xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_priv.h (revision b830f94f)
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 <linux/sysfs.h>
39 #include <kgd_kfd_interface.h>
40 
41 #include "amd_shared.h"
42 
43 #define KFD_MAX_RING_ENTRY_SIZE	8
44 
45 #define KFD_SYSFS_FILE_MODE 0444
46 
47 /* GPU ID hash width in bits */
48 #define KFD_GPU_ID_HASH_WIDTH 16
49 
50 /* Use upper bits of mmap offset to store KFD driver specific information.
51  * BITS[63:62] - Encode MMAP type
52  * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to
53  * BITS[45:0]  - MMAP offset value
54  *
55  * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these
56  *  defines are w.r.t to PAGE_SIZE
57  */
58 #define KFD_MMAP_TYPE_SHIFT	(62 - PAGE_SHIFT)
59 #define KFD_MMAP_TYPE_MASK	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
60 #define KFD_MMAP_TYPE_DOORBELL	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
61 #define KFD_MMAP_TYPE_EVENTS	(0x2ULL << KFD_MMAP_TYPE_SHIFT)
62 #define KFD_MMAP_TYPE_RESERVED_MEM	(0x1ULL << KFD_MMAP_TYPE_SHIFT)
63 #define KFD_MMAP_TYPE_MMIO	(0x0ULL << KFD_MMAP_TYPE_SHIFT)
64 
65 #define KFD_MMAP_GPU_ID_SHIFT (46 - PAGE_SHIFT)
66 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \
67 				<< KFD_MMAP_GPU_ID_SHIFT)
68 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\
69 				& KFD_MMAP_GPU_ID_MASK)
70 #define KFD_MMAP_GPU_ID_GET(offset)    ((offset & KFD_MMAP_GPU_ID_MASK) \
71 				>> KFD_MMAP_GPU_ID_SHIFT)
72 
73 #define KFD_MMAP_OFFSET_VALUE_MASK	(0x3FFFFFFFFFFFULL >> PAGE_SHIFT)
74 #define KFD_MMAP_OFFSET_VALUE_GET(offset) (offset & KFD_MMAP_OFFSET_VALUE_MASK)
75 
76 /*
77  * When working with cp scheduler we should assign the HIQ manually or via
78  * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot
79  * definitions for Kaveri. In Kaveri only the first ME queues participates
80  * in the cp scheduling taking that in mind we set the HIQ slot in the
81  * second ME.
82  */
83 #define KFD_CIK_HIQ_PIPE 4
84 #define KFD_CIK_HIQ_QUEUE 0
85 
86 /* Macro for allocating structures */
87 #define kfd_alloc_struct(ptr_to_struct)	\
88 	((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
89 
90 #define KFD_MAX_NUM_OF_PROCESSES 512
91 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
92 
93 /*
94  * Size of the per-process TBA+TMA buffer: 2 pages
95  *
96  * The first page is the TBA used for the CWSR ISA code. The second
97  * page is used as TMA for daisy changing a user-mode trap handler.
98  */
99 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
100 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE
101 
102 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE		\
103 	(KFD_MAX_NUM_OF_PROCESSES *			\
104 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
105 
106 #define KFD_KERNEL_QUEUE_SIZE 2048
107 
108 #define KFD_UNMAP_LATENCY_MS	(4000)
109 
110 /*
111  * 512 = 0x200
112  * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
113  * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.
114  * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC
115  * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in
116  * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.
117  */
118 #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
119 
120 
121 /*
122  * Kernel module parameter to specify maximum number of supported queues per
123  * device
124  */
125 extern int max_num_of_queues_per_device;
126 
127 
128 /* Kernel module parameter to specify the scheduling policy */
129 extern int sched_policy;
130 
131 /*
132  * Kernel module parameter to specify the maximum process
133  * number per HW scheduler
134  */
135 extern int hws_max_conc_proc;
136 
137 extern int cwsr_enable;
138 
139 /*
140  * Kernel module parameter to specify whether to send sigterm to HSA process on
141  * unhandled exception
142  */
143 extern int send_sigterm;
144 
145 /*
146  * This kernel module is used to simulate large bar machine on non-large bar
147  * enabled machines.
148  */
149 extern int debug_largebar;
150 
151 /*
152  * Ignore CRAT table during KFD initialization, can be used to work around
153  * broken CRAT tables on some AMD systems
154  */
155 extern int ignore_crat;
156 
157 /*
158  * Set sh_mem_config.retry_disable on Vega10
159  */
160 extern int amdgpu_noretry;
161 
162 /*
163  * Halt if HWS hang is detected
164  */
165 extern int halt_if_hws_hang;
166 
167 /*
168  * Whether MEC FW support GWS barriers
169  */
170 extern bool hws_gws_support;
171 
172 /*
173  * Queue preemption timeout in ms
174  */
175 extern int queue_preemption_timeout_ms;
176 
177 enum cache_policy {
178 	cache_policy_coherent,
179 	cache_policy_noncoherent
180 };
181 
182 #define KFD_IS_VI(chip) ((chip) >= CHIP_CARRIZO && (chip) <= CHIP_POLARIS11)
183 #define KFD_IS_DGPU(chip) (((chip) >= CHIP_TONGA && \
184 			   (chip) <= CHIP_NAVI10) || \
185 			   (chip) == CHIP_HAWAII)
186 #define KFD_IS_SOC15(chip) ((chip) >= CHIP_VEGA10)
187 
188 struct kfd_event_interrupt_class {
189 	bool (*interrupt_isr)(struct kfd_dev *dev,
190 			const uint32_t *ih_ring_entry, uint32_t *patched_ihre,
191 			bool *patched_flag);
192 	void (*interrupt_wq)(struct kfd_dev *dev,
193 			const uint32_t *ih_ring_entry);
194 };
195 
196 struct kfd_device_info {
197 	enum amd_asic_type asic_family;
198 	const struct kfd_event_interrupt_class *event_interrupt_class;
199 	unsigned int max_pasid_bits;
200 	unsigned int max_no_of_hqd;
201 	unsigned int doorbell_size;
202 	size_t ih_ring_entry_size;
203 	uint8_t num_of_watch_points;
204 	uint16_t mqd_size_aligned;
205 	bool supports_cwsr;
206 	bool needs_iommu_device;
207 	bool needs_pci_atomics;
208 	unsigned int num_sdma_engines;
209 	unsigned int num_xgmi_sdma_engines;
210 	unsigned int num_sdma_queues_per_engine;
211 };
212 
213 struct kfd_mem_obj {
214 	uint32_t range_start;
215 	uint32_t range_end;
216 	uint64_t gpu_addr;
217 	uint32_t *cpu_ptr;
218 	void *gtt_mem;
219 };
220 
221 struct kfd_vmid_info {
222 	uint32_t first_vmid_kfd;
223 	uint32_t last_vmid_kfd;
224 	uint32_t vmid_num_kfd;
225 };
226 
227 struct kfd_dev {
228 	struct kgd_dev *kgd;
229 
230 	const struct kfd_device_info *device_info;
231 	struct pci_dev *pdev;
232 
233 	unsigned int id;		/* topology stub index */
234 
235 	phys_addr_t doorbell_base;	/* Start of actual doorbells used by
236 					 * KFD. It is aligned for mapping
237 					 * into user mode
238 					 */
239 	size_t doorbell_id_offset;	/* Doorbell offset (from KFD doorbell
240 					 * to HW doorbell, GFX reserved some
241 					 * at the start)
242 					 */
243 	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
244 					   * page used by kernel queue
245 					   */
246 
247 	struct kgd2kfd_shared_resources shared_resources;
248 	struct kfd_vmid_info vm_info;
249 
250 	const struct kfd2kgd_calls *kfd2kgd;
251 	struct mutex doorbell_mutex;
252 	DECLARE_BITMAP(doorbell_available_index,
253 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
254 
255 	void *gtt_mem;
256 	uint64_t gtt_start_gpu_addr;
257 	void *gtt_start_cpu_ptr;
258 	void *gtt_sa_bitmap;
259 	struct mutex gtt_sa_lock;
260 	unsigned int gtt_sa_chunk_size;
261 	unsigned int gtt_sa_num_of_chunks;
262 
263 	/* Interrupts */
264 	struct kfifo ih_fifo;
265 	struct workqueue_struct *ih_wq;
266 	struct work_struct interrupt_work;
267 	spinlock_t interrupt_lock;
268 
269 	/* QCM Device instance */
270 	struct device_queue_manager *dqm;
271 
272 	bool init_complete;
273 	/*
274 	 * Interrupts of interest to KFD are copied
275 	 * from the HW ring into a SW ring.
276 	 */
277 	bool interrupts_active;
278 
279 	/* Debug manager */
280 	struct kfd_dbgmgr *dbgmgr;
281 
282 	/* Firmware versions */
283 	uint16_t mec_fw_version;
284 	uint16_t sdma_fw_version;
285 
286 	/* Maximum process number mapped to HW scheduler */
287 	unsigned int max_proc_per_quantum;
288 
289 	/* CWSR */
290 	bool cwsr_enabled;
291 	const void *cwsr_isa;
292 	unsigned int cwsr_isa_size;
293 
294 	/* xGMI */
295 	uint64_t hive_id;
296 
297 	bool pci_atomic_requested;
298 
299 	/* SRAM ECC flag */
300 	atomic_t sram_ecc_flag;
301 
302 	/* Compute Profile ref. count */
303 	atomic_t compute_profile;
304 
305 	/* Global GWS resource shared b/t processes*/
306 	void *gws;
307 };
308 
309 enum kfd_mempool {
310 	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
311 	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
312 	KFD_MEMPOOL_FRAMEBUFFER = 3,
313 };
314 
315 /* Character device interface */
316 int kfd_chardev_init(void);
317 void kfd_chardev_exit(void);
318 struct device *kfd_chardev(void);
319 
320 /**
321  * enum kfd_unmap_queues_filter
322  *
323  * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
324  *
325  * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
326  *						running queues list.
327  *
328  * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
329  *						specific process.
330  *
331  */
332 enum kfd_unmap_queues_filter {
333 	KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
334 	KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
335 	KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
336 	KFD_UNMAP_QUEUES_FILTER_BY_PASID
337 };
338 
339 /**
340  * enum kfd_queue_type
341  *
342  * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
343  *
344  * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
345  *
346  * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
347  *
348  * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
349  */
350 enum kfd_queue_type  {
351 	KFD_QUEUE_TYPE_COMPUTE,
352 	KFD_QUEUE_TYPE_SDMA,
353 	KFD_QUEUE_TYPE_HIQ,
354 	KFD_QUEUE_TYPE_DIQ,
355 	KFD_QUEUE_TYPE_SDMA_XGMI
356 };
357 
358 enum kfd_queue_format {
359 	KFD_QUEUE_FORMAT_PM4,
360 	KFD_QUEUE_FORMAT_AQL
361 };
362 
363 enum KFD_QUEUE_PRIORITY {
364 	KFD_QUEUE_PRIORITY_MINIMUM = 0,
365 	KFD_QUEUE_PRIORITY_MAXIMUM = 15
366 };
367 
368 /**
369  * struct queue_properties
370  *
371  * @type: The queue type.
372  *
373  * @queue_id: Queue identifier.
374  *
375  * @queue_address: Queue ring buffer address.
376  *
377  * @queue_size: Queue ring buffer size.
378  *
379  * @priority: Defines the queue priority relative to other queues in the
380  * process.
381  * This is just an indication and HW scheduling may override the priority as
382  * necessary while keeping the relative prioritization.
383  * the priority granularity is from 0 to f which f is the highest priority.
384  * currently all queues are initialized with the highest priority.
385  *
386  * @queue_percent: This field is partially implemented and currently a zero in
387  * this field defines that the queue is non active.
388  *
389  * @read_ptr: User space address which points to the number of dwords the
390  * cp read from the ring buffer. This field updates automatically by the H/W.
391  *
392  * @write_ptr: Defines the number of dwords written to the ring buffer.
393  *
394  * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
395  * the queue ring buffer. This field should be similar to write_ptr and the
396  * user should update this field after he updated the write_ptr.
397  *
398  * @doorbell_off: The doorbell offset in the doorbell pci-bar.
399  *
400  * @is_interop: Defines if this is a interop queue. Interop queue means that
401  * the queue can access both graphics and compute resources.
402  *
403  * @is_evicted: Defines if the queue is evicted. Only active queues
404  * are evicted, rendering them inactive.
405  *
406  * @is_active: Defines if the queue is active or not. @is_active and
407  * @is_evicted are protected by the DQM lock.
408  *
409  * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
410  * of the queue.
411  *
412  * This structure represents the queue properties for each queue no matter if
413  * it's user mode or kernel mode queue.
414  *
415  */
416 struct queue_properties {
417 	enum kfd_queue_type type;
418 	enum kfd_queue_format format;
419 	unsigned int queue_id;
420 	uint64_t queue_address;
421 	uint64_t  queue_size;
422 	uint32_t priority;
423 	uint32_t queue_percent;
424 	uint32_t *read_ptr;
425 	uint32_t *write_ptr;
426 	void __iomem *doorbell_ptr;
427 	uint32_t doorbell_off;
428 	bool is_interop;
429 	bool is_evicted;
430 	bool is_active;
431 	/* Not relevant for user mode queues in cp scheduling */
432 	unsigned int vmid;
433 	/* Relevant only for sdma queues*/
434 	uint32_t sdma_engine_id;
435 	uint32_t sdma_queue_id;
436 	uint32_t sdma_vm_addr;
437 	/* Relevant only for VI */
438 	uint64_t eop_ring_buffer_address;
439 	uint32_t eop_ring_buffer_size;
440 	uint64_t ctx_save_restore_area_address;
441 	uint32_t ctx_save_restore_area_size;
442 	uint32_t ctl_stack_size;
443 	uint64_t tba_addr;
444 	uint64_t tma_addr;
445 	/* Relevant for CU */
446 	uint32_t cu_mask_count; /* Must be a multiple of 32 */
447 	uint32_t *cu_mask;
448 };
449 
450 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 &&	\
451 			    (q).queue_address != 0 &&	\
452 			    (q).queue_percent > 0 &&	\
453 			    !(q).is_evicted)
454 
455 /**
456  * struct queue
457  *
458  * @list: Queue linked list.
459  *
460  * @mqd: The queue MQD.
461  *
462  * @mqd_mem_obj: The MQD local gpu memory object.
463  *
464  * @gart_mqd_addr: The MQD gart mc address.
465  *
466  * @properties: The queue properties.
467  *
468  * @mec: Used only in no cp scheduling mode and identifies to micro engine id
469  *	 that the queue should be execute on.
470  *
471  * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
472  *	  id.
473  *
474  * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
475  *
476  * @process: The kfd process that created this queue.
477  *
478  * @device: The kfd device that created this queue.
479  *
480  * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
481  * otherwise.
482  *
483  * This structure represents user mode compute queues.
484  * It contains all the necessary data to handle such queues.
485  *
486  */
487 
488 struct queue {
489 	struct list_head list;
490 	void *mqd;
491 	struct kfd_mem_obj *mqd_mem_obj;
492 	uint64_t gart_mqd_addr;
493 	struct queue_properties properties;
494 
495 	uint32_t mec;
496 	uint32_t pipe;
497 	uint32_t queue;
498 
499 	unsigned int sdma_id;
500 	unsigned int doorbell_id;
501 
502 	struct kfd_process	*process;
503 	struct kfd_dev		*device;
504 	void *gws;
505 };
506 
507 /*
508  * Please read the kfd_mqd_manager.h description.
509  */
510 enum KFD_MQD_TYPE {
511 	KFD_MQD_TYPE_COMPUTE = 0,	/* for no cp scheduling */
512 	KFD_MQD_TYPE_HIQ,		/* for hiq */
513 	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
514 	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
515 	KFD_MQD_TYPE_DIQ,		/* for diq */
516 	KFD_MQD_TYPE_MAX
517 };
518 
519 enum KFD_PIPE_PRIORITY {
520 	KFD_PIPE_PRIORITY_CS_LOW = 0,
521 	KFD_PIPE_PRIORITY_CS_MEDIUM,
522 	KFD_PIPE_PRIORITY_CS_HIGH
523 };
524 
525 struct scheduling_resources {
526 	unsigned int vmid_mask;
527 	enum kfd_queue_type type;
528 	uint64_t queue_mask;
529 	uint64_t gws_mask;
530 	uint32_t oac_mask;
531 	uint32_t gds_heap_base;
532 	uint32_t gds_heap_size;
533 };
534 
535 struct process_queue_manager {
536 	/* data */
537 	struct kfd_process	*process;
538 	struct list_head	queues;
539 	unsigned long		*queue_slot_bitmap;
540 };
541 
542 struct qcm_process_device {
543 	/* The Device Queue Manager that owns this data */
544 	struct device_queue_manager *dqm;
545 	struct process_queue_manager *pqm;
546 	/* Queues list */
547 	struct list_head queues_list;
548 	struct list_head priv_queue_list;
549 
550 	unsigned int queue_count;
551 	unsigned int vmid;
552 	bool is_debug;
553 	unsigned int evicted; /* eviction counter, 0=active */
554 
555 	/* This flag tells if we should reset all wavefronts on
556 	 * process termination
557 	 */
558 	bool reset_wavefronts;
559 
560 	/*
561 	 * All the memory management data should be here too
562 	 */
563 	uint64_t gds_context_area;
564 	/* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
565 	uint64_t page_table_base;
566 	uint32_t sh_mem_config;
567 	uint32_t sh_mem_bases;
568 	uint32_t sh_mem_ape1_base;
569 	uint32_t sh_mem_ape1_limit;
570 	uint32_t gds_size;
571 	uint32_t num_gws;
572 	uint32_t num_oac;
573 	uint32_t sh_hidden_private_base;
574 
575 	/* CWSR memory */
576 	void *cwsr_kaddr;
577 	uint64_t cwsr_base;
578 	uint64_t tba_addr;
579 	uint64_t tma_addr;
580 
581 	/* IB memory */
582 	uint64_t ib_base;
583 	void *ib_kaddr;
584 
585 	/* doorbell resources per process per device */
586 	unsigned long *doorbell_bitmap;
587 };
588 
589 /* KFD Memory Eviction */
590 
591 /* Approx. wait time before attempting to restore evicted BOs */
592 #define PROCESS_RESTORE_TIME_MS 100
593 /* Approx. back off time if restore fails due to lack of memory */
594 #define PROCESS_BACK_OFF_TIME_MS 100
595 /* Approx. time before evicting the process again */
596 #define PROCESS_ACTIVE_TIME_MS 10
597 
598 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
599  * idr_handle in the least significant 4 bytes
600  */
601 #define MAKE_HANDLE(gpu_id, idr_handle) \
602 	(((uint64_t)(gpu_id) << 32) + idr_handle)
603 #define GET_GPU_ID(handle) (handle >> 32)
604 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
605 
606 enum kfd_pdd_bound {
607 	PDD_UNBOUND = 0,
608 	PDD_BOUND,
609 	PDD_BOUND_SUSPENDED,
610 };
611 
612 /* Data that is per-process-per device. */
613 struct kfd_process_device {
614 	/*
615 	 * List of all per-device data for a process.
616 	 * Starts from kfd_process.per_device_data.
617 	 */
618 	struct list_head per_device_list;
619 
620 	/* The device that owns this data. */
621 	struct kfd_dev *dev;
622 
623 	/* The process that owns this kfd_process_device. */
624 	struct kfd_process *process;
625 
626 	/* per-process-per device QCM data structure */
627 	struct qcm_process_device qpd;
628 
629 	/*Apertures*/
630 	uint64_t lds_base;
631 	uint64_t lds_limit;
632 	uint64_t gpuvm_base;
633 	uint64_t gpuvm_limit;
634 	uint64_t scratch_base;
635 	uint64_t scratch_limit;
636 
637 	/* VM context for GPUVM allocations */
638 	struct file *drm_file;
639 	void *vm;
640 
641 	/* GPUVM allocations storage */
642 	struct idr alloc_idr;
643 
644 	/* Flag used to tell the pdd has dequeued from the dqm.
645 	 * This is used to prevent dev->dqm->ops.process_termination() from
646 	 * being called twice when it is already called in IOMMU callback
647 	 * function.
648 	 */
649 	bool already_dequeued;
650 
651 	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
652 	enum kfd_pdd_bound bound;
653 };
654 
655 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
656 
657 /* Process data */
658 struct kfd_process {
659 	/*
660 	 * kfd_process are stored in an mm_struct*->kfd_process*
661 	 * hash table (kfd_processes in kfd_process.c)
662 	 */
663 	struct hlist_node kfd_processes;
664 
665 	/*
666 	 * Opaque pointer to mm_struct. We don't hold a reference to
667 	 * it so it should never be dereferenced from here. This is
668 	 * only used for looking up processes by their mm.
669 	 */
670 	void *mm;
671 
672 	struct kref ref;
673 	struct work_struct release_work;
674 
675 	struct mutex mutex;
676 
677 	/*
678 	 * In any process, the thread that started main() is the lead
679 	 * thread and outlives the rest.
680 	 * It is here because amd_iommu_bind_pasid wants a task_struct.
681 	 * It can also be used for safely getting a reference to the
682 	 * mm_struct of the process.
683 	 */
684 	struct task_struct *lead_thread;
685 
686 	/* We want to receive a notification when the mm_struct is destroyed */
687 	struct mmu_notifier mmu_notifier;
688 
689 	/* Use for delayed freeing of kfd_process structure */
690 	struct rcu_head	rcu;
691 
692 	unsigned int pasid;
693 	unsigned int doorbell_index;
694 
695 	/*
696 	 * List of kfd_process_device structures,
697 	 * one for each device the process is using.
698 	 */
699 	struct list_head per_device_data;
700 
701 	struct process_queue_manager pqm;
702 
703 	/*Is the user space process 32 bit?*/
704 	bool is_32bit_user_mode;
705 
706 	/* Event-related data */
707 	struct mutex event_mutex;
708 	/* Event ID allocator and lookup */
709 	struct idr event_idr;
710 	/* Event page */
711 	struct kfd_signal_page *signal_page;
712 	size_t signal_mapped_size;
713 	size_t signal_event_count;
714 	bool signal_event_limit_reached;
715 
716 	/* Information used for memory eviction */
717 	void *kgd_process_info;
718 	/* Eviction fence that is attached to all the BOs of this process. The
719 	 * fence will be triggered during eviction and new one will be created
720 	 * during restore
721 	 */
722 	struct dma_fence *ef;
723 
724 	/* Work items for evicting and restoring BOs */
725 	struct delayed_work eviction_work;
726 	struct delayed_work restore_work;
727 	/* seqno of the last scheduled eviction */
728 	unsigned int last_eviction_seqno;
729 	/* Approx. the last timestamp (in jiffies) when the process was
730 	 * restored after an eviction
731 	 */
732 	unsigned long last_restore_timestamp;
733 
734 	/* Kobj for our procfs */
735 	struct kobject *kobj;
736 	struct attribute attr_pasid;
737 };
738 
739 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
740 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
741 extern struct srcu_struct kfd_processes_srcu;
742 
743 /**
744  * Ioctl function type.
745  *
746  * \param filep pointer to file structure.
747  * \param p amdkfd process pointer.
748  * \param data pointer to arg that was copied from user.
749  */
750 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
751 				void *data);
752 
753 struct amdkfd_ioctl_desc {
754 	unsigned int cmd;
755 	int flags;
756 	amdkfd_ioctl_t *func;
757 	unsigned int cmd_drv;
758 	const char *name;
759 };
760 bool kfd_dev_is_large_bar(struct kfd_dev *dev);
761 
762 int kfd_process_create_wq(void);
763 void kfd_process_destroy_wq(void);
764 struct kfd_process *kfd_create_process(struct file *filep);
765 struct kfd_process *kfd_get_process(const struct task_struct *);
766 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
767 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
768 void kfd_unref_process(struct kfd_process *p);
769 int kfd_process_evict_queues(struct kfd_process *p);
770 int kfd_process_restore_queues(struct kfd_process *p);
771 void kfd_suspend_all_processes(void);
772 int kfd_resume_all_processes(void);
773 
774 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
775 			       struct file *drm_file);
776 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
777 						struct kfd_process *p);
778 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
779 							struct kfd_process *p);
780 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
781 							struct kfd_process *p);
782 
783 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
784 			  struct vm_area_struct *vma);
785 
786 /* KFD process API for creating and translating handles */
787 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
788 					void *mem);
789 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
790 					int handle);
791 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
792 					int handle);
793 
794 /* Process device data iterator */
795 struct kfd_process_device *kfd_get_first_process_device_data(
796 							struct kfd_process *p);
797 struct kfd_process_device *kfd_get_next_process_device_data(
798 						struct kfd_process *p,
799 						struct kfd_process_device *pdd);
800 bool kfd_has_process_device_data(struct kfd_process *p);
801 
802 /* PASIDs */
803 int kfd_pasid_init(void);
804 void kfd_pasid_exit(void);
805 bool kfd_set_pasid_limit(unsigned int new_limit);
806 unsigned int kfd_get_pasid_limit(void);
807 unsigned int kfd_pasid_alloc(void);
808 void kfd_pasid_free(unsigned int pasid);
809 
810 /* Doorbells */
811 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
812 int kfd_doorbell_init(struct kfd_dev *kfd);
813 void kfd_doorbell_fini(struct kfd_dev *kfd);
814 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
815 		      struct vm_area_struct *vma);
816 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
817 					unsigned int *doorbell_off);
818 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
819 u32 read_kernel_doorbell(u32 __iomem *db);
820 void write_kernel_doorbell(void __iomem *db, u32 value);
821 void write_kernel_doorbell64(void __iomem *db, u64 value);
822 unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
823 					struct kfd_process *process,
824 					unsigned int doorbell_id);
825 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
826 					struct kfd_process *process);
827 int kfd_alloc_process_doorbells(struct kfd_process *process);
828 void kfd_free_process_doorbells(struct kfd_process *process);
829 
830 /* GTT Sub-Allocator */
831 
832 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
833 			struct kfd_mem_obj **mem_obj);
834 
835 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
836 
837 extern struct device *kfd_device;
838 
839 /* KFD's procfs */
840 void kfd_procfs_init(void);
841 void kfd_procfs_shutdown(void);
842 
843 /* Topology */
844 int kfd_topology_init(void);
845 void kfd_topology_shutdown(void);
846 int kfd_topology_add_device(struct kfd_dev *gpu);
847 int kfd_topology_remove_device(struct kfd_dev *gpu);
848 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
849 						uint32_t proximity_domain);
850 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
851 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
852 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
853 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd);
854 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
855 int kfd_numa_node_to_apic_id(int numa_node_id);
856 
857 /* Interrupts */
858 int kfd_interrupt_init(struct kfd_dev *dev);
859 void kfd_interrupt_exit(struct kfd_dev *dev);
860 bool enqueue_ih_ring_entry(struct kfd_dev *kfd,	const void *ih_ring_entry);
861 bool interrupt_is_wanted(struct kfd_dev *dev,
862 				const uint32_t *ih_ring_entry,
863 				uint32_t *patched_ihre, bool *flag);
864 
865 /* amdkfd Apertures */
866 int kfd_init_apertures(struct kfd_process *process);
867 
868 /* Queue Context Management */
869 int init_queue(struct queue **q, const struct queue_properties *properties);
870 void uninit_queue(struct queue *q);
871 void print_queue_properties(struct queue_properties *q);
872 void print_queue(struct queue *q);
873 
874 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
875 		struct kfd_dev *dev);
876 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
877 		struct kfd_dev *dev);
878 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
879 		struct kfd_dev *dev);
880 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
881 		struct kfd_dev *dev);
882 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
883 		struct kfd_dev *dev);
884 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
885 		struct kfd_dev *dev);
886 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
887 void device_queue_manager_uninit(struct device_queue_manager *dqm);
888 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
889 					enum kfd_queue_type type);
890 void kernel_queue_uninit(struct kernel_queue *kq);
891 int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid);
892 
893 /* Process Queue Manager */
894 struct process_queue_node {
895 	struct queue *q;
896 	struct kernel_queue *kq;
897 	struct list_head process_queue_list;
898 };
899 
900 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
901 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
902 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
903 void pqm_uninit(struct process_queue_manager *pqm);
904 int pqm_create_queue(struct process_queue_manager *pqm,
905 			    struct kfd_dev *dev,
906 			    struct file *f,
907 			    struct queue_properties *properties,
908 			    unsigned int *qid);
909 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
910 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
911 			struct queue_properties *p);
912 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
913 			struct queue_properties *p);
914 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
915 			void *gws);
916 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
917 						unsigned int qid);
918 int pqm_get_wave_state(struct process_queue_manager *pqm,
919 		       unsigned int qid,
920 		       void __user *ctl_stack,
921 		       u32 *ctl_stack_used_size,
922 		       u32 *save_area_used_size);
923 
924 int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
925 			      unsigned int fence_value,
926 			      unsigned int timeout_ms);
927 
928 /* Packet Manager */
929 
930 #define KFD_FENCE_COMPLETED (100)
931 #define KFD_FENCE_INIT   (10)
932 
933 struct packet_manager {
934 	struct device_queue_manager *dqm;
935 	struct kernel_queue *priv_queue;
936 	struct mutex lock;
937 	bool allocated;
938 	struct kfd_mem_obj *ib_buffer_obj;
939 	unsigned int ib_size_bytes;
940 	bool is_over_subscription;
941 
942 	const struct packet_manager_funcs *pmf;
943 };
944 
945 struct packet_manager_funcs {
946 	/* Support ASIC-specific packet formats for PM4 packets */
947 	int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
948 			struct qcm_process_device *qpd);
949 	int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
950 			uint64_t ib, size_t ib_size_in_dwords, bool chain);
951 	int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
952 			struct scheduling_resources *res);
953 	int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
954 			struct queue *q, bool is_static);
955 	int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
956 			enum kfd_queue_type type,
957 			enum kfd_unmap_queues_filter mode,
958 			uint32_t filter_param, bool reset,
959 			unsigned int sdma_engine);
960 	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
961 			uint64_t fence_address,	uint32_t fence_value);
962 	int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
963 
964 	/* Packet sizes */
965 	int map_process_size;
966 	int runlist_size;
967 	int set_resources_size;
968 	int map_queues_size;
969 	int unmap_queues_size;
970 	int query_status_size;
971 	int release_mem_size;
972 };
973 
974 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
975 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
976 extern const struct packet_manager_funcs kfd_v10_pm_funcs;
977 
978 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
979 void pm_uninit(struct packet_manager *pm);
980 int pm_send_set_resources(struct packet_manager *pm,
981 				struct scheduling_resources *res);
982 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
983 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
984 				uint32_t fence_value);
985 
986 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
987 			enum kfd_unmap_queues_filter mode,
988 			uint32_t filter_param, bool reset,
989 			unsigned int sdma_engine);
990 
991 void pm_release_ib(struct packet_manager *pm);
992 
993 /* Following PM funcs can be shared among VI and AI */
994 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
995 int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer,
996 			struct scheduling_resources *res);
997 
998 
999 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
1000 
1001 /* Events */
1002 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1003 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1004 
1005 extern const struct kfd_device_global_init_class device_global_init_class_cik;
1006 
1007 void kfd_event_init_process(struct kfd_process *p);
1008 void kfd_event_free_process(struct kfd_process *p);
1009 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1010 int kfd_wait_on_events(struct kfd_process *p,
1011 		       uint32_t num_events, void __user *data,
1012 		       bool all, uint32_t user_timeout_ms,
1013 		       uint32_t *wait_result);
1014 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
1015 				uint32_t valid_id_bits);
1016 void kfd_signal_iommu_event(struct kfd_dev *dev,
1017 		unsigned int pasid, unsigned long address,
1018 		bool is_write_requested, bool is_execute_requested);
1019 void kfd_signal_hw_exception_event(unsigned int pasid);
1020 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1021 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1022 int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
1023 		       uint64_t size);
1024 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1025 		     uint32_t event_type, bool auto_reset, uint32_t node_id,
1026 		     uint32_t *event_id, uint32_t *event_trigger_data,
1027 		     uint64_t *event_page_offset, uint32_t *event_slot_index);
1028 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1029 
1030 void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid,
1031 				struct kfd_vm_fault_info *info);
1032 
1033 void kfd_signal_reset_event(struct kfd_dev *dev);
1034 
1035 void kfd_flush_tlb(struct kfd_process_device *pdd);
1036 
1037 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
1038 
1039 bool kfd_is_locked(void);
1040 
1041 /* Compute profile */
1042 void kfd_inc_compute_active(struct kfd_dev *dev);
1043 void kfd_dec_compute_active(struct kfd_dev *dev);
1044 
1045 /* Debugfs */
1046 #if defined(CONFIG_DEBUG_FS)
1047 
1048 void kfd_debugfs_init(void);
1049 void kfd_debugfs_fini(void);
1050 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1051 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1052 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1053 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1054 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1055 int pm_debugfs_runlist(struct seq_file *m, void *data);
1056 
1057 int kfd_debugfs_hang_hws(struct kfd_dev *dev);
1058 int pm_debugfs_hang_hws(struct packet_manager *pm);
1059 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm);
1060 
1061 #else
1062 
1063 static inline void kfd_debugfs_init(void) {}
1064 static inline void kfd_debugfs_fini(void) {}
1065 
1066 #endif
1067 
1068 #endif
1069