xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_priv.h (revision a3a02a52bcfcbcc4a637d4b68bf1bc391c9fad02)
1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #ifndef KFD_PRIV_H_INCLUDED
25 #define KFD_PRIV_H_INCLUDED
26 
27 #include <linux/hashtable.h>
28 #include <linux/mmu_notifier.h>
29 #include <linux/memremap.h>
30 #include <linux/mutex.h>
31 #include <linux/types.h>
32 #include <linux/atomic.h>
33 #include <linux/workqueue.h>
34 #include <linux/spinlock.h>
35 #include <linux/kfd_ioctl.h>
36 #include <linux/idr.h>
37 #include <linux/kfifo.h>
38 #include <linux/seq_file.h>
39 #include <linux/kref.h>
40 #include <linux/sysfs.h>
41 #include <linux/device_cgroup.h>
42 #include <drm/drm_file.h>
43 #include <drm/drm_drv.h>
44 #include <drm/drm_device.h>
45 #include <drm/drm_ioctl.h>
46 #include <kgd_kfd_interface.h>
47 #include <linux/swap.h>
48 
49 #include "amd_shared.h"
50 #include "amdgpu.h"
51 
52 #define KFD_MAX_RING_ENTRY_SIZE	8
53 
54 #define KFD_SYSFS_FILE_MODE 0444
55 
56 /* GPU ID hash width in bits */
57 #define KFD_GPU_ID_HASH_WIDTH 16
58 
59 /* Use upper bits of mmap offset to store KFD driver specific information.
60  * BITS[63:62] - Encode MMAP type
61  * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to
62  * BITS[45:0]  - MMAP offset value
63  *
64  * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these
65  *  defines are w.r.t to PAGE_SIZE
66  */
67 #define KFD_MMAP_TYPE_SHIFT	62
68 #define KFD_MMAP_TYPE_MASK	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
69 #define KFD_MMAP_TYPE_DOORBELL	(0x3ULL << KFD_MMAP_TYPE_SHIFT)
70 #define KFD_MMAP_TYPE_EVENTS	(0x2ULL << KFD_MMAP_TYPE_SHIFT)
71 #define KFD_MMAP_TYPE_RESERVED_MEM	(0x1ULL << KFD_MMAP_TYPE_SHIFT)
72 #define KFD_MMAP_TYPE_MMIO	(0x0ULL << KFD_MMAP_TYPE_SHIFT)
73 
74 #define KFD_MMAP_GPU_ID_SHIFT 46
75 #define KFD_MMAP_GPU_ID_MASK (((1ULL << KFD_GPU_ID_HASH_WIDTH) - 1) \
76 				<< KFD_MMAP_GPU_ID_SHIFT)
77 #define KFD_MMAP_GPU_ID(gpu_id) ((((uint64_t)gpu_id) << KFD_MMAP_GPU_ID_SHIFT)\
78 				& KFD_MMAP_GPU_ID_MASK)
79 #define KFD_MMAP_GET_GPU_ID(offset)    ((offset & KFD_MMAP_GPU_ID_MASK) \
80 				>> KFD_MMAP_GPU_ID_SHIFT)
81 
82 /*
83  * When working with cp scheduler we should assign the HIQ manually or via
84  * the amdgpu driver to a fixed hqd slot, here are the fixed HIQ hqd slot
85  * definitions for Kaveri. In Kaveri only the first ME queues participates
86  * in the cp scheduling taking that in mind we set the HIQ slot in the
87  * second ME.
88  */
89 #define KFD_CIK_HIQ_PIPE 4
90 #define KFD_CIK_HIQ_QUEUE 0
91 
92 /* Macro for allocating structures */
93 #define kfd_alloc_struct(ptr_to_struct)	\
94 	((typeof(ptr_to_struct)) kzalloc(sizeof(*ptr_to_struct), GFP_KERNEL))
95 
96 #define KFD_MAX_NUM_OF_PROCESSES 512
97 #define KFD_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
98 
99 /*
100  * Size of the per-process TBA+TMA buffer: 2 pages
101  *
102  * The first chunk is the TBA used for the CWSR ISA code. The second
103  * chunk is used as TMA for user-mode trap handler setup in daisy-chain mode.
104  */
105 #define KFD_CWSR_TBA_TMA_SIZE (PAGE_SIZE * 2)
106 #define KFD_CWSR_TMA_OFFSET (PAGE_SIZE + 2048)
107 
108 #define KFD_MAX_NUM_OF_QUEUES_PER_DEVICE		\
109 	(KFD_MAX_NUM_OF_PROCESSES *			\
110 			KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
111 
112 #define KFD_KERNEL_QUEUE_SIZE 2048
113 
114 #define KFD_UNMAP_LATENCY_MS	(4000)
115 
116 #define KFD_MAX_SDMA_QUEUES	128
117 
118 /*
119  * 512 = 0x200
120  * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the
121  * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA.
122  * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC
123  * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in
124  * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE.
125  */
126 #define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512
127 
128 /**
129  * enum kfd_ioctl_flags - KFD ioctl flags
130  * Various flags that can be set in &amdkfd_ioctl_desc.flags to control how
131  * userspace can use a given ioctl.
132  */
133 enum kfd_ioctl_flags {
134 	/*
135 	 * @KFD_IOC_FLAG_CHECKPOINT_RESTORE:
136 	 * Certain KFD ioctls such as AMDKFD_IOC_CRIU_OP can potentially
137 	 * perform privileged operations and load arbitrary data into MQDs and
138 	 * eventually HQD registers when the queue is mapped by HWS. In order to
139 	 * prevent this we should perform additional security checks.
140 	 *
141 	 * This is equivalent to callers with the CHECKPOINT_RESTORE capability.
142 	 *
143 	 * Note: Since earlier versions of docker do not support CHECKPOINT_RESTORE,
144 	 * we also allow ioctls with SYS_ADMIN capability.
145 	 */
146 	KFD_IOC_FLAG_CHECKPOINT_RESTORE = BIT(0),
147 };
148 /*
149  * Kernel module parameter to specify maximum number of supported queues per
150  * device
151  */
152 extern int max_num_of_queues_per_device;
153 
154 
155 /* Kernel module parameter to specify the scheduling policy */
156 extern int sched_policy;
157 
158 /*
159  * Kernel module parameter to specify the maximum process
160  * number per HW scheduler
161  */
162 extern int hws_max_conc_proc;
163 
164 extern int cwsr_enable;
165 
166 /*
167  * Kernel module parameter to specify whether to send sigterm to HSA process on
168  * unhandled exception
169  */
170 extern int send_sigterm;
171 
172 /*
173  * This kernel module is used to simulate large bar machine on non-large bar
174  * enabled machines.
175  */
176 extern int debug_largebar;
177 
178 /* Set sh_mem_config.retry_disable on GFX v9 */
179 extern int amdgpu_noretry;
180 
181 /* Halt if HWS hang is detected */
182 extern int halt_if_hws_hang;
183 
184 /* Whether MEC FW support GWS barriers */
185 extern bool hws_gws_support;
186 
187 /* Queue preemption timeout in ms */
188 extern int queue_preemption_timeout_ms;
189 
190 /*
191  * Don't evict process queues on vm fault
192  */
193 extern int amdgpu_no_queue_eviction_on_vm_fault;
194 
195 /* Enable eviction debug messages */
196 extern bool debug_evictions;
197 
198 extern struct mutex kfd_processes_mutex;
199 
200 enum cache_policy {
201 	cache_policy_coherent,
202 	cache_policy_noncoherent
203 };
204 
205 #define KFD_GC_VERSION(dev) (amdgpu_ip_version((dev)->adev, GC_HWIP, 0))
206 #define KFD_IS_SOC15(dev)   ((KFD_GC_VERSION(dev)) >= (IP_VERSION(9, 0, 1)))
207 #define KFD_SUPPORT_XNACK_PER_PROCESS(dev)\
208 	((KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2)) ||	\
209 	 (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) ||	\
210 	 (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 4)))
211 
212 struct kfd_node;
213 
214 struct kfd_event_interrupt_class {
215 	bool (*interrupt_isr)(struct kfd_node *dev,
216 			const uint32_t *ih_ring_entry, uint32_t *patched_ihre,
217 			bool *patched_flag);
218 	void (*interrupt_wq)(struct kfd_node *dev,
219 			const uint32_t *ih_ring_entry);
220 };
221 
222 struct kfd_device_info {
223 	uint32_t gfx_target_version;
224 	const struct kfd_event_interrupt_class *event_interrupt_class;
225 	unsigned int max_pasid_bits;
226 	unsigned int max_no_of_hqd;
227 	unsigned int doorbell_size;
228 	size_t ih_ring_entry_size;
229 	uint8_t num_of_watch_points;
230 	uint16_t mqd_size_aligned;
231 	bool supports_cwsr;
232 	bool needs_pci_atomics;
233 	uint32_t no_atomic_fw_version;
234 	unsigned int num_sdma_queues_per_engine;
235 	unsigned int num_reserved_sdma_queues_per_engine;
236 	DECLARE_BITMAP(reserved_sdma_queues_bitmap, KFD_MAX_SDMA_QUEUES);
237 };
238 
239 unsigned int kfd_get_num_sdma_engines(struct kfd_node *kdev);
240 unsigned int kfd_get_num_xgmi_sdma_engines(struct kfd_node *kdev);
241 
242 struct kfd_mem_obj {
243 	uint32_t range_start;
244 	uint32_t range_end;
245 	uint64_t gpu_addr;
246 	uint32_t *cpu_ptr;
247 	void *gtt_mem;
248 };
249 
250 struct kfd_vmid_info {
251 	uint32_t first_vmid_kfd;
252 	uint32_t last_vmid_kfd;
253 	uint32_t vmid_num_kfd;
254 };
255 
256 #define MAX_KFD_NODES	8
257 
258 struct kfd_dev;
259 
260 struct kfd_node {
261 	unsigned int node_id;
262 	struct amdgpu_device *adev;     /* Duplicated here along with keeping
263 					 * a copy in kfd_dev to save a hop
264 					 */
265 	const struct kfd2kgd_calls *kfd2kgd; /* Duplicated here along with
266 					      * keeping a copy in kfd_dev to
267 					      * save a hop
268 					      */
269 	struct kfd_vmid_info vm_info;
270 	unsigned int id;                /* topology stub index */
271 	uint32_t xcc_mask; /* Instance mask of XCCs present */
272 	struct amdgpu_xcp *xcp;
273 
274 	/* Interrupts */
275 	struct kfifo ih_fifo;
276 	struct workqueue_struct *ih_wq;
277 	struct work_struct interrupt_work;
278 	spinlock_t interrupt_lock;
279 
280 	/*
281 	 * Interrupts of interest to KFD are copied
282 	 * from the HW ring into a SW ring.
283 	 */
284 	bool interrupts_active;
285 	uint32_t interrupt_bitmap; /* Only used for GFX 9.4.3 */
286 
287 	/* QCM Device instance */
288 	struct device_queue_manager *dqm;
289 
290 	/* Global GWS resource shared between processes */
291 	void *gws;
292 	bool gws_debug_workaround;
293 
294 	/* Clients watching SMI events */
295 	struct list_head smi_clients;
296 	spinlock_t smi_lock;
297 	uint32_t reset_seq_num;
298 
299 	/* SRAM ECC flag */
300 	atomic_t sram_ecc_flag;
301 
302 	/*spm process id */
303 	unsigned int spm_pasid;
304 
305 	/* Maximum process number mapped to HW scheduler */
306 	unsigned int max_proc_per_quantum;
307 
308 	unsigned int compute_vmid_bitmap;
309 
310 	struct kfd_local_mem_info local_mem_info;
311 
312 	struct kfd_dev *kfd;
313 };
314 
315 struct kfd_dev {
316 	struct amdgpu_device *adev;
317 
318 	struct kfd_device_info device_info;
319 
320 	u32 __iomem *doorbell_kernel_ptr; /* This is a pointer for a doorbells
321 					   * page used by kernel queue
322 					   */
323 
324 	struct kgd2kfd_shared_resources shared_resources;
325 
326 	const struct kfd2kgd_calls *kfd2kgd;
327 	struct mutex doorbell_mutex;
328 
329 	void *gtt_mem;
330 	uint64_t gtt_start_gpu_addr;
331 	void *gtt_start_cpu_ptr;
332 	void *gtt_sa_bitmap;
333 	struct mutex gtt_sa_lock;
334 	unsigned int gtt_sa_chunk_size;
335 	unsigned int gtt_sa_num_of_chunks;
336 
337 	bool init_complete;
338 
339 	/* Firmware versions */
340 	uint16_t mec_fw_version;
341 	uint16_t mec2_fw_version;
342 	uint16_t sdma_fw_version;
343 
344 	/* CWSR */
345 	bool cwsr_enabled;
346 	const void *cwsr_isa;
347 	unsigned int cwsr_isa_size;
348 
349 	/* xGMI */
350 	uint64_t hive_id;
351 
352 	bool pci_atomic_requested;
353 
354 	/* Compute Profile ref. count */
355 	atomic_t compute_profile;
356 
357 	struct ida doorbell_ida;
358 	unsigned int max_doorbell_slices;
359 
360 	int noretry;
361 
362 	struct kfd_node *nodes[MAX_KFD_NODES];
363 	unsigned int num_nodes;
364 
365 	/* Track per device allocated watch points */
366 	uint32_t alloc_watch_ids;
367 	spinlock_t watch_points_lock;
368 
369 	/* Kernel doorbells for KFD device */
370 	struct amdgpu_bo *doorbells;
371 
372 	/* bitmap for dynamic doorbell allocation from doorbell object */
373 	unsigned long *doorbell_bitmap;
374 };
375 
376 enum kfd_mempool {
377 	KFD_MEMPOOL_SYSTEM_CACHEABLE = 1,
378 	KFD_MEMPOOL_SYSTEM_WRITECOMBINE = 2,
379 	KFD_MEMPOOL_FRAMEBUFFER = 3,
380 };
381 
382 /* Character device interface */
383 int kfd_chardev_init(void);
384 void kfd_chardev_exit(void);
385 
386 /**
387  * enum kfd_unmap_queues_filter - Enum for queue filters.
388  *
389  * @KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES: Preempts all queues in the
390  *						running queues list.
391  *
392  * @KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES: Preempts all non-static queues
393  *						in the run list.
394  *
395  * @KFD_UNMAP_QUEUES_FILTER_BY_PASID: Preempts queues that belongs to
396  *						specific process.
397  *
398  */
399 enum kfd_unmap_queues_filter {
400 	KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES = 1,
401 	KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES = 2,
402 	KFD_UNMAP_QUEUES_FILTER_BY_PASID = 3
403 };
404 
405 /**
406  * enum kfd_queue_type - Enum for various queue types.
407  *
408  * @KFD_QUEUE_TYPE_COMPUTE: Regular user mode queue type.
409  *
410  * @KFD_QUEUE_TYPE_SDMA: SDMA user mode queue type.
411  *
412  * @KFD_QUEUE_TYPE_HIQ: HIQ queue type.
413  *
414  * @KFD_QUEUE_TYPE_DIQ: DIQ queue type.
415  *
416  * @KFD_QUEUE_TYPE_SDMA_XGMI: Special SDMA queue for XGMI interface.
417  */
418 enum kfd_queue_type  {
419 	KFD_QUEUE_TYPE_COMPUTE,
420 	KFD_QUEUE_TYPE_SDMA,
421 	KFD_QUEUE_TYPE_HIQ,
422 	KFD_QUEUE_TYPE_DIQ,
423 	KFD_QUEUE_TYPE_SDMA_XGMI
424 };
425 
426 enum kfd_queue_format {
427 	KFD_QUEUE_FORMAT_PM4,
428 	KFD_QUEUE_FORMAT_AQL
429 };
430 
431 enum KFD_QUEUE_PRIORITY {
432 	KFD_QUEUE_PRIORITY_MINIMUM = 0,
433 	KFD_QUEUE_PRIORITY_MAXIMUM = 15
434 };
435 
436 /**
437  * struct queue_properties
438  *
439  * @type: The queue type.
440  *
441  * @queue_id: Queue identifier.
442  *
443  * @queue_address: Queue ring buffer address.
444  *
445  * @queue_size: Queue ring buffer size.
446  *
447  * @priority: Defines the queue priority relative to other queues in the
448  * process.
449  * This is just an indication and HW scheduling may override the priority as
450  * necessary while keeping the relative prioritization.
451  * the priority granularity is from 0 to f which f is the highest priority.
452  * currently all queues are initialized with the highest priority.
453  *
454  * @queue_percent: This field is partially implemented and currently a zero in
455  * this field defines that the queue is non active.
456  *
457  * @read_ptr: User space address which points to the number of dwords the
458  * cp read from the ring buffer. This field updates automatically by the H/W.
459  *
460  * @write_ptr: Defines the number of dwords written to the ring buffer.
461  *
462  * @doorbell_ptr: Notifies the H/W of new packet written to the queue ring
463  * buffer. This field should be similar to write_ptr and the user should
464  * update this field after updating the write_ptr.
465  *
466  * @doorbell_off: The doorbell offset in the doorbell pci-bar.
467  *
468  * @is_interop: Defines if this is a interop queue. Interop queue means that
469  * the queue can access both graphics and compute resources.
470  *
471  * @is_evicted: Defines if the queue is evicted. Only active queues
472  * are evicted, rendering them inactive.
473  *
474  * @is_active: Defines if the queue is active or not. @is_active and
475  * @is_evicted are protected by the DQM lock.
476  *
477  * @is_gws: Defines if the queue has been updated to be GWS-capable or not.
478  * @is_gws should be protected by the DQM lock, since changing it can yield the
479  * possibility of updating DQM state on number of GWS queues.
480  *
481  * @vmid: If the scheduling mode is no cp scheduling the field defines the vmid
482  * of the queue.
483  *
484  * This structure represents the queue properties for each queue no matter if
485  * it's user mode or kernel mode queue.
486  *
487  */
488 
489 struct queue_properties {
490 	enum kfd_queue_type type;
491 	enum kfd_queue_format format;
492 	unsigned int queue_id;
493 	uint64_t queue_address;
494 	uint64_t  queue_size;
495 	uint32_t priority;
496 	uint32_t queue_percent;
497 	uint32_t *read_ptr;
498 	uint32_t *write_ptr;
499 	void __iomem *doorbell_ptr;
500 	uint32_t doorbell_off;
501 	bool is_interop;
502 	bool is_evicted;
503 	bool is_suspended;
504 	bool is_being_destroyed;
505 	bool is_active;
506 	bool is_gws;
507 	uint32_t pm4_target_xcc;
508 	bool is_dbg_wa;
509 	bool is_user_cu_masked;
510 	/* Not relevant for user mode queues in cp scheduling */
511 	unsigned int vmid;
512 	/* Relevant only for sdma queues*/
513 	uint32_t sdma_engine_id;
514 	uint32_t sdma_queue_id;
515 	uint32_t sdma_vm_addr;
516 	/* Relevant only for VI */
517 	uint64_t eop_ring_buffer_address;
518 	uint32_t eop_ring_buffer_size;
519 	uint64_t ctx_save_restore_area_address;
520 	uint32_t ctx_save_restore_area_size;
521 	uint32_t ctl_stack_size;
522 	uint64_t tba_addr;
523 	uint64_t tma_addr;
524 	uint64_t exception_status;
525 };
526 
527 #define QUEUE_IS_ACTIVE(q) ((q).queue_size > 0 &&	\
528 			    (q).queue_address != 0 &&	\
529 			    (q).queue_percent > 0 &&	\
530 			    !(q).is_evicted &&		\
531 			    !(q).is_suspended)
532 
533 enum mqd_update_flag {
534 	UPDATE_FLAG_DBG_WA_ENABLE = 1,
535 	UPDATE_FLAG_DBG_WA_DISABLE = 2,
536 	UPDATE_FLAG_IS_GWS = 4, /* quirk for gfx9 IP */
537 };
538 
539 struct mqd_update_info {
540 	union {
541 		struct {
542 			uint32_t count; /* Must be a multiple of 32 */
543 			uint32_t *ptr;
544 		} cu_mask;
545 	};
546 	enum mqd_update_flag update_flag;
547 };
548 
549 /**
550  * struct queue
551  *
552  * @list: Queue linked list.
553  *
554  * @mqd: The queue MQD (memory queue descriptor).
555  *
556  * @mqd_mem_obj: The MQD local gpu memory object.
557  *
558  * @gart_mqd_addr: The MQD gart mc address.
559  *
560  * @properties: The queue properties.
561  *
562  * @mec: Used only in no cp scheduling mode and identifies to micro engine id
563  *	 that the queue should be executed on.
564  *
565  * @pipe: Used only in no cp scheduling mode and identifies the queue's pipe
566  *	  id.
567  *
568  * @queue: Used only in no cp scheduliong mode and identifies the queue's slot.
569  *
570  * @process: The kfd process that created this queue.
571  *
572  * @device: The kfd device that created this queue.
573  *
574  * @gws: Pointing to gws kgd_mem if this is a gws control queue; NULL
575  * otherwise.
576  *
577  * This structure represents user mode compute queues.
578  * It contains all the necessary data to handle such queues.
579  *
580  */
581 
582 struct queue {
583 	struct list_head list;
584 	void *mqd;
585 	struct kfd_mem_obj *mqd_mem_obj;
586 	uint64_t gart_mqd_addr;
587 	struct queue_properties properties;
588 
589 	uint32_t mec;
590 	uint32_t pipe;
591 	uint32_t queue;
592 
593 	unsigned int sdma_id;
594 	unsigned int doorbell_id;
595 
596 	struct kfd_process	*process;
597 	struct kfd_node		*device;
598 	void *gws;
599 
600 	/* procfs */
601 	struct kobject kobj;
602 
603 	void *gang_ctx_bo;
604 	uint64_t gang_ctx_gpu_addr;
605 	void *gang_ctx_cpu_ptr;
606 
607 	struct amdgpu_bo *wptr_bo;
608 };
609 
610 enum KFD_MQD_TYPE {
611 	KFD_MQD_TYPE_HIQ = 0,		/* for hiq */
612 	KFD_MQD_TYPE_CP,		/* for cp queues and diq */
613 	KFD_MQD_TYPE_SDMA,		/* for sdma queues */
614 	KFD_MQD_TYPE_DIQ,		/* for diq */
615 	KFD_MQD_TYPE_MAX
616 };
617 
618 enum KFD_PIPE_PRIORITY {
619 	KFD_PIPE_PRIORITY_CS_LOW = 0,
620 	KFD_PIPE_PRIORITY_CS_MEDIUM,
621 	KFD_PIPE_PRIORITY_CS_HIGH
622 };
623 
624 struct scheduling_resources {
625 	unsigned int vmid_mask;
626 	enum kfd_queue_type type;
627 	uint64_t queue_mask;
628 	uint64_t gws_mask;
629 	uint32_t oac_mask;
630 	uint32_t gds_heap_base;
631 	uint32_t gds_heap_size;
632 };
633 
634 struct process_queue_manager {
635 	/* data */
636 	struct kfd_process	*process;
637 	struct list_head	queues;
638 	unsigned long		*queue_slot_bitmap;
639 };
640 
641 struct qcm_process_device {
642 	/* The Device Queue Manager that owns this data */
643 	struct device_queue_manager *dqm;
644 	struct process_queue_manager *pqm;
645 	/* Queues list */
646 	struct list_head queues_list;
647 	struct list_head priv_queue_list;
648 
649 	unsigned int queue_count;
650 	unsigned int vmid;
651 	bool is_debug;
652 	unsigned int evicted; /* eviction counter, 0=active */
653 
654 	/* This flag tells if we should reset all wavefronts on
655 	 * process termination
656 	 */
657 	bool reset_wavefronts;
658 
659 	/* This flag tells us if this process has a GWS-capable
660 	 * queue that will be mapped into the runlist. It's
661 	 * possible to request a GWS BO, but not have the queue
662 	 * currently mapped, and this changes how the MAP_PROCESS
663 	 * PM4 packet is configured.
664 	 */
665 	bool mapped_gws_queue;
666 
667 	/* All the memory management data should be here too */
668 	uint64_t gds_context_area;
669 	/* Contains page table flags such as AMDGPU_PTE_VALID since gfx9 */
670 	uint64_t page_table_base;
671 	uint32_t sh_mem_config;
672 	uint32_t sh_mem_bases;
673 	uint32_t sh_mem_ape1_base;
674 	uint32_t sh_mem_ape1_limit;
675 	uint32_t gds_size;
676 	uint32_t num_gws;
677 	uint32_t num_oac;
678 	uint32_t sh_hidden_private_base;
679 
680 	/* CWSR memory */
681 	struct kgd_mem *cwsr_mem;
682 	void *cwsr_kaddr;
683 	uint64_t cwsr_base;
684 	uint64_t tba_addr;
685 	uint64_t tma_addr;
686 
687 	/* IB memory */
688 	struct kgd_mem *ib_mem;
689 	uint64_t ib_base;
690 	void *ib_kaddr;
691 
692 	/* doorbells for kfd process */
693 	struct amdgpu_bo *proc_doorbells;
694 
695 	/* bitmap for dynamic doorbell allocation from the bo */
696 	unsigned long *doorbell_bitmap;
697 };
698 
699 /* KFD Memory Eviction */
700 
701 /* Approx. wait time before attempting to restore evicted BOs */
702 #define PROCESS_RESTORE_TIME_MS 100
703 /* Approx. back off time if restore fails due to lack of memory */
704 #define PROCESS_BACK_OFF_TIME_MS 100
705 /* Approx. time before evicting the process again */
706 #define PROCESS_ACTIVE_TIME_MS 10
707 
708 /* 8 byte handle containing GPU ID in the most significant 4 bytes and
709  * idr_handle in the least significant 4 bytes
710  */
711 #define MAKE_HANDLE(gpu_id, idr_handle) \
712 	(((uint64_t)(gpu_id) << 32) + idr_handle)
713 #define GET_GPU_ID(handle) (handle >> 32)
714 #define GET_IDR_HANDLE(handle) (handle & 0xFFFFFFFF)
715 
716 enum kfd_pdd_bound {
717 	PDD_UNBOUND = 0,
718 	PDD_BOUND,
719 	PDD_BOUND_SUSPENDED,
720 };
721 
722 #define MAX_SYSFS_FILENAME_LEN 15
723 
724 /*
725  * SDMA counter runs at 100MHz frequency.
726  * We display SDMA activity in microsecond granularity in sysfs.
727  * As a result, the divisor is 100.
728  */
729 #define SDMA_ACTIVITY_DIVISOR  100
730 
731 /* Data that is per-process-per device. */
732 struct kfd_process_device {
733 	/* The device that owns this data. */
734 	struct kfd_node *dev;
735 
736 	/* The process that owns this kfd_process_device. */
737 	struct kfd_process *process;
738 
739 	/* per-process-per device QCM data structure */
740 	struct qcm_process_device qpd;
741 
742 	/*Apertures*/
743 	uint64_t lds_base;
744 	uint64_t lds_limit;
745 	uint64_t gpuvm_base;
746 	uint64_t gpuvm_limit;
747 	uint64_t scratch_base;
748 	uint64_t scratch_limit;
749 
750 	/* VM context for GPUVM allocations */
751 	struct file *drm_file;
752 	void *drm_priv;
753 
754 	/* GPUVM allocations storage */
755 	struct idr alloc_idr;
756 
757 	/* Flag used to tell the pdd has dequeued from the dqm.
758 	 * This is used to prevent dev->dqm->ops.process_termination() from
759 	 * being called twice when it is already called in IOMMU callback
760 	 * function.
761 	 */
762 	bool already_dequeued;
763 	bool runtime_inuse;
764 
765 	/* Is this process/pasid bound to this device? (amd_iommu_bind_pasid) */
766 	enum kfd_pdd_bound bound;
767 
768 	/* VRAM usage */
769 	uint64_t vram_usage;
770 	struct attribute attr_vram;
771 	char vram_filename[MAX_SYSFS_FILENAME_LEN];
772 
773 	/* SDMA activity tracking */
774 	uint64_t sdma_past_activity_counter;
775 	struct attribute attr_sdma;
776 	char sdma_filename[MAX_SYSFS_FILENAME_LEN];
777 
778 	/* Eviction activity tracking */
779 	uint64_t last_evict_timestamp;
780 	atomic64_t evict_duration_counter;
781 	struct attribute attr_evict;
782 
783 	struct kobject *kobj_stats;
784 
785 	/*
786 	 * @cu_occupancy: Reports occupancy of Compute Units (CU) of a process
787 	 * that is associated with device encoded by "this" struct instance. The
788 	 * value reflects CU usage by all of the waves launched by this process
789 	 * on this device. A very important property of occupancy parameter is
790 	 * that its value is a snapshot of current use.
791 	 *
792 	 * Following is to be noted regarding how this parameter is reported:
793 	 *
794 	 *  The number of waves that a CU can launch is limited by couple of
795 	 *  parameters. These are encoded by struct amdgpu_cu_info instance
796 	 *  that is part of every device definition. For GFX9 devices this
797 	 *  translates to 40 waves (simd_per_cu * max_waves_per_simd) when waves
798 	 *  do not use scratch memory and 32 waves (max_scratch_slots_per_cu)
799 	 *  when they do use scratch memory. This could change for future
800 	 *  devices and therefore this example should be considered as a guide.
801 	 *
802 	 *  All CU's of a device are available for the process. This may not be true
803 	 *  under certain conditions - e.g. CU masking.
804 	 *
805 	 *  Finally number of CU's that are occupied by a process is affected by both
806 	 *  number of CU's a device has along with number of other competing processes
807 	 */
808 	struct attribute attr_cu_occupancy;
809 
810 	/* sysfs counters for GPU retry fault and page migration tracking */
811 	struct kobject *kobj_counters;
812 	struct attribute attr_faults;
813 	struct attribute attr_page_in;
814 	struct attribute attr_page_out;
815 	uint64_t faults;
816 	uint64_t page_in;
817 	uint64_t page_out;
818 
819 	/* Exception code status*/
820 	uint64_t exception_status;
821 	void *vm_fault_exc_data;
822 	size_t vm_fault_exc_data_size;
823 
824 	/* Tracks debug per-vmid request settings */
825 	uint32_t spi_dbg_override;
826 	uint32_t spi_dbg_launch_mode;
827 	uint32_t watch_points[4];
828 	uint32_t alloc_watch_ids;
829 
830 	/*
831 	 * If this process has been checkpointed before, then the user
832 	 * application will use the original gpu_id on the
833 	 * checkpointed node to refer to this device.
834 	 */
835 	uint32_t user_gpu_id;
836 
837 	void *proc_ctx_bo;
838 	uint64_t proc_ctx_gpu_addr;
839 	void *proc_ctx_cpu_ptr;
840 };
841 
842 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
843 
844 struct svm_range_list {
845 	struct mutex			lock;
846 	struct rb_root_cached		objects;
847 	struct list_head		list;
848 	struct work_struct		deferred_list_work;
849 	struct list_head		deferred_range_list;
850 	struct list_head                criu_svm_metadata_list;
851 	spinlock_t			deferred_list_lock;
852 	atomic_t			evicted_ranges;
853 	atomic_t			drain_pagefaults;
854 	struct delayed_work		restore_work;
855 	DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE);
856 	struct task_struct		*faulting_task;
857 };
858 
859 /* Process data */
860 struct kfd_process {
861 	/*
862 	 * kfd_process are stored in an mm_struct*->kfd_process*
863 	 * hash table (kfd_processes in kfd_process.c)
864 	 */
865 	struct hlist_node kfd_processes;
866 
867 	/*
868 	 * Opaque pointer to mm_struct. We don't hold a reference to
869 	 * it so it should never be dereferenced from here. This is
870 	 * only used for looking up processes by their mm.
871 	 */
872 	void *mm;
873 
874 	struct kref ref;
875 	struct work_struct release_work;
876 
877 	struct mutex mutex;
878 
879 	/*
880 	 * In any process, the thread that started main() is the lead
881 	 * thread and outlives the rest.
882 	 * It is here because amd_iommu_bind_pasid wants a task_struct.
883 	 * It can also be used for safely getting a reference to the
884 	 * mm_struct of the process.
885 	 */
886 	struct task_struct *lead_thread;
887 
888 	/* We want to receive a notification when the mm_struct is destroyed */
889 	struct mmu_notifier mmu_notifier;
890 
891 	u32 pasid;
892 
893 	/*
894 	 * Array of kfd_process_device pointers,
895 	 * one for each device the process is using.
896 	 */
897 	struct kfd_process_device *pdds[MAX_GPU_INSTANCE];
898 	uint32_t n_pdds;
899 
900 	struct process_queue_manager pqm;
901 
902 	/*Is the user space process 32 bit?*/
903 	bool is_32bit_user_mode;
904 
905 	/* Event-related data */
906 	struct mutex event_mutex;
907 	/* Event ID allocator and lookup */
908 	struct idr event_idr;
909 	/* Event page */
910 	u64 signal_handle;
911 	struct kfd_signal_page *signal_page;
912 	size_t signal_mapped_size;
913 	size_t signal_event_count;
914 	bool signal_event_limit_reached;
915 
916 	/* Information used for memory eviction */
917 	void *kgd_process_info;
918 	/* Eviction fence that is attached to all the BOs of this process. The
919 	 * fence will be triggered during eviction and new one will be created
920 	 * during restore
921 	 */
922 	struct dma_fence __rcu *ef;
923 
924 	/* Work items for evicting and restoring BOs */
925 	struct delayed_work eviction_work;
926 	struct delayed_work restore_work;
927 	/* seqno of the last scheduled eviction */
928 	unsigned int last_eviction_seqno;
929 	/* Approx. the last timestamp (in jiffies) when the process was
930 	 * restored after an eviction
931 	 */
932 	unsigned long last_restore_timestamp;
933 
934 	/* Indicates device process is debug attached with reserved vmid. */
935 	bool debug_trap_enabled;
936 
937 	/* per-process-per device debug event fd file */
938 	struct file *dbg_ev_file;
939 
940 	/* If the process is a kfd debugger, we need to know so we can clean
941 	 * up at exit time.  If a process enables debugging on itself, it does
942 	 * its own clean-up, so we don't set the flag here.  We track this by
943 	 * counting the number of processes this process is debugging.
944 	 */
945 	atomic_t debugged_process_count;
946 
947 	/* If the process is a debugged, this is the debugger process */
948 	struct kfd_process *debugger_process;
949 
950 	/* Kobj for our procfs */
951 	struct kobject *kobj;
952 	struct kobject *kobj_queues;
953 	struct attribute attr_pasid;
954 
955 	/* Keep track cwsr init */
956 	bool has_cwsr;
957 
958 	/* Exception code enable mask and status */
959 	uint64_t exception_enable_mask;
960 	uint64_t exception_status;
961 
962 	/* Used to drain stale interrupts */
963 	wait_queue_head_t wait_irq_drain;
964 	bool irq_drain_is_open;
965 
966 	/* shared virtual memory registered by this process */
967 	struct svm_range_list svms;
968 
969 	bool xnack_enabled;
970 
971 	/* Work area for debugger event writer worker. */
972 	struct work_struct debug_event_workarea;
973 
974 	/* Tracks debug per-vmid request for debug flags */
975 	u32 dbg_flags;
976 
977 	atomic_t poison;
978 	/* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */
979 	bool queues_paused;
980 
981 	/* Tracks runtime enable status */
982 	struct semaphore runtime_enable_sema;
983 	bool is_runtime_retry;
984 	struct kfd_runtime_info runtime_info;
985 };
986 
987 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
988 extern DECLARE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
989 extern struct srcu_struct kfd_processes_srcu;
990 
991 /**
992  * typedef amdkfd_ioctl_t - typedef for ioctl function pointer.
993  *
994  * @filep: pointer to file structure.
995  * @p: amdkfd process pointer.
996  * @data: pointer to arg that was copied from user.
997  *
998  * Return: returns ioctl completion code.
999  */
1000 typedef int amdkfd_ioctl_t(struct file *filep, struct kfd_process *p,
1001 				void *data);
1002 
1003 struct amdkfd_ioctl_desc {
1004 	unsigned int cmd;
1005 	int flags;
1006 	amdkfd_ioctl_t *func;
1007 	unsigned int cmd_drv;
1008 	const char *name;
1009 };
1010 bool kfd_dev_is_large_bar(struct kfd_node *dev);
1011 
1012 int kfd_process_create_wq(void);
1013 void kfd_process_destroy_wq(void);
1014 void kfd_cleanup_processes(void);
1015 struct kfd_process *kfd_create_process(struct task_struct *thread);
1016 struct kfd_process *kfd_get_process(const struct task_struct *task);
1017 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid);
1018 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm);
1019 
1020 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id);
1021 int kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
1022 				uint32_t *gpuid, uint32_t *gpuidx);
1023 static inline int kfd_process_gpuid_from_gpuidx(struct kfd_process *p,
1024 				uint32_t gpuidx, uint32_t *gpuid) {
1025 	return gpuidx < p->n_pdds ? p->pdds[gpuidx]->dev->id : -EINVAL;
1026 }
1027 static inline struct kfd_process_device *kfd_process_device_from_gpuidx(
1028 				struct kfd_process *p, uint32_t gpuidx) {
1029 	return gpuidx < p->n_pdds ? p->pdds[gpuidx] : NULL;
1030 }
1031 
1032 void kfd_unref_process(struct kfd_process *p);
1033 int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger);
1034 int kfd_process_restore_queues(struct kfd_process *p);
1035 void kfd_suspend_all_processes(void);
1036 int kfd_resume_all_processes(void);
1037 
1038 struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *process,
1039 							 uint32_t gpu_id);
1040 
1041 int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id);
1042 
1043 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1044 			       struct file *drm_file);
1045 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,
1046 						struct kfd_process *p);
1047 struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,
1048 							struct kfd_process *p);
1049 struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,
1050 							struct kfd_process *p);
1051 
1052 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported);
1053 
1054 int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,
1055 			  struct vm_area_struct *vma);
1056 
1057 /* KFD process API for creating and translating handles */
1058 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1059 					void *mem);
1060 void *kfd_process_device_translate_handle(struct kfd_process_device *p,
1061 					int handle);
1062 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1063 					int handle);
1064 struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid);
1065 
1066 /* PASIDs */
1067 int kfd_pasid_init(void);
1068 void kfd_pasid_exit(void);
1069 bool kfd_set_pasid_limit(unsigned int new_limit);
1070 unsigned int kfd_get_pasid_limit(void);
1071 u32 kfd_pasid_alloc(void);
1072 void kfd_pasid_free(u32 pasid);
1073 
1074 /* Doorbells */
1075 size_t kfd_doorbell_process_slice(struct kfd_dev *kfd);
1076 int kfd_doorbell_init(struct kfd_dev *kfd);
1077 void kfd_doorbell_fini(struct kfd_dev *kfd);
1078 int kfd_doorbell_mmap(struct kfd_node *dev, struct kfd_process *process,
1079 		      struct vm_area_struct *vma);
1080 void __iomem *kfd_get_kernel_doorbell(struct kfd_dev *kfd,
1081 					unsigned int *doorbell_off);
1082 void kfd_release_kernel_doorbell(struct kfd_dev *kfd, u32 __iomem *db_addr);
1083 u32 read_kernel_doorbell(u32 __iomem *db);
1084 void write_kernel_doorbell(void __iomem *db, u32 value);
1085 void write_kernel_doorbell64(void __iomem *db, u64 value);
1086 unsigned int kfd_get_doorbell_dw_offset_in_bar(struct kfd_dev *kfd,
1087 					struct kfd_process_device *pdd,
1088 					unsigned int doorbell_id);
1089 phys_addr_t kfd_get_process_doorbells(struct kfd_process_device *pdd);
1090 int kfd_alloc_process_doorbells(struct kfd_dev *kfd,
1091 				struct kfd_process_device *pdd);
1092 void kfd_free_process_doorbells(struct kfd_dev *kfd,
1093 				struct kfd_process_device *pdd);
1094 /* GTT Sub-Allocator */
1095 
1096 int kfd_gtt_sa_allocate(struct kfd_node *node, unsigned int size,
1097 			struct kfd_mem_obj **mem_obj);
1098 
1099 int kfd_gtt_sa_free(struct kfd_node *node, struct kfd_mem_obj *mem_obj);
1100 
1101 extern struct device *kfd_device;
1102 
1103 /* KFD's procfs */
1104 void kfd_procfs_init(void);
1105 void kfd_procfs_shutdown(void);
1106 int kfd_procfs_add_queue(struct queue *q);
1107 void kfd_procfs_del_queue(struct queue *q);
1108 
1109 /* Topology */
1110 int kfd_topology_init(void);
1111 void kfd_topology_shutdown(void);
1112 int kfd_topology_add_device(struct kfd_node *gpu);
1113 int kfd_topology_remove_device(struct kfd_node *gpu);
1114 struct kfd_topology_device *kfd_topology_device_by_proximity_domain(
1115 						uint32_t proximity_domain);
1116 struct kfd_topology_device *kfd_topology_device_by_proximity_domain_no_lock(
1117 						uint32_t proximity_domain);
1118 struct kfd_topology_device *kfd_topology_device_by_id(uint32_t gpu_id);
1119 struct kfd_node *kfd_device_by_id(uint32_t gpu_id);
1120 struct kfd_node *kfd_device_by_pci_dev(const struct pci_dev *pdev);
1121 static inline bool kfd_irq_is_from_node(struct kfd_node *node, uint32_t node_id,
1122 					uint32_t vmid)
1123 {
1124 	return (node->interrupt_bitmap & (1 << node_id)) != 0 &&
1125 	       (node->compute_vmid_bitmap & (1 << vmid)) != 0;
1126 }
1127 static inline struct kfd_node *kfd_node_by_irq_ids(struct amdgpu_device *adev,
1128 					uint32_t node_id, uint32_t vmid) {
1129 	struct kfd_dev *dev = adev->kfd.dev;
1130 	uint32_t i;
1131 
1132 	if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3) &&
1133 	    KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 4))
1134 		return dev->nodes[0];
1135 
1136 	for (i = 0; i < dev->num_nodes; i++)
1137 		if (kfd_irq_is_from_node(dev->nodes[i], node_id, vmid))
1138 			return dev->nodes[i];
1139 
1140 	return NULL;
1141 }
1142 int kfd_topology_enum_kfd_devices(uint8_t idx, struct kfd_node **kdev);
1143 int kfd_numa_node_to_apic_id(int numa_node_id);
1144 
1145 /* Interrupts */
1146 #define	KFD_IRQ_FENCE_CLIENTID	0xff
1147 #define	KFD_IRQ_FENCE_SOURCEID	0xff
1148 #define	KFD_IRQ_IS_FENCE(client, source)				\
1149 				((client) == KFD_IRQ_FENCE_CLIENTID &&	\
1150 				(source) == KFD_IRQ_FENCE_SOURCEID)
1151 int kfd_interrupt_init(struct kfd_node *dev);
1152 void kfd_interrupt_exit(struct kfd_node *dev);
1153 bool enqueue_ih_ring_entry(struct kfd_node *kfd, const void *ih_ring_entry);
1154 bool interrupt_is_wanted(struct kfd_node *dev,
1155 				const uint32_t *ih_ring_entry,
1156 				uint32_t *patched_ihre, bool *flag);
1157 int kfd_process_drain_interrupts(struct kfd_process_device *pdd);
1158 void kfd_process_close_interrupt_drain(unsigned int pasid);
1159 
1160 /* amdkfd Apertures */
1161 int kfd_init_apertures(struct kfd_process *process);
1162 
1163 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1164 				  uint64_t tba_addr,
1165 				  uint64_t tma_addr);
1166 void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
1167 				     bool enabled);
1168 
1169 /* CWSR initialization */
1170 int kfd_process_init_cwsr_apu(struct kfd_process *process, struct file *filep);
1171 
1172 /* CRIU */
1173 /*
1174  * Need to increment KFD_CRIU_PRIV_VERSION each time a change is made to any of the CRIU private
1175  * structures:
1176  * kfd_criu_process_priv_data
1177  * kfd_criu_device_priv_data
1178  * kfd_criu_bo_priv_data
1179  * kfd_criu_queue_priv_data
1180  * kfd_criu_event_priv_data
1181  * kfd_criu_svm_range_priv_data
1182  */
1183 
1184 #define KFD_CRIU_PRIV_VERSION 1
1185 
1186 struct kfd_criu_process_priv_data {
1187 	uint32_t version;
1188 	uint32_t xnack_mode;
1189 };
1190 
1191 struct kfd_criu_device_priv_data {
1192 	/* For future use */
1193 	uint64_t reserved;
1194 };
1195 
1196 struct kfd_criu_bo_priv_data {
1197 	uint64_t user_addr;
1198 	uint32_t idr_handle;
1199 	uint32_t mapped_gpuids[MAX_GPU_INSTANCE];
1200 };
1201 
1202 /*
1203  * The first 4 bytes of kfd_criu_queue_priv_data, kfd_criu_event_priv_data,
1204  * kfd_criu_svm_range_priv_data is the object type
1205  */
1206 enum kfd_criu_object_type {
1207 	KFD_CRIU_OBJECT_TYPE_QUEUE,
1208 	KFD_CRIU_OBJECT_TYPE_EVENT,
1209 	KFD_CRIU_OBJECT_TYPE_SVM_RANGE,
1210 };
1211 
1212 struct kfd_criu_svm_range_priv_data {
1213 	uint32_t object_type;
1214 	uint64_t start_addr;
1215 	uint64_t size;
1216 	/* Variable length array of attributes */
1217 	struct kfd_ioctl_svm_attribute attrs[];
1218 };
1219 
1220 struct kfd_criu_queue_priv_data {
1221 	uint32_t object_type;
1222 	uint64_t q_address;
1223 	uint64_t q_size;
1224 	uint64_t read_ptr_addr;
1225 	uint64_t write_ptr_addr;
1226 	uint64_t doorbell_off;
1227 	uint64_t eop_ring_buffer_address;
1228 	uint64_t ctx_save_restore_area_address;
1229 	uint32_t gpu_id;
1230 	uint32_t type;
1231 	uint32_t format;
1232 	uint32_t q_id;
1233 	uint32_t priority;
1234 	uint32_t q_percent;
1235 	uint32_t doorbell_id;
1236 	uint32_t gws;
1237 	uint32_t sdma_id;
1238 	uint32_t eop_ring_buffer_size;
1239 	uint32_t ctx_save_restore_area_size;
1240 	uint32_t ctl_stack_size;
1241 	uint32_t mqd_size;
1242 };
1243 
1244 struct kfd_criu_event_priv_data {
1245 	uint32_t object_type;
1246 	uint64_t user_handle;
1247 	uint32_t event_id;
1248 	uint32_t auto_reset;
1249 	uint32_t type;
1250 	uint32_t signaled;
1251 
1252 	union {
1253 		struct kfd_hsa_memory_exception_data memory_exception_data;
1254 		struct kfd_hsa_hw_exception_data hw_exception_data;
1255 	};
1256 };
1257 
1258 int kfd_process_get_queue_info(struct kfd_process *p,
1259 			       uint32_t *num_queues,
1260 			       uint64_t *priv_data_sizes);
1261 
1262 int kfd_criu_checkpoint_queues(struct kfd_process *p,
1263 			 uint8_t __user *user_priv_data,
1264 			 uint64_t *priv_data_offset);
1265 
1266 int kfd_criu_restore_queue(struct kfd_process *p,
1267 			   uint8_t __user *user_priv_data,
1268 			   uint64_t *priv_data_offset,
1269 			   uint64_t max_priv_data_size);
1270 
1271 int kfd_criu_checkpoint_events(struct kfd_process *p,
1272 			 uint8_t __user *user_priv_data,
1273 			 uint64_t *priv_data_offset);
1274 
1275 int kfd_criu_restore_event(struct file *devkfd,
1276 			   struct kfd_process *p,
1277 			   uint8_t __user *user_priv_data,
1278 			   uint64_t *priv_data_offset,
1279 			   uint64_t max_priv_data_size);
1280 /* CRIU - End */
1281 
1282 /* Queue Context Management */
1283 int init_queue(struct queue **q, const struct queue_properties *properties);
1284 void uninit_queue(struct queue *q);
1285 void print_queue_properties(struct queue_properties *q);
1286 void print_queue(struct queue *q);
1287 
1288 struct mqd_manager *mqd_manager_init_cik(enum KFD_MQD_TYPE type,
1289 		struct kfd_node *dev);
1290 struct mqd_manager *mqd_manager_init_vi(enum KFD_MQD_TYPE type,
1291 		struct kfd_node *dev);
1292 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
1293 		struct kfd_node *dev);
1294 struct mqd_manager *mqd_manager_init_v10(enum KFD_MQD_TYPE type,
1295 		struct kfd_node *dev);
1296 struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type,
1297 		struct kfd_node *dev);
1298 struct mqd_manager *mqd_manager_init_v12(enum KFD_MQD_TYPE type,
1299 		struct kfd_node *dev);
1300 struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev);
1301 void device_queue_manager_uninit(struct device_queue_manager *dqm);
1302 struct kernel_queue *kernel_queue_init(struct kfd_node *dev,
1303 					enum kfd_queue_type type);
1304 void kernel_queue_uninit(struct kernel_queue *kq);
1305 int kfd_dqm_evict_pasid(struct device_queue_manager *dqm, u32 pasid);
1306 
1307 /* Process Queue Manager */
1308 struct process_queue_node {
1309 	struct queue *q;
1310 	struct kernel_queue *kq;
1311 	struct list_head process_queue_list;
1312 };
1313 
1314 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd);
1315 void kfd_process_dequeue_from_all_devices(struct kfd_process *p);
1316 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p);
1317 void pqm_uninit(struct process_queue_manager *pqm);
1318 int pqm_create_queue(struct process_queue_manager *pqm,
1319 			    struct kfd_node *dev,
1320 			    struct file *f,
1321 			    struct queue_properties *properties,
1322 			    unsigned int *qid,
1323 			    struct amdgpu_bo *wptr_bo,
1324 			    const struct kfd_criu_queue_priv_data *q_data,
1325 			    const void *restore_mqd,
1326 			    const void *restore_ctl_stack,
1327 			    uint32_t *p_doorbell_offset_in_process);
1328 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid);
1329 int pqm_update_queue_properties(struct process_queue_manager *pqm, unsigned int qid,
1330 			struct queue_properties *p);
1331 int pqm_update_mqd(struct process_queue_manager *pqm, unsigned int qid,
1332 			struct mqd_update_info *minfo);
1333 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
1334 			void *gws);
1335 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
1336 						unsigned int qid);
1337 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
1338 						unsigned int qid);
1339 int pqm_get_wave_state(struct process_queue_manager *pqm,
1340 		       unsigned int qid,
1341 		       void __user *ctl_stack,
1342 		       u32 *ctl_stack_used_size,
1343 		       u32 *save_area_used_size);
1344 int pqm_get_queue_snapshot(struct process_queue_manager *pqm,
1345 			   uint64_t exception_clear_mask,
1346 			   void __user *buf,
1347 			   int *num_qss_entries,
1348 			   uint32_t *entry_size);
1349 
1350 int amdkfd_fence_wait_timeout(struct device_queue_manager *dqm,
1351 			      uint64_t fence_value,
1352 			      unsigned int timeout_ms);
1353 
1354 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,
1355 				  unsigned int qid,
1356 				  u32 *mqd_size,
1357 				  u32 *ctl_stack_size);
1358 /* Packet Manager */
1359 
1360 #define KFD_FENCE_COMPLETED (100)
1361 #define KFD_FENCE_INIT   (10)
1362 
1363 struct packet_manager {
1364 	struct device_queue_manager *dqm;
1365 	struct kernel_queue *priv_queue;
1366 	struct mutex lock;
1367 	bool allocated;
1368 	struct kfd_mem_obj *ib_buffer_obj;
1369 	unsigned int ib_size_bytes;
1370 	bool is_over_subscription;
1371 
1372 	const struct packet_manager_funcs *pmf;
1373 };
1374 
1375 struct packet_manager_funcs {
1376 	/* Support ASIC-specific packet formats for PM4 packets */
1377 	int (*map_process)(struct packet_manager *pm, uint32_t *buffer,
1378 			struct qcm_process_device *qpd);
1379 	int (*runlist)(struct packet_manager *pm, uint32_t *buffer,
1380 			uint64_t ib, size_t ib_size_in_dwords, bool chain);
1381 	int (*set_resources)(struct packet_manager *pm, uint32_t *buffer,
1382 			struct scheduling_resources *res);
1383 	int (*map_queues)(struct packet_manager *pm, uint32_t *buffer,
1384 			struct queue *q, bool is_static);
1385 	int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer,
1386 			enum kfd_unmap_queues_filter mode,
1387 			uint32_t filter_param, bool reset);
1388 	int (*set_grace_period)(struct packet_manager *pm, uint32_t *buffer,
1389 			uint32_t grace_period);
1390 	int (*query_status)(struct packet_manager *pm, uint32_t *buffer,
1391 			uint64_t fence_address,	uint64_t fence_value);
1392 	int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer);
1393 
1394 	/* Packet sizes */
1395 	int map_process_size;
1396 	int runlist_size;
1397 	int set_resources_size;
1398 	int map_queues_size;
1399 	int unmap_queues_size;
1400 	int set_grace_period_size;
1401 	int query_status_size;
1402 	int release_mem_size;
1403 };
1404 
1405 extern const struct packet_manager_funcs kfd_vi_pm_funcs;
1406 extern const struct packet_manager_funcs kfd_v9_pm_funcs;
1407 extern const struct packet_manager_funcs kfd_aldebaran_pm_funcs;
1408 
1409 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm);
1410 void pm_uninit(struct packet_manager *pm);
1411 int pm_send_set_resources(struct packet_manager *pm,
1412 				struct scheduling_resources *res);
1413 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues);
1414 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
1415 				uint64_t fence_value);
1416 
1417 int pm_send_unmap_queue(struct packet_manager *pm,
1418 			enum kfd_unmap_queues_filter mode,
1419 			uint32_t filter_param, bool reset);
1420 
1421 void pm_release_ib(struct packet_manager *pm);
1422 
1423 int pm_update_grace_period(struct packet_manager *pm, uint32_t grace_period);
1424 
1425 /* Following PM funcs can be shared among VI and AI */
1426 unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size);
1427 
1428 uint64_t kfd_get_number_elems(struct kfd_dev *kfd);
1429 
1430 /* Events */
1431 extern const struct kfd_event_interrupt_class event_interrupt_class_cik;
1432 extern const struct kfd_event_interrupt_class event_interrupt_class_v9;
1433 extern const struct kfd_event_interrupt_class event_interrupt_class_v9_4_3;
1434 extern const struct kfd_event_interrupt_class event_interrupt_class_v10;
1435 extern const struct kfd_event_interrupt_class event_interrupt_class_v11;
1436 
1437 extern const struct kfd_device_global_init_class device_global_init_class_cik;
1438 
1439 int kfd_event_init_process(struct kfd_process *p);
1440 void kfd_event_free_process(struct kfd_process *p);
1441 int kfd_event_mmap(struct kfd_process *process, struct vm_area_struct *vma);
1442 int kfd_wait_on_events(struct kfd_process *p,
1443 		       uint32_t num_events, void __user *data,
1444 		       bool all, uint32_t *user_timeout_ms,
1445 		       uint32_t *wait_result);
1446 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
1447 				uint32_t valid_id_bits);
1448 void kfd_signal_hw_exception_event(u32 pasid);
1449 int kfd_set_event(struct kfd_process *p, uint32_t event_id);
1450 int kfd_reset_event(struct kfd_process *p, uint32_t event_id);
1451 int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset);
1452 
1453 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
1454 		     uint32_t event_type, bool auto_reset, uint32_t node_id,
1455 		     uint32_t *event_id, uint32_t *event_trigger_data,
1456 		     uint64_t *event_page_offset, uint32_t *event_slot_index);
1457 
1458 int kfd_get_num_events(struct kfd_process *p);
1459 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id);
1460 
1461 void kfd_signal_vm_fault_event(struct kfd_node *dev, u32 pasid,
1462 				struct kfd_vm_fault_info *info,
1463 				struct kfd_hsa_memory_exception_data *data);
1464 
1465 void kfd_signal_reset_event(struct kfd_node *dev);
1466 
1467 void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid);
1468 
1469 static inline void kfd_flush_tlb(struct kfd_process_device *pdd,
1470 				 enum TLB_FLUSH_TYPE type)
1471 {
1472 	struct amdgpu_device *adev = pdd->dev->adev;
1473 	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1474 
1475 	amdgpu_vm_flush_compute_tlb(adev, vm, type, pdd->dev->xcc_mask);
1476 }
1477 
1478 static inline bool kfd_flush_tlb_after_unmap(struct kfd_dev *dev)
1479 {
1480 	return KFD_GC_VERSION(dev) >= IP_VERSION(9, 4, 2) ||
1481 	       (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && dev->sdma_fw_version >= 18) ||
1482 	       KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0);
1483 }
1484 
1485 int kfd_send_exception_to_runtime(struct kfd_process *p,
1486 				unsigned int queue_id,
1487 				uint64_t error_reason);
1488 bool kfd_is_locked(void);
1489 
1490 /* Compute profile */
1491 void kfd_inc_compute_active(struct kfd_node *dev);
1492 void kfd_dec_compute_active(struct kfd_node *dev);
1493 
1494 /* Cgroup Support */
1495 /* Check with device cgroup if @kfd device is accessible */
1496 static inline int kfd_devcgroup_check_permission(struct kfd_node *node)
1497 {
1498 #if defined(CONFIG_CGROUP_DEVICE) || defined(CONFIG_CGROUP_BPF)
1499 	struct drm_device *ddev;
1500 
1501 	if (node->xcp)
1502 		ddev = node->xcp->ddev;
1503 	else
1504 		ddev = adev_to_drm(node->adev);
1505 
1506 	return devcgroup_check_permission(DEVCG_DEV_CHAR, DRM_MAJOR,
1507 					  ddev->render->index,
1508 					  DEVCG_ACC_WRITE | DEVCG_ACC_READ);
1509 #else
1510 	return 0;
1511 #endif
1512 }
1513 
1514 static inline bool kfd_is_first_node(struct kfd_node *node)
1515 {
1516 	return (node == node->kfd->nodes[0]);
1517 }
1518 
1519 /* Debugfs */
1520 #if defined(CONFIG_DEBUG_FS)
1521 
1522 void kfd_debugfs_init(void);
1523 void kfd_debugfs_fini(void);
1524 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data);
1525 int pqm_debugfs_mqds(struct seq_file *m, void *data);
1526 int kfd_debugfs_hqds_by_device(struct seq_file *m, void *data);
1527 int dqm_debugfs_hqds(struct seq_file *m, void *data);
1528 int kfd_debugfs_rls_by_device(struct seq_file *m, void *data);
1529 int pm_debugfs_runlist(struct seq_file *m, void *data);
1530 
1531 int kfd_debugfs_hang_hws(struct kfd_node *dev);
1532 int pm_debugfs_hang_hws(struct packet_manager *pm);
1533 int dqm_debugfs_hang_hws(struct device_queue_manager *dqm);
1534 
1535 #else
1536 
1537 static inline void kfd_debugfs_init(void) {}
1538 static inline void kfd_debugfs_fini(void) {}
1539 
1540 #endif
1541 
1542 #endif
1543