1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 */
22
23 #ifndef KFD_PRIV_H_INCLUDED
24 #define KFD_PRIV_H_INCLUDED
25
26 #include <linux/hashtable.h>
27 #include <linux/mmu_notifier.h>
28 #include <linux/mutex.h>
29 #include <linux/types.h>
30 #include <linux/atomic.h>
31 #include <linux/workqueue.h>
32 #include <linux/spinlock.h>
33 #include <linux/kfd_ioctl.h>
34 #include <linux/idr.h>
35 #include <linux/kfifo.h>
36 #include <linux/seq_file.h>
37 #include <linux/kref.h>
38 #include <kgd_kfd_interface.h>
39
40 #include "amd_shared.h"
41
42 #define KFD_MAX_RING_ENTRY_SIZE 8
43
44 #define KFD_SYSFS_FILE_MODE 0444
45
46 /* GPU ID hash width in bits */
47 #define KFD_GPU_ID_HASH_WIDTH 16
48
49 /* Use upper bits of mmap offset to store KFD driver specific information.
50 * BITS[63:62] - Encode MMAP type
51 * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to
52 * BITS[45:0] - MMAP offset value
53 *
54 * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these
55 * defines are w.r.t to PAGE_SIZE
56 */
57 #define KFD_MMAP_TYPE_SHIFT (62 - PAGE_SHIFT)
58 #define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT)
59 #define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT)
60 #define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT)
61 #define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT)
62
63 #define KFD_MMAP_GPU_ID_SHIFT (46 - PAGE_SHIFT)
64 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \
65 << KFD_MMAP_GPU_ID_SHIFT)
66 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\
67 & KFD_MMAP_GPU_ID_MASK)
68 #define KFD_MMAP_GPU_ID_GET(offset) ((offset & KFD_MMAP_GPU_ID_MASK) \
69 >> KFD_MMAP_GPU_ID_SHIFT)
70
71 #define KFD_MMAP_OFFSET_VALUE_MASK (0x3FFFFFFFFFFFULL >> PAGE_SHIFT)
72 #define KFD_MMAP_OFFSET_VALUE_GET(offset) (offset & KFD_MMAP_OFFSET_VALUE_MASK)
73
74 /*
75 * When working with cp scheduler we should assign the HIQ manually or via
76 * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot
77 * definitions for Kaveri. In Kaveri only the first ME queues participates
78 * in the cp scheduling taking that in mind we set the HIQ slot in the
79 * second ME.
80 */
81 #define KFD_CIK_HIQ_PIPE 4
82 #define KFD_CIK_HIQ_QUEUE 0
83
84 /* Macro for allocating structures */
85 #define kfd_alloc_struct(ptr_to_struct) \
86 ((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
87
88 #define KFD_MAX_NUM_OF_PROCESSES 512
89 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
90
91 /*
92 * Size of the per-process TBA+TMA buffer: 2 pages
93 *
94 * The first page is the TBA used for the CWSR ISA code. The second
95 * page is used as TMA for daisy changing a user-mode trap handler.
96 */
97 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
98 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE
99
100 /*
101 * Kernel module parameter to specify maximum number of supported queues per
102 * device
103 */
104 extern int max_num_of_queues_per_device;
105
106 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE_DEFAULT 4096
107 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \
108 (KFD_MAX_NUM_OF_PROCESSES * \
109 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
110
111 #define KFD_KERNEL_QUEUE_SIZE 2048
112
113 /* Kernel module parameter to specify the scheduling policy */
114 extern int sched_policy;
115
116 /*
117 * Kernel module parameter to specify the maximum process
118 * number per HW scheduler
119 */
120 extern int hws_max_conc_proc;
121
122 extern int cwsr_enable;
123
124 /*
125 * Kernel module parameter to specify whether to send sigterm to HSA process on
126 * unhandled exception
127 */
128 extern int send_sigterm;
129
130 /*
131 * This kernel module is used to simulate large bar machine on non-large bar
132 * enabled machines.
133 */
134 extern int debug_largebar;
135
136 /*
137 * Ignore CRAT table during KFD initialization, can be used to work around
138 * broken CRAT tables on some AMD systems
139 */
140 extern int ignore_crat;
141
142 /*
143 * Set sh_mem_config.retry_disable on Vega10
144 */
145 extern int noretry;
146
147 /*
148 * Halt if HWS hang is detected
149 */
150 extern int halt_if_hws_hang;
151
152 /**
153 * enum kfd_sched_policy
154 *
155 * @KFD_SCHED_POLICY_HWS: H/W scheduling policy known as command processor (cp)
156 * scheduling. In this scheduling mode we're using the firmware code to
157 * schedule the user mode queues and kernel queues such as HIQ and DIQ.
158 * the HIQ queue is used as a special queue that dispatches the configuration
159 * to the cp and the user mode queues list that are currently running.
160 * the DIQ queue is a debugging queue that dispatches debugging commands to the
161 * firmware.
162 * in this scheduling mode user mode queues over subscription feature is
163 * enabled.
164 *
165 * @KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION: The same as above but the over
166 * subscription feature disabled.
167 *
168 * @KFD_SCHED_POLICY_NO_HWS: no H/W scheduling policy is a mode which directly
169 * set the command processor registers and sets the queues "manually". This
170 * mode is used *ONLY* for debugging proposes.
171 *
172 */
173 enum kfd_sched_policy {
174 KFD_SCHED_POLICY_HWS = 0,
175 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION,
176 KFD_SCHED_POLICY_NO_HWS
177 };
178
179 enum cache_policy {
180 cache_policy_coherent,
181 cache_policy_noncoherent
182 };
183
184 #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10)
185
186 struct kfd_event_interrupt_class {
187 bool (*interrupt_isr)(struct kfd_dev *dev,
188 const uint32_t *ih_ring_entry, uint32_t *patched_ihre,
189 bool *patched_flag);
190 void (*interrupt_wq)(struct kfd_dev *dev,
191 const uint32_t *ih_ring_entry);
192 };
193
194 struct kfd_device_info {
195 enum amd_asic_type asic_family;
196 const struct kfd_event_interrupt_class *event_interrupt_class;
197 unsigned int max_pasid_bits;
198 unsigned int max_no_of_hqd;
199 unsigned int doorbell_size;
200 size_t ih_ring_entry_size;
201 uint8_t num_of_watch_points;
202 uint16_t mqd_size_aligned;
203 bool supports_cwsr;
204 bool needs_iommu_device;
205 bool needs_pci_atomics;
206 unsigned int num_sdma_engines;
207 };
208
209 struct kfd_mem_obj {
210 uint32_t range_start;
211 uint32_t range_end;
212 uint64_t gpu_addr;
213 uint32_t *cpu_ptr;
214 void *gtt_mem;
215 };
216
217 struct kfd_vmid_info {
218 uint32_t first_vmid_kfd;
219 uint32_t last_vmid_kfd;
220 uint32_t vmid_num_kfd;
221 };
222
223 struct kfd_dev {
224 struct kgd_dev *kgd;
225
226 const struct kfd_device_info *device_info;
227 struct pci_dev *pdev;
228
229 unsigned int id; /* topology stub index */
230
231 phys_addr_t doorbell_base; /* Start of actual doorbells used by
232 * KFD. It is aligned for mapping
233 * into user mode
234 */
235 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell
236 * to HW doorbell, GFX reserved some
237 * at the start)
238 */
239 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
240 * page used by kernel queue
241 */
242
243 struct kgd2kfd_shared_resources shared_resources;
244 struct kfd_vmid_info vm_info;
245
246 const struct kfd2kgd_calls *kfd2kgd;
247 struct mutex doorbell_mutex;
248 DECLARE_BITMAP(doorbell_available_index,
249 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
250
251 void *gtt_mem;
252 uint64_t gtt_start_gpu_addr;
253 void *gtt_start_cpu_ptr;
254 void *gtt_sa_bitmap;
255 struct mutex gtt_sa_lock;
256 unsigned int gtt_sa_chunk_size;
257 unsigned int gtt_sa_num_of_chunks;
258
259 /* Interrupts */
260 struct kfifo ih_fifo;
261 struct workqueue_struct *ih_wq;
262 struct work_struct interrupt_work;
263 spinlock_t interrupt_lock;
264
265 /* QCM Device instance */
266 struct device_queue_manager *dqm;
267
268 bool init_complete;
269 /*
270 * Interrupts of interest to KFD are copied
271 * from the HW ring into a SW ring.
272 */
273 bool interrupts_active;
274
275 /* Debug manager */
276 struct kfd_dbgmgr *dbgmgr;
277
278 /* Maximum process number mapped to HW scheduler */
279 unsigned int max_proc_per_quantum;
280
281 /* CWSR */
282 bool cwsr_enabled;
283 const void *cwsr_isa;
284 unsigned int cwsr_isa_size;
285 };
286
287 /* KGD2KFD callbacks */
288 void kgd2kfd_exit(void);
289 struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd,
290 struct pci_dev *pdev, const struct kfd2kgd_calls *f2g);
291 bool kgd2kfd_device_init(struct kfd_dev *kfd,
292 const struct kgd2kfd_shared_resources *gpu_resources);
293 void kgd2kfd_device_exit(struct kfd_dev *kfd);
294
295 enum kfd_mempool {
296 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
297 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
298 KFD_MEMPOOL_FRAMEBUFFER = 3,
299 };
300
301 /* Character device interface */
302 int kfd_chardev_init(void);
303 void kfd_chardev_exit(void);
304 struct device *kfd_chardev(void);
305
306 /**
307 * enum kfd_unmap_queues_filter
308 *
309 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
310 *
311 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
312 * running queues list.
313 *
314 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
315 * specific process.
316 *
317 */
318 enum kfd_unmap_queues_filter {
319 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
320 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
321 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
322 KFD_UNMAP_QUEUES_FILTER_BY_PASID
323 };
324
325 /**
326 * enum kfd_queue_type
327 *
328 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
329 *
330 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
331 *
332 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
333 *
334 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
335 */
336 enum kfd_queue_type {
337 KFD_QUEUE_TYPE_COMPUTE,
338 KFD_QUEUE_TYPE_SDMA,
339 KFD_QUEUE_TYPE_HIQ,
340 KFD_QUEUE_TYPE_DIQ
341 };
342
343 enum kfd_queue_format {
344 KFD_QUEUE_FORMAT_PM4,
345 KFD_QUEUE_FORMAT_AQL
346 };
347
348 /**
349 * struct queue_properties
350 *
351 * @type: The queue type.
352 *
353 * @queue_id: Queue identifier.
354 *
355 * @queue_address: Queue ring buffer address.
356 *
357 * @queue_size: Queue ring buffer size.
358 *
359 * @priority: Defines the queue priority relative to other queues in the
360 * process.
361 * This is just an indication and HW scheduling may override the priority as
362 * necessary while keeping the relative prioritization.
363 * the priority granularity is from 0 to f which f is the highest priority.
364 * currently all queues are initialized with the highest priority.
365 *
366 * @queue_percent: This field is partially implemented and currently a zero in
367 * this field defines that the queue is non active.
368 *
369 * @read_ptr: User space address which points to the number of dwords the
370 * cp read from the ring buffer. This field updates automatically by the H/W.
371 *
372 * @write_ptr: Defines the number of dwords written to the ring buffer.
373 *
374 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
375 * the queue ring buffer. This field should be similar to write_ptr and the
376 * user should update this field after he updated the write_ptr.
377 *
378 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
379 *
380 * @is_interop: Defines if this is a interop queue. Interop queue means that
381 * the queue can access both graphics and compute resources.
382 *
383 * @is_evicted: Defines if the queue is evicted. Only active queues
384 * are evicted, rendering them inactive.
385 *
386 * @is_active: Defines if the queue is active or not. @is_active and
387 * @is_evicted are protected by the DQM lock.
388 *
389 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
390 * of the queue.
391 *
392 * This structure represents the queue properties for each queue no matter if
393 * it's user mode or kernel mode queue.
394 *
395 */
396 struct queue_properties {
397 enum kfd_queue_type type;
398 enum kfd_queue_format format;
399 unsigned int queue_id;
400 uint64_t queue_address;
401 uint64_t queue_size;
402 uint32_t priority;
403 uint32_t queue_percent;
404 uint32_t *read_ptr;
405 uint32_t *write_ptr;
406 void __iomem *doorbell_ptr;
407 uint32_t doorbell_off;
408 bool is_interop;
409 bool is_evicted;
410 bool is_active;
411 /* Not relevant for user mode queues in cp scheduling */
412 unsigned int vmid;
413 /* Relevant only for sdma queues*/
414 uint32_t sdma_engine_id;
415 uint32_t sdma_queue_id;
416 uint32_t sdma_vm_addr;
417 /* Relevant only for VI */
418 uint64_t eop_ring_buffer_address;
419 uint32_t eop_ring_buffer_size;
420 uint64_t ctx_save_restore_area_address;
421 uint32_t ctx_save_restore_area_size;
422 uint32_t ctl_stack_size;
423 uint64_t tba_addr;
424 uint64_t tma_addr;
425 /* Relevant for CU */
426 uint32_t cu_mask_count; /* Must be a multiple of 32 */
427 uint32_t *cu_mask;
428 };
429
430 /**
431 * struct queue
432 *
433 * @list: Queue linked list.
434 *
435 * @mqd: The queue MQD.
436 *
437 * @mqd_mem_obj: The MQD local gpu memory object.
438 *
439 * @gart_mqd_addr: The MQD gart mc address.
440 *
441 * @properties: The queue properties.
442 *
443 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
444 * that the queue should be execute on.
445 *
446 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
447 * id.
448 *
449 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
450 *
451 * @process: The kfd process that created this queue.
452 *
453 * @device: The kfd device that created this queue.
454 *
455 * This structure represents user mode compute queues.
456 * It contains all the necessary data to handle such queues.
457 *
458 */
459
460 struct queue {
461 struct list_head list;
462 void *mqd;
463 struct kfd_mem_obj *mqd_mem_obj;
464 uint64_t gart_mqd_addr;
465 struct queue_properties properties;
466
467 uint32_t mec;
468 uint32_t pipe;
469 uint32_t queue;
470
471 unsigned int sdma_id;
472 unsigned int doorbell_id;
473
474 struct kfd_process *process;
475 struct kfd_dev *device;
476 };
477
478 /*
479 * Please read the kfd_mqd_manager.h description.
480 */
481 enum KFD_MQD_TYPE {
482 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */
483 KFD_MQD_TYPE_HIQ, /* for hiq */
484 KFD_MQD_TYPE_CP, /* for cp queues and diq */
485 KFD_MQD_TYPE_SDMA, /* for sdma queues */
486 KFD_MQD_TYPE_MAX
487 };
488
489 struct scheduling_resources {
490 unsigned int vmid_mask;
491 enum kfd_queue_type type;
492 uint64_t queue_mask;
493 uint64_t gws_mask;
494 uint32_t oac_mask;
495 uint32_t gds_heap_base;
496 uint32_t gds_heap_size;
497 };
498
499 struct process_queue_manager {
500 /* data */
501 struct kfd_process *process;
502 struct list_head queues;
503 unsigned long *queue_slot_bitmap;
504 };
505
506 struct qcm_process_device {
507 /* The Device Queue Manager that owns this data */
508 struct device_queue_manager *dqm;
509 struct process_queue_manager *pqm;
510 /* Queues list */
511 struct list_head queues_list;
512 struct list_head priv_queue_list;
513
514 unsigned int queue_count;
515 unsigned int vmid;
516 bool is_debug;
517 unsigned int evicted; /* eviction counter, 0=active */
518
519 /* This flag tells if we should reset all wavefronts on
520 * process termination
521 */
522 bool reset_wavefronts;
523
524 /*
525 * All the memory management data should be here too
526 */
527 uint64_t gds_context_area;
528 uint32_t sh_mem_config;
529 uint32_t sh_mem_bases;
530 uint32_t sh_mem_ape1_base;
531 uint32_t sh_mem_ape1_limit;
532 uint32_t page_table_base;
533 uint32_t gds_size;
534 uint32_t num_gws;
535 uint32_t num_oac;
536 uint32_t sh_hidden_private_base;
537
538 /* CWSR memory */
539 void *cwsr_kaddr;
540 uint64_t cwsr_base;
541 uint64_t tba_addr;
542 uint64_t tma_addr;
543
544 /* IB memory */
545 uint64_t ib_base;
546 void *ib_kaddr;
547
548 /* doorbell resources per process per device */
549 unsigned long *doorbell_bitmap;
550 };
551
552 /* KFD Memory Eviction */
553
554 /* Approx. wait time before attempting to restore evicted BOs */
555 #define PROCESS_RESTORE_TIME_MS 100
556 /* Approx. back off time if restore fails due to lack of memory */
557 #define PROCESS_BACK_OFF_TIME_MS 100
558 /* Approx. time before evicting the process again */
559 #define PROCESS_ACTIVE_TIME_MS 10
560
561 int kgd2kfd_quiesce_mm(struct mm_struct *mm);
562 int kgd2kfd_resume_mm(struct mm_struct *mm);
563 int kgd2kfd_schedule_evict_and_restore_process(struct mm_struct *mm,
564 struct dma_fence *fence);
565
566 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
567 * idr_handle in the least significant 4 bytes
568 */
569 #define MAKE_HANDLE(gpu_id, idr_handle) \
570 (((uint64_t)(gpu_id) << 32) + idr_handle)
571 #define GET_GPU_ID(handle) (handle >> 32)
572 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
573
574 enum kfd_pdd_bound {
575 PDD_UNBOUND = 0,
576 PDD_BOUND,
577 PDD_BOUND_SUSPENDED,
578 };
579
580 /* Data that is per-process-per device. */
581 struct kfd_process_device {
582 /*
583 * List of all per-device data for a process.
584 * Starts from kfd_process.per_device_data.
585 */
586 struct list_head per_device_list;
587
588 /* The device that owns this data. */
589 struct kfd_dev *dev;
590
591 /* The process that owns this kfd_process_device. */
592 struct kfd_process *process;
593
594 /* per-process-per device QCM data structure */
595 struct qcm_process_device qpd;
596
597 /*Apertures*/
598 uint64_t lds_base;
599 uint64_t lds_limit;
600 uint64_t gpuvm_base;
601 uint64_t gpuvm_limit;
602 uint64_t scratch_base;
603 uint64_t scratch_limit;
604
605 /* VM context for GPUVM allocations */
606 struct file *drm_file;
607 void *vm;
608
609 /* GPUVM allocations storage */
610 struct idr alloc_idr;
611
612 /* Flag used to tell the pdd has dequeued from the dqm.
613 * This is used to prevent dev->dqm->ops.process_termination() from
614 * being called twice when it is already called in IOMMU callback
615 * function.
616 */
617 bool already_dequeued;
618
619 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
620 enum kfd_pdd_bound bound;
621 };
622
623 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
624
625 /* Process data */
626 struct kfd_process {
627 /*
628 * kfd_process are stored in an mm_struct*->kfd_process*
629 * hash table (kfd_processes in kfd_process.c)
630 */
631 struct hlist_node kfd_processes;
632
633 /*
634 * Opaque pointer to mm_struct. We don't hold a reference to
635 * it so it should never be dereferenced from here. This is
636 * only used for looking up processes by their mm.
637 */
638 void *mm;
639
640 struct kref ref;
641 struct work_struct release_work;
642
643 struct mutex mutex;
644
645 /*
646 * In any process, the thread that started main() is the lead
647 * thread and outlives the rest.
648 * It is here because amd_iommu_bind_pasid wants a task_struct.
649 * It can also be used for safely getting a reference to the
650 * mm_struct of the process.
651 */
652 struct task_struct *lead_thread;
653
654 /* We want to receive a notification when the mm_struct is destroyed */
655 struct mmu_notifier mmu_notifier;
656
657 /* Use for delayed freeing of kfd_process structure */
658 struct rcu_head rcu;
659
660 unsigned int pasid;
661 unsigned int doorbell_index;
662
663 /*
664 * List of kfd_process_device structures,
665 * one for each device the process is using.
666 */
667 struct list_head per_device_data;
668
669 struct process_queue_manager pqm;
670
671 /*Is the user space process 32 bit?*/
672 bool is_32bit_user_mode;
673
674 /* Event-related data */
675 struct mutex event_mutex;
676 /* Event ID allocator and lookup */
677 struct idr event_idr;
678 /* Event page */
679 struct kfd_signal_page *signal_page;
680 size_t signal_mapped_size;
681 size_t signal_event_count;
682 bool signal_event_limit_reached;
683
684 /* Information used for memory eviction */
685 void *kgd_process_info;
686 /* Eviction fence that is attached to all the BOs of this process. The
687 * fence will be triggered during eviction and new one will be created
688 * during restore
689 */
690 struct dma_fence *ef;
691
692 /* Work items for evicting and restoring BOs */
693 struct delayed_work eviction_work;
694 struct delayed_work restore_work;
695 /* seqno of the last scheduled eviction */
696 unsigned int last_eviction_seqno;
697 /* Approx. the last timestamp (in jiffies) when the process was
698 * restored after an eviction
699 */
700 unsigned long last_restore_timestamp;
701 };
702
703 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
704 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
705 extern struct srcu_struct kfd_processes_srcu;
706
707 /**
708 * Ioctl function type.
709 *
710 * \param filep pointer to file structure.
711 * \param p amdkfd process pointer.
712 * \param data pointer to arg that was copied from user.
713 */
714 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
715 void *data);
716
717 struct amdkfd_ioctl_desc {
718 unsigned int cmd;
719 int flags;
720 amdkfd_ioctl_t *func;
721 unsigned int cmd_drv;
722 const char *name;
723 };
724
725 int kfd_process_create_wq(void);
726 void kfd_process_destroy_wq(void);
727 struct kfd_process *kfd_create_process(struct file *filep);
728 struct kfd_process *kfd_get_process(const struct task_struct *);
729 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
730 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
731 void kfd_unref_process(struct kfd_process *p);
732 int kfd_process_evict_queues(struct kfd_process *p);
733 int kfd_process_restore_queues(struct kfd_process *p);
734 void kfd_suspend_all_processes(void);
735 int kfd_resume_all_processes(void);
736
737 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
738 struct file *drm_file);
739 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
740 struct kfd_process *p);
741 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
742 struct kfd_process *p);
743 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
744 struct kfd_process *p);
745
746 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
747 struct vm_area_struct *vma);
748
749 /* KFD process API for creating and translating handles */
750 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
751 void *mem);
752 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
753 int handle);
754 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
755 int handle);
756
757 /* Process device data iterator */
758 struct kfd_process_device *kfd_get_first_process_device_data(
759 struct kfd_process *p);
760 struct kfd_process_device *kfd_get_next_process_device_data(
761 struct kfd_process *p,
762 struct kfd_process_device *pdd);
763 bool kfd_has_process_device_data(struct kfd_process *p);
764
765 /* PASIDs */
766 int kfd_pasid_init(void);
767 void kfd_pasid_exit(void);
768 bool kfd_set_pasid_limit(unsigned int new_limit);
769 unsigned int kfd_get_pasid_limit(void);
770 unsigned int kfd_pasid_alloc(void);
771 void kfd_pasid_free(unsigned int pasid);
772
773 /* Doorbells */
774 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
775 int kfd_doorbell_init(struct kfd_dev *kfd);
776 void kfd_doorbell_fini(struct kfd_dev *kfd);
777 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
778 struct vm_area_struct *vma);
779 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
780 unsigned int *doorbell_off);
781 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
782 u32 read_kernel_doorbell(u32 __iomem *db);
783 void write_kernel_doorbell(void __iomem *db, u32 value);
784 void write_kernel_doorbell64(void __iomem *db, u64 value);
785 unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
786 struct kfd_process *process,
787 unsigned int doorbell_id);
788 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
789 struct kfd_process *process);
790 int kfd_alloc_process_doorbells(struct kfd_process *process);
791 void kfd_free_process_doorbells(struct kfd_process *process);
792
793 /* GTT Sub-Allocator */
794
795 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
796 struct kfd_mem_obj **mem_obj);
797
798 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
799
800 extern struct device *kfd_device;
801
802 /* Topology */
803 int kfd_topology_init(void);
804 void kfd_topology_shutdown(void);
805 int kfd_topology_add_device(struct kfd_dev *gpu);
806 int kfd_topology_remove_device(struct kfd_dev *gpu);
807 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
808 uint32_t proximity_domain);
809 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
810 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
811 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
812 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
813 int kfd_numa_node_to_apic_id(int numa_node_id);
814
815 /* Interrupts */
816 int kfd_interrupt_init(struct kfd_dev *dev);
817 void kfd_interrupt_exit(struct kfd_dev *dev);
818 void kgd2kfd_interrupt(struct kfd_dev *kfd, const void *ih_ring_entry);
819 bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry);
820 bool interrupt_is_wanted(struct kfd_dev *dev,
821 const uint32_t *ih_ring_entry,
822 uint32_t *patched_ihre, bool *flag);
823
824 /* Power Management */
825 void kgd2kfd_suspend(struct kfd_dev *kfd);
826 int kgd2kfd_resume(struct kfd_dev *kfd);
827
828 /* GPU reset */
829 int kgd2kfd_pre_reset(struct kfd_dev *kfd);
830 int kgd2kfd_post_reset(struct kfd_dev *kfd);
831
832 /* amdkfd Apertures */
833 int kfd_init_apertures(struct kfd_process *process);
834
835 /* Queue Context Management */
836 int init_queue(struct queue **q, const struct queue_properties *properties);
837 void uninit_queue(struct queue *q);
838 void print_queue_properties(struct queue_properties *q);
839 void print_queue(struct queue *q);
840
841 struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
842 struct kfd_dev *dev);
843 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
844 struct kfd_dev *dev);
845 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
846 struct kfd_dev *dev);
847 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
848 struct kfd_dev *dev);
849 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
850 struct kfd_dev *dev);
851 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
852 struct kfd_dev *dev);
853 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
854 void device_queue_manager_uninit(struct device_queue_manager *dqm);
855 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
856 enum kfd_queue_type type);
857 void kernel_queue_uninit(struct kernel_queue *kq);
858 int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid);
859
860 /* Process Queue Manager */
861 struct process_queue_node {
862 struct queue *q;
863 struct kernel_queue *kq;
864 struct list_head process_queue_list;
865 };
866
867 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
868 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
869 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
870 void pqm_uninit(struct process_queue_manager *pqm);
871 int pqm_create_queue(struct process_queue_manager *pqm,
872 struct kfd_dev *dev,
873 struct file *f,
874 struct queue_properties *properties,
875 unsigned int *qid);
876 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
877 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
878 struct queue_properties *p);
879 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
880 struct queue_properties *p);
881 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
882 unsigned int qid);
883
884 int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
885 unsigned int fence_value,
886 unsigned int timeout_ms);
887
888 /* Packet Manager */
889
890 #define KFD_FENCE_COMPLETED (100)
891 #define KFD_FENCE_INIT (10)
892
893 struct packet_manager {
894 struct device_queue_manager *dqm;
895 struct kernel_queue *priv_queue;
896 struct mutex lock;
897 bool allocated;
898 struct kfd_mem_obj *ib_buffer_obj;
899 unsigned int ib_size_bytes;
900
901 const struct packet_manager_funcs *pmf;
902 };
903
904 struct packet_manager_funcs {
905 /* Support ASIC-specific packet formats for PM4 packets */
906 int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
907 struct qcm_process_device *qpd);
908 int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
909 uint64_t ib, size_t ib_size_in_dwords, bool chain);
910 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
911 struct scheduling_resources *res);
912 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
913 struct queue *q, bool is_static);
914 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
915 enum kfd_queue_type type,
916 enum kfd_unmap_queues_filter mode,
917 uint32_t filter_param, bool reset,
918 unsigned int sdma_engine);
919 int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
920 uint64_t fence_address, uint32_t fence_value);
921 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
922
923 /* Packet sizes */
924 int map_process_size;
925 int runlist_size;
926 int set_resources_size;
927 int map_queues_size;
928 int unmap_queues_size;
929 int query_status_size;
930 int release_mem_size;
931 };
932
933 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
934 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
935
936 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
937 void pm_uninit(struct packet_manager *pm);
938 int pm_send_set_resources(struct packet_manager *pm,
939 struct scheduling_resources *res);
940 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
941 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
942 uint32_t fence_value);
943
944 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
945 enum kfd_unmap_queues_filter mode,
946 uint32_t filter_param, bool reset,
947 unsigned int sdma_engine);
948
949 void pm_release_ib(struct packet_manager *pm);
950
951 /* Following PM funcs can be shared among VI and AI */
952 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
953 int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer,
954 struct scheduling_resources *res);
955
956 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
957
958 /* Events */
959 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
960 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
961
962 extern const struct kfd_device_global_init_class device_global_init_class_cik;
963
964 void kfd_event_init_process(struct kfd_process *p);
965 void kfd_event_free_process(struct kfd_process *p);
966 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
967 int kfd_wait_on_events(struct kfd_process *p,
968 uint32_t num_events, void __user *data,
969 bool all, uint32_t user_timeout_ms,
970 uint32_t *wait_result);
971 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
972 uint32_t valid_id_bits);
973 void kfd_signal_iommu_event(struct kfd_dev *dev,
974 unsigned int pasid, unsigned long address,
975 bool is_write_requested, bool is_execute_requested);
976 void kfd_signal_hw_exception_event(unsigned int pasid);
977 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
978 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
979 int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
980 uint64_t size);
981 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
982 uint32_t event_type, bool auto_reset, uint32_t node_id,
983 uint32_t *event_id, uint32_t *event_trigger_data,
984 uint64_t *event_page_offset, uint32_t *event_slot_index);
985 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
986
987 void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid,
988 struct kfd_vm_fault_info *info);
989
990 void kfd_signal_reset_event(struct kfd_dev *dev);
991
992 void kfd_flush_tlb(struct kfd_process_device *pdd);
993
994 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
995
996 bool kfd_is_locked(void);
997
998 /* Debugfs */
999 #if defined(CONFIG_DEBUG_FS)
1000
1001 void kfd_debugfs_init(void);
1002 void kfd_debugfs_fini(void);
1003 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1004 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1005 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1006 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1007 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1008 int pm_debugfs_runlist(struct seq_file *m, void *data);
1009
1010 int kfd_debugfs_hang_hws(struct kfd_dev *dev);
1011 int pm_debugfs_hang_hws(struct packet_manager *pm);
1012 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm);
1013
1014 #else
1015
kfd_debugfs_init(void)1016 static inline void kfd_debugfs_init(void) {}
kfd_debugfs_fini(void)1017 static inline void kfd_debugfs_fini(void) {}
1018
1019 #endif
1020
1021 #endif
1022
