab42014-07-17 00:18:51 +0300321 uint32_t doorbell_off;
322 bool is_interop;
323 bool is_active;
324 /* Not relevant for user mode queues in cp scheduling */
325 unsigned int vmid;
Ben Goz77669eb2015-01-03 22:12:31 +0200326 /* Relevant only for sdma queues*/
327 uint32_t sdma_engine_id;
328 uint32_t sdma_queue_id;
329 uint32_t sdma_vm_addr;
Ben Gozff3d04a2015-01-04 10:37:18 +0200330 /* Relevant only for VI */
331 uint64_t eop_ring_buffer_address;
332 uint32_t eop_ring_buffer_size;
333 uint64_t ctx_save_restore_area_address;
334 uint32_t ctx_save_restore_area_size;
Ben Gozed8aab42014-07-17 00:18:51 +0300335};
336
337/**
338 * struct queue
339 *
340 * @list: Queue linked list.
341 *
342 * @mqd: The queue MQD.
343 *
344 * @mqd_mem_obj: The MQD local gpu memory object.
345 *
346 * @gart_mqd_addr: The MQD gart mc address.
347 *
348 * @properties: The queue properties.
349 *
350 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
Kent Russell8eabaf52017-08-15 23:00:04 -0400351 * that the queue should be execute on.
Ben Gozed8aab42014-07-17 00:18:51 +0300352 *
Kent Russell8eabaf52017-08-15 23:00:04 -0400353 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
354 * id.
Ben Gozed8aab42014-07-17 00:18:51 +0300355 *
356 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
357 *
358 * @process: The kfd process that created this queue.
359 *
360 * @device: The kfd device that created this queue.
361 *
362 * This structure represents user mode compute queues.
363 * It contains all the necessary data to handle such queues.
364 *
365 */
366
367struct queue {
368 struct list_head list;
369 void *mqd;
370 struct kfd_mem_obj *mqd_mem_obj;
371 uint64_t gart_mqd_addr;
372 struct queue_properties properties;
373
374 uint32_t mec;
375 uint32_t pipe;
376 uint32_t queue;
377
Ben Goz77669eb2015-01-03 22:12:31 +0200378 unsigned int sdma_id;
379
Ben Gozed8aab42014-07-17 00:18:51 +0300380 struct kfd_process *process;
381 struct kfd_dev *device;
382};
383
Ben Goz6e99df52014-07-17 00:36:17 +0300384/*
385 * Please read the kfd_mqd_manager.h description.
386 */
387enum KFD_MQD_TYPE {
Ben Goz85d258f2015-01-04 10:36:30 +0200388 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */
389 KFD_MQD_TYPE_HIQ, /* for hiq */
390 KFD_MQD_TYPE_CP, /* for cp queues and diq */
391 KFD_MQD_TYPE_SDMA, /* for sdma queues */
Ben Goz6e99df52014-07-17 00:36:17 +0300392 KFD_MQD_TYPE_MAX
393};
394
Ben Goz241f24f2014-07-17 00:55:28 +0300395struct scheduling_resources {
396 unsigned int vmid_mask;
397 enum kfd_queue_type type;
398 uint64_t queue_mask;
399 uint64_t gws_mask;
400 uint32_t oac_mask;
401 uint32_t gds_heap_base;
402 uint32_t gds_heap_size;
403};
404
405struct process_queue_manager {
406 /* data */
407 struct kfd_process *process;
408 unsigned int num_concurrent_processes;
409 struct list_head queues;
410 unsigned long *queue_slot_bitmap;
411};
412
413struct qcm_process_device {
414 /* The Device Queue Manager that owns this data */
415 struct device_queue_manager *dqm;
416 struct process_queue_manager *pqm;
Ben Goz241f24f2014-07-17 00:55:28 +0300417 /* Queues list */
418 struct list_head queues_list;
419 struct list_head priv_queue_list;
420
421 unsigned int queue_count;
422 unsigned int vmid;
423 bool is_debug;
Felix Kuehling9fd3f1bf2017-09-27 00:09:52 -0400424
425 /* This flag tells if we should reset all wavefronts on
426 * process termination
427 */
428 bool reset_wavefronts;
429
Ben Goz241f24f2014-07-17 00:55:28 +0300430 /*
431 * All the memory management data should be here too
432 */
433 uint64_t gds_context_area;
434 uint32_t sh_mem_config;
435 uint32_t sh_mem_bases;
436 uint32_t sh_mem_ape1_base;
437 uint32_t sh_mem_ape1_limit;
438 uint32_t page_table_base;
439 uint32_t gds_size;
440 uint32_t num_gws;
441 uint32_t num_oac;
Moses Reuben6a1c9512017-08-15 23:00:20 -0400442 uint32_t sh_hidden_private_base;
Ben Goz241f24f2014-07-17 00:55:28 +0300443};
444
Yong Zhao733fa1f2017-09-20 18:10:14 -0400445
446enum kfd_pdd_bound {
447 PDD_UNBOUND = 0,
448 PDD_BOUND,
449 PDD_BOUND_SUSPENDED,
450};
451
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300452/* Data that is per-process-per device. */
453struct kfd_process_device {
454 /*
455 * List of all per-device data for a process.
456 * Starts from kfd_process.per_device_data.
457 */
458 struct list_head per_device_list;
459
460 /* The device that owns this data. */
461 struct kfd_dev *dev;
462
Felix Kuehling9fd3f1bf2017-09-27 00:09:52 -0400463 /* The process that owns this kfd_process_device. */
464 struct kfd_process *process;
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300465
Ben Goz45102042014-07-17 01:04:10 +0300466 /* per-process-per device QCM data structure */
467 struct qcm_process_device qpd;
468
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300469 /*Apertures*/
470 uint64_t lds_base;
471 uint64_t lds_limit;
472 uint64_t gpuvm_base;
473 uint64_t gpuvm_limit;
474 uint64_t scratch_base;
475 uint64_t scratch_limit;
476
477 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
Yong Zhao733fa1f2017-09-20 18:10:14 -0400478 enum kfd_pdd_bound bound;
Ben Goza82918f2015-03-25 13:12:20 +0200479
Felix Kuehling9fd3f1bf2017-09-27 00:09:52 -0400480 /* Flag used to tell the pdd has dequeued from the dqm.
481 * This is used to prevent dev->dqm->ops.process_termination() from
482 * being called twice when it is already called in IOMMU callback
483 * function.
Ben Goza82918f2015-03-25 13:12:20 +0200484 */
Felix Kuehling9fd3f1bf2017-09-27 00:09:52 -0400485 bool already_dequeued;
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300486};
487
Alexey Skidanov52a5fdc2014-11-19 17:07:00 +0200488#define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
489
Oded Gabbay4a488a72014-07-16 21:08:55 +0300490/* Process data */
491struct kfd_process {
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300492 /*
493 * kfd_process are stored in an mm_struct*->kfd_process*
494 * hash table (kfd_processes in kfd_process.c)
495 */
496 struct hlist_node kfd_processes;
497
498 struct mm_struct *mm;
499
500 struct mutex mutex;
501
502 /*
503 * In any process, the thread that started main() is the lead
504 * thread and outlives the rest.
505 * It is here because amd_iommu_bind_pasid wants a task_struct.
506 */
507 struct task_struct *lead_thread;
508
509 /* We want to receive a notification when the mm_struct is destroyed */
510 struct mmu_notifier mmu_notifier;
511
512 /* Use for delayed freeing of kfd_process structure */
513 struct rcu_head rcu;
514
515 unsigned int pasid;
Felix Kuehlinga91e70e2017-08-26 02:00:57 -0400516 unsigned int doorbell_index;
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300517
518 /*
519 * List of kfd_process_device structures,
520 * one for each device the process is using.
521 */
522 struct list_head per_device_data;
523
Ben Goz45102042014-07-17 01:04:10 +0300524 struct process_queue_manager pqm;
525
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300526 /* The process's queues. */
527 size_t queue_array_size;
528
529 /* Size is queue_array_size, up to MAX_PROCESS_QUEUES. */
530 struct kfd_queue **queues;
531
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300532 /*Is the user space process 32 bit?*/
533 bool is_32bit_user_mode;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300534
535 /* Event-related data */
536 struct mutex event_mutex;
537 /* All events in process hashed by ID, linked on kfd_event.events. */
538 DECLARE_HASHTABLE(events, 4);
Kent Russell8eabaf52017-08-15 23:00:04 -0400539 /* struct slot_page_header.event_pages */
540 struct list_head signal_event_pages;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300541 u32 next_nonsignal_event_id;
542 size_t signal_event_count;
Felix Kuehlingc9861692017-09-20 18:10:22 -0400543 bool signal_event_limit_reached;
Oded Gabbay4a488a72014-07-16 21:08:55 +0300544};
545
Oded Gabbay76baee62014-12-29 14:20:05 +0200546/**
547 * Ioctl function type.
548 *
549 * \param filep pointer to file structure.
550 * \param p amdkfd process pointer.
551 * \param data pointer to arg that was copied from user.
552 */
553typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
554 void *data);
555
556struct amdkfd_ioctl_desc {
557 unsigned int cmd;
558 int flags;
559 amdkfd_ioctl_t *func;
560 unsigned int cmd_drv;
561 const char *name;
562};
563
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300564void kfd_process_create_wq(void);
565void kfd_process_destroy_wq(void);
566struct kfd_process *kfd_create_process(const struct task_struct *);
567struct kfd_process *kfd_get_process(const struct task_struct *);
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300568struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300569
Ben Goz64c7f8c2014-07-17 01:27:00 +0300570struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
Yong Zhao733fa1f2017-09-20 18:10:14 -0400571 struct kfd_process *p);
572int kfd_bind_processes_to_device(struct kfd_dev *dev);
573void kfd_unbind_processes_from_device(struct kfd_dev *dev);
574void kfd_process_iommu_unbind_callback(struct kfd_dev *dev, unsigned int pasid);
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300575struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
Alexey Skidanov093c7d82014-11-18 14:00:04 +0200576 struct kfd_process *p);
577struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
578 struct kfd_process *p);
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300579
Alexey Skidanov775921e2014-07-17 01:49:36 +0300580/* Process device data iterator */
Kent Russell8eabaf52017-08-15 23:00:04 -0400581struct kfd_process_device *kfd_get_first_process_device_data(
582 struct kfd_process *p);
583struct kfd_process_device *kfd_get_next_process_device_data(
584 struct kfd_process *p,
Alexey Skidanov775921e2014-07-17 01:49:36 +0300585 struct kfd_process_device *pdd);
586bool kfd_has_process_device_data(struct kfd_process *p);
587
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300588/* PASIDs */
589int kfd_pasid_init(void);
590void kfd_pasid_exit(void);
591bool kfd_set_pasid_limit(unsigned int new_limit);
592unsigned int kfd_get_pasid_limit(void);
593unsigned int kfd_pasid_alloc(void);
594void kfd_pasid_free(unsigned int pasid);
595
596/* Doorbells */
Felix Kuehling735df2b2017-08-15 23:00:10 -0400597int kfd_doorbell_init(struct kfd_dev *kfd);
598void kfd_doorbell_fini(struct kfd_dev *kfd);
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300599int kfd_doorbell_mmap(struct kfd_process *process, struct vm_area_struct *vma);
600u32 __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
601 unsigned int *doorbell_off);
602void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
603u32 read_kernel_doorbell(u32 __iomem *db);
604void write_kernel_doorbell(u32 __iomem *db, u32 value);
605unsigned int kfd_queue_id_to_doorbell(struct kfd_dev *kfd,
606 struct kfd_process *process,
607 unsigned int queue_id);
Felix Kuehlinga91e70e2017-08-26 02:00:57 -0400608phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
609 struct kfd_process *process);
610int kfd_alloc_process_doorbells(struct kfd_process *process);
611void kfd_free_process_doorbells(struct kfd_process *process);
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300612
Oded Gabbay6e810902014-10-27 14:36:07 +0200613/* GTT Sub-Allocator */
614
615int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
616 struct kfd_mem_obj **mem_obj);
617
618int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
619
Oded Gabbay4a488a72014-07-16 21:08:55 +0300620extern struct device *kfd_device;
621
Evgeny Pinchuk5b5c4e42014-07-16 21:22:32 +0300622/* Topology */
623int kfd_topology_init(void);
624void kfd_topology_shutdown(void);
625int kfd_topology_add_device(struct kfd_dev *gpu);
626int kfd_topology_remove_device(struct kfd_dev *gpu);
627struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
628struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
629struct kfd_dev *kfd_topology_enum_kfd_devices(uint8_t idx);
630
Oded Gabbay4a488a72014-07-16 21:08:55 +0300631/* Interrupts */
Andrew Lewycky2249d552014-07-17 01:37:30 +0300632int kfd_interrupt_init(struct kfd_dev *dev);
633void kfd_interrupt_exit(struct kfd_dev *dev);
Andrew Lewyckyb3f5e6b2014-07-17 01:37:30 +0300634void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
Andrew Lewycky2249d552014-07-17 01:37:30 +0300635bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry);
636bool interrupt_is_wanted(struct kfd_dev *dev, const uint32_t *ih_ring_entry);
Oded Gabbay4a488a72014-07-16 21:08:55 +0300637
638/* Power Management */
Andrew Lewyckyb3f5e6b2014-07-17 01:37:30 +0300639void kgd2kfd_suspend(struct kfd_dev *kfd);
640int kgd2kfd_resume(struct kfd_dev *kfd);
Oded Gabbay4a488a72014-07-16 21:08:55 +0300641
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300642/* amdkfd Apertures */
643int kfd_init_apertures(struct kfd_process *process);
644
Ben Gozed6e6a32014-07-17 00:45:35 +0300645/* Queue Context Management */
Ben Goz77669eb2015-01-03 22:12:31 +0200646struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd);
Ben Goz241f24f2014-07-17 00:55:28 +0300647
Edward O'Callaghane88a6142016-09-17 15:01:45 +1000648int init_queue(struct queue **q, const struct queue_properties *properties);
Ben Gozed6e6a32014-07-17 00:45:35 +0300649void uninit_queue(struct queue *q);
Ben Goz45102042014-07-17 01:04:10 +0300650void print_queue_properties(struct queue_properties *q);
Ben Gozed6e6a32014-07-17 00:45:35 +0300651void print_queue(struct queue *q);
652
Ben Goz64c7f8c2014-07-17 01:27:00 +0300653struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
654 struct kfd_dev *dev);
Ben Goz4b8f5892015-01-04 11:24:25 +0200655struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
656 struct kfd_dev *dev);
657struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
658 struct kfd_dev *dev);
Ben Goz64c7f8c2014-07-17 01:27:00 +0300659struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
660void device_queue_manager_uninit(struct device_queue_manager *dqm);
Ben Goz241f24f2014-07-17 00:55:28 +0300661struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
662 enum kfd_queue_type type);
663void kernel_queue_uninit(struct kernel_queue *kq);
664
Ben Goz45102042014-07-17 01:04:10 +0300665/* Process Queue Manager */
666struct process_queue_node {
667 struct queue *q;
668 struct kernel_queue *kq;
669 struct list_head process_queue_list;
670};
671
Felix Kuehling9fd3f1bf2017-09-27 00:09:52 -0400672void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
673void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
Ben Goz45102042014-07-17 01:04:10 +0300674int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
675void pqm_uninit(struct process_queue_manager *pqm);
676int pqm_create_queue(struct process_queue_manager *pqm,
677 struct kfd_dev *dev,
678 struct file *f,
679 struct queue_properties *properties,
680 unsigned int flags,
681 enum kfd_queue_type type,
682 unsigned int *qid);
683int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
684int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
685 struct queue_properties *p);
Yair Shacharfbeb6612015-05-20 13:48:26 +0300686struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
687 unsigned int qid);
Ben Goz45102042014-07-17 01:04:10 +0300688
Yair Shachar788bf832015-05-20 13:58:12 +0300689int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
690 unsigned int fence_value,
Yong Zhao8c72c3d2017-09-20 18:10:15 -0400691 unsigned int timeout_ms);
Yair Shachar788bf832015-05-20 13:58:12 +0300692
Ben Gozed6e6a32014-07-17 00:45:35 +0300693/* Packet Manager */
694
Ben Goz64c7f8c2014-07-17 01:27:00 +0300695#define KFD_FENCE_COMPLETED (100)
696#define KFD_FENCE_INIT (10)
Ben Goz241f24f2014-07-17 00:55:28 +0300697
Ben Gozed6e6a32014-07-17 00:45:35 +0300698struct packet_manager {
699 struct device_queue_manager *dqm;
700 struct kernel_queue *priv_queue;
701 struct mutex lock;
702 bool allocated;
703 struct kfd_mem_obj *ib_buffer_obj;
704};
705
Ben Goz64c7f8c2014-07-17 01:27:00 +0300706int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
707void pm_uninit(struct packet_manager *pm);
708int pm_send_set_resources(struct packet_manager *pm,
709 struct scheduling_resources *res);
710int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
711int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
712 uint32_t fence_value);
713
714int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
Yong Zhao7da2bcf2017-09-27 00:09:48 -0400715 enum kfd_unmap_queues_filter mode,
Ben Goz64c7f8c2014-07-17 01:27:00 +0300716 uint32_t filter_param, bool reset,
717 unsigned int sdma_engine);
718
Ben Goz241f24f2014-07-17 00:55:28 +0300719void pm_release_ib(struct packet_manager *pm);
720
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300721uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
Oded Gabbay19f6d2a2014-07-16 23:25:31 +0300722
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300723/* Events */
724extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
Alexey Skidanov930c5ff2014-11-25 10:34:31 +0200725extern const struct kfd_device_global_init_class device_global_init_class_cik;
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300726
727enum kfd_event_wait_result {
728 KFD_WAIT_COMPLETE,
729 KFD_WAIT_TIMEOUT,
730 KFD_WAIT_ERROR
731};
732
733void kfd_event_init_process(struct kfd_process *p);
734void kfd_event_free_process(struct kfd_process *p);
735int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
736int kfd_wait_on_events(struct kfd_process *p,
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300737 uint32_t num_events, void __user *data,
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300738 bool all, uint32_t user_timeout_ms,
739 enum kfd_event_wait_result *wait_result);
740void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
741 uint32_t valid_id_bits);
Alexey Skidanov59d3e8b2015-04-14 18:05:49 +0300742void kfd_signal_iommu_event(struct kfd_dev *dev,
743 unsigned int pasid, unsigned long address,
744 bool is_write_requested, bool is_execute_requested);
Alexey Skidanov930c5ff2014-11-25 10:34:31 +0200745void kfd_signal_hw_exception_event(unsigned int pasid);
Andrew Lewyckyf3a39812015-05-10 12:15:46 +0300746int kfd_set_event(struct kfd_process *p, uint32_t event_id);
747int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
748int kfd_event_create(struct file *devkfd, struct kfd_process *p,
749 uint32_t event_type, bool auto_reset, uint32_t node_id,
750 uint32_t *event_id, uint32_t *event_trigger_data,
751 uint64_t *event_page_offset, uint32_t *event_slot_index);
752int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
753
Ben Gozc3447e82015-05-20 18:05:44 +0300754int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
755
Oded Gabbay4a488a72014-07-16 21:08:55 +0300756#endif