xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_priv.h (revision 9a6b55ac)
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 <linux/device_cgroup.h>
40 #include <drm/drm_file.h>
41 #include <drm/drm_drv.h>
42 #include <drm/drm_device.h>
43 #include <kgd_kfd_interface.h>
44 
45 #include "amd_shared.h"
46 
47 #define KFD_MAX_RING_ENTRY_SIZE	8
48 
49 #define KFD_SYSFS_FILE_MODE 0444
50 
51 /* GPU ID hash width in bits */
52 #define KFD_GPU_ID_HASH_WIDTH 16
53 
54 /* Use upper bits of mmap offset to store KFD driver specific information.
55  * BITS[63:62] - Encode MMAP type
56  * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to
57  * BITS[45:0]  - MMAP offset value
58  *
59  * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these
60  *  defines are w.r.t to PAGE_SIZE
61  */
62 #define KFD_MMAP_TYPE_SHIFT	(62 - PAGE_SHIFT)
63 #define KFD_MMAP_TYPE_MASK	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
64 #define KFD_MMAP_TYPE_DOORBELL	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
65 #define KFD_MMAP_TYPE_EVENTS	(0x2ULL << KFD_MMAP_TYPE_SHIFT)
66 #define KFD_MMAP_TYPE_RESERVED_MEM	(0x1ULL << KFD_MMAP_TYPE_SHIFT)
67 #define KFD_MMAP_TYPE_MMIO	(0x0ULL << KFD_MMAP_TYPE_SHIFT)
68 
69 #define KFD_MMAP_GPU_ID_SHIFT (46 - PAGE_SHIFT)
70 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \
71 				<< KFD_MMAP_GPU_ID_SHIFT)
72 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\
73 				& KFD_MMAP_GPU_ID_MASK)
74 #define KFD_MMAP_GPU_ID_GET(offset)    ((offset & KFD_MMAP_GPU_ID_MASK) \
75 				>> KFD_MMAP_GPU_ID_SHIFT)
76 
77 #define KFD_MMAP_OFFSET_VALUE_MASK	(0x3FFFFFFFFFFFULL >> PAGE_SHIFT)
78 #define KFD_MMAP_OFFSET_VALUE_GET(offset) (offset & KFD_MMAP_OFFSET_VALUE_MASK)
79 
80 /*
81  * When working with cp scheduler we should assign the HIQ manually or via
82  * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot
83  * definitions for Kaveri. In Kaveri only the first ME queues participates
84  * in the cp scheduling taking that in mind we set the HIQ slot in the
85  * second ME.
86  */
87 #define KFD_CIK_HIQ_PIPE 4
88 #define KFD_CIK_HIQ_QUEUE 0
89 
90 /* Macro for allocating structures */
91 #define kfd_alloc_struct(ptr_to_struct)	\
92 	((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
93 
94 #define KFD_MAX_NUM_OF_PROCESSES 512
95 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
96 
97 /*
98  * Size of the per-process TBA+TMA buffer: 2 pages
99  *
100  * The first page is the TBA used for the CWSR ISA code. The second
101  * page is used as TMA for daisy changing a user-mode trap handler.
102  */
103 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
104 #define KFD_CWSR_TMA_OFFSET PAGE_SIZE
105 
106 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE		\
107 	(KFD_MAX_NUM_OF_PROCESSES *			\
108 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
109 
110 #define KFD_KERNEL_QUEUE_SIZE 2048
111 
112 #define KFD_UNMAP_LATENCY_MS	(4000)
113 
114 /*
115  * 512 = 0x200
116  * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
117  * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.
118  * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC
119  * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in
120  * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.
121  */
122 #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
123 
124 
125 /*
126  * Kernel module parameter to specify maximum number of supported queues per
127  * device
128  */
129 extern int max_num_of_queues_per_device;
130 
131 
132 /* Kernel module parameter to specify the scheduling policy */
133 extern int sched_policy;
134 
135 /*
136  * Kernel module parameter to specify the maximum process
137  * number per HW scheduler
138  */
139 extern int hws_max_conc_proc;
140 
141 extern int cwsr_enable;
142 
143 /*
144  * Kernel module parameter to specify whether to send sigterm to HSA process on
145  * unhandled exception
146  */
147 extern int send_sigterm;
148 
149 /*
150  * This kernel module is used to simulate large bar machine on non-large bar
151  * enabled machines.
152  */
153 extern int debug_largebar;
154 
155 /*
156  * Ignore CRAT table during KFD initialization, can be used to work around
157  * broken CRAT tables on some AMD systems
158  */
159 extern int ignore_crat;
160 
161 /*
162  * Set sh_mem_config.retry_disable on Vega10
163  */
164 extern int amdgpu_noretry;
165 
166 /*
167  * Halt if HWS hang is detected
168  */
169 extern int halt_if_hws_hang;
170 
171 /*
172  * Whether MEC FW support GWS barriers
173  */
174 extern bool hws_gws_support;
175 
176 /*
177  * Queue preemption timeout in ms
178  */
179 extern int queue_preemption_timeout_ms;
180 
181 enum cache_policy {
182 	cache_policy_coherent,
183 	cache_policy_noncoherent
184 };
185 
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 char *asic_name;
199 	const struct kfd_event_interrupt_class *event_interrupt_class;
200 	unsigned int max_pasid_bits;
201 	unsigned int max_no_of_hqd;
202 	unsigned int doorbell_size;
203 	size_t ih_ring_entry_size;
204 	uint8_t num_of_watch_points;
205 	uint16_t mqd_size_aligned;
206 	bool supports_cwsr;
207 	bool needs_iommu_device;
208 	bool needs_pci_atomics;
209 	unsigned int num_sdma_engines;
210 	unsigned int num_xgmi_sdma_engines;
211 	unsigned int num_sdma_queues_per_engine;
212 };
213 
214 struct kfd_mem_obj {
215 	uint32_t range_start;
216 	uint32_t range_end;
217 	uint64_t gpu_addr;
218 	uint32_t *cpu_ptr;
219 	void *gtt_mem;
220 };
221 
222 struct kfd_vmid_info {
223 	uint32_t first_vmid_kfd;
224 	uint32_t last_vmid_kfd;
225 	uint32_t vmid_num_kfd;
226 };
227 
228 struct kfd_dev {
229 	struct kgd_dev *kgd;
230 
231 	const struct kfd_device_info *device_info;
232 	struct pci_dev *pdev;
233 	struct drm_device *ddev;
234 
235 	unsigned int id;		/* topology stub index */
236 
237 	phys_addr_t doorbell_base;	/* Start of actual doorbells used by
238 					 * KFD. It is aligned for mapping
239 					 * into user mode
240 					 */
241 	size_t doorbell_id_offset;	/* Doorbell offset (from KFD doorbell
242 					 * to HW doorbell, GFX reserved some
243 					 * at the start)
244 					 */
245 	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
246 					   * page used by kernel queue
247 					   */
248 
249 	struct kgd2kfd_shared_resources shared_resources;
250 	struct kfd_vmid_info vm_info;
251 
252 	const struct kfd2kgd_calls *kfd2kgd;
253 	struct mutex doorbell_mutex;
254 	DECLARE_BITMAP(doorbell_available_index,
255 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
256 
257 	void *gtt_mem;
258 	uint64_t gtt_start_gpu_addr;
259 	void *gtt_start_cpu_ptr;
260 	void *gtt_sa_bitmap;
261 	struct mutex gtt_sa_lock;
262 	unsigned int gtt_sa_chunk_size;
263 	unsigned int gtt_sa_num_of_chunks;
264 
265 	/* Interrupts */
266 	struct kfifo ih_fifo;
267 	struct workqueue_struct *ih_wq;
268 	struct work_struct interrupt_work;
269 	spinlock_t interrupt_lock;
270 
271 	/* QCM Device instance */
272 	struct device_queue_manager *dqm;
273 
274 	bool init_complete;
275 	/*
276 	 * Interrupts of interest to KFD are copied
277 	 * from the HW ring into a SW ring.
278 	 */
279 	bool interrupts_active;
280 
281 	/* Debug manager */
282 	struct kfd_dbgmgr *dbgmgr;
283 
284 	/* Firmware versions */
285 	uint16_t mec_fw_version;
286 	uint16_t sdma_fw_version;
287 
288 	/* Maximum process number mapped to HW scheduler */
289 	unsigned int max_proc_per_quantum;
290 
291 	/* CWSR */
292 	bool cwsr_enabled;
293 	const void *cwsr_isa;
294 	unsigned int cwsr_isa_size;
295 
296 	/* xGMI */
297 	uint64_t hive_id;
298 
299 	bool pci_atomic_requested;
300 
301 	/* SRAM ECC flag */
302 	atomic_t sram_ecc_flag;
303 
304 	/* Compute Profile ref. count */
305 	atomic_t compute_profile;
306 
307 	/* Global GWS resource shared b/t processes*/
308 	void *gws;
309 };
310 
311 enum kfd_mempool {
312 	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
313 	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
314 	KFD_MEMPOOL_FRAMEBUFFER = 3,
315 };
316 
317 /* Character device interface */
318 int kfd_chardev_init(void);
319 void kfd_chardev_exit(void);
320 struct device *kfd_chardev(void);
321 
322 /**
323  * enum kfd_unmap_queues_filter
324  *
325  * @KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE: Preempts single queue.
326  *
327  * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
328  *						running queues list.
329  *
330  * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
331  *						specific process.
332  *
333  */
334 enum kfd_unmap_queues_filter {
335 	KFD_UNMAP_QUEUES_FILTER_SINGLE_QUEUE,
336 	KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
337 	KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
338 	KFD_UNMAP_QUEUES_FILTER_BY_PASID
339 };
340 
341 /**
342  * enum kfd_queue_type
343  *
344  * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
345  *
346  * @KFD_QUEUE_TYPE_SDMA: Sdma user mode queue type.
347  *
348  * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
349  *
350  * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
351  */
352 enum kfd_queue_type  {
353 	KFD_QUEUE_TYPE_COMPUTE,
354 	KFD_QUEUE_TYPE_SDMA,
355 	KFD_QUEUE_TYPE_HIQ,
356 	KFD_QUEUE_TYPE_DIQ,
357 	KFD_QUEUE_TYPE_SDMA_XGMI
358 };
359 
360 enum kfd_queue_format {
361 	KFD_QUEUE_FORMAT_PM4,
362 	KFD_QUEUE_FORMAT_AQL
363 };
364 
365 enum KFD_QUEUE_PRIORITY {
366 	KFD_QUEUE_PRIORITY_MINIMUM = 0,
367 	KFD_QUEUE_PRIORITY_MAXIMUM = 15
368 };
369 
370 /**
371  * struct queue_properties
372  *
373  * @type: The queue type.
374  *
375  * @queue_id: Queue identifier.
376  *
377  * @queue_address: Queue ring buffer address.
378  *
379  * @queue_size: Queue ring buffer size.
380  *
381  * @priority: Defines the queue priority relative to other queues in the
382  * process.
383  * This is just an indication and HW scheduling may override the priority as
384  * necessary while keeping the relative prioritization.
385  * the priority granularity is from 0 to f which f is the highest priority.
386  * currently all queues are initialized with the highest priority.
387  *
388  * @queue_percent: This field is partially implemented and currently a zero in
389  * this field defines that the queue is non active.
390  *
391  * @read_ptr: User space address which points to the number of dwords the
392  * cp read from the ring buffer. This field updates automatically by the H/W.
393  *
394  * @write_ptr: Defines the number of dwords written to the ring buffer.
395  *
396  * @doorbell_ptr: This field aim is to notify the H/W of new packet written to
397  * the queue ring buffer. This field should be similar to write_ptr and the
398  * user should update this field after he updated the write_ptr.
399  *
400  * @doorbell_off: The doorbell offset in the doorbell pci-bar.
401  *
402  * @is_interop: Defines if this is a interop queue. Interop queue means that
403  * the queue can access both graphics and compute resources.
404  *
405  * @is_evicted: Defines if the queue is evicted. Only active queues
406  * are evicted, rendering them inactive.
407  *
408  * @is_active: Defines if the queue is active or not. @is_active and
409  * @is_evicted are protected by the DQM lock.
410  *
411  * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
412  * of the queue.
413  *
414  * This structure represents the queue properties for each queue no matter if
415  * it's user mode or kernel mode queue.
416  *
417  */
418 struct queue_properties {
419 	enum kfd_queue_type type;
420 	enum kfd_queue_format format;
421 	unsigned int queue_id;
422 	uint64_t queue_address;
423 	uint64_t  queue_size;
424 	uint32_t priority;
425 	uint32_t queue_percent;
426 	uint32_t *read_ptr;
427 	uint32_t *write_ptr;
428 	void __iomem *doorbell_ptr;
429 	uint32_t doorbell_off;
430 	bool is_interop;
431 	bool is_evicted;
432 	bool is_active;
433 	/* Not relevant for user mode queues in cp scheduling */
434 	unsigned int vmid;
435 	/* Relevant only for sdma queues*/
436 	uint32_t sdma_engine_id;
437 	uint32_t sdma_queue_id;
438 	uint32_t sdma_vm_addr;
439 	/* Relevant only for VI */
440 	uint64_t eop_ring_buffer_address;
441 	uint32_t eop_ring_buffer_size;
442 	uint64_t ctx_save_restore_area_address;
443 	uint32_t ctx_save_restore_area_size;
444 	uint32_t ctl_stack_size;
445 	uint64_t tba_addr;
446 	uint64_t tma_addr;
447 	/* Relevant for CU */
448 	uint32_t cu_mask_count; /* Must be a multiple of 32 */
449 	uint32_t *cu_mask;
450 };
451 
452 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 &&	\
453 			    (q).queue_address != 0 &&	\
454 			    (q).queue_percent > 0 &&	\
455 			    !(q).is_evicted)
456 
457 /**
458  * struct queue
459  *
460  * @list: Queue linked list.
461  *
462  * @mqd: The queue MQD.
463  *
464  * @mqd_mem_obj: The MQD local gpu memory object.
465  *
466  * @gart_mqd_addr: The MQD gart mc address.
467  *
468  * @properties: The queue properties.
469  *
470  * @mec: Used only in no cp scheduling mode and identifies to micro engine id
471  *	 that the queue should be execute on.
472  *
473  * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
474  *	  id.
475  *
476  * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
477  *
478  * @process: The kfd process that created this queue.
479  *
480  * @device: The kfd device that created this queue.
481  *
482  * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
483  * otherwise.
484  *
485  * This structure represents user mode compute queues.
486  * It contains all the necessary data to handle such queues.
487  *
488  */
489 
490 struct queue {
491 	struct list_head list;
492 	void *mqd;
493 	struct kfd_mem_obj *mqd_mem_obj;
494 	uint64_t gart_mqd_addr;
495 	struct queue_properties properties;
496 
497 	uint32_t mec;
498 	uint32_t pipe;
499 	uint32_t queue;
500 
501 	unsigned int sdma_id;
502 	unsigned int doorbell_id;
503 
504 	struct kfd_process	*process;
505 	struct kfd_dev		*device;
506 	void *gws;
507 };
508 
509 /*
510  * Please read the kfd_mqd_manager.h description.
511  */
512 enum KFD_MQD_TYPE {
513 	KFD_MQD_TYPE_COMPUTE = 0,	/* for no cp scheduling */
514 	KFD_MQD_TYPE_HIQ,		/* for hiq */
515 	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
516 	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
517 	KFD_MQD_TYPE_DIQ,		/* for diq */
518 	KFD_MQD_TYPE_MAX
519 };
520 
521 enum KFD_PIPE_PRIORITY {
522 	KFD_PIPE_PRIORITY_CS_LOW = 0,
523 	KFD_PIPE_PRIORITY_CS_MEDIUM,
524 	KFD_PIPE_PRIORITY_CS_HIGH
525 };
526 
527 struct scheduling_resources {
528 	unsigned int vmid_mask;
529 	enum kfd_queue_type type;
530 	uint64_t queue_mask;
531 	uint64_t gws_mask;
532 	uint32_t oac_mask;
533 	uint32_t gds_heap_base;
534 	uint32_t gds_heap_size;
535 };
536 
537 struct process_queue_manager {
538 	/* data */
539 	struct kfd_process	*process;
540 	struct list_head	queues;
541 	unsigned long		*queue_slot_bitmap;
542 };
543 
544 struct qcm_process_device {
545 	/* The Device Queue Manager that owns this data */
546 	struct device_queue_manager *dqm;
547 	struct process_queue_manager *pqm;
548 	/* Queues list */
549 	struct list_head queues_list;
550 	struct list_head priv_queue_list;
551 
552 	unsigned int queue_count;
553 	unsigned int vmid;
554 	bool is_debug;
555 	unsigned int evicted; /* eviction counter, 0=active */
556 
557 	/* This flag tells if we should reset all wavefronts on
558 	 * process termination
559 	 */
560 	bool reset_wavefronts;
561 
562 	/*
563 	 * All the memory management data should be here too
564 	 */
565 	uint64_t gds_context_area;
566 	/* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
567 	uint64_t page_table_base;
568 	uint32_t sh_mem_config;
569 	uint32_t sh_mem_bases;
570 	uint32_t sh_mem_ape1_base;
571 	uint32_t sh_mem_ape1_limit;
572 	uint32_t gds_size;
573 	uint32_t num_gws;
574 	uint32_t num_oac;
575 	uint32_t sh_hidden_private_base;
576 
577 	/* CWSR memory */
578 	void *cwsr_kaddr;
579 	uint64_t cwsr_base;
580 	uint64_t tba_addr;
581 	uint64_t tma_addr;
582 
583 	/* IB memory */
584 	uint64_t ib_base;
585 	void *ib_kaddr;
586 
587 	/* doorbell resources per process per device */
588 	unsigned long *doorbell_bitmap;
589 };
590 
591 /* KFD Memory Eviction */
592 
593 /* Approx. wait time before attempting to restore evicted BOs */
594 #define PROCESS_RESTORE_TIME_MS 100
595 /* Approx. back off time if restore fails due to lack of memory */
596 #define PROCESS_BACK_OFF_TIME_MS 100
597 /* Approx. time before evicting the process again */
598 #define PROCESS_ACTIVE_TIME_MS 10
599 
600 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
601  * idr_handle in the least significant 4 bytes
602  */
603 #define MAKE_HANDLE(gpu_id, idr_handle) \
604 	(((uint64_t)(gpu_id) << 32) + idr_handle)
605 #define GET_GPU_ID(handle) (handle >> 32)
606 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
607 
608 enum kfd_pdd_bound {
609 	PDD_UNBOUND = 0,
610 	PDD_BOUND,
611 	PDD_BOUND_SUSPENDED,
612 };
613 
614 /* Data that is per-process-per device. */
615 struct kfd_process_device {
616 	/*
617 	 * List of all per-device data for a process.
618 	 * Starts from kfd_process.per_device_data.
619 	 */
620 	struct list_head per_device_list;
621 
622 	/* The device that owns this data. */
623 	struct kfd_dev *dev;
624 
625 	/* The process that owns this kfd_process_device. */
626 	struct kfd_process *process;
627 
628 	/* per-process-per device QCM data structure */
629 	struct qcm_process_device qpd;
630 
631 	/*Apertures*/
632 	uint64_t lds_base;
633 	uint64_t lds_limit;
634 	uint64_t gpuvm_base;
635 	uint64_t gpuvm_limit;
636 	uint64_t scratch_base;
637 	uint64_t scratch_limit;
638 
639 	/* VM context for GPUVM allocations */
640 	struct file *drm_file;
641 	void *vm;
642 
643 	/* GPUVM allocations storage */
644 	struct idr alloc_idr;
645 
646 	/* Flag used to tell the pdd has dequeued from the dqm.
647 	 * This is used to prevent dev->dqm->ops.process_termination() from
648 	 * being called twice when it is already called in IOMMU callback
649 	 * function.
650 	 */
651 	bool already_dequeued;
652 
653 	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
654 	enum kfd_pdd_bound bound;
655 };
656 
657 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
658 
659 /* Process data */
660 struct kfd_process {
661 	/*
662 	 * kfd_process are stored in an mm_struct*->kfd_process*
663 	 * hash table (kfd_processes in kfd_process.c)
664 	 */
665 	struct hlist_node kfd_processes;
666 
667 	/*
668 	 * Opaque pointer to mm_struct. We don't hold a reference to
669 	 * it so it should never be dereferenced from here. This is
670 	 * only used for looking up processes by their mm.
671 	 */
672 	void *mm;
673 
674 	struct kref ref;
675 	struct work_struct release_work;
676 
677 	struct mutex mutex;
678 
679 	/*
680 	 * In any process, the thread that started main() is the lead
681 	 * thread and outlives the rest.
682 	 * It is here because amd_iommu_bind_pasid wants a task_struct.
683 	 * It can also be used for safely getting a reference to the
684 	 * mm_struct of the process.
685 	 */
686 	struct task_struct *lead_thread;
687 
688 	/* We want to receive a notification when the mm_struct is destroyed */
689 	struct mmu_notifier mmu_notifier;
690 
691 	uint16_t pasid;
692 	unsigned int doorbell_index;
693 
694 	/*
695 	 * List of kfd_process_device structures,
696 	 * one for each device the process is using.
697 	 */
698 	struct list_head per_device_data;
699 
700 	struct process_queue_manager pqm;
701 
702 	/*Is the user space process 32 bit?*/
703 	bool is_32bit_user_mode;
704 
705 	/* Event-related data */
706 	struct mutex event_mutex;
707 	/* Event ID allocator and lookup */
708 	struct idr event_idr;
709 	/* Event page */
710 	struct kfd_signal_page *signal_page;
711 	size_t signal_mapped_size;
712 	size_t signal_event_count;
713 	bool signal_event_limit_reached;
714 
715 	/* Information used for memory eviction */
716 	void *kgd_process_info;
717 	/* Eviction fence that is attached to all the BOs of this process. The
718 	 * fence will be triggered during eviction and new one will be created
719 	 * during restore
720 	 */
721 	struct dma_fence *ef;
722 
723 	/* Work items for evicting and restoring BOs */
724 	struct delayed_work eviction_work;
725 	struct delayed_work restore_work;
726 	/* seqno of the last scheduled eviction */
727 	unsigned int last_eviction_seqno;
728 	/* Approx. the last timestamp (in jiffies) when the process was
729 	 * restored after an eviction
730 	 */
731 	unsigned long last_restore_timestamp;
732 
733 	/* Kobj for our procfs */
734 	struct kobject *kobj;
735 	struct attribute attr_pasid;
736 };
737 
738 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
739 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
740 extern struct srcu_struct kfd_processes_srcu;
741 
742 /**
743  * Ioctl function type.
744  *
745  * \param filep pointer to file structure.
746  * \param p amdkfd process pointer.
747  * \param data pointer to arg that was copied from user.
748  */
749 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
750 				void *data);
751 
752 struct amdkfd_ioctl_desc {
753 	unsigned int cmd;
754 	int flags;
755 	amdkfd_ioctl_t *func;
756 	unsigned int cmd_drv;
757 	const char *name;
758 };
759 bool kfd_dev_is_large_bar(struct kfd_dev *dev);
760 
761 int kfd_process_create_wq(void);
762 void kfd_process_destroy_wq(void);
763 struct kfd_process *kfd_create_process(struct file *filep);
764 struct kfd_process *kfd_get_process(const struct task_struct *);
765 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid);
766 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
767 void kfd_unref_process(struct kfd_process *p);
768 int kfd_process_evict_queues(struct kfd_process *p);
769 int kfd_process_restore_queues(struct kfd_process *p);
770 void kfd_suspend_all_processes(void);
771 int kfd_resume_all_processes(void);
772 
773 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
774 			       struct file *drm_file);
775 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
776 						struct kfd_process *p);
777 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
778 							struct kfd_process *p);
779 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
780 							struct kfd_process *p);
781 
782 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
783 			  struct vm_area_struct *vma);
784 
785 /* KFD process API for creating and translating handles */
786 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
787 					void *mem);
788 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
789 					int handle);
790 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
791 					int handle);
792 
793 /* Process device data iterator */
794 struct kfd_process_device *kfd_get_first_process_device_data(
795 							struct kfd_process *p);
796 struct kfd_process_device *kfd_get_next_process_device_data(
797 						struct kfd_process *p,
798 						struct kfd_process_device *pdd);
799 bool kfd_has_process_device_data(struct kfd_process *p);
800 
801 /* PASIDs */
802 int kfd_pasid_init(void);
803 void kfd_pasid_exit(void);
804 bool kfd_set_pasid_limit(unsigned int new_limit);
805 unsigned int kfd_get_pasid_limit(void);
806 unsigned int kfd_pasid_alloc(void);
807 void kfd_pasid_free(unsigned int pasid);
808 
809 /* Doorbells */
810 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
811 int kfd_doorbell_init(struct kfd_dev *kfd);
812 void kfd_doorbell_fini(struct kfd_dev *kfd);
813 int kfd_doorbell_mmap(struct kfd_dev *dev, struct kfd_process *process,
814 		      struct vm_area_struct *vma);
815 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
816 					unsigned int *doorbell_off);
817 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
818 u32 read_kernel_doorbell(u32 __iomem *db);
819 void write_kernel_doorbell(void __iomem *db, u32 value);
820 void write_kernel_doorbell64(void __iomem *db, u64 value);
821 unsigned int kfd_doorbell_id_to_offset(struct kfd_dev *kfd,
822 					struct kfd_process *process,
823 					unsigned int doorbell_id);
824 phys_addr_t kfd_get_process_doorbells(struct kfd_dev *dev,
825 					struct kfd_process *process);
826 int kfd_alloc_process_doorbells(struct kfd_process *process);
827 void kfd_free_process_doorbells(struct kfd_process *process);
828 
829 /* GTT Sub-Allocator */
830 
831 int kfd_gtt_sa_allocate(struct kfd_dev *kfd, unsigned int size,
832 			struct kfd_mem_obj **mem_obj);
833 
834 int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj);
835 
836 extern struct device *kfd_device;
837 
838 /* KFD's procfs */
839 void kfd_procfs_init(void);
840 void kfd_procfs_shutdown(void);
841 
842 /* Topology */
843 int kfd_topology_init(void);
844 void kfd_topology_shutdown(void);
845 int kfd_topology_add_device(struct kfd_dev *gpu);
846 int kfd_topology_remove_device(struct kfd_dev *gpu);
847 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
848 						uint32_t proximity_domain);
849 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
850 struct kfd_dev *kfd_device_by_id(uint32_t gpu_id);
851 struct kfd_dev *kfd_device_by_pci_dev(const struct pci_dev *pdev);
852 struct kfd_dev *kfd_device_by_kgd(const struct kgd_dev *kgd);
853 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_dev **kdev);
854 int kfd_numa_node_to_apic_id(int numa_node_id);
855 
856 /* Interrupts */
857 int kfd_interrupt_init(struct kfd_dev *dev);
858 void kfd_interrupt_exit(struct kfd_dev *dev);
859 bool enqueue_ih_ring_entry(struct kfd_dev *kfd,	const void *ih_ring_entry);
860 bool interrupt_is_wanted(struct kfd_dev *dev,
861 				const uint32_t *ih_ring_entry,
862 				uint32_t *patched_ihre, bool *flag);
863 
864 /* amdkfd Apertures */
865 int kfd_init_apertures(struct kfd_process *process);
866 
867 /* Queue Context Management */
868 int init_queue(struct queue **q, const struct queue_properties *properties);
869 void uninit_queue(struct queue *q);
870 void print_queue_properties(struct queue_properties *q);
871 void print_queue(struct queue *q);
872 
873 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
874 		struct kfd_dev *dev);
875 struct mqd_manager *mqd_manager_init_cik_hawaii(enum KFD_MQD_TYPE type,
876 		struct kfd_dev *dev);
877 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
878 		struct kfd_dev *dev);
879 struct mqd_manager *mqd_manager_init_vi_tonga(enum KFD_MQD_TYPE type,
880 		struct kfd_dev *dev);
881 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
882 		struct kfd_dev *dev);
883 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
884 		struct kfd_dev *dev);
885 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev);
886 void device_queue_manager_uninit(struct device_queue_manager *dqm);
887 struct kernel_queue *kernel_queue_init(struct kfd_dev *dev,
888 					enum kfd_queue_type type);
889 void kernel_queue_uninit(struct kernel_queue *kq);
890 int kfd_process_vm_fault(struct device_queue_manager *dqm, unsigned int pasid);
891 
892 /* Process Queue Manager */
893 struct process_queue_node {
894 	struct queue *q;
895 	struct kernel_queue *kq;
896 	struct list_head process_queue_list;
897 };
898 
899 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
900 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
901 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
902 void pqm_uninit(struct process_queue_manager *pqm);
903 int pqm_create_queue(struct process_queue_manager *pqm,
904 			    struct kfd_dev *dev,
905 			    struct file *f,
906 			    struct queue_properties *properties,
907 			    unsigned int *qid);
908 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
909 int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
910 			struct queue_properties *p);
911 int pqm_set_cu_mask(struct process_queue_manager *pqm, unsigned int qid,
912 			struct queue_properties *p);
913 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
914 			void *gws);
915 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
916 						unsigned int qid);
917 int pqm_get_wave_state(struct process_queue_manager *pqm,
918 		       unsigned int qid,
919 		       void __user *ctl_stack,
920 		       u32 *ctl_stack_used_size,
921 		       u32 *save_area_used_size);
922 
923 int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
924 			      unsigned int fence_value,
925 			      unsigned int timeout_ms);
926 
927 /* Packet Manager */
928 
929 #define KFD_FENCE_COMPLETED (100)
930 #define KFD_FENCE_INIT   (10)
931 
932 struct packet_manager {
933 	struct device_queue_manager *dqm;
934 	struct kernel_queue *priv_queue;
935 	struct mutex lock;
936 	bool allocated;
937 	struct kfd_mem_obj *ib_buffer_obj;
938 	unsigned int ib_size_bytes;
939 	bool is_over_subscription;
940 
941 	const struct packet_manager_funcs *pmf;
942 };
943 
944 struct packet_manager_funcs {
945 	/* Support ASIC-specific packet formats for PM4 packets */
946 	int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
947 			struct qcm_process_device *qpd);
948 	int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
949 			uint64_t ib, size_t ib_size_in_dwords, bool chain);
950 	int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
951 			struct scheduling_resources *res);
952 	int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
953 			struct queue *q, bool is_static);
954 	int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
955 			enum kfd_queue_type type,
956 			enum kfd_unmap_queues_filter mode,
957 			uint32_t filter_param, bool reset,
958 			unsigned int sdma_engine);
959 	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
960 			uint64_t fence_address,	uint32_t fence_value);
961 	int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
962 
963 	/* Packet sizes */
964 	int map_process_size;
965 	int runlist_size;
966 	int set_resources_size;
967 	int map_queues_size;
968 	int unmap_queues_size;
969 	int query_status_size;
970 	int release_mem_size;
971 };
972 
973 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
974 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
975 extern const struct packet_manager_funcs kfd_v10_pm_funcs;
976 
977 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
978 void pm_uninit(struct packet_manager *pm);
979 int pm_send_set_resources(struct packet_manager *pm,
980 				struct scheduling_resources *res);
981 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
982 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
983 				uint32_t fence_value);
984 
985 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
986 			enum kfd_unmap_queues_filter mode,
987 			uint32_t filter_param, bool reset,
988 			unsigned int sdma_engine);
989 
990 void pm_release_ib(struct packet_manager *pm);
991 
992 /* Following PM funcs can be shared among VI and AI */
993 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
994 int pm_set_resources_vi(struct packet_manager *pm, uint32_t *buffer,
995 			struct scheduling_resources *res);
996 
997 
998 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
999 
1000 /* Events */
1001 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1002 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1003 
1004 extern const struct kfd_device_global_init_class device_global_init_class_cik;
1005 
1006 void kfd_event_init_process(struct kfd_process *p);
1007 void kfd_event_free_process(struct kfd_process *p);
1008 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1009 int kfd_wait_on_events(struct kfd_process *p,
1010 		       uint32_t num_events, void __user *data,
1011 		       bool all, uint32_t user_timeout_ms,
1012 		       uint32_t *wait_result);
1013 void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
1014 				uint32_t valid_id_bits);
1015 void kfd_signal_iommu_event(struct kfd_dev *dev,
1016 		unsigned int pasid, unsigned long address,
1017 		bool is_write_requested, bool is_execute_requested);
1018 void kfd_signal_hw_exception_event(unsigned int pasid);
1019 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1020 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1021 int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
1022 		       uint64_t size);
1023 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1024 		     uint32_t event_type, bool auto_reset, uint32_t node_id,
1025 		     uint32_t *event_id, uint32_t *event_trigger_data,
1026 		     uint64_t *event_page_offset, uint32_t *event_slot_index);
1027 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1028 
1029 void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid,
1030 				struct kfd_vm_fault_info *info);
1031 
1032 void kfd_signal_reset_event(struct kfd_dev *dev);
1033 
1034 void kfd_flush_tlb(struct kfd_process_device *pdd);
1035 
1036 int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p);
1037 
1038 bool kfd_is_locked(void);
1039 
1040 /* Compute profile */
1041 void kfd_inc_compute_active(struct kfd_dev *dev);
1042 void kfd_dec_compute_active(struct kfd_dev *dev);
1043 
1044 /* Cgroup Support */
1045 /* Check with device cgroup if @kfd device is accessible */
1046 static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)
1047 {
1048 #if defined(CONFIG_CGROUP_DEVICE)
1049 	struct drm_device *ddev = kfd->ddev;
1050 
1051 	return devcgroup_check_permission(DEVCG_DEV_CHAR, ddev->driver->major,
1052 					  ddev->render->index,
1053 					  DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1054 #else
1055 	return 0;
1056 #endif
1057 }
1058 
1059 /* Debugfs */
1060 #if defined(CONFIG_DEBUG_FS)
1061 
1062 void kfd_debugfs_init(void);
1063 void kfd_debugfs_fini(void);
1064 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1065 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1066 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1067 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1068 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1069 int pm_debugfs_runlist(struct seq_file *m, void *data);
1070 
1071 int kfd_debugfs_hang_hws(struct kfd_dev *dev);
1072 int pm_debugfs_hang_hws(struct packet_manager *pm);
1073 int dqm_debugfs_execute_queues(struct device_queue_manager *dqm);
1074 
1075 #else
1076 
1077 static inline void kfd_debugfs_init(void) {}
1078 static inline void kfd_debugfs_fini(void) {}
1079 
1080 #endif
1081 
1082 #endif
1083