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#define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE \
101 (KFD_MAX_NUM_OF_PROCESSES * \
102 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
103
104#define KFD_KERNEL_QUEUE_SIZE 2048
105
106/*
107 * 512 = 0x200
108 * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
109 * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.
110 * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC
111 * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in
112 * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.
113 */
114#define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
115
116
117/*
118 * Kernel module parameter to specify maximum number of supported queues per
119 * device
120 */
121extern int max_num_of_queues_per_device;
122
123
124/* Kernel module parameter to specify the scheduling policy */
125extern int sched_policy;
126
127/*
128 * Kernel module parameter to specify the maximum process
129 * number per HW scheduler
130 */
131extern int hws_max_conc_proc;
132
133extern int cwsr_enable;
134
135/*
136 * Kernel module parameter to specify whether to send sigterm to HSA process on
137 * unhandled exception
138 */
139extern int send_sigterm;
140
141/*
142 * This kernel module is used to simulate large bar machine on non-large bar
143 * enabled machines.
144 */
145extern int debug_largebar;
146
147/*
148 * Ignore CRAT table during KFD initialization, can be used to work around
149 * broken CRAT tables on some AMD systems
150 */
151extern int ignore_crat;
152
153/*
154 * Set sh_mem_config.retry_disable on Vega10
155 */
156extern int noretry;
157
158/*
159 * Halt if HWS hang is detected
160 */
161extern int halt_if_hws_hang;
162
163enum cache_policy {
164 cache_policy_coherent,
165 cache_policy_noncoherent
166};
167
168#define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10)
169
170struct kfd_event_interrupt_class {
171 bool (*interrupt_isr)(struct kfd_dev *dev,
172 const uint32_t *ih_ring_entry, uint32_t *patched_ihre,
173 bool *patched_flag);
174 void (*interrupt_wq)(struct kfd_dev *dev,
175 const uint32_t *ih_ring_entry);
176};
177
178struct kfd_device_info {
179 enum amd_asic_type asic_family;
180 const struct kfd_event_interrupt_class *event_interrupt_class;
181 unsigned int max_pasid_bits;
182 unsigned int max_no_of_hqd;
183 unsigned int doorbell_size;
184 size_t ih_ring_entry_size;
185 uint8_t num_of_watch_points;
186 uint16_t mqd_size_aligned;
187 bool supports_cwsr;
188 bool needs_iommu_device;
189 bool needs_pci_atomics;
190 unsigned int num_sdma_engines;
191 unsigned int num_sdma_queues_per_engine;
192};
193
194struct kfd_mem_obj {
195 uint32_t range_start;
196 uint32_t range_end;
197 uint64_t gpu_addr;
198 uint32_t *cpu_ptr;
199 void *gtt_mem;
200};
201
202struct kfd_vmid_info {
203 uint32_t first_vmid_kfd;
204 uint32_t last_vmid_kfd;
205 uint32_t vmid_num_kfd;
206};
207
208struct kfd_dev {
209 struct kgd_dev *kgd;
210
211 const struct kfd_device_info *device_info;
212 struct pci_dev *pdev;
213
214 unsigned int id; /* topology stub index */
215
216 phys_addr_t doorbell_base; /* Start of actual doorbells used by
217 * KFD. It is aligned for mapping
218 * into user mode
219 */
220 size_t doorbell_id_offset; /* Doorbell offset (from KFD doorbell
221 * to HW doorbell, GFX reserved some
222 * at the start)
223 */
224 u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
225 * page used by kernel queue
226 */
227
228 struct kgd2kfd_shared_resources shared_resources;
229 struct kfd_vmid_info vm_info;
230
231 const struct kfd2kgd_calls *kfd2kgd;
232 struct mutex doorbell_mutex;
233 DECLARE_BITMAP(doorbell_available_index,
234 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
235
236 void *gtt_mem;
237 uint64_t gtt_start_gpu_addr;
238 void *gtt_start_cpu_ptr;
239 void *gtt_sa_bitmap;
240 struct mutex gtt_sa_lock;
241 unsigned int gtt_sa_chunk_size;
242 unsigned int gtt_sa_num_of_chunks;
243
244 /* Interrupts */
245 struct kfifo ih_fifo;
246 struct workqueue_struct *ih_wq;
247 struct work_struct interrupt_work;
248 spinlock_t interrupt_lock;
249
250 /* QCM Device instance */
251 struct device_queue_manager *dqm;
252
253 bool init_complete;
254 /*
255 * Interrupts of interest to KFD are copied
256 * from the HW ring into a SW ring.
257 */
258 bool interrupts_active;
259
260 /* Debug manager */
261 struct kfd_dbgmgr *dbgmgr;
262
263 /* Firmware versions */
264 uint16_t mec_fw_version;
265 uint16_t sdma_fw_version;
266
267 /* Maximum process number mapped to HW scheduler */
268 unsigned int max_proc_per_quantum;
269
270 /* CWSR */
271 bool cwsr_enabled;
272 const void *cwsr_isa;
273 unsigned int cwsr_isa_size;
274
275 /* xGMI */
276 uint64_t hive_id;
277
278 bool pci_atomic_requested;
279};
280
281enum kfd_mempool {
282 KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
283 KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
284 KFD_MEMPOOL_FRAMEBUFFER = 3,
285};
286
287/* Character device interface */
288int kfd_chardev_init(void);
289void kfd_chardev_exit(void);
290struct device *kfd_chardev(void);
291
292/**
293 * enum kfd_unmap_queues_filter
294 *
295 * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
296 *
297 * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
298 * running queues list.
299 *
300 * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
301 * specific process.
302 *
303 */
304enum kfd_unmap_queues_filter {
305 KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
306 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
307 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
308 KFD_UNMAP_QUEUES_FILTER_BY_PASID
309};
310
311/**
312 * enum kfd_queue_type
313 *
314 * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
315 *
316 * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
317 *
318 * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
319 *
320 * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
321 */
322enum kfd_queue_type {
323 KFD_QUEUE_TYPE_COMPUTE,
324 KFD_QUEUE_TYPE_SDMA,
325 KFD_QUEUE_TYPE_HIQ,
326 KFD_QUEUE_TYPE_DIQ
327};
328
329enum kfd_queue_format {
330 KFD_QUEUE_FORMAT_PM4,
331 KFD_QUEUE_FORMAT_AQL
332};
333
334/**
335 * struct queue_properties
336 *
337 * @type: The queue type.
338 *
339 * @queue_id: Queue identifier.
340 *
341 * @queue_address: Queue ring buffer address.
342 *
343 * @queue_size: Queue ring buffer size.
344 *
345 * @priority: Defines the queue priority relative to other queues in the
346 * process.
347 * This is just an indication and HW scheduling may override the priority as
348 * necessary while keeping the relative prioritization.
349 * the priority granularity is from 0 to f which f is the highest priority.
350 * currently all queues are initialized with the highest priority.
351 *
352 * @queue_percent: This field is partially implemented and currently a zero in
353 * this field defines that the queue is non active.
354 *
355 * @read_ptr: User space address which points to the number of dwords the
356 * cp read from the ring buffer. This field updates automatically by the H/W.
357 *
358 * @write_ptr: Defines the number of dwords written to the ring buffer.
359 *
360 * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
361 * the queue ring buffer. This field should be similar to write_ptr and the
362 * user should update this field after he updated the write_ptr.
363 *
364 * @doorbell_off: The doorbell offset in the doorbell pci-bar.
365 *
366 * @is_interop: Defines if this is a interop queue. Interop queue means that
367 * the queue can access both graphics and compute resources.
368 *
369 * @is_evicted: Defines if the queue is evicted. Only active queues
370 * are evicted, rendering them inactive.
371 *
372 * @is_active: Defines if the queue is active or not. @is_active and
373 * @is_evicted are protected by the DQM lock.
374 *
375 * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
376 * of the queue.
377 *
378 * This structure represents the queue properties for each queue no matter if
379 * it's user mode or kernel mode queue.
380 *
381 */
382struct queue_properties {
383 enum kfd_queue_type type;
384 enum kfd_queue_format format;
385 unsigned int queue_id;
386 uint64_t queue_address;
387 uint64_t queue_size;
388 uint32_t priority;
389 uint32_t queue_percent;
390 uint32_t *read_ptr;
391 uint32_t *write_ptr;
392 void __iomem *doorbell_ptr;
393 uint32_t doorbell_off;
394 bool is_interop;
395 bool is_evicted;
396 bool is_active;
397 /* Not relevant for user mode queues in cp scheduling */
398 unsigned int vmid;
399 /* Relevant only for sdma queues*/
400 uint32_t sdma_engine_id;
401 uint32_t sdma_queue_id;
402 uint32_t sdma_vm_addr;
403 /* Relevant only for VI */
404 uint64_t eop_ring_buffer_address;
405 uint32_t eop_ring_buffer_size;
406 uint64_t ctx_save_restore_area_address;
407 uint32_t ctx_save_restore_area_size;
408 uint32_t ctl_stack_size;
409 uint64_t tba_addr;
410 uint64_t tma_addr;
411 /* Relevant for CU */
412 uint32_t cu_mask_count; /* Must be a multiple of 32 */
413 uint32_t *cu_mask;
414};
415
416/**
417 * struct queue
418 *
419 * @list: Queue linked list.
420 *
421 * @mqd: The queue MQD.
422 *
423 * @mqd_mem_obj: The MQD local gpu memory object.
424 *
425 * @gart_mqd_addr: The MQD gart mc address.
426 *
427 * @properties: The queue properties.
428 *
429 * @mec: Used only in no cp scheduling mode and identifies to micro engine id
430 * that the queue should be execute on.
431 *
432 * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
433 * id.
434 *
435 * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
436 *
437 * @process: The kfd process that created this queue.
438 *
439 * @device: The kfd device that created this queue.
440 *
441 * This structure represents user mode compute queues.
442 * It contains all the necessary data to handle such queues.
443 *
444 */
445
446struct queue {
447 struct list_head list;
448 void *mqd;
449 struct kfd_mem_obj *mqd_mem_obj;
450 uint64_t gart_mqd_addr;
451 struct queue_properties properties;
452
453 uint32_t mec;
454 uint32_t pipe;
455 uint32_t queue;
456
457 unsigned int sdma_id;
458 unsigned int doorbell_id;
459
460 struct kfd_process *process;
461 struct kfd_dev *device;
462};
463
464/*
465 * Please read the kfd_mqd_manager.h description.
466 */
467enum KFD_MQD_TYPE {
468 KFD_MQD_TYPE_COMPUTE = 0, /* for no cp scheduling */
469 KFD_MQD_TYPE_HIQ, /* for hiq */
470 KFD_MQD_TYPE_CP, /* for cp queues and diq */
471 KFD_MQD_TYPE_SDMA, /* for sdma queues */
472 KFD_MQD_TYPE_MAX
473};
474
475struct scheduling_resources {
476 unsigned int vmid_mask;
477 enum kfd_queue_type type;
478 uint64_t queue_mask;
479 uint64_t gws_mask;
480 uint32_t oac_mask;
481 uint32_t gds_heap_base;
482 uint32_t gds_heap_size;
483};
484
485struct process_queue_manager {
486 /* data */
487 struct kfd_process *process;
488 struct list_head queues;
489 unsigned long *queue_slot_bitmap;
490};
491
492struct qcm_process_device {
493 /* The Device Queue Manager that owns this data */
494 struct device_queue_manager *dqm;
495 struct process_queue_manager *pqm;
496 /* Queues list */
497 struct list_head queues_list;
498 struct list_head priv_queue_list;
499
500 unsigned int queue_count;
501 unsigned int vmid;
502 bool is_debug;
503 unsigned int evicted; /* eviction counter, 0=active */
504
505 /* This flag tells if we should reset all wavefronts on
506 * process termination
507 */
508 bool reset_wavefronts;
509
510 /*
511 * All the memory management data should be here too
512 */
513 uint64_t gds_context_area;
514 /* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
515 uint64_t page_table_base;
516 uint32_t sh_mem_config;
517 uint32_t sh_mem_bases;
518 uint32_t sh_mem_ape1_base;
519 uint32_t sh_mem_ape1_limit;
520 uint32_t gds_size;
521 uint32_t num_gws;
522 uint32_t num_oac;
523 uint32_t sh_hidden_private_base;
524
525 /* CWSR memory */
526 void *cwsr_kaddr;
527 uint64_t cwsr_base;
528 uint64_t tba_addr;
529 uint64_t tma_addr;
530
531 /* IB memory */
532 uint64_t ib_base;
533 void *ib_kaddr;
534
535 /* doorbell resources per process per device */
536 unsigned long *doorbell_bitmap;
537};
538
539/* KFD Memory Eviction */
540
541/* Approx. wait time before attempting to restore evicted BOs */
542#define PROCESS_RESTORE_TIME_MS 100
543/* Approx. back off time if restore fails due to lack of memory */
544#define PROCESS_BACK_OFF_TIME_MS 100
545/* Approx. time before evicting the process again */
546#define PROCESS_ACTIVE_TIME_MS 10
547
548/* 8 byte handle containing GPU ID in the most significant 4 bytes and
549 * idr_handle in the least significant 4 bytes
550 */
551#define MAKE_HANDLE(gpu_id, idr_handle) \
552 (((uint64_t)(gpu_id) << 32) + idr_handle)
553#define GET_GPU_ID(handle) (handle >> 32)
554#define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
555
556enum kfd_pdd_bound {
557 PDD_UNBOUND = 0,
558 PDD_BOUND,
559 PDD_BOUND_SUSPENDED,
560};
561
562/* Data that is per-process-per device. */
563struct kfd_process_device {
564 /*
565 * List of all per-device data for a process.
566 * Starts from kfd_process.per_device_data.
567 */
568 struct list_head per_device_list;
569
570 /* The device that owns this data. */
571 struct kfd_dev *dev;
572
573 /* The process that owns this kfd_process_device. */
574 struct kfd_process *process;
575
576 /* per-process-per device QCM data structure */
577 struct qcm_process_device qpd;
578
579 /*Apertures*/
580 uint64_t lds_base;
581 uint64_t lds_limit;
582 uint64_t gpuvm_base;
583 uint64_t gpuvm_limit;
584 uint64_t scratch_base;
585 uint64_t scratch_limit;
586
587 /* VM context for GPUVM allocations */
588 struct file *drm_file;
589 void *vm;
590
591 /* GPUVM allocations storage */
592 struct idr alloc_idr;
593
594 /* Flag used to tell the pdd has dequeued from the dqm.
595 * This is used to prevent dev->dqm->ops.process_termination() from
596 * being called twice when it is already called in IOMMU callback
597 * function.
598 */
599 bool already_dequeued;
600
601 /* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
602 enum kfd_pdd_bound bound;
603};
604
605#define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
606
607/* Process data */
608struct kfd_process {
609 /*
610 * kfd_process are stored in an mm_struct*->kfd_process*
611 * hash table (kfd_processes in kfd_process.c)
612 */
613 struct hlist_node kfd_processes;
614
615 /*
616 * Opaque pointer to mm_struct. We don't hold a reference to
617 * it so it should never be dereferenced from here. This is
618 * only used for looking up processes by their mm.
619 */
620 void *mm;
621
622 struct kref ref;
623 struct work_struct release_work;
624
625 struct mutex mutex;
626
627 /*
628 * In any process, the thread that started main() is the lead
629 * thread and outlives the rest.
630 * It is here because amd_iommu_bind_pasid wants a task_struct.
631 * It can also be used for safely getting a reference to the
632 * mm_struct of the process.
633 */
634 struct task_struct *lead_thread;
635
636 /* We want to receive a notification when the mm_struct is destroyed */
637 struct mmu_notifier mmu_notifier;
638
639 /* Use for delayed freeing of kfd_process structure */
640 struct rcu_head rcu;
641
642 unsigned int pasid;
643 unsigned int doorbell_index;
644
645 /*
646 * List of kfd_process_device structures,
647 * one for each device the process is using.
648 */
649 struct list_head per_device_data;
650
651 struct process_queue_manager pqm;
652
653 /*Is the user space process 32 bit?*/
654 bool is_32bit_user_mode;
655
656 /* Event-related data */
657 struct mutex event_mutex;
658 /* Event ID allocator and lookup */
659 struct idr event_idr;
660 /* Event page */
661 struct kfd_signal_page *signal_page;
662 size_t signal_mapped_size;
663 size_t signal_event_count;
664 bool signal_event_limit_reached;
665
666 /* Information used for memory eviction */
667 void *kgd_process_info;
668 /* Eviction fence that is attached to all the BOs of this process. The
669 * fence will be triggered during eviction and new one will be created
670 * during restore
671 */
672 struct dma_fence *ef;
673
674 /* Work items for evicting and restoring BOs */
675 struct delayed_work eviction_work;
676 struct delayed_work restore_work;
677 /* seqno of the last scheduled eviction */
678 unsigned int last_eviction_seqno;
679 /* Approx. the last timestamp (in jiffies) when the process was
680 * restored after an eviction
681 */
682 unsigned long last_restore_timestamp;
683};
684
685#define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
686extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
687extern struct srcu_struct kfd_processes_srcu;
688
689/**
690 * Ioctl function type.
691 *
692 * \param filep pointer to file structure.
693 * \param p amdkfd process pointer.
694 * \param data pointer to arg that was copied from user.
695 */
696typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
697 void *data);
698
699struct amdkfd_ioctl_desc {
700 unsigned int cmd;
701 int flags;
702 amdkfd_ioctl_t *func;
703 unsigned int cmd_drv;
704 const char *name;
705};
706bool kfd_dev_is_large_bar(struct kfd_dev *dev);
707
708int kfd_process_create_wq(void);
709void kfd_process_destroy_wq(void);
710struct kfd_process *kfd_create_process(struct file *filep);
711struct kfd_process *kfd_get_process(const struct task_struct *);
712struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
713struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
714void kfd_unref_process(struct kfd_process *p);
715int kfd_process_evict_queues(struct kfd_process *p);
716int kfd_process_restore_queues(struct kfd_process *p);
717void kfd_suspend_all_processes(void);
718int kfd_resume_all_processes(void);
719
720int kfd_process_device_init_vm(struct kfd_process_device *pdd,
721 struct file *drm_file);
722struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
723 struct kfd_process *p);
724struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
725 struct kfd_process *p);
726struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
727 struct kfd_process *p);
728
729int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
730 struct vm_area_struct *vma);
731
732/* KFD process API for creating and translating handles */
733int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
734 void *mem);
735void *kfd_process_device_translate_handle(struct kfd_process_device *p,
736 int handle);
737void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
738 int handle);
739
740/* Process device data iterator */
741struct kfd_process_device *kfd_get_first_process_device_data(
742 struct kfd_process *p);
743struct kfd_process_device *kfd_get_next_process_device_data(
744 struct kfd_process *p,
745 struct kfd_process_device *pdd);
746bool kfd_has_process_device_data(struct kfd_process *p);
747
748/* PASIDs */
749int kfd_pasid_init(void);
750void kfd_pasid_exit(void);
751bool kfd_set_pasid_limit(unsigned int new_limit);
752unsigned int kfd_get_pasid_limit(void);
753unsigned int kfd_pasid_alloc(void);
754void kfd_pasid_free(unsigned int pasid);
755
756/* Doorbells */
757size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
758int kfd_doorbell_init(struct kfd_dev *kfd);
759void kfd_doorbell_fini(struct kfd_dev *kfd);
760int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
761 struct vm_area_struct *vma);
762void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
763 unsigned int *doorbell_off);
764void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
765u32 read_kernel_doorbell(u32 __iomem *db);
766void write_kernel_doorbell(void __iomem *db, u32 value);
767void write_kernel_doorbell64(void __iomem *db, u64 value);
768unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
769 struct kfd_process *process,
770 unsigned int doorbell_id);
771phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
772 struct kfd_process *process);
773int kfd_alloc_process_doorbells(struct kfd_process *process);
774void kfd_free_process_doorbells(struct kfd_process *process);
775
776/* GTT Sub-Allocator */
777
778int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
779 struct kfd_mem_obj **mem_obj);
780
781int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
782
783extern struct device *kfd_device;
784
785/* Topology */
786int kfd_topology_init(void);
787void kfd_topology_shutdown(void);
788int kfd_topology_add_device(struct kfd_dev *gpu);
789int kfd_topology_remove_device(struct kfd_dev *gpu);
790struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
791 uint32_t proximity_domain);
792struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
793struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
794struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
795struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd);
796int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
797int kfd_numa_node_to_apic_id(int numa_node_id);
798
799/* Interrupts */
800int kfd_interrupt_init(struct kfd_dev *dev);
801void kfd_interrupt_exit(struct kfd_dev *dev);
802bool enqueue_ih_ring_entry(struct kfd_dev *kfd, const void *ih_ring_entry);
803bool interrupt_is_wanted(struct kfd_dev *dev,
804 const uint32_t *ih_ring_entry,
805 uint32_t *patched_ihre, bool *flag);
806
807/* amdkfd Apertures */
808int kfd_init_apertures(struct kfd_process *process);
809
810/* Queue Context Management */
811int init_queue(struct queue **q, const struct queue_properties *properties);
812void uninit_queue(struct queue *q);
813void print_queue_properties(struct queue_properties *q);
814void print_queue(struct queue *q);
815
816struct mqd_manager *mqd_manager_init(enum KFD_MQD_TYPE type,
817 struct kfd_dev *dev);
818struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
819 struct kfd_dev *dev);
820struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
821 struct kfd_dev *dev);
822struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
823 struct kfd_dev *dev);
824struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
825 struct kfd_dev *dev);
826struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
827 struct kfd_dev *dev);
828struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
829void device_queue_manager_uninit(struct device_queue_manager *dqm);
830struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
831 enum kfd_queue_type type);
832void kernel_queue_uninit(struct kernel_queue *kq);
833int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid);
834
835/* Process Queue Manager */
836struct process_queue_node {
837 struct queue *q;
838 struct kernel_queue *kq;
839 struct list_head process_queue_list;
840};
841
842void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
843void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
844int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
845void pqm_uninit(struct process_queue_manager *pqm);
846int pqm_create_queue(struct process_queue_manager *pqm,
847 struct kfd_dev *dev,
848 struct file *f,
849 struct queue_properties *properties,
850 unsigned int *qid);
851int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
852int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
853 struct queue_properties *p);
854int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
855 struct queue_properties *p);
856struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
857 unsigned int qid);
858int pqm_get_wave_state(struct process_queue_manager *pqm,
859 unsigned int qid,
860 void __user *ctl_stack,
861 u32 *ctl_stack_used_size,
862 u32 *save_area_used_size);
863
864int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
865 unsigned int fence_value,
866 unsigned int timeout_ms);
867
868/* Packet Manager */
869
870#define KFD_FENCE_COMPLETED (100)
871#define KFD_FENCE_INIT (10)
872
873struct packet_manager {
874 struct device_queue_manager *dqm;
875 struct kernel_queue *priv_queue;
876 struct mutex lock;
877 bool allocated;
878 struct kfd_mem_obj *ib_buffer_obj;
879 unsigned int ib_size_bytes;
880
881 const struct packet_manager_funcs *pmf;
882};
883
884struct packet_manager_funcs {
885 /* Support ASIC-specific packet formats for PM4 packets */
886 int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
887 struct qcm_process_device *qpd);
888 int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
889 uint64_t ib, size_t ib_size_in_dwords, bool chain);
890 int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
891 struct scheduling_resources *res);
892 int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
893 struct queue *q, bool is_static);
894 int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
895 enum kfd_queue_type type,
896 enum kfd_unmap_queues_filter mode,
897 uint32_t filter_param, bool reset,
898 unsigned int sdma_engine);
899 int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
900 uint64_t fence_address, uint32_t fence_value);
901 int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
902
903 /* Packet sizes */
904 int map_process_size;
905 int runlist_size;
906 int set_resources_size;
907 int map_queues_size;
908 int unmap_queues_size;
909 int query_status_size;
910 int release_mem_size;
911};
912
913extern const struct packet_manager_funcs kfd_vi_pm_funcs;
914extern const struct packet_manager_funcs kfd_v9_pm_funcs;
915
916int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
917void pm_uninit(struct packet_manager *pm);
918int pm_send_set_resources(struct packet_manager *pm,
919 struct scheduling_resources *res);
920int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
921int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
922 uint32_t fence_value);
923
924int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
925 enum kfd_unmap_queues_filter mode,
926 uint32_t filter_param, bool reset,
927 unsigned int sdma_engine);
928
929void pm_release_ib(struct packet_manager *pm);
930
931/* Following PM funcs can be shared among VI and AI */
932unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
933int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer,
934 struct scheduling_resources *res);
935
936uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
937
938/* Events */
939extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
940extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
941
942extern const struct kfd_device_global_init_class device_global_init_class_cik;
943
944void kfd_event_init_process(struct kfd_process *p);
945void kfd_event_free_process(struct kfd_process *p);
946int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
947int kfd_wait_on_events(struct kfd_process *p,
948 uint32_t num_events, void __user *data,
949 bool all, uint32_t user_timeout_ms,
950 uint32_t *wait_result);
951void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
952 uint32_t valid_id_bits);
953void kfd_signal_iommu_event(struct kfd_dev *dev,
954 unsigned int pasid, unsigned long address,
955 bool is_write_requested, bool is_execute_requested);
956void kfd_signal_hw_exception_event(unsigned int pasid);
957int kfd_set_event(struct kfd_process *p, uint32_t event_id);
958int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
959int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
960 uint64_t size);
961int kfd_event_create(struct file *devkfd, struct kfd_process *p,
962 uint32_t event_type, bool auto_reset, uint32_t node_id,
963 uint32_t *event_id, uint32_t *event_trigger_data,
964 uint64_t *event_page_offset, uint32_t *event_slot_index);
965int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
966
967void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid,
968 struct kfd_vm_fault_info *info);
969
970void kfd_signal_reset_event(struct kfd_dev *dev);
971
972void kfd_flush_tlb(struct kfd_process_device *pdd);
973
974int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
975
976bool kfd_is_locked(void);
977
978/* Debugfs */
979#if defined(CONFIG_DEBUG_FS)
980
981void kfd_debugfs_init(void);
982void kfd_debugfs_fini(void);
983int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
984int pqm_debugfs_mqds(struct seq_file *m, void *data);
985int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
986int dqm_debugfs_hqds(struct seq_file *m, void *data);
987int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
988int pm_debugfs_runlist(struct seq_file *m, void *data);
989
990int kfd_debugfs_hang_hws(struct kfd_dev *dev);
991int pm_debugfs_hang_hws(struct packet_manager *pm);
992int dqm_debugfs_execute_queues(struct device_queue_manager *dqm);
993
994#else
995
996static inline void kfd_debugfs_init(void) {}
997static inline void kfd_debugfs_fini(void) {}
998
999#endif
1000
1001#endif
1002